Abstract
The rapid emergence of global antimicrobial resistance (AMR) has outpaced the development of new antimicrobial agents, necessitating transformative approaches such as AI-driven in-silico drug discovery. Antimicrobial peptides (AMPs) are promising alternatives due to their broad-spectrum bactericidal activity and limited susceptibility to resistance. However, the transition from bench to bedside remains constrained by certain challenges, such as instability, potency, and cytotoxicity. Here, well-integrated, multi-model AI-driven pipeline strategies for the mining and discovery of novel AMPs are designed to address these concerns through simultaneous multi-purpose optimization. The proposed architectural frameworks combine discriminative AMP classifiers, quantitative potency, and cytotoxicity screening filters to prioritize AMPs with good efficacy and safety profiles. To ensure novelty, the pipeline integrates multi-layer sequential and genomic screening by adopting alignment and profile-based approaches. Structural refinements are achieved through advanced molecular docking (MD) and protein folding approaches, providing mechanistic insights regarding peptide–target interactions. In parallel, enzymatic susceptibility and stability prediction models are incorporated to optimize AMP pharmacokinetic potentials. Notably, the multi-objective pipeline operates within iterative optimization loops; discriminative and generative AI models and sequential redesign strategies refine candidates based on multi-purpose closed feedback loops across stability, novelty, efficacy, and toxicity outcomes. These systematic and integrated approaches overcome key bottlenecks associated with traditional linear drug discovery, potentially reducing late-stage attrition and accelerating the transformation from in-silico predictions to wet-lab validation. Collectively, this review provides reproducible and generalizable blueprints for next-generation antimicrobial agents, demonstrating the computational potentials of AI-driven, multi-model architectural frameworks to tackle the global AMR crisis and favour precision design of AMPs with better therapeutic indices.
1 Introduction
In the 21st century, antimicrobial resistance (AMR) has evolved as one of the most serious global health concerns, undermining decades of progress in modern medicine (; ). Recent global AMR trends have indicated that millions of deaths are directly attributable to bacterial AMR annually and continue to rise across both low-and high-income countries (; ). The World Health Organization (WHO) highlights that these trends are increasing exponentially across critical priority pathogens, thereby compromising the effectiveness of last-resort antimicrobial therapies and jeopardizing clinical practices such as chemotherapy and regular surgical and transplantation procedures (; ). Without transformative interventional approaches, AMR is going to impose serious health and economic burdens, potentially exceeding morbidity, and mortality outcomes within the coming decades ().
Despite these escalating threats, traditional antibiotic development has remained critically fragile (). At present, few antimicrobial candidates are in clinical development, and few of these use novel mechanisms of action () (Figure 1 and 2). Most current antibiotic discovery paradigms are dependent on incremental structural modifications, vigorous empirical screening, and testing. These often fail to explore the vast chemical and biological spaces that may be open for new development, which are essential to identify next-generation candidates due to high costs (; ) (Figure 2). Consequently, there is an urgent need to transition from traditional discovery. This is where integrative and data-oriented approaches can help to accelerate drug innovation strategies (; ).
Figure 1
Figure 2
Artificial intelligence (AI) has emerged as a transformative paradigm for antibiotic discovery, enabling the systematic interrogation of available large-scale chemical and biological datasets (). AI-driven approaches, such as machine learning (ML) and deep learning (DL), generative modelling, and large language models (LLMs), have shown the capabilities to mine and identify novel candidates, predict their physicochemical properties, and optimize molecular designs with better efficacy and safety profiles (; ; ). Notably, studies have leveraged LLM architectural frameworks to construct end-to-end pipelines capable of generating new antibiotics from vast sequential spaces, achieving higher potency and limited cytotoxicity during wet-lab validation (; ). These advancements have underscored the paradigm shifts from linear and hypothesis-driven drug discovery toward iterative and AI-oriented designs and build–test cycles (; ).
Amongst various emerging modalities, antimicrobial peptides (AMPs) represent one of the most promising candidates, having rapid bactericidal mechanisms and broad-spectrum activity with limited propensity for developing resistance relative to conventional antimicrobial agents (; ; ). However, their clinical translation remains limited by challenges such as suboptimal pharmacokinetics, proteolytic instability, and cytotoxicity. Addressing these limitations requires holistic frameworks, integrating multi-dimensional predictive and generative modelling alongside structural and functional optimization (; ; ).
The current review proposes multi-model AI-driven strategies for AMP mining and optimization, designed to bridge the gap between in-silico prediction and translational applicability (Figure 3). This multifaceted framework integrates advanced discriminative AMP classifiers for activity prediction and generative models for de novo AMP design and generation (; ; ; ) (Figure 3); in-silico potency filters to predict antimicrobial efficacy against proposed pathogens (; ; ) (Figure 3); haemolysis and cytotoxicity predicted modules to ensure host safety (; ; ) (Figure 3); novelty models to enhance chemical and structural diversity, both at the genomic and protein sequence levels () (Figure 3); structural alignments and folding predictions to limit three-dimensional conformations (Figure 3); proteolytic-resistant mappings to optimize peptide stability and in-vivo persistence (Figure 3).
Figure 3
The incorporation of different AI-driven architectural frameworks within a unified pipeline strategy optimizes multi-purpose outcomes across novelty, stability, efficacy, and safety, addressing key bottlenecks which historically limit AMP’s clinical translation (; ) (Figure 3). Importantly, the integration of structural, functional, and proteolytic constraints at initial stages demonstrates critical advancements over pre-existing pipelines, which prioritize activity-based predictions in isolation only (; ).
This review highlights recent progress in AI-enabled AMP discovery and contextualizes the adoption of multi-model strategies within the broader landscape for developing next-generation novel therapeutics. Such integrative frameworks are critical for navigating the complexities associated with AMP designs and for accelerating the clinical translation of computational discoveries to combat drug-resistant pathogens (; ). This review proposes the adaptation of multi-model advanced AI-driven approaches with context for novel AMP discovery, tailored for computational biologists, microbiologists, and clinicians, showing interdisciplinary translational aspects. By doing so, this highlights that the future goals must be directed towards precision-driven antimicrobial development, which will be capable of addressing the global AMR crisis.
1.1 Instability and non-selectivity as pre-design constraints, not post-screening failures
Traditional antibiotic discovery paradigms, which progress linearly from drug design and synthesis to in-vitro and in-vivo validation, may have historically undermined antimicrobial development () (Figure 4). Whilst these conventional frameworks have enabled the discovery of several first-generation antibiotics, they are increasingly unable to address the many new challenges (; ). One of the key limitations is their inherently low throughputs, where the majority of compounds fail at later validation stages due to poor efficacy and drug-related cytotoxic events (; ). Such iterative approaches usually require extensive financial resources and experimental efforts alongside substantial time, often extending across a decade from bench to bedside () (Figure 4). Consequently, these pipelines are economically unsustainable, and the probability of success remains low; therefore, investments for antibiotics are comparatively limited both from the public and private sectors (; ) (Figure 4).
Figure 4
Recently, the WHO warned of the paucity of antibiotics currently in clinical development, and the low numbers of these representing truly novel therapeutic agents (). Mostly, candidates are derived from pre-existing scaffolds, thereby offering accretive improvements over transformative outcomes (; ). This stagnation is compounded by technical bottlenecks; the difficulty of identifying compounds that penetrate bacterial membranes, evade efflux pumps, and retain bactericidal activities against drug-resistant pathogens (; ). Traditional pipelines have failed to tackle the rapid evolutionary resistance approaches, creating widening gaps between therapeutic availability and clinical needs (; ).
The sequential nature of the conventional workflows has introduced a critical structural inefficiency: decision-making is suspended until later-stage wet-lab validation. Instability, non-selectivity, and cytotoxicity issues are usually detected during drug synthesis and biological testing, leading to candidate elimination due to cost attrition (). These “fail-late” paradigms are particularly problematic for peptide-based therapeutics, as protease instability and host cytotoxic events are context-dependent and challenging to streamline through pragmatic iterations (). Altogether, these hindrances underpin fundamental bottlenecks associated with antibiotic development, necessitating a shift toward predicted and multi-purpose design frameworks which are capable of addressing stability, efficacy, and safety simultaneously at the initial stages of drug discovery ().
1.2 AI, as the first stage of stability, efficacy, and safety validation
1.2.1 Discriminative AMP classifiers to predict AMP mining and discovery
Discriminative ML models are considered the computational backbone for early AMP discovery, enabling them to mine and classify candidates based on their likelihood of demonstrating adequate bactericidal activities (; ) (Figure 5). Classical ML methods, such as Support Vector Machines (SVM), Random Forests (RF), k-Nearest Neighbours (kNN), and gradient-boosting frameworks, leverage engineered characteristics based on parameters, i.e., amino acid composition, physicochemical properties, and evolutionary descriptors, to differentiate AMPs from non-functional candidates (; ) (Table 1). These approaches are compatible with high-dimensional biological datasets for capturing non-linear relationships between sequential features and antibacterial properties (; ).
Figure 5
Table 1
| Models & year | Architectural framework | Performance metrics | Key features & experimental validations |
|---|---|---|---|
| a. Classical ML (RF, SVM, GBMs) | |||
| AntiBP2 & AntiBP3 2017/2024 | SVM | AntiBP2: Accuracy (92.1%), MCC (0.84); AntiBP3: AUC (0.93–0.98), (MCC up to 0.86) | SVM-based predictors used balanced +ve and -ve training datasets using residue contact maps to integrate with in silico toxicity screening. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Gabere and Noble, 2017; Bajiya et al., 2024). |
| AmpClass 2024 | Ensemble ML (XGBoost, RF, NN, DT) | Accuracy (93.2%) | Classification & regression approaches to predict novel AMPs. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (). |
| AmpGram 2020 | N-gram encoding & stacked random forests. | AUROC (0.98) | Predicted longer peptides (>10 A.A), used for high-throughput proteomics-based AMPs. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Burdukiewicz et al., 2020). |
| AmPEP 2018 | RF classifiers | Accuracy (96%), MCC (0.90), AUC (0.99) | Provided distribution patterns of amino acid features, with higher accuracy, simplicity, & reduction capability. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (). |
| AmPEPpy 2021 | Random Forest (RF) | Accuracy (91%) | It predicted plant peptides using amino acid composition features. Experimental Validation: No, computational validation only (Lawrence et al., 2021; ). |
| CalcAMP 2023 | LightGBM, ensemble& tree algorithms | Accuracy 86–90% (best RF model 90%) | It used a curated dataset of validated short AMPs and classified them by spectrum. Predicted antimicrobial activity. Experimental Validation: Yes, in-vitro validation (Bournez et al., 2023). |
| CAMP-R3/R4 2016/2023 | ML classifiers (SVM, RF, DA, HMM) | R3: Accuracy (90.5%), R4: AUROC (0.93) | Multi-model approaches showed improvement in performance, i.e., area under the receiver operating characteristic curve. Predicted antimicrobial activity. Experimental Validation: N/A (Waghu et al., 2014; Waghu et al., 2016; Gabere and Noble, 2017; Gawde et al., 2022; Liu et al., 2022; Musa, 2024). |
| iAMPpred (AMPredict) 2021 | SVM | Accuracy (74–88% depending on class) | It used physicochemical descriptors like hydrophobicity & charge. Predicted antimicrobial activity. Experimental Validation: Unclear (Donaghy et al., 2020; Norvell et al., 2021). |
| iAMP-2L 2013 | 2-layered FKNN with PseAAC | Overall accuracy (87.6%) | The multi-model approach predicted five function categories. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Xiao et al., 2013). |
| MLAMP (Multi-Label AMP predictor) 2016 | Ensemble of different ML, i.e., SMOTE & PseAAC | Micro-F1 (0.78) | Multi-label peptide functional models identified AMPs & their biological roles. Experimental Validation: No, computational validation only (Lin and Xu, 2016). |
| MLBP 2022 | Multi-scale ML alongside CNN & BiGRU | Accuracy (86%) | It processed raw sequential vectors without having LLM inputs. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Li et al., 2022; Tang et al., 2022). |
| PEPred-Suite 2019 | Different RF models | Accuracy (89%), AUC (0.92) | Using sequence-based descriptors & adaptive learning to predict efficacy & safety profiles. Experimental Validation: No, computational validation only (Wei et al., 2019). |
| Target-AMP 2022 | RF, SVM, KNNs | Accuracy (93.8%) | Used evolutionary data & composition features with multiple classifiers to predict AMP. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Jan et al., 2022). |
| b. Deep learning (CNN, GNN, RNN, Transformers, others) | |||
| AI4AMP 2021 | CNN, LSTM, DNNs | Accuracy (91.7%) | Used PC6 & autocovariance to predict novel AMPs. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Lin et al., 2021). |
| AMPlify 2022 | Bi-LSTM, Multi-head attention using Word2Vec tokens | AUROC (0.984), AUPRC (0.986). F1 (0.94) | An attention-based model to validate against the WHO priority pathogens & used ensemble learning to improve robustness. Experimental Validation: Yes, in-vitro validation (Li et al., 2022; Li et al., 2023). |
| AMPpred-CNN 2023 | 1D CNNs | Accuracy (92%) | Peptide sequences were encoded using CTD descriptors to predict AMPs. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Mitra and Singh, 2023). |
| AMP Scanner AMP Scanner v2 2018 | CNN, RNN, deep neural network | Accuracy (92%, MCC (0.85) | Predicted efficacy based on physicochemical properties. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Veltri et al., 2018; ). |
| APIN 2019 | CNN with embedded layers | Accuracy (94%) | Predicted AMPs using convolutional architectures directly from sequences. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Su et al., 2019). |
| CG-AMP 2025 | ProtT5 pre-trained protein LM embeddings. Supervised contrastive learning. Gated-CNN with Gated Linear Units (GLU) for AMP classification | AMPlify benchmark: Accuracy (0.9497) F1 (0.9508). DAMP benchmark: Accuracy (0.9403) F1-score (0.9392) | Discriminative AMP representations using contrastive learning, integrated sequence semantics, evolutionary information and physicochemical properties for classification. Experimental Validation: No, computational validation only (Li et al., 2025). |
| deepAMPNet 2024 | Pretrained Bi-LSTM, GNNs, & structural graphs | AUROC (0.97) | Alongside structure and language-derived encodings, it predicted delivering accuracy and biological insights. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Zhao et al., 2024). |
| Deep-AmPEP30 2020 | CNN trained on PseKRAAC for shorter peptides | Accuracy (92.6%) | Optimized & predicted shorter peptides. Experimental Validation: No, computational validation only (Yan et al., 2020; Lobanov et al., 2024; Nguyen et al., 2024). |
| DMAMP 2024 | CNN, Residual CNN Blocks, PSSM | Accuracy (91.3%) | A multi-task predictive model used CNN–residual architecture & evolutionary features fusion for robust and accurate prediction. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Meng et al., 2024). |
| GAC-BiTCNN-AMP 2026 | Hybrid GANs, capsule networks, bidirectional Temporal CNN alongside UniRep, ProtTrans-T5, ESM-2 and PsePSSM-DCT | Accuracy (97.42%) F1 (97.83%) AUC (99.54%) MCC (0.923) | Predicted AMPs with latent-feature analysis. Used as a discriminative & generative-augmentation hybrid model. Experimental Validation: No, computational validation only (Ali et al., 2026). |
| HDM-AMP 2021 | ESM-1b, Deep Forest | Accuracy (89.5%) | It was interpreted and predicted using an ensemble of decision trees. Experimental Validation: No, computational validation only (Yu et al., 2021). |
| iAMPCN 2023 | CNN | Accuracy (93.4%) | It used handcrafted features & classifies using a CNN without LLM embeddings. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Xu et al., 2023; Han et al., 2025). |
| iAMP-CA2L | CNN, Bi-LSTM, & SVM | Accuracy (91.2%) | It used hybrid SVM models for final classification with a dual-task focus. Experimental Validation: No, computational validation only (Xiao et al., 2021; Ruffolo and Madani, 2024). |
| LMPred 2021 | ProtTrans, CNN classifiers, T5, & XLNet | Accuracy (92–93%) | Bridged protein sequence understanding & learned pattern detection. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Dee, 2022). |
| MBC-attention 2023 | CNN, Attention, ML | MCC (0.81) | It focused on critical residual motifs causing membrane disruption using attention models. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Yan et al., 2023; Dong et al., 2025a). |
| sAMPpred-GAT 2022 | Graph Neural Network (GAT) & ML | AUROC (0.95) | Graph attention networks leverage sequence-to-graph conversion. Experimental Validation: No, computational validation only (Lv et al., 2022). |
| SGAC 2026 | Spatial GNN-based AMP classifiers used Cα-based peptide graphs | F1 (94.83 0.77) MCC (93.02 0.56) | Predicted AMPs from peptide sequences alongside inferred/predicted 3D structural information. Highly relevant for metagenomic AMP mining. Experimental Validation: No, computational validation only (Wang et al., 2026). |
| TP-LMMSG 2024 | Deep learning GNN on proteins, LM residues | Accuracy (94%) | Protein-LM, based on each residue, provided multi-scaled structural information. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Chen et al., 2024). |
| c. Ensemble models with hybrid frameworks | |||
| AMP-BRET 2023 | RF, ProtBERT Transformer with fine-tuning for regression | Accuracy (92.1%) | Demonstrated high precision & transfer learning from protein corpora. Predicted antimicrobial activity. Experimental Validation: Unclear (Lee et al., 2023). |
| AMPpred-EL 2022 | Ensemble ML (Logistic Regression & LightGBM) | Accuracy (93.8%) | It combined multiple ML components for stronger AMPs prediction. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Lv et al., 2022). |
| AMPpredMFA 2024 | LSTM, CNN, Attention, & MLP | AUROC (0.97) | It integrated local (CNN) and long-range (LSTM) sequence features to predict features. Experimental Validation: No, computational validation only (Li et al., 2024). |
| AMP-META 2024 | Light GBM (LGBM) comprises different AMP tools | Accuracy (95%) | It predicted physicochemical descriptors using larger datasets. Predicted antimicrobial activity. Experimental Validation: Yes, in-vitro validation (Jorge et al., 2012; Tsai et al., 2024). |
| E-CLEAP 2024 | Ensemble of four MLP classifiers | Accuracy (97.3%) (AAC features), Accuracy 84.0% (PseAAC features), F1 (0.93) | Ensemble neural classifiers performed high-accuracy AMP classification. Predicted antimicrobial activity. Experimental Validation: No, computational validation only (Wang, 2024). |
| d. Transformers-based models | |||
| AMPTrans 2023 | LSTM, Transformer, RF, SVM with adaptive QSAR | Accuracy (93%) | QSAR-enabled designer facilitated sequence guidance & novelty. Experimental Validation: No, computational validation only (Mao et al., 2023). |
| AMP-ProtBERT 2025 | Fine-tuned ProtBERT | AUROC (0.98)5 (ProtBERT AMP classifier), Accuracy (94%) | iAMP-bert (ESM-2) pretrained models outperformed for AMPs prediction and antimicrobial activity. Experimental Validation: No, computational validation only (Lee et al., 2023; Abbasi et al., 2025; Georgoulis et al., 2025). |
| e. Specialized supportive models | |||
| ESKAPEE-MICpred 2024 | LSTM, CNN, & MLP | R² = 0.82(Species-specific MIC regression model | It used sequence-derived descriptors to predict activity. Experimental Validation: No, computational validation only (Sharma et al., 2024). |
| EnDL-HemoLyt 2024 | LSTM, CNN, & MLP | AUROC (0.97), MCC (0.80) | It optimized the therapeutic index by predicting haemolysis. Experimental Validation: No, computational validation only (Maasch et al., 2023). |
| StaBle-ABPpred 2022 | BiLSTM | AUROC (0.95); Accuracy (0.91), MCC (0.82), AUPRC (0.97) | It predicted peptides’ stability and activity using Word2Vec embeddings. Experimental Validation: No, computational validation only (Singh et al., 2022). |
| SMEP/SAMP 2023/2024 | LSTM, XGBoost | Accuracy (90%), F1 (0.89) | Used different libraries such as nonapeptide, heptapeptide, and octapeptide to predict efficacy & safety. Experimental Validation: No, computational validation only (Huang et al., 2023; Feng et al., 2024; ). |
| f. Data-centric ML | |||
| GMSC-mapper 2024 | Modified version of Prodigal, i.e., RF | Accuracy (89%) | It identified and annotated smaller proteins using microbial (meta)genomes. Experimental Validation: No, computational validation only (Duan et al., 2024; Neville and Lawley, 2025; ). |
| Macrel 2025 | RF with metagenomics mining, LP models | AUROC (0.97–0.99 depending on datasets) | It predicted sequence-derived descriptors using proteome, genome, & transcriptome to show efficacy & safety. Experimental Validation: Yes, validation reported for predicted AMPs (Gonzalez-de la Rosa et al., 2025; Wu et al., 2025). |
| g. Emerging automated ML models | |||
| AutoPeptideML 2024 | AutoML-based peptide classifier with evolutionary features | Accuracy (90%) | It automated feature engineering from evolutionary indices to predict antimicrobial activity. Experimental Validation: No, computational validation only (Fernández-Díaz et al., 2024). |
| APEX 2024 | RNN, ATT, & MLP | Accuracy (92%) | Deep learning aided by Molecular de-extinction to show in-vitro & in-vivo properties. Experimental Validation: Yes, in-vitro & in-vivo validation (Wan et al., 2024). |
| PrMFTP 2022 | Multiscale CNN, BiLSTM, Multiheaded self-attention approach | Accuracy (93%), F1 (0.92) | It combined architectural models, i.e., CNN, BiLSTM, and self-attention, which were not derived from LLM. Experimental Validation: No, computational validation only (Yan et al., 2022). |
| h. Other ML/LLM models | |||
| AMPActiPred 2024 | Multi-class ML functional | Accuracy (91%) | A three-stage framework employed peptide descriptors to capture compositional and physicochemical properties & activity. Experimental Validation: No, computational validation only (Yao et al., 2024). |
| AMP-Distillation 2026 | Knowledge-distillation frameworks for AMP prediction | NA | AMP classification for high-throughput screening. Experimental Validation: No, computational validation only (Khorramfard et al., 2026). |
| Ansari & White 2023/2024 | RNN, LSTM | AUROC (0.93) | Semi-supervised method predicted peptide efficacy 7 safety via positive-unlabelled learning (PU). Experimental Validation: No, computational validation only (Ansari and White, 2023; Ansari and White, 2024). |
| Biochemical-driven ML pipeline 2026 | RF classifier alongside a biochemical-knowledge-based ML for AMP screening | NA | Predicted potent AMPs using biochemical constraints. 18.5% of generated AMPs showed antimicrobial activity (MIC ≤10 μg/mL), and 38.9% had MIC ≤50 μg/mL. Experimental Validation: Yes, in-vitro validation only (Yang et al., 2026). |
| Capecchi et al. 2021 | RNN, GRU, SVM; MLP | AUC (0.95) | ML used DBAASP to predict short non-haemolytic & microbial activity of AMPs. Experimental Validation: Yes, in-vitro validation (Capecchi et al., 2021). |
| PepGraphormer 2026 | ESM-2, Graph Attention Network (GAT) | Matched state-of-the-art AMP predictors across seven benchmark datasets | Predicted AMP activity without requiring 3D structures & combined protein LLMs semantics with graph relational learning. Experimental Validation: No, computational validation (Lin et al., 2026). |
| Zhuang et al. 2024 | QSVM (Quantum Support Vector Machine) | Accuracy (94%) data-dependent | It predicted sequence-derived descriptors to predict safety profiles. Experimental Validation: No, computational validation (Zhuang et al., 2024). |
Different discriminative AMP classifiers (machine & deep learning approaches) to predict AMP mining and discovery.
Experimental validation referred to wet-lab testing of the model-predicted model. Computational cross-validation and retrospective database benchmarking were not considered experimental validation. “Unclear” indicated that experimental confirmation could not be established unequivocally from the cited study; N/A denoted databases or computational resources for which experimental validation is not directly applicable.
Apart from binary AMP classification, discriminative models have provided support for high-throughput screening of diverse datasets, i.e., metagenomics, proteomics, and other commercially available datasets (; ). Architectural Frameworks, i.e., RF -based AmPEP and gradient-boosting architectures (e.g., XGBoost, LightGBM), have prioritized novel AMPs, subsequently experimentally validated against clinically relevant pathogens (Escherichia coli and Staphylococcus aureus) (; ). Importantly, ensemble frameworks have integrated multiple AMP classifiers to limit individual model bias and have improved generalization by achieving predictive accuracies around 95% (; ). Altogether, these strategies illustrate the critical role of discriminative models for narrowing vast sequential spaces into experimentally tractable candidates for clinical translation (; ).
Recent studies have highlighted the scope of discriminative modelling by integrating DL architectures, i.e., convolutional neural networks (CNNs), recurrent neural networks (RNNs), and long short-term memory (LSTM) frameworks (; ). These models have captured local motifs and longer-range dependencies by analysing sequences as analogous to linguistic inputs, with improved representations for sequence–function relationships (; ). Deep classifiers, which are trained on larger AMP databases, have shown better performance in comparison to conventional feature-oriented models, having accuracies exceeding 90% and stronger correlations with in-vitro validated antimicrobial activities (; Zhang et al., 2023). Hybrid architectural frameworks combining physicochemical features alongside sequential embeddings have upgraded predicted robustness, especially supported by transfer learning outputs from large protein corpora (Wang, 2024; Wang et al., 2025c).
Transformer models represent greater evolutionary changes for discriminative AMP predictions, leveraging attention mechanistic approaches to capture global sequential contexts and complex residue–residue interactions (Rives et al., 2021; Lee et al., 2023; Wang et al., 2025c). Fine-tuned protein-based LLMs, BERT-derived architectures, have given better outcomes for both AMP classification and regression tasks by predicting MICs (minimum inhibitory concentrations) against resistant pathogens (Stokes et al., 2020; Rives et al., 2021; Ma et al., 2022). Collectively, these models have reduced dependency on handcrafted features by learning directly from raw sequential datasets, enabling screening of datasets containing millions of candidates (Stokes et al., 2020; Rives et al., 2021; Ma et al., 2022). However, their performances remain contingent on the availability of larger, diverse training datasets with careful validations to mitigate bias and overfitting (Stokes et al., 2020; Rives et al., 2021; Ma et al., 2022).
Despite having higher predictive accuracy, discriminative AMP classifiers have key limitations for translation (Stokes et al., 2020). Most models are trained on overlapping benchmark datasets, limiting their generalizability (Stokes et al., 2020). Furthermore, classification accuracy is not sufficient to capture clinical endpoints, such as stability, potency, and toxicity (; ; Zhang et al., 2025). Consequently, discriminative models are considered essential for AMP discovery and rapid filtering to prioritize candidates, but they must be incorporated alongside multi-purpose, AI-driven pipelines for downstream predictions of pharmacological stability, efficacy, and safety profiles (; ; Zhang et al., 2025). This multifaceted integration can mark necessary progressions from isolated predictive models toward holistic and AI-based frameworks for AMP design and prediction (; ; Zhang et al., 2025). Despite reported higher classification metrics, direct comparison amongst discriminative models remained constrained by differences in dataset provenance, validation strategies and negative-set construction; these limitations are critically discussed in section 1.3.3.
1.2.2 Generative AI models for de novo AMP design
Discriminative models are used to score and prioritize pre-existing peptide sequences, while generative frameworks demonstrate fundamental shifts toward de novo design and generation of entirely new AMPs by learning underlying sequential distributions (Lin et al., 2023; Wang et al., 2024a; ). Deep generative architectural frameworks, such as diffusion models, generative adversarial networks (GANs), large language models (LLMs), and variational autoencoders (VAEs), have enabled exploration of previously inaccessible and vast sequential spaces (Van Oort et al., 2021; Lin et al., 2023; ; Guan et al., 2025) (Figure 6, Table 2). These models are highly effective for addressing the traditional discovery pipelines’ limitations by incorporating multi-purpose outputs, such as stability, potency, and limited toxicity, directly within the design processes.
Figure 6
Table 2
| Generative methods for AMPs discovery | |||
|---|---|---|---|
| Models & year | Architectural frameworks | Control generation | Key features & experimental validations |
| a. GAN architectural frameworks | |||
| AMP-GAN 2021 | GAN (Generator & Discriminator) | Random latent vectors without explicit conditioning | Antimicrobial & cytotoxicity assays. Experimental Validation: Yes, in-vitro validation (Van Oort et al., 2021). |
| AMPGAN v2 2021 | BiCGAN (Bidirectional Conditional GAN) | Binary vectors for targeting microbes & mechanisms | Metrics: Validity (95%), Novelty (94%), Uniqueness (100%). Experimental Validation: No, computational validation (Van Oort et al., 2021). |
| dsAMP & dsAMPGAN 2024 | CNN Attention, BiLSTM, transfer learning models | AMP prediction. Metrics: Accuracy (95%), F1 (0.94). Experimental Validation: No, computational validation (). | |
| FBGAN 2024 | GAN, & ESM-2 | Controlled & conditioned generation | Antimicrobial, haemolytic and cytotoxicity assays. Metrics: AUROC (0.92).Experimental Validation: Unclear (Chung et al., 2024; ). |
| GAC-BiTCNN-AMP 2026 | GAN alongside capsule networks, bidirectional temporal CNNs with ProtTrans-T5, ESM-2 embeddings & UniRep. PsePSSM-DCT & SHAP | Primarily generative augmentations alongside AMP predictions | Accuracy 97.42%, MCC 0.92 in cross-validation alongside independent accuracy 95.32% and MCC 0.91. Experimental Validation: No, computational validation (Ali et al., 2026). |
| WGAN-GP 2023 | WGAN-GP | AI4AMP & classifiers for in-silico | Predicted novel peptides using methods like PC6, based on physicochemical properties. Experimental Validation: Unclear (Lin et al., 2023). |
| Multi-CGAN 2024 | cGAN | Conditional generation | Antimicrobial & cytotoxicity assays. Experimental Validation: No, computational validation (Yu et al., 2024). |
| b. VAE & latent spaces-based frameworks | |||
| CLaSS (Controlled Latent Attribute Space Sampling) 2023 | WAE | Discriminator-guiding filtering | In-vivo models using Antimicrobial, haemolytic and cytotoxicity assays. Metrics: Precision (90%) for desired sampling. Experimental Validation: Yes, in-vitro & in-vivo validation (Wang et al., 2023). |
| LSSAMP 2023/2025 | Vector quantized VAE | Latent space sampling | Predict Antimicrobial, haemolytic and cytotoxicity assays. Metrics: Accuracy (91.7%). Experimental Validation: Yes, in-vitro validation only (Wang et al., 2023; Yan et al., 2025). |
| Pathogen-targeted programmable AMPs 2025 | Conditional VAEs, conditional diffusion models & pathogen-specific MIC predictors | Predicted physicochemical properties, targeting E. coli & S. aureus | Pathogen-specific efficacy & identified two “star” AMPs per target species with predicted antibacterial activity, toxicity & haemolysis profiles. Experimental Validation (limited): Yes, in-vitro validation (Zhao et al., 2026). |
| PepVAE 2021 | VAE | Latent space sampling | Microbial activity. Metrics: Validity (>95%), Novelty (80%. Experimental Validation: Yes, in-vitro validation only (Dean and Walper, 2020). |
| c. Diffusion-based framework | |||
| AMP-Diffusion 2025/2024 | Structurally guided diffusion model | Positive learning, using discriminator-guiding filtering | Microbial & cytotoxicity assays. Sequence validity (97%). Experimental Validation: Yes, in-vivo validation (Torres et al., 2024; Luo et al., 2025). |
| Diff-AMPs 2024 | Diffusion | Discriminator-guiding filtering | Metrics: AUROC (0.94). Experimental Validation: No, computational validation (Wang et al., 2024). |
| ProDCARL 2026 (pre-print) | RL-aligned diffusion model having EvoDiff OA-DM 38M backbone | Predicted AMP activity and low toxicity using classifier-derived rewards | Mean predicted score (0.081 → 0.178), hit rate (6.3% using pAMP >0.7 and pTox <0.3), diversity retained with 1 − mean pairwise identity = 0.929. Experimental Validation: No, computational validation (Sheng et al., 2026). |
| ProT-Dif 2024 | Protein language diffusion | Condition generation with discriminator-guided filtering. Positive-only learning | De novo generation of novel AMP sequences. Metrics: Validity (98.3%), Novelty (99%).Experimental Validation: Yes, in-vitro validation only (Wang et al., 2024). |
| MMCD 2024 | Diffusion model based on discriminators | Conditional generation, contrastive learning | Predict microbial, haemolytic &cytotoxicity assays. Experimental Validation: No, computational validation (Wang et al., 2024b; Lai et al., 2025). |
| d. Multi-objective evolutionary or genetic optimization models | |||
| AMPEMO 2023/2025 | Genetic algorithm | Discriminator-guiding filtering | Antimicrobial activity. Experimental Validation: Unclear (Liu et al., 2023; Wang et al., 2025b). |
| MODAN 2023 | GAN & RL | Bayesian optimization | Antimicrobial, haemolytic assays. Multi-objective score improvement >30% over baseline. Experimental Validation: Unclear (Murakami et al., 2023). |
| MOQA Multi-CGAN QMO (Multi-Objective Quantum Annealing) 2024 | Binary VAE, Multi-generator CGAN | D-Wave quantum annealer, with conditional generation | In-vivo models to predict Antimicrobial, haemolytic and cytotoxicity assays. Experimental Validation: Unclear (Dongrui et al., 2024). |
| M3-CAD 2024 | cVAE | Conditional generation using discriminator-guided filtering | In-vivo models to predict Antimicrobial, haemolytic and cytotoxicity assays. Experimental Validation: Unclear (Wang et al., 2024a). |
| QMO 2022 | WAE | Zero-order optimization with gradient descent | Optimized materials, i.e., drug-likeness & solubility. Experimental Validation: No, computational validation (Hoffman et al., 2022). |
| e. Transformers or an RNN-based hybrid framework | |||
| AMPGen 2025 | Autoregressive diffusion model, XGBoost discriminator, & STM | MSA-conditional generation | Metrics: Validity (94%), Novelty (96%). Experimental Validation: Yes, in-vitro validation only (Jin et al., 2025). |
| AMPTrans-LSTM 2023 | LSTM & transformers | Learning using protein databases | Antimicrobial activity. Experimental Validation: No, computational validation (Mao et al., 2023). |
| CoLPAT-AMP 2026 | Transformer-based AMP generators | Property-aware | Designed for novel AMPs with desired properties. Experimental Validation: No, computational validation (Salimi and Lee, 2026). |
| HydrAMP 2023 | cVAE-GAN hybrid | Conditional generation | Antimicrobial, haemolytic assays. Metrics: AUROC (0.93). Experimental Validation: Yes, in-vitro validation (Szymczak et al., 2023). |
| f. Large language models (LLMs) | |||
| CFlowAMP 2026 | ESM-2 protein language model | Property-based generation in latent peptide spaces | Designed for property-controllable de novo AMP generation. Experimental Validation: No, computational validation (Chen et al., 2026). |
| ChatGPT prompt-oriented multifunctional AMPs 2026 | Prompt-engineered LLMs | Prompt-based AMPs having single-, dual-, triple- and quadruple-functional outputs | In-silico accuracy >0.9 for single, dual-functional AMPs and 0.8 for triple & quadruple designs. Experimental Validation: No, computational validation (Sowers et al., 2026). |
| MAC-AMP 2026 | Closed-loop multi-agent LLMs | Multi-objective activity, AMP-likeness, toxicity, novelty, and structure | Autonomous task-to-design frameworks, and reported AMP generators across activity, toxicity, novelty, and structural reliability. Experimental Validation: No, computational validation (Zhou et al., 2026). |
| PepGen-FB 2026 | Feedback-driven generative frameworks alongside curriculum learning-guided multi-purpose optimization | Multi-faceted AMP generation alongside feedback optimization | Data-centric, feedback-driven frameworks for multi-purpose AMP generation for activity, safety & functional optimization. Experimental Validation: No, computational validation (Xiao et al., 2026). |
| g. Other generative models | |||
| Buehler et al. 2023 | GNN | Conditional generation | Physicochemical properties prediction. Experimental Validation: No, computational validation (Buehler, 2023). |
| Cao et al. 2020/2021 | GAN | Discriminator-guiding filtering | Antimicrobial activity. Experimental Validation: Yes, in-vitro validation (Cao et al., 2023). |
| Capecchi et al. 2023 | RNN | Positive learning, using discriminator-guided filtering | Antimicrobial and haemolysis assays. Experimental Validation: Yes, in-vitro validation (Capecchi et al., 2021). |
| Dean et al. 2021 | VAE | Latent space sampling | Antimicrobial activity. Experimental Validation: Yes, in-vitro validation (Dean and Walper, 2020). |
| Ghorbani et al. 2022 (pre-print) | VAE | Standard VAE without variational attention | AMP prediction. Metrics: AUROC 0.90.Experimental Validation: No, computational validation (Ghorbani et al., 2022). |
| Jain et al. 2022 | GFlowNets & active learning | Active learning | AMP prediction. Experimental Validation: No, computational validation (Jain et al., 2022). |
| Pandi et al. 2023 | VAE | Discriminator-guiding filtering | Antimicrobial, haemolytic and cytotoxicity assays. Metrics: Validity 96%. Experimental Validation: Yes, in-vitro validation only (Pandi et al., 2023). |
| Renaud & Mansbach 2023 | VAE | Latent space sampling | Physicochemical properties prediction. Experimental Validation: No, computational validation (Renaud and Mansbach, 2023). |
| Zeng et al. 2024 (pre-print) | PLM & BERT | Discriminator-guiding filtering | Metrics: Accuracy 92%. Experimental Validation: No, computational validation (Zeng et al., 2024). |
Generative-AI models for de novo design and generation of novel AMPs.
Amongst these, GAN-based methods have good capabilities for the de novo generation of structurally and functionally novel and active AMPs (Dong et al., 2025b; Lai et al., 2025). Bidirectional and conditional GAN architectural frameworks have enabled iterative feedback loops between generators and discriminators, incorporating auxiliary classifiers and regressors to implement physicochemical constraints alongside antimicrobial properties (Vasconcelos et al., 2024; ; Dong et al., 2025b; Lai et al., 2025). These approaches have produced novel AMPs having broad-spectrum activity against multidrug-resistant pathogens exhibiting lower MICs and minimal haemolytic activity (Vasconcelos et al., 2024; ; Lai et al., 2025). However, GAN models are sensitive to training instability and usually require careful fine-tuning, necessitating incorporation alongside downstream predictive models for guiding sequential optimization (; ; Lai et al., 2025).
Variational autoencoders (VAEs) have allowed complementary approaches by encoding sequences into continual latent spaces, enabling controlled interpolations for property-guided AMP design (Zhao et al., 2025). Conditional VAEs can incorporate physicochemical features, i.e., charge, molecular weight, and hydrophobicity parameters, into the design and generation processes, with targeted structural and functional profiles (Zhao et al., 2025). VAEs have provided interpretable and stable latent frameworks compared to GANs; their outputs usually require filtering and validation to ensure their biological relevance (Xie et al., 2024; Zhao et al., 2025). Diffusion-based approaches have recently evolved as stronger generative tools, iteratively processing random noises into biologically relevant AMPs (Wang et al., 2024; Wang et al., 2024).
These models, in combination with pretrained protein-based LLM embeddings, have demonstrated higher success rates for generating potent and diverse AMPs with better in-vivo efficacy and safety profiles (Wang et al., 2024; Wang et al., 2024). Generative performance metrics must be interpreted cautiously because computational outputs such as validity, novelty and predicted bioactivity cannot establish biological and translational generalizability, especially when reward functions and discriminative filters inherit overlapping available AMP datasets; these concerns are thoroughly discussed further in section 1.3.3.
Collectively, generative AI frameworks have enabled the transition from passive screening to active mining and optimization of sequential spaces, accelerating the novel AMPs discovery (Lai et al., 2025; Lai et al., 2025). However, the translation impacts depend on their incorporation of discriminative inputs, structural verification, and Iterative experimentation (Lai et al., 2025; Lai et al., 2025).
Most efficient strategies strongly rely on multi-model and hybrid pipelines, where generative AI models de novo design and generate non-natural and novel AMPs, learn distributions within peptide sequential spaces and propose previously unseen sequences, thereby shifting computational discovery towards de novo AMP designs (Lai et al., 2025). Generative models are increasingly being combined alongside hybrid frameworks, where AMPs are evaluated using structural, novelty, bioactivity, stability and toxicity predictors (Lai et al., 2025) (Table 3). Therefore, hybrid approaches can offer attractive routes towards multi-purpose AMP optimization, although they require independent validations, model calibration, followed by experimentally validated feedback (Lai et al., 2025) (Table 3).
Table 3
| AI paradigm | key strengths | Key limitations | AMP applications | Role within multimodel pipelines |
|---|---|---|---|---|
| Discriminative models | High-throughput ranking and classification of pre-existing sequences (; ). It requires straightforward training and benchmarking, which exploits evolutionary, physicochemical, and structural representations. These models are suitable for binary, multilabel and multiclass regression tasks, which are computationally efficient once trained (; ). | These models are inherently unable to explore sequential spaces beyond supplied candidates and are dependent on training-set quality and negative-set construction (; ; Zhang et al., 2025). They are susceptible to sequence redundancy, class imbalance and data leakage; therefore, their performance may deteriorate on taxonomically and structurally distant AMPs (; ; Zhang et al., 2025). | Binary AMP classification; stability/protease susceptibility prediction; MIC estimation; cytotoxicity and haemolysis screening (; ). | Decision and screening layer, which helps to filter larger datasets and assigns safety, bioactivity, potency, and developability scores before wet-lab evaluation (; ). |
| Generative models | Generate de novo AMPs beyond known libraries and explore larger sequence spaces and permit property-aware generation (Van Oort et al., 2021; Ghorbani et al., 2022; Lin et al., 2023; ; Guan et al., 2025). They also enable sequential diversification and scaffold redesigns (Van Oort et al., 2021; Ghorbani et al., 2022; Lin et al., 2023; ; Guan et al., 2025). Latest diffusion and LLMs learn complex sequence distributions without explicating handcrafted rules (Van Oort et al., 2021; Ghorbani et al., 2022; Lin et al., 2023; ; Guan et al., 2025). | Generated sequences might not be active, safe, and novel (Van Oort et al., 2021; Ghorbani et al., 2022; Lin et al., 2023; ; Guan et al., 2025). These models are prone to training-data bias, mode collapse or memorization; property control can be imperfect and computational metrics frequently exceed the depth of wet-lab validation (Van Oort et al., 2021; Ghorbani et al., 2022; Lin et al., 2023; ; Guan et al., 2025). | De novo generation and diversification of known AMP scaffolds. Conditional AMP generation based on charge, length and bioactivity (pathogen-directed designs), generation of broad-spectrum candidates by exploring previously untested sequence spaces (Van Oort et al., 2021; Ghorbani et al., 2022; Lin et al., 2023; ; Guan et al., 2025). | The exploration layer will help to propose new AMPs that are passed through filters for novelty, structure, bioactivity, cytotoxicity, and stability (Van Oort et al., 2021; Ghorbani et al., 2022; Lin et al., 2023; ; Guan et al., 2025). |
| Hybrid AI approaches | These approaches combine the exploratory capacity of generative AI with the selectivity of discriminative prediction, enabling simultaneous optimisation (e.g., novelty, stability, potency, toxicity) and reducing the number of generated AMPs that require experimental testing, thereby best aligning with end-to-end precise AMP engineering (; ). | These are computationally and architecturally complex, and prediction errors can propagate between modules (; ). Performance metrics depend on compatibility and calibration of individual models. Moreover, closed-loop optimization can exploit model biases, and prospective wet-lab validation remains limited (; ). | Suitable for multi-purpose AMP designing. Activity–toxicity trade-off optimization by adapting pathogen-specific AMP engineering (; ). Optimization of stability and selectivity by iteratively design–predict–refine workflows (; ). | Introduction of an integration and optimization layer will link sequential generation, discriminative screening, and structural evaluations into iterative decision frameworks before biological validation (; ). |
Comparative characteristics of discriminative, generative and hybrid AI approaches for AMP discovery and optimization.
On the contrary, discriminative models assess whether predicted peptides possess antimicrobial properties from pre-existing resources and learn quantitative relationships between predefined biological endpoints and peptide representations, therefore suitable for classification, regression and AMP prioritization (Lai et al., 2025) (Table 3). Convergence of both generative and predictive paradigms demonstrates critical steps toward precision-driven AMPs design and generation to address the accelerating global AMR burden (; ; Wang et al., 2025; Zervou et al., 2025).
1.2.3 Dataset overlapping, negative-set construction and over-generalization
One of the key limitations associated with discriminative models; independent predictors are developed and evaluated using overlapping databases (Agüero-Chapin et al., 2023). Mostly, positive AMPs are repeatedly extracted from commercially available repositories, followed by diverse combinations of sequence-length filtering, redundancy reduction and train–test partitioning (Agüero-Chapin et al., 2023). Consequently, comparisons amongst models cannot be interpreted as evaluations based on independent biological information (Agüero-Chapin et al., 2023).
Importantly, random splitting places identical and evolutionarily related AMPs across test and training partitions, exploiting sequential similarities rather than learning determinants of bioactivity (Zhao et al., 2026). They can inflate performance metrics, explaining why models report excellent internal benchmark performances but can generalize less effectively to AMPs from previously unexplored families and ecological environments (Zhao et al., 2026). Thus, complex architectures must not be interpreted as biologically superior when sequential overlapping, benchmark provenance, and homology between evaluation and training datasets have not been controlled (Wang et al., 2025).
Negative class has represented one of the equally important sources of bias (Yang and Xu, 2024). Experimentally validated inactive AMPs are scarce because unsuccessful AMPs have not been reported systematically (Yang and Xu, 2024). Fewer AMP classifiers have been constructed using negative datasets from computationally convenient and reliable sources (Yang and Xu, 2024). Absence of AMP annotations does not support experimental evidence of microbial inactivity. If artificially negative outputs differ from positive AMPs depending on amino-acid composition, length, charge, and cellular origin, AMP classifiers can learn these dataset-construction artefacts rather than biologically active determinants (Zhao et al., 2026). These concerns have been extended to specialized safety classifiers, where incomplete non-haemolytic sequences and heterogeneous assay conditions can influence labels, thereby apparently affecting model performance (Yang and Xu, 2024).
Performance metrics across AMP prediction models must not be interpreted as directly comparable measures for superiority (Agüero-Chapin et al., 2023). Although several studies have reported F1-score, accuracy, AUROC, and MCC values exceeding 90%, these outcomes are derived from heterogeneous datasets, sequence-redundancy thresholds, negative-set definitions, and validation strategies (Zhao et al., 2026). Moreover, these performance metrics capture different aspects of predictive behaviour; therefore, they cannot be ranked interchangeably across studies. Accordingly, the performance metrics summarized in Table 1–9 must be interpreted as within-study evidence of model performance only. Direct superiority claims can be justified when competing models have been evaluated on the same independent datasets using redundancy control, identical preprocessing, decision thresholds and evaluation metrics (Zhao et al., 2026).
Table 4
| Models & year | Architectural framework | Performance metrics | Key findings & experimental-based evidence |
|---|---|---|---|
| AMPDeep 2022 | Transfer learning & DL with protein LLM. | Accuracy =90%, AUC =0.94 | Crucial step toward representation learning in AMP safety & used pretrained embeddings. Experimental Validation: No, computational validation (Salem et al., 2022). |
| CNN-based AMP haemolysis predictor 2024 | DL with CNNs. | Accuracy =88–92% | Reinforced the end-to-end sequence models to improve haemolysis prediction. Experimental Validation: No, computational validation (Abdelbaky et al., 2024). |
| ConsAMPHemo 2025 | Classical & hybrid workflows with binary classification alongside concentration-regression frameworks. | AUC =0.95 | Both classification and haemolysis concentration prediction. Experimental Validation: No, computational validation (Xie et al., 2025). |
| Enhanced DL haemolysis predictor 2024 | DL sequence-analysis frameworks using learned peptide representations. | Accuracy 94.60%, Precision 94.10%, Recall 95.20%, F1-score 94.65%, ROC-AUC 0.973 | No new peptide synthesis reported. Experimental Validation: No, computational validation (Abdelbaky et al., 2024). |
| HemoPred 2017 | Classical ML webserver. | Accuracy=95%, MCC reported | Early benchmark haemolysis classifier. Experimental Validation: No, computational validation (Win et al., 2017). |
| HLPpred-Fuse 2020 | Multi-feature fusion ML. | AUC =0.94 | Improved robustness over earlier single-model tools. Experimental Validation: No, computational validation (Hasan et al., 2020). |
| HAPPENN 2020 | DL with neural-network classifiers. | Accuracy =85–90% | Strengthened sequence-only deep classification. Experimental Validation: No, computational validation (Timmons and Hewage, 2020). |
| Haemolytic potency prediction for chemically modified peptides 2020 | Classical ML, built from 2D/3D descriptors and fingerprints. | Quantitative HC50 predictions. | Prediction toward potency-aware haemolysis modelling in modified peptide spaces. Experimental Validation: No, computational validation (Kumar et al., 2020). |
| HemoNet 2021 | DL with CNNs & RNNs. | AUC =0.93 | Reported a stronger AUC-ROC than HemoPI/HemoPred. Experimental Validation: No, computational validation (Kumar et al., 2020). |
| HemPepPred 2025 | Ensemble ML & PLM-derived features, using ML with protein-LLM. | State-of-the-art performances. | Next-generation trend of combining PLM embeddings. Experimental Validation: No, computational validation (Li et al., 2025). |
| HemoDL 2024 | Ensemble DL with a double-ensemble learning strategy. | AUC =0.95 | Engineered ensemble designs even in the deep-learning era. Experimental Validation: No, computational validation (Yang and Xu, 2024). |
| HemoFuse 2024 | Multi-feature fusion networks with multi-head cross-attention. | Accuracy 95.33%, Sensitivity 94.00%, Specificity 96.67%, MCC 0.907, AUC 0.982 | No direct wet-lab validation reported. Experimental Validation: No, computational validation (Zhao et al., 2024). |
| LLM toxicity prediction for AMPs 2024 | LLMs, transformers & GPT prompting | Emerging benchmark study. | General-purpose LLMs for peptide toxicity prediction. Experimental Validation: No, computational validation (Orsi and Reymond, 2024). |
| ML-guided non-haemolytic peptide discovery 2020 | ML & design workflow. | Experimental validation | Linked prediction to prospective peptide design. Experimental Validation: No, computational validation (Plisson et al., 2020). |
| Quantitative regression model 2024 | Regression ML, predicting hemolytic concentrations. | RMSE/R² reported | Moved beyond binary labels toward quantitative haemolysis modelling. Experimental Validation: No, computational validation (Karasev et al., 2024). |
| Prediction of hemolytic peptides and HC50 2025 | Hybrid ML, regression plus classification for HC50 prediction. | Classification alongside HC50 | HC50-level quantitative prediction. Experimental Validation: No, computational validation (Rathore et al., 2025). |
| tAMPer 2024 | Transformer-based safety model | High AUC reported. | Structure-aware peptide toxicity prediction to improve toxicity estimation. Experimental Validation: No, computational validation (Ebrahimikondori et al., 2024). |
| ToxGIN 2024 | Graphs DL. | AUC =0.94 | General peptide toxicity model. Experimental Validation: No, computational validation (Yu et al., 2024). |
In-silico haemolytic predicted models that assess the safety of novel AMPs.
Table 5
| Models & year | Architectural framework | Performance metrics | Key findings & Prospective experimental validation |
|---|---|---|---|
| ATSE 2021 | DL (GNN, attention), structural, evolutionary & sequence integrations. | Accuracy=93.7%, MCC=0.87 AUC=0.96 | Enhanced prediction using graph-based representations. Experimental Validation: Computational benchmarking only (Wei et al., 2021). |
| ClanTox 2019 | ML Sequence-based classifiers. | AUC =0.90 Accuracy=85–90% | Applied to animal toxin classification. Experimental Validation: Computational benchmarking only (Cole and Brewer, 2019). |
| DeepTox 2016–ongoing | DL, multilayer neural networks. | ROC-AUC=0.84 | Widely validated for chemical toxicity. Experimental Validation: Computational benchmarking only (Mayr et al., 2016). |
| LLM toxicity prediction for AMPs 2024 | GPT & LLM prompting. | Competitive with classical ML | GPT-3 performed not reproducibly for activity prediction and haemolysis, taken as a proxy for toxicity. Experimental Validation: Computational benchmarking only (Orsi and Reymond, 2024). |
| NNTox 2019 | ML (Neural Network) with Gene Ontology and functional annotations. | Accuracy=84.5% AUC=0.91 | Improved functional toxicity prediction. Experimental Validation: Computational benchmarking only (Jain and Kihara, 2019). |
| ProTox-II 2018–ongoing | ML, Descriptor-based toxicity predictions. | Accuracy=76% ROC-AUC=0.84 | Extensively validated for drug candidates. Experimental Validation: Computational benchmarking only (Banerjee et al., 2018). |
| Structure-aware peptide toxicity model/tAMPer 2024 | DL & structure-aware embeddings | AUC=0.94 | Experimental Validation: In-vitro validation using pig RBCs (Ebrahimikondori et al., 2024). |
| ToxinPred 2013 | ML (SVM) AAC, dipeptide composition, motif-based features. | Accuracy =89.5%, MCC =0.80 AUC=0.94 | Widely validated on peptide datasets; foundational toxicity predictor. Experimental Validation: Computational benchmarking only (Gupta et al., 2013). |
| ToxClassifier 2016 | ML, feature-based classifiers for venom toxins | Accuracy =95% | Differentiates venom vs non-toxic proteins. Experimental Validation: Computational benchmarking only (Gacesa et al., 2016). |
| TOXIFY 2019 | DL, sequence-based neural network | Accuracy=95.0% F1-score=0.93 AUC=0.97 | Validated on venom protein datasets. Experimental Validation: Computational benchmarking only (Cole and Brewer, 2019). |
| ToxDL 2020 | DL, sequence & domain embeddings (DL architecture) | ROC-AUC=0.83–0.88 | Demonstrated improved classification of protein toxicity. Experimental Validation: Computational benchmarking only (Villegas-Morcillo et al., 2020). |
| ToxinPred2 2022 | Hybrid ML, BLAST, motifs & ML ensemble (RF, SVM, XGBoost) | Accuracy=96.5% MCC=0.93 AUC=0.98 | AUC=0.99; validated on large SwissProt datasets. Experimental Validation: Computational benchmarking only (Sharma et al., 2022). |
In-silico models that assess the cytotoxicity profiles of predicted novel AMPs.
Table 6
| Models & year | Architectural framework | Performance metrics | Key findings & experimental evidence |
|---|---|---|---|
| ANIA 2026 | Inception & attention MIC predictions. | High MIC prediction accuracy. | Not reported. Experimental Validation: Computational benchmarking only (Chiu et al., 2026). |
| Amino acid–based activity prediction 2014 | Classical ML with residual contribution scorings & sequence optimization. | MIC hit enrichment reported. | Experimental Validation: In-vitro & in-vivo validations (Wu et al., 2014). |
| AMP regression benchmarking 2021 | ML + DL comparison CNN, RF, XGBoost, Elastic Net for MIC predictions. | R²=0.45–0.72 RMSE improved over baselines. | Demonstrated GB/XGBoost superiority for MIC regression. Experimental Validation: In-vitro validation (Dean et al., 2021). |
| AMPredictor & GAN pipeline 2025 | Generative; graph DL GAN generator & GCN MIC regressors. | RMSE 0.5348 Higher hit rates. | Experimental validation: 24 peptides validated via in-vivo efficacy (Dong et al., 2025b). |
| BERT-AmPEP60 2025 | Transformers with fine-tuned ProtBERT & regression heads. | Accuracy=95%, AUC=0.97 MSE=0.266 (E. coli) Pearson=0.796 | Strong generalization. Experimental Validation: Computational benchmarking only (Cai et al., 2025). |
| BroadAMP-GPT 2025 | Foundation & GPT model | Broad-spectrum activity. | Predicted antimicrobial activity. Experimental validation: In-vitro/potentially in-vivo, requires paper-specific qualification (). |
| CNN-based AMP activity prediction 2024 | DL with CNN-based sequence & structural features. | Accuracy=92–95%, AUC=0.95 | Improved AMP activity prediction vs ML baselines. Experimental Validation: Unclear (Mitra and Singh, 2023). |
| DeepPIE/ProtBERT-based MIC frameworks 2024–2025 | Transformer DL with PLM embeddings & neural regression layers. | R²=0.70–0.82 MAE improved vs classical QSAR | Improved MIC prediction using contextual embeddings. Experimental Validation: Unclear (Wang et al., 2025). |
| ESKAPEE-MICpred 2024 | Transfer learning & ensemble pathogen-aware predictions integrating sequence & genomic features | RMSE low Pearson r=>0.80 | Designed for WHO-priority pathogens; improved generalization. Experimental Validation: Computational benchmarking only (Sharma et al., 2024). |
| Ensemble MIC predictor 2024 | Deep ensemble with BiLSTM, CNN, multi-branch & genomic features. | AUC=0.96 R²=0.79 Pearson 0.756–0.802 | Predicted antimicrobial activity. Experimental Validation: Computational benchmarking only (Chung et al., 2024). |
| LLM foundation AMP model 2025 | LLM peptide generators. | Potent candidates against resistant pathogens | De novo designing, generation and antimicrobial activity prediction. Experimental Validation: In-vitro & in-vivo validations (). |
| MBC-Attention 2023 | DL with multi-branch CNN & attention mechanisms. | R²=0.84 PCC=0.92 Pearson 0.775 RMSE 0.533 | E. coli MIC prediction. Experimental Validation: Computational benchmarking only (Yan et al., 2023). |
| PepVAE 2021 | Generative DL, ML regression, VAE & downstream regressors (GB, RF). | R2 = 0.73 RMSE=0.50 | Experimentally validate AMPs. Experimental Validation: In-vitro validation (Dean et al., 2021). |
| QSAR-based AMP activity models 2006–2010 | Classical ML with physicochemical descriptors & regression (QSAR). | R² 0.55–0.75 | Early quantitative activity predictions with limited datasets. Experimental Validation: Unclear (Rajakumar et al., 2006). |
In-silico MIC regression models to predict efficacy against drug-resistant pathogens.
Table 7
| Models & year | Architectural framework | Performance metrics | Key findings & experimental validation |
|---|---|---|---|
| BLAST/BLASTP/ tBLASTn 2006-2009 | Heuristic local alignment with seed-and-extend architecture. Protein-to-translated-nucleotide local alignment across six frames. | High sensitivity with standard benchmark references & faster than Smith-Waterman with strong recall. | Foundational sequence-similarity baseline with sensitive protein novelty checks. Improved translated-search reliability via composition-based statistics & suited to genomic novelty mapping. Experimental Validation: Computational benchmarking only (Gertz et al., 2006; Camacho et al., 2009). |
| CD-HIT 2012 | Greedy incremental clustering with word-count filtering. | More than 100 times faster than all clustering. | Fast redundancy reduction & useful to collapse near-identical candidates before novelty analysis. Experimental Validation: Computational benchmarking only (Fu et al., 2012). |
| DIAMOND v2 2021 | Double indexing, spaced seeds, vectorized extension, HPC-distributed protein alignments. | Up to 20,000× faster than BLASTX. | BLAST-like sensitivity with 80-fold BLASTP speedup in ultra-sensitive modes & 18 h nr-vs-UniRef50 on 520 nodes. Experimental Validation: Computational benchmarking only (Buchfink et al., 2021). |
| Foldseek 2024 | Fast structural search using structure-to-sequence style encoding. | >10,000× faster than structural aligners, high sensitivity comparable to TM-align/DALI class tools. | Enabled structural novelty screening when sequence similarity was weak, but fold similarity remained high. Experimental Validation: Computational benchmarking only (van Kempen et al., 2024). |
| HMMER 2011 | Profile-HMM–based remote homology search. | Higher remote-homology sensitivity than pairwise alignment with trusted E-value frameworks. | Widely used for remote family assignment beyond pairwise sequence similarity. Experimental Validation: Computational benchmarking only (Finn et al., 2011). |
| InterProScan 5 2014 | Integrated protein-signature and domain-scanning frameworks. | Broad functional annotation coverage with millions of sequences. | Genome-scale functional & domain classification helped reject candidates from known families. Experimental Validation: Computational benchmarking only (Jones et al., 2014). |
| MMseqs2 2017 | Sensitive, indexed protein search optimized for massive datasets. | 100–400× faster than BLAST, high sensitivity close to PSI-BLAST. | Reported sensitivity better than PSI-BLAST at >400× its speed. Experimental Validation: Computational benchmarking only (Steinegger and Söding, 2017). |
In-silico novelty filters that assess novelty at the genomic and protein-sequence levels.
Table 8
| Models & year | Architectural framework | Performance metrics | Key findings & experimental validation |
|---|---|---|---|
| DeepCleave 2020 | Deep learning protease cleavage prediction. | AUC=0.93 | Outperformed in predicting caspase and matrix metalloprotease substrate-cleavage sites. Experimental Validation: Computational benchmarking only (Li et al., 2020a). |
| MPCutter 2025/26 | Multi-protease cleavage prediction model. | Outperformed existing general tools on independent tests. | High-throughput prediction of protease-specific substrates and exploration of protease proteolytic events. Experimental Validation: Computational benchmarking only (Wang et al., 2025). |
| PROSPER 2012 | ML has an integrated feature-based cleavage-site prediction using sequence-derived features across multiple proteases. | AUC=0.84–0.96 Accuracy =80–90% | Reported comprehensive multi-protease cleavage prediction & established ML-based substrate prediction as a practical workflow. Experimental Validation: Computational benchmarking only (Song et al., 2012). |
| PROSPERous 2018 | ML has a logistic-regression-based high-throughput prediction using multiple scoring functions and pairwise feature combinations for 90 proteases. | AUC=0.87–0.98 MCC improved over PROSPER. | Improved accuracy and broader protease coverage than earlier tools & state-of-the-art high-throughput cleavage prediction at publication. Experimental Validation: Computational benchmarking only (Song et al., 2018). |
| ProsperousPlus 2023 | Ensemble ML & AutoML cleavage prediction platform. | State-of-the-art prediction across 110 protease types | Predicted substrate cleavage sites & facilitated the use of the rapidly accumulating substrate cleavage data to train, empirically assess and deploy predictive models. Experimental Validation: Computational benchmarking only (Li et al., 2023). |
| Procleave 2020 | ML with protease-specific prediction combining sequence & structural information with LOWESS-smoothed structural determinants. | AUC=0.90 Accuracy=85–92% | Demonstrated improved cleavage-site prediction by incorporating 3D structure, which is directly relevant to peptide stability design. Experimental Validation: Computational benchmarking only (Li et al., 2020b). |
| PlifePred 2018 | ML has an SVM-based peptide blood half-life prediction using chemical descriptors from experimentally measured natural and modified peptides. | MCC=0.80, Accuracy =85% AUC =0.91 | Built from peptides with experimentally determined half-lives in mammalian blood & useful for pre-screening pharmacological persistence. Experimental Validation: Computational benchmarking only (Mathur et al., 2018). |
| panCleave 2023 | ML with a pan-protease random forest model for computational proteolysis and encrypted-peptide discovery. | AUC=0.90 | Used to mine encrypted antimicrobial fragments with downstream peptides that showed antimicrobial activity with experimental and mouse-model validation in the discovery workflow. Experimental Validation: In-vitro & in-vivo validations (Maasch et al., 2023). |
| PeptideCutter 2025 | Rule-based cleavage-site mapping for proteases and chemicals. | Deterministic cleavage mapping. | Long-standing practical screen for identifying cleavage hotspots in peptide sequences. Experimental Validation: Computational benchmarking only (Turalić et al., 2025). |
| UniZyme 2025 | Transformer & PLM-based unified protein cleavage predictor with enzyme active sites | Strong generalization | High accuracy for predicting cleavage sites across a range of proteolytic enzymes. Experimental Validation: Computational benchmarking only (Li et al., 2025). |
In-silico models that evaluate stability against non-enzymatic and enzymatic degradation.
Table 9
| Model & year | Architectural framework | Performance metrics | Experimental validation |
|---|---|---|---|
| Classical physics-oriented molecular or ligand docking | |||
| AutoDock 4/Vina 2004–2010 | It is physics & grid-based docking, having empirical scoring & local optimization. | Vina is 100× faster than AD4 & improved pose prediction accuracy. | Widely validated & improved speed and accuracy for binding pose prediction. Experimental Validation: Computational benchmarking only (Morris et al., 2009). |
| AutoDock-GPU 2022 | GPU-accelerated AutoDock. | Massive runtime reduction vs CPU AD4. | Facilitated large-scale virtual screening. Experimental Validation: Computational benchmarking only (Santos-Martins et al., 2021). |
| DOCK6 2009-2024 | Anchor-and-grow incremental docking alongside grid-based scorings. | Robust enrichments for virtual screening benchmarks. | Validated for ligand discovery campaigns. Experimental Validation: Computational benchmarking only (Lang et al., 2009; Allen et al., 2015; Balius et al., 2024). |
| FlexAID 2015 | Partial (heuristic AI). Flexible docking with surface complementarity scoring. | Competitive pose predictions & robust for induced-fit cases. | Improved performance for flexible binding sites. Experimental Validation: Computational benchmarking only (Alhossary et al., 2015). |
| Glide (Schrödinger) 2004–present | It is physics-based, having hierarchical docking & empirical scoring functions. | Top-tier enrichment with 65–70% success in comparative studies. | High accuracy docking benchmark performance vs other tools. Experimental Validation: Computational benchmarking only (Friesner et al., 2004). |
| GOLD 2003-2006 | Genetic algorithm–based flexible docking. | High enrichment factors alongside reproducible docking accuracy. | Pharmaceutical structure-based drug discovery. Experimental Validation: Computational benchmarking only (Verdonk et al., 2003; Joy et al., 2006). |
| Surflex-Dock 2012 | Incremental fragment placements alongside empirical Hammerhead scorings. | Pose predictions and enrichments during redocking studies. | Benchmarked against Glide and GOLD in datasets. Experimental Validation: Computational benchmarking only (Spitzer and Jain, 2012). |
| Smina 2013 | Enhanced AutoDock Vina forks with customizable minimization & scoring. | Improved scoring and benchmarking utility over Vina. | ML-assisted docking workflows, including GNINA. Experimental Validation: Computational benchmarking only (Koes et al., 2013). |
| Physics-based structural optimization | |||
| FlexPepDock 2017 | Rosetta Monte Carlo minimization alongside peptide backbone flexibility. | Achieved near-native peptide conformations (<2 Å RMSD). | Validated experimentally based peptide–protein complexes. Experimental Validation: Computational benchmarking only (Alam et al., 2017). |
| RosettaLigand 2006–2015 | It is physics-based. Monte Carlo sampling, energy minimisation & flexible docking. | Strong local refinement with accurate near-native poses in benchmark sets. | Accurate binding pose prediction; used in peptide–protein docking studies. Experimental Validation: Computational benchmarking only (Davis and Baker, 2009; Lemmon and Meiler, 2012; DeLuca et al., 2015). |
| Structure-based foundational docking | |||
| AlphaFold2 2021 | It is deep learning (transformer-based protein structure prediction). | Median backbone accuracy at the atomic level. | Near-experimental accuracy for protein structures is widely adopted in docking pipelines. Experimental Validation: Computational benchmarking only (Jumper et al., 2021). |
| AlphaFold3 2024 | Multimodal deep learning for protein–ligands & co-complex prediction. | Great improvement for protein–ligands & complex prediction over AF2. | Outperforms classical docking in some binding predictions. Experimental Validation: Computational benchmarking only (Abramson et al., 2024). |
| RoseTTAFold All-Atom 2011-2021 | Unified all-atom DL frameworks for proteins, ligands, nucleic acids, & complexes | All-atom biomolecular complex predictions. | Validated across protein–ligand and macromolecular benchmark datasets. Experimental Validation: Computational benchmarking only (London et al., 2011; Raveh et al., 2011; Alam et al., 2017; Badaczewska-Dawid et al., 2021). |
| Diffusion/Generative molecular docking | |||
| ArtiDock 2026 | It is an ML-based docking method, having deep learning scoring & structural refinement. | High enrichment and accurate pose ranking for screening-scale throughput. | Improved RMSD and binding accuracy vs classical docking. Experimental Validation: Computational benchmarking only (Voitsitskyi et al., 2026). |
| DiffDock 2022-26 | Generative DL & diffusion model for ligand pose generation. | RMSD <2 Å success =38–45% depending on benchmark | 38% success rate vs 23% traditional docking on benchmarks. Experimental Validation: Computational benchmarking only (Corso et al., 2022). |
| LigDockTailor 2025 | It is an ML meta-model with the selection of optimal docking tools. | Improved selection accuracy & meta-docking performance gains. | Improves docking success by matching ligand properties to tools. Experimental Validation: Computational benchmarking only (Zhang et al., 2025). |
| AI-assisted molecular docking | |||
| EquiBind 2022-2023 | SE ()-equivariant geometric DL for direct pose predictions. | Faster orders-of-magnitude & inference than classical docking. | Benchmarked on PDBBind datasets. Experimental Validation: Computational benchmarking only (Zhang et al., 2022; Li et al., 2023). |
| GNINA 2021-25 | CNN rescoring & docking. | Outperformed Vina on pose ranking. | Better performance in both virtual screening of active ligands and re-docking steps of co-crystallized ligands. Experimental Validation: Computational benchmarking only (Buccheri and Rescifina, 2025). |
| NeuralPLexer 2024 | End-to-end neural complex predictions integrating ligand and protein structural representations. | Competitive complex predictions and docking accuracies. | Demonstrated performance across protein–ligand and biomolecular complexes. Experimental Validation: Computational benchmarking only (Qiao et al., 2024). |
| TankBind 2022 | Geometric deep learning docking. | Competitive AI docking accuracy. | Integrated trigonometric constraints and protein binding-site block attention to outperform existing physics-based and DL methods in protein–ligand pose and affinity prediction. Experimental Validation: Computational benchmarking only (Lu et al., 2022). |
| Uni-Mol Docking 2023-2026 | Transformer-based molecular representations integrated 3D geometry. | State-of-the-art binding pose & docking predictions. | Validated across molecular and docking tasks. Experimental Validation: Computational benchmarking only (Alcaide et al., 2026). |
| Protein-peptide molecular docking | |||
| CABS-dock 2015 | Coarse-grained flexible peptides docking using CABS protein modelling alongside Replica Exchange Monte Carlo. | Median ligand RMSD often <5 Å without prior binding-site knowledge. | Validated on larger datasets for flexible peptide docking. Experimental Validation: Computational benchmarking only (Kurcinski et al., 2015). |
| GalaxyPepDock 2015 | Template-based docking using interaction energy & similarity optimization. | Outperformed several docking methods in benchmark datasets. | Reproduced experimental peptide–protein complexes. Experimental Validation: Computational benchmarking only (Lee et al., 2015). |
| HADDOCK 2010-2024 | Information-driven docking integrated biophysical & biochemical restraints alongside flexible refinements in explicit solvents. | Ranked amongst the top CAPRI predictors having strong peptide–protein docking performances. | Protein–protein and peptide–protein complex validations, also including antibody–antigen systems. Experimental Validation: Computational benchmarking only (De Vries et al., 2010; Honorato et al., 2024). |
| HPEPDOCK 2018 | Hierarchical FFT-oriented blind peptide molecular docking alongside flexible peptide refinements. | Up to 33.3% success rates, using unbound benchmark complexes. | Benchmarked against pepATTRACT and other docking tools. Experimental Validation: Computational benchmarking only (Zhou et al., 2018). |
| pepATTRACT 2017-2020 | Coarse-grained rigid-body docking. | Competitive blind peptide docking performances & used benchmark baselines. | Validated across protein–peptide benchmark datasets. Experimental Validation: Computational benchmarking only (Schindler et al., 2015; de Vries et al., 2017; Weng et al., 2020). |
| PIPER-FlexPepDock 2017-2026 | Global rigid-body docking followed by Rosetta FlexPepDock refinements. | High-resolution refinements alongside sub-angstrom peptide interface accuracy. | Strong CAPRI-style benchmark performances. Experimental Validation: Computational benchmarking only (Alam et al., 2017; Tsaban et al., 2022; Martins et al., 2026). |
Molecular docking to assess and evaluate mechanisms by providing sequence-based outcomes before wet-lab validation of AMPs.
Many discriminative models are constructed using benchmark families assembled from commercially available repositories and usually supplemented with UniProt/Swiss-Prot-derived negative sequential spaces (Zhao et al., 2026). Consequently, test and training datasets contain near-identical and evolutionarily related AMPs (Zhao et al., 2026). Moreover, highly cross-validated performance metrics partly reflect recognition of AMPs, which are represented during training rather than generalization to novel AMPs (Renaud and Mansbach, 2023). Independent test datasets can provide stronger and novel outputs only when their sequences are not included in model development and are separate from training data; labelling them as “independent datasets” cannot support biological independence (Renaud and Mansbach, 2023). Predicted models must be supported by cross-validation, and benchmark families must be interpreted cautiously and evaluated on prospective metagenomics, rigorously held-out external datasets and experimentally validated candidates ().
Future benchmarking must prioritize wet-lab validated both negative and positive AMPs, explicitly mention assay conditions and database provenance, limit cross-partition sequential redundancy using cluster-oriented splitting, and assess models on temporally separated and external datasets (). Performance metrics must be reported against family-held-out and experimental tests, alongside traditional random splits; this would provide stronger measures of whether AI-predicted models can generalize to the novel AMP sequential spaces that are intended to discover ().
1.2.4 In-silico safety screening for therapeutic-index modelling
The key barriers to the clinical translation of AMPs are host-cell toxicity, reflecting their membrane-disruptive properties (Rathore et al., 2025). Consequently, cytotoxicity-predicting models have emerged as equal objectives alongside better efficacy profiles within modern computational pipeline strategies (Rathore et al., 2025). Haemolysis and cytotoxicity are major factors in later-stage AMP therapeutic failures, highlighting the need to incorporate safety-predicted models at the initial stages of AMP design rather than as downstream filters (Rathore et al., 2025). This transition has accelerated the development of AI-driven cytotoxicity models by mapping peptide sequence spaces and structural features that link them to toxicity endpoints, enabling rapid in-silico triaging of novel AMPs before synthesis and wet-lab validation.
1.2.4.1 In-silico haemolytic filters
Classical ML methods used physicochemical descriptors; however, recent approaches have increasingly leveraged DL and protein-based LLM models for capturing higher-order sequential structural–cytotoxicity relationships (Abdelbaky et al., 2024) (Figure 7; Table 4). For instance, CNNs and RNNs demonstrated improved performance for predicting haemolytic activity directly by learning local motifs and from sequential representations associated with membrane disruption and toxicity properties (Abdelbaky et al., 2024). The latest architectural frameworks, i.e. tAMPer, integrated sequential embeddings from protein-LLMs (ESM-2) alongside structural features, improving peptide toxicity predictions by incorporating 3D contexts alongside primary sequential information (Ebrahimikondori et al., 2024). These advancements reflect broader transitions toward representation learning and reduce reliance on handcrafted features, improving generalisation across diverse datasets.
Figure 7
Haemolytic predictions have evolved from binary outcomes toward quantitative modelling. ConsAMPHemo and the latest regression-based models predict haemolytic concentrations (HC50), the peptide concentrations causing 50% lysis of red blood cells, and higher HC50 demonstrates greater blood-cell compatibility due to lower haemolytic toxicity, providing both mechanistic insights and predictive capabilities (Xie et al., 2025). Therefore, integrated pipelines have combined efficacy and safety profiling alongside binary predicted outcomes, highlighting the importance of multi-purpose learning for optimizing therapeutic indices rather than single outputs (Bhatnagar et al., 2024; Xie et al., 2025). However, certain challenges remain within dataset bias, a lack of negative data demonstrating those AMPs having haemolytic effects compared with successful, safer ones, cross-species generalization, and assay variability, underscoring the need for wet-lab validated datasets and standardized benchmarks (Bhatnagar et al., 2024; Xie et al., 2025). Discriminative AMP classifiers reporting performance of haemolysis-predicted models must be interpreted in the context of overlapping datasets and heterogeneous experimental outputs (Bhatnagar et al., 2024; Xie et al., 2025).
AMP haemolytic models have progressed from descriptor-driven binary classifiers, i.e. HemoPred with neural-network and ensemble systems (HemoNet, HLPpred-Fuse and HAPPENN) (Win et al., 2017; Hasan et al., 2020; Timmons and Hewage, 2020), to advanced graph and transfer-based learning, and multi-purpose structural-aware predictors (AMPDeep, tAMPer and ToxGIN) (Salem et al., 2022; Ebrahimikondori et al., 2024; Yu et al., 2024). These latest models not only identify haemolytic AMPs, but also estimate quantitative toxicity outputs (haemolytic concentrations), more relevant for prioritizing AMPs based on therapeutic index (the ratio of antibacterial activity to toxicity) than mere binary labels (Win et al., 2017; Hasan et al., 2020; Timmons and Hewage, 2020). These evolutionary approaches strongly support the incorporation of cytotoxicity-predicted models as front-end design constraints within AMP mining and discovery pipelines. However, reported performance metrics can be inflated by datasets overlapping with limited external validation.
1.2.4.2 In-silico cytotoxicity filters for AMP therapeutics
Traditional wet-lab approaches for toxicity assessment are labour-intensive, costly and may rely on in-vivo models, reducing their scalability during early-phase discovery. Consequently, there have been sustained shifts over the past years towards computational-based toxicity predictions, leveraging (Gupta et al., 2013). Recently, DL has enabled rapid in-silico screening of novel AMPs (Gupta et al., 2013). Early efforts are primarily focused on similarity-driven approaches, but are limited by poor generalization (Sharma et al., 2022) (Figure 8; Table 5). ToxinPred2 has illustrated that there is a transition toward hybrid frameworks incorporating motif discovery, sequence similarity (BLAST), and ML scoring, achieving better predictive outcomes (AUC = 0.99) across various protein datasets (Sharma et al., 2022).
Figure 8
Toxicity-predicting models have evolved from feature-engineered classical ML (SVM using motif features and amino acid composition) to representation-oriented learning frameworks (Jain and Kihara, 2019). ToxinPred demonstrated sequence-derived features that effectively distinguished toxic from non-toxic AMPs, but was limited to constrained datasets and shorter peptides (Gupta et al., 2013). Advanced models (ToxDL and TOXIFY) integrated evolutionary information, gene ontology, and deep neural architectural frameworks to enhance generalization across diverse protein families (Villegas-Morcillo et al., 2020). Notably, DL models, i.e., ToxDL and TOXIFY, incorporated end-to-end sequential learning to reduce dependency on handcrafted features and enabled adequate capture of latent toxicity outputs across diverse biological contexts (Villegas-Morcillo et al., 2020).
Multimodel architectural frameworks have integrated structural features, evolutionary profiles (PSSM), and protein LLM-based embeddings. ATSE demonstrated how attention mechanisms, graph-based learning, and transfer learning can enhance toxicity predictions by integrating both sequential and structural contexts (Wei et al., 2021). Notably, contemporary approaches have increasingly adopted hybrid and ensemble strategies, combining motif analysis, similarity searches, and ML predictions to reduce false positive outcomes by improving robustness (Sharma et al., 2022). Such integrative architectural frameworks have outperformed individual approaches by balancing specificity and sensitivity across diverse datasets (Sharma et al., 2022). Altogether, these are paradigm shifts toward multi-layered AI-driven predictive strategies, where toxicity must be embedded as core design constraints within AMP discovery pipelines.
1.2.5 In-silico potency screening
One of the key limitations of early AMP models was that they performed binary classification, which was not sufficient for AMP translational prioritization. In practice, the key useful endpoints are quantitative efficacy outcomes, measured as MICs (lowest concentration of an antimicrobial agents which can inhibit visible growth of pathogens following incubation for a certain time period), and it favours triage, ranking, and optimization against specific drug-resistant pathogens (Wu et al., 2014; Chiu et al., 2026). In recent years, the computational approaches have shifted from sequence-based statistical screenings toward species-based MIC regressions, where the goals are to estimate how strongly novel AMPs inhibit bacterial activities against WHO priority pathogens (Wu et al., 2014; Chiu et al., 2026) (Figure 9; Table 6).
Figure 9
Traditional potency-predicted models mainly relied on handcrafted sequential descriptors (Wu et al., 2014). An amino-acid activity contribution and linked substitution-based sequential optimization and wet-lab validation have demonstrated that computationally guided AMPs could achieve better potency both in-vitro and in-vivo (Wu et al., 2014). This has been followed by explicit MIC-prediction architectural frameworks, i.e., pseudo-amino-acid compositions alongside Gaussian kernel regression models for E. coli, establishing MIC predictions as distinctive computational tasks (Chiu et al., 2026). These early approaches demonstrated that antimicrobial potency outcomes were not only predictable but also empirically validated (Chiu et al., 2026).
Incorporation of DL regression architectural frameworks has improved predictive potency and accuracy (Dean et al., 2021). PepVAE coupled VAEs alongside activity-predicted models showed that gradient boosting performed better against E. coli; median R2 = 0.73 and root mean square error (RMSE)= 0.50. MBC-Attention approaches, combining multi-branch CNNs along with attention for E. coli MIC regression, demonstrated Pearson correlation = 0.775 and RMSE = 0.533, respectively (Dean et al., 2021). Recently, an ensemble model integrated genomic and sequence-derived features against WHO-priority pathogens (Pearson correlations of 0.756 for S. aureus, 0.781 for E. coli, and 0.802 for P. aeruginosa, respectively), marking the transition from generic activity screening towards pathogen-based quantitative potency predicted outcomes (Chung et al., 2024).
BERT-AmPEP60, fine-tuned ProtBERT for species-based MIC predictions and reported mean squared errors (MSEs) against E. coli (0.266) and S. aureus (0.303) and Pearson correlations of 0.796 and 0.753, respectively (Chung et al., 2024). ESKAPEE-MICpred incorporated two-input transfer-learning and ensemble architectural frameworks for pathogen-specific MIC predictions (Sharma et al., 2024). A recent pipeline (AMPredictor) has integrated generative designs alongside potency scorings: bifunctional AMP frameworks linked GAN generators to graph-convolutional MIC regressors, achieved RMSE = 0.5348 and Pearson= 0.7072, and validated 24 novel AMPs in an animal infection model (Dong et al., 2025b). Collectively, these architectural frameworks show how potency predictions have become a central layer within multi-purpose AMP pipelines to jointly optimize antimicrobial activity and novelty against resistant pathogens (Chung et al., 2024).
1.2.6 In-silico novelty screening
In silico novelty screening must be considered one of the key components of the pipeline to limit the rediscovery of known AMPs. Novelty must be evaluated at two stages: first, against well-known peptide sequential spaces and proteins to limit homology; second, against genomics and transcriptomics spaces to determine whether the sequence maps back to untranslated regions, known genes and unannotated ORFs in pre-existing assemblies (Gertz et al., 2006; Camacho et al., 2009; Buchfink et al., 2021) (Table 7). In the past few years, it has relied on profile-based or alignment retrieval systems instead of traditional AI models (Gertz et al., 2006; Camacho et al., 2009). tBLASTn served as a standard translated-search-based method for mapping protein and peptide queries to nucleotide databases against all six reading frames (Gertz et al., 2006; Camacho et al., 2009) (Figure 10). tBLASTn has upgraded the reliability for nucleotide homologous searches by providing composition-oriented statistical outcomes, making it more effective for genome-level novelty outputs rather than naive translational searches (Gertz et al., 2006; Camacho et al., 2009) (Figure 10).
Figure 10
At the protein sequence level, novelty screening requires searching for candidates against commercially available large databases (Fu et al., 2012; Steinegger and Söding, 2017; Buchfink et al., 2021). MMseqs2 and DIAMOND have become crucial in recent times, as they conserve BLAST-like sensitivity and limit the time costs for large-scale datasets (Buchfink et al., 2021). DIAMOND has improved based-matched BLAST-like sensitivity and delivers approximately 80-fold speedup over BLASTP, as well as completing an nr versus UniRef50 tree-of-life scaling (Buchfink et al., 2021) (Figure 11). MMseqs2 demonstrated higher sensitivity than PSI-BLAST at far higher speeds; therefore, it can be run iteratively for larger workflows (Steinegger and Söding, 2017). To limit redundancy, CD-HIT considers practical upstream filters allowing it to quickly cluster highly similar sequences and limit inflated novelty due to near-duplicate inputs (Fu et al., 2012; Steinegger and Söding, 2017; Buchfink et al., 2021).
Figure 11
Pairwise local alignments alone can miss remote homologies for peptides that are short and compositionally biased towards precursor proteins (van Kempen et al., 2024). Many higher-confidence novelty workflows add domain-based or profile-HMM screenings. HMMER tools (jackhammer and phmmer) have provided profile-based sequential similarities that are considered more sensitive than simple pairwise methods for remote relationships (Finn et al., 2011). InterProScan has helped to determine whether a candidate is within an already known protein and motif architectural landscapes (Jones et al., 2014). Foldseek has extended novelty screenings beyond primary sequences by enabling fast structural searches; useful when sequential identities are lower and fold-based similarities suggest that candidates belong to pre-existing structural families rather than representing truly novel scaffolds (van Kempen et al., 2024).
For the AMP pipeline, multi-layer novelty screening is essential and requires removal of redundancy with CD-HITs. Peptides and proteins must be searched against well-known databases with BLASTP, MMseqs2 and DIAMOND, respectively, and promising AMPs must be backed onto genomic assemblies and transcript catalogues with tBLASTn to check mapping with known homologous regions (Fu et al., 2012; Steinegger and Söding, 2017; Buchfink et al., 2021). Later, InterProScan and HMMER can be applied for remote family assignments, and, when structural predictions are considered as an essential part of the workflows (Finn et al., 2011; Jones et al., 2014). Foldseek can be used to exclude structurally familiar scaffolds to limit rediscovery risks before downstream wet-lab validation (van Kempen et al., 2024).
1.2.7 In-silico stability screening
Poor stability associated with rapid degradation by host and microbial proteases is one of the major translational barriers (Mathur et al., 2018). Therefore, in-silico AMP stability screening has become a valuable tool, with two key goals: firstly, to recognize protease-sensitive cleavage motifs that affect peptide half-life; secondly, to estimate circulation half-life as one of the proxies (Mathur et al., 2018) (Figure 12; Table 8). In recent times, advanced ML models incorporate evolutionary, sequential, structural, and physicochemical parameters, which are essential to identify early triage layers to flag unstable AMPs before synthesis and redesign to make protease resistance (Song et al., 2012; Mathur et al., 2018).
Figure 12
PeptideCutter contains knowledge-based cleavage simulators and maps potential non-enzymatic and enzymatic cleavage sites, as well as providing rapid first-pass stability assessments by highlighting liability sites (Turalić et al., 2025). The protease-specific ML PROSPER can predict integrated feature-based cleavage sites across various proteases, and when followed by PROSPERous, can improve high-throughput predictions by expanding the coverage across 90 proteases (Song et al., 2012; Song et al., 2018). These tools are essential for AMP optimization, as shorter AMPs are sensitive to single cleavage events. Therefore, they provide advanced information regarding the peptide’s engineering, such as cyclization, residual substitutions, and D-amino-acid design strategies (Song et al., 2012; Song et al., 2018; Turalić et al., 2025).
Procleave has incorporated both sequential and 3D structural information to improve protease-based cleavage-site prediction, reflecting that local conformation often determines whether nominal motifs are actually cleaved (Li et al., 2020b). On the contrary, panCleave possesses pan-protease RF models for computational-based proteolysis and is used to mine encrypted AMPs from extinct human proteomes, leading to wet-lab validations of antimicrobial fragments, highlighting that cleavage predictions must be used constructively for AMP discovery (Kelley et al., 2024). PlifePred, a complementary cleavage-site tool, has addressed other stability dimensions by predicting peptide circulation half-life from experimentally validated data (Mathur et al., 2018). Altogether, these approaches highlight the need for adapting a multi-model workflow in which AMPs are screened for novelty, pharmacokinetic stability, potency, and toxicity before peptide synthesis.
1.2.8 In-silico molecular docking approaches for structural optimization
Molecular docking (MD) is considered the principal component of structure-based AMP optimization, which predicts the binding modes and target specificity at atomic resolution. Unlike sequential-based predicted models, MD models ligand–receptor interactions to rationalise how AMPs interact with microbial membranes, enzymes, and intracellular targets (Davis and Baker, 2009; Lemmon and Meiler, 2012; DeLuca et al., 2015) (Figure 13; Table 9). In recent years, MD has advanced from rigid-body, physics-oriented scoring to hybrid architectural frameworks incorporating diverse ML, structural predictions, and flexible docking, improving performance metrics (Davis and Baker, 2009; Lemmon and Meiler, 2012; DeLuca et al., 2015). Traditional MD tools are widely used due to their interpretability and robustness, particularly during the initial stages of virtual screening (Morris et al., 2009) (Figure 13, Table 9).
Figure 13
AutoDock and AutoDock Vina have been foundational for peptide-protein docking. AutoDock Vina demonstrated better improvements in terms of speed and accuracy due to multi-threading and effective scoring functions, making them ideal for larger-scale virtual screening (Morris et al., 2009). Rosetta-based MD frameworks (RosettaLigand) have incorporated conformational sampling alongside physics-based energy functions to predict binding poses and optimized peptide-protein interactions, suitable for flexible ligands (Davis and Baker, 2009; Lemmon and Meiler, 2012; DeLuca et al., 2015). Notably, tools have extensively validated structure-based drug discovery and identified novel bioactive compounds. However, traditional MD approaches struggle with solvent effects, protein flexibility, and free-energy estimations (Davis and Baker, 2009; Lemmon and Meiler, 2012; DeLuca et al., 2015).
Recently, MD approaches have been transformed by integrating AI and structure-predicted models (Jumper et al., 2021; Abramson et al., 2024). AlphaFold has drastically increased the availability of high-quality protein structures, removing key bottlenecks for docking workflows (Jumper et al., 2021). AlphaFold3 can directly model protein-peptide and protein-ligand complexes, outperforming by predicting binding conformations (Abramson et al., 2024). Diffusion-based models (DiffDock) have reframed docking as a generative issue and significantly improved hit rates by correctly predicting bindings (Corso et al., 2022). Collectively, it highlights paradigm shifts toward data-oriented structural modelling by learning representations that complement and replace physics-based scoring.
MD approaches to identify novel bacterial targets against predicted AMPs require intensive computational resources. Latest MD frameworks support closed-loop optimization and larger-scale virtual screenings (Zhang et al., 2025; Voitsitskyi et al., 2026). Ultra-large-scale approaches screen billions of candidates against targeted structures, ensuring in-silico high-affinity binders and novel scaffold discovery before wet-lab validation (Zhang et al., 2025; Voitsitskyi et al., 2026). Hybrid systems have integrated docking with ML–based selection and scoring, e.g., LigDockTailor, to optimise performance outputs based on ligand properties (Zhang et al., 2025). Altogether, MD critically bridges sequence-oriented AMP design and structure-based optimization, enabling refinement for binding affinity and specificity within an advanced multi-model AMP pipeline.
2 Current AI-based approaches for mining, discovery, and optimization of AMPs
Evolution from classical ML to DL approaches and LLMs has revolutionized the computational toolkits for AMP mining and discovery in recent times (Fernandes et al., 2023; ). However, increasing architectural frameworks and complexities do not guarantee better performance metrics (Fernandes et al., 2023; ). There is no single AI framework that is optimal for AMP mining, identification, and characterization; therefore, it is dependent on the multitask models and available data resources (Fernandes et al., 2023; ).
Multitask-matched models will be more beneficial than automatic progression based on their complex architectural frameworks. Classical models, RF, SVM, and XGBoost are effective for imbalanced and smaller datasets for rapid binary classification, while DL models, such as transformers, CNNs, and LSTMs, have outperformed for larger and labelled sequential datasets by learning long-range residual dependencies (Moretta et al., 2020; ; Yu et al., 2025). For structural tasks, 3D-aware geometric DL approaches are required for receptor-based designing and membrane-interactive modelling (Fernandes et al., 2023). Depending on the available computational resources, decisions regarding selecting various architectural frameworks vary with modality and predictive outcomes.
Models, such as AMPGen and EBAMP, have demonstrated that an architecture framework-based selection can be task-driven, i.e., RL-based and diffusion models were used for de novo AMP generation, while transformer-oriented MIC regressors predicted efficacy, and ensemble models favoured safety outcomes with multi-purpose screening when stability and interpretability were critical (Sharma et al., 2024; ; Zhao et al., 2025). Earlier evolutionary and genetic strategies demonstrated better multi-objective outputs regarding optimization, but their performances were hampered by the noisy biological data and limited datasets (Shin et al., 2022; Agüero-Chapin et al., 2023).
AMP workflows can span heterogeneous objectives, i.e., structure-based classification and designs, target-based specificity, MICs regression, and toxicity prediction (Wang et al., 2025; ; ). Multi-model evolution can expand the AMP discovery toolkit, depending on whether architectural choices are aligned with data quality and biological contexts, not complexity alone (Cao et al., 2023; Wang et al., 2025b). Therefore, a multi-model AI-driven pipeline strategy, as reviewed in the current paper, for predicting diverse architectural frameworks, aligning with the best practical implications for AMP discovery and optimization, is essential (Cao et al., 2023; Wang et al., 2025b) (Figure 14).
Figure 14
Multi-model AI-driven pipelines can facilitate mechanistic-informed routes to accelerate AMP discovery by progressing from wet-lab explorations to computation-based triage (Li et al., 2024; ) (Figure 14). Therefore, mining larger metagenomic datasets using single-model screenings will be insufficient, as AMP discovery has intrinsic multi-objectives, i.e., predicted potency must be balanced against novelty, stability, bioactivity, and safety (Figure 15). Multi-model AI pipelines can address these concerns by combining foundation models to encode peptide grammar to enable targeted prioritization, discriminative predicted models for pathogen-specific potency, and safety filters to predict cytotoxicity and haemolysis (Li et al., 2024). Recent studies have explicitly demonstrated that multi-stage AI approaches have delivered validated AMPs against Gram-negative bacteria inside compressed design and test cycles, including pathogen-targeted outputs (Li et al., 2024; ).
Figure 15
A multi-model AI-driven pipeline can mine and discover novel AMPs from diverse resources, i.e., natural, genomics, metagenomics, and de novo generative spaces, to predict efficacy and safety profile before wet-lab validations, thereby limiting the peptides that must be synthesized and assayed to achieve clinical relevance (Santos-Júnior et al., 2024; Li et al., 2025). It has been supported by recent ML and DL frameworks by identifying AMPs from microbiomes, subsequently in-vitro validation, and generating de novo AMPs with higher “hit rates” coupled with discriminative and safety models (Jin et al., 2025). MIC-based models against Gram-negative strains further prioritize WHO-listed pathogens to accelerate mechanism-driven selection (Bae et al., 2025).
Stronger computational and preclinical performances cannot guarantee downstream therapeutic successes. AMPGen provided contemporary examples of attrition at the computational-experimental interfaces (Jin et al., 2025). Although its XGBoost discriminators, diffusion-driven generators and LSTM-based target-specific scorer prioritized 40 de novo AMPs (Jin et al., 2025). 38 were synthesized and 31 of these met antimicrobial activity criteria (success rates of 81.58%) and were target-dependent: 19/20 synthesized AMPs targeting E. coli were active compared with only 8/18 targeting S. aureus (Jin et al., 2025). After multimodel computational filtering, a subset of highly prioritized AMPs had failed experiments, illustrating in-silico outputs alone cannot be considered as substitutes for biological validation (Jin et al., 2025).
At later translational stages, pexiganan (MSI-78) has provided a historical example (Lipsky et al., 2008). Despite having strong potency and advancement to phase III trials for managing infected diabetic foot ulcers, the clinical outcomes did not translate into regulatory approvals; the studies failed to meet their primary efficacy objectives (Lipsky et al., 2008). Iseganan, broad-spectrum protegrin analogues with better antimicrobial activity, failed to reduce ulcerative oral mucositis in a larger and randomized Phase III trial recruiting 545 patients (Trotti et al., 2004). Together, these studies have illustrated a continuum of attrition starting from computational predictions to wet-lab bioactivity, final clinical outcomes, and reinforce the urgent need for incorporating multimodel AMP pipelines to demonstrate pharmacological (PK/PD) and physiological constraints before clinical progression (Trotti et al., 2004; Lipsky et al., 2008).
Table 10 summarizes different multi-model ML approaches with reported wet-lab, in-vitro, and in-vivo validation by highlighting their predicted outcomes, essential to demonstrate translational impacts. Prioritizing multi-model AI approaches can offer real-world performance for antimicrobial outcomes, not just in-silico validity.
Table 10
| Multi-model AI | Architectural frameworks | Experimental-based evidence |
|---|---|---|
| a. Hybrid methods with ensemble frameworks | ||
| AMPredictor | SVM, ESM-2, and MLP. | Predicted residual physicochemical descriptors using attention maps (Dong et al., 2025b; ). |
| AMP-BRET | RF and ProtBERT Transformers fine-tuned for regression. | Transfer learning and high precision from protein corpus to predict antimicrobial activity (Lee et al., 2023). |
| AMP-EF (AMP– Ensemble Framework) | Bi-LSTM, XGBoost, alongside attention layers. | Provided higher performance and generalization to predict antimicrobial activity (Zhang et al., 2023). |
| AMPpred-EL | ML ensemble with LightGBM and logistic Regression. | Combined various ML components to predict AMPs with antimicrobial activity (Lv et al., 2022). |
| AMPpred-DLFF | CNN and ESM-2-based feature extractions. | Synergized graphical attentions, convolutional features and protein-language embeddings for AMP bioactivity predictions (Chen et al., 2024). |
| AMP-META | LGBM with different AMP tools. | Predicted physicochemical descriptors and bioactivity using larger datasets (Jorge et al., 2012; Tsai et al., 2024). |
| AMPpredMFA | LSTM, CNN, Attention, & MLP. | Integrated CNN and LSTM to predict features (Li et al., 2024). |
| Ensemble MIC regression | Ensemble DL regression (multi-inputs). | MIC validation against E. coli and P. aeruginosa (Chung et al., 2024). |
| E-CLEAP | Ensemble of four MLP classifiers. | Ensemble neural classifiers for binary classification and antimicrobial activity predictions (Wang, 2024). |
| PepMultiFinder | Multi-filter ML approaches. | Predicted efficacy & safety (Duque-Salazar et al., 2020; Monsalve et al., 2024). |
| b. Hybrid protein language-based frameworks | ||
| FusPB-ESM2 | Neural networks with ProBERT and ESM-2 embeddings. | Fused two LLM embeddings for multi-purpose representations and microbial predictions (Wu et al., 2024; Wang et al., 2025a). |
| PGAT-ABPp | GAT with ProtT5 embeddings. | Integrated geometric DL for 3D structural representations to predict efficacy & safety profile (Hao et al., 2024). |
| c. Hybrid DL approaches based on deep representations & fusion features | ||
| AFP-MFL | MLP with co-attention mechanistic approaches. | BLOSUM62 & ProtT5 predicted physicochemical properties, whereas MLP with co-attention showed multi-purpose fusion & antibacterial activity (Fang et al., 2023; Sultan et al., 2024). |
| AMPFinder | MLP, ProtTrans and OntoProtein. | Used protein sequences for functional annotations to predict pathogen-specific properties (Yang et al., 2023). |
| FSLSME | MLP and ESM-1. | Used different libraries for mining AMPs and to predict bioactivity (). |
| Pang’s Approach | MLP with pre-trained BERT. | Predicted regression-oriented MICs and safety outputs (Pang et al., 2022). |
| UniproLcad | Bi-LSTM with attention methods, ESM-2, 1D-CNN, UniRep, ProtBERT. | Multi-purpose fusion models alongside competitive interpretability and accuracy (Wang et al., 2024). |
| UniDL4Biopep | CNNs with ESM-2 embeddings. | Pretrained model, with CNNs spatial features extracted from embeddings to predict antimicrobial & stability activity (Du et al., 2023). |
| d. Large language-based hybrid model | ||
| AMP-Designer | LLM, foundation model with multi-purpose design components. | Validated AMPs’ bioactivity against Gram-negative bacteria, including in-vivo murine lung infection (Wang et al., 2025). |
| ProteoGPT | Pertained protein LLM (ProteoGPT) with sub-LLMs. | Validation against clinical Gram-negative superbugs (). |
| GPT-3-AMP | GPT-3 (generative) & SVM/RF (discriminative). | Uses GPT-3 to generate candidate peptides, using SVM/RF filters for antimicrobial activity (Wang et al., 2021; ). |
| deepAMP | DL predictors alongside in-silico optimization (deepAMP-POM). | MIC testing against Gram-negative MDR pathogens (Li et al., 2024). |
| AMPGen | Diffusion-based generators, i.e., autoregressive diffusion alongside attention, XGBoost discriminator and LSTM scorer. | Target-specific, designed with in-vitro validation (Jin et al., 2025). |
| e. Hybrid models with CNNs, RNNs, or transformer-based frameworks | ||
| AMPDeep | MLP with BERT, i.e., ProtBert. | Optimized bioavailability and safety profiles (Salem et al., 2022; Ramos-Llorens et al., 2024). |
| AI4AMP | Deep neural networks, i.e., CNNs, LSTM. | PC6 encodings to map sequences into physicochemical vectors to predict bioactivity (Lin et al., 2021). |
| AMPGen | STM, autoregressive diffusion models with XGBoost discriminators. | Predicted antimicrobial activity (Jin et al., 2025). |
| AMPTrans-LSTM | Transformers and LSTMs. | Protein databases to predict bioactivity (Mao et al., 2023). |
| sAMP-VGG16 | VGG-based CNNs. | Optimized adaptive layers by incorporating LLB & deep CNNs for binary classification (Pandey and Srivastava, 2025; Pandey et al., 2025). |
| Capecchi et al. | Sequential GAN model, transformers & RNN classifiers. | Non-hemolytic AMPs targeting Gram-negative pathogens (Capecchi et al., 2021). |
| HydrAMP | cVAE-GAN hybrid. | Predicted efficacy and safety profiles (Szymczak et al., 2023). |
| LLAMP | Transformer-based protein LMs (ESM-2) for species-based potency. | MICs against Gram-negative and Gram-positive panels (Bae et al., 2025). |
| TACaPe-based (2025) | Transformer-based GAN models. | In-vitro MIC testing against P. aeruginosa (de Albernaz et al., 2025). |
Studies using multi-model hybrid-based frameworks for AMP predictions alongside experimental validation.
2.1 Current barriers and limitations to robust AI-driven AMP discovery
Despite rapid advances in ML, LLMs and generative AI, the predicted performance reported for AMP models must be interpreted cautiously. Major challenges arise from dataset bias and poor representation of biodiversity. Currently available AMP databases are enriched with previously characterized, cationic and relatively shorter membrane-active AMPs, whereas inactive AMP sequences have not been reported systematically (; ). Mostly negative datasets are assembled computationally from different commercial resources, rather than experimentally validated AMPs (). Such approaches can create differences and will allow models to adapt to dataset-specific characteristics rather than biologically relevant AMPs. Homologous AMPs and sequence redundancy distributed across test and training sets can cause information leakage, thereby overestimating the generalization (Wang, 2023; Wang et al., 2026). Therefore, future benchmarking requires similarity-aware train–test separations, redundancy-controlled partitions and evaluation against robust and independent sequence families and organism sources (Wang, 2023; Wang et al., 2026).
There is limited availability and heterogeneity of experimentally validated datasets, and only a smaller number of AMPs are characterized under universal experimental conditions. Activity annotations originating from heterogeneous assays may differ in microbial strains, medium compositions and endpoint definitions (). MIC values obtained for the same AMPs may not be directly comparable across different studies. Moreover, safety endpoints are fragmented across cell lines, erythrocyte species and exposure conditions (; ). These concerns are particularly consequential for quantitative stability, haemolysis, potency, cytotoxicity, and pharmacokinetic models, where inconsistent experimental metadata can cause label noise (; ). Therefore, databases should evolve from sequence and bioactivity repositories towards assay-aware multimodel platforms having universal experimental protocols including pathogen-based and concentration-dependent information alongside serum binding, protease stability, and clinically-related safety endpoints (; ).
Model interpretability has represented an additional barrier to translational confidence. Different AI-oriented architectural frameworks can provide strong discriminative predictions with limited insights about sequence motifs and physicochemical interactions (; ). Post hoc approaches (e.g. attention mappings and residue-level saliency) can enhance interpretability, but do not propose biological mechanisms, particularly for generative models, where highly predicted AMP probabilities do not establish plausible mechanisms, microbial and resistance profiles (). Therefore, explainable AI must be coupled with mechanistic assays and experimental perturbations, so computationally identified determinants can be tested independently.
Methodological transparency and reproducibility have remained inconsistent across AMP modelling studies; differences in databases, redundancy thresholds, sequential preprocessing, hyperparameters, random seeds and data splitting can produce variations amongst comparable methods (). In some studies, training sequences may occur accidentally within benchmark sets, while unavailable source code and pretrained weights can limit independent reproduction (; ).
Absence of universally accepted benchmarking protocols for AMP classifiers, and differences within the benchmark compositions have made direct comparison problematic. This is amplified by class imbalances, where higher AUROC may provide incomplete performance descriptions (). Generative models can be assessed using divergent measures, such as validity, novelty and predicted activity, while quantitative MIC models can be evaluated using heterogeneous transformations, regression metrics and pathogen subsets (). Temporal validations where models can be trained on available data, subsequently tested on reported AMPs can provide especially rigorous measures of prospective performance ().
Computational performance cannot equate with therapeutic performance. Most AI-generated AMPs have been evaluated in-silico, and only smaller proportions have progressed through synthesis, serum stability assessments and in-vivo validations (MIC determination and cell toxicity) (; ). Physiological factors (ionic strength, serum and protein binding, proteolysis, aggregation, biodistribution and immune interactions) can substantially alter activities (; ). Multimodel pipelines must be viewed as prioritization platforms for the next phases of AI-driven AMP discovery, requiring prospective design-build and test-learn cycles, where standardized experiments will update computational models, essential for moving from retrospective predictions towards clinically-oriented AMP engineering.
2.2 Future directives
Convergence of advances within structural biology, sequential modelling, and high-throughput computational approaches has favoured a shift from fragmented predicted workflows toward a well-integrated and multi-purpose strategy for AMP discovery (). Whilst traditional linear approaches evaluate stability, activity and toxicity sequentially, the proposed methodologies adapt closed-loop, multi-model architectural frameworks, where AMPs are iteratively refined across orthogonal predictive multi-layers (Wang et al., 2025). This advanced AI-driven design has enabled simultaneous optimization of structural and proteolytic compatibility, novelty, potency, and host safety, thereby addressing the key bottlenecks historically limited to AMPs’ clinical translation ().
Multi-model AI-driven exhibits a hierarchical screening strategy. Candidate sequences can be derived from metagenomics or curated libraries and are evaluated using discriminative AMP classifiers, which later pass through quantitative MIC regression frameworks to predict and prioritize AMPs against resistant pathogens (Guan et al., 2025). Parallelly, hemolytic and cytotoxicity models serve as initial safety filters to exclude AMPs predicted to have higher cytotoxicity (Xie et al., 2025). Early-stage incorporation of potency and safety filters demonstrates key departures from traditional drug discovery pipelines, where cytotoxicity is evaluated following wet-lab validation (Xie et al., 2025).
Following efficacy and safety filtering, shortlisted AMPs must undergo novelty assessments; genomic search and profile-based homology detection must translate with rapid sequential alignment tools (DIAMOND, MMseqs2) to identify peptides that are active and distinctive from known AMPs and conserved protein domains (Steinegger and Söding, 2017; Buchfink et al., 2021). Later, they must be structurally characterized using sophisticated structure prediction models (AlphaFold2/3) and downstream molecular interactions against novel targets, providing insights into binding orientations and potential mechanistic approaches facilitating rational optimization (Jumper et al., 2021; Voitsitskyi et al., 2026).
Integration of degradation-aware and stability tools must be used to predict proteolytic susceptibility and AMP half-life to exclude unstable candidates (Song et al., 2012; Li et al., 2020b). Peptide modifications, such as cyclization, residual substitution, and incorporation of non-natural components, can enhance stability without impairing bactericidal activity, enabling iterative closed-loop optimization cycles to improve diverse objectives simultaneously (Song et al., 2018).
At present, the absence of standardized datasets demonstrating the serum protein binding effects on antimicrobial activity is one of the major obstacles associated with AI-driven drug discovery (Wang, 2023; Wang et al., 2026). Most AMPs possess better potency outcomes under laboratory conditions, but exhibit reduced efficacy within physiological environments due to non-specific interactions with circulating serum proteins, i.e., albumin, globulin, and lipoproteins (Wang, 2023; Wang et al., 2026).
There is a need to develop larger-scale and wet-lab-validated serum-binding databases for systematically quantifying AMP in the presence of human serum. Therefore, integration of serum-binding data resources alongside ML, DL and generative AI frameworks would enable the identification, prediction and optimization of AMPs’ performance under biologically relevant conditions, thereby improving the clinical translation of next-generation AMP from bench to bedside (Wang, 2023; Wang et al., 2026).
Despite rapid advancements in AI-assisted AMP discovery, clinical translation depends on establishing regulatory and manufacturing frameworks (Zhao et al., 2026). Regulatory assessments require demonstration of higher predictive accuracies; therefore, developers must provide documentation of model provenance, predefined decision thresholds and to what extent computational predictions will contribute towards AMP selection and sequence modification (Zhao et al., 2026). Additional scrutiny is required when models can introduce unusual sequence motifs and chemical modifications for which pharmacological evidence is limited. It must converge with conventional chemistry, control and manufacturing requirements (Zhao et al., 2026).
At present, manufacturing has become challenging for highly modified, cyclized and non-canonical AMPs, where solid-phase synthesis, analytical characterization and purification have increased production cost and complicated scale-up (; ). Future pipelines must incorporate developability and manufacturability as one of explicit design objectives, so that AMPs characterized and optimized using computational screenings will be compatible with scalable manufacturing production and subsequent regulatory evaluation (Zhang et al., 2025).
From retrospective validation to larger prospective experimental studies, AI-designed AMPs must be evaluated against several clinical and resistant strains using standardized in-vitro assays. Clinical translation must demonstrate across genetically and phenotypically diverse pathogens, followed by incorporation of protein-binding and protease stability, cytotoxicity, and pharmacodynamic–pharmacokinetic relationships, before advancing towards animal models (). External validations across different laboratories will be mandatory to determine model performance beyond the datasets (Zhao et al., 2026). Most credible pipelines must route towards clinical applications possessing prospective design-build and test-learn frameworks, where AI-predicted AMPs must be subjected to diverse experimental filters and the resulting biological data must be returned to the computational pipelines. Such iterative validations will enable AMP discovery by progressing from high-performing in-silico drug design towards reproducible AMPs with better therapeutic potential ().
2.3 Conclusion
An integrated, multi-model, advanced AI-driven pipeline demonstrates transitions toward systematic AMP design and optimization, where different predicted modalities operate in a concerted manner rather than in isolation. Therefore, by embedding functional, structural, and pharmacological constraints within a unified and well-integrated framework, this approach favours effective navigation of peptide design spacing to reduce later-stage attrition. These approaches are well-suited for high-throughput computing, where larger-scale AMP mining, generation and evaluation can be performed in parallel to support rapid translation from in-silico discovery to wet-lab validation for clinical translation. This multi-model strategy provides generalizable blueprints for developing next-generation antimicrobial agents to combat the global AMR crisis in the era of computational biology.
Statements
Author contributions
NS: Writing – review & editing, Investigation, Data curation, Software, Resources, Conceptualization, Writing – original draft, Visualization, Validation, Formal analysis. EE-O: Supervision, Writing – review & editing, Investigation. NK: Writing – review & editing, Investigation, Supervision. MW: Writing – review & editing, Investigation, Supervision, Conceptualization. X-TJ: Investigation, Writing – review & editing, Supervision.
Funding
The author(s) declared financial support was received for this work and/or its publication. NS is supported by a University International Postgraduate Award (UIPA) scholarship from UNSW Sydney, Australia, for their research program.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was used in the creation of this manuscript. All the images were created with BioRender (https://www.biorender.com), alongside assistance from ChatGPT (GPT-5.5) for conceptual illustrations, and scientifically curated, edited and reviewed by the authors. All scientific content, interpretation, verification of references and final manuscript approval were performed by the authors, who take full responsibility for the integrity of the content
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Abbreviations
A. B, Acinetobacter baumannii; AI, Artificial Intelligence; AMPs, Antimicrobial peptides; AMR, Antimicrobial resistance; ATT, Attention mechanism; AUROC, Area Under the Curve; BERT, Bidirectional Encoder Representations from Transformers; Bi-LSTM, Bidirectional Long Short-Term Memory; CNN, Convolutional Neural Network; DF, Deep Forest; DL, Deep Learning; E.C., Escherichia coli; GAN, Generative Adversarial Network; GNN, Graph Neural Network; GPT-3, Generative Pre-trained Transformer-3; K. P, Klebsiella pneumoniae; LGBM, Light Gradient-Boosting Machine; LLMs, Large Language Models; LSTM, Long Short-Term Memory; MD, Molecular docking; MDR, Multidrug-resistant; MIC, Minimum Inhibitory Concentration; ML, Machine learning; P. A, Pseudomonas aeruginosa;, PSEAAC, Pseudo Amino Acid Composition; QSVM, Quantum Support Vector Machine; RF, Random Forest; RMSE, Root Mean Square Error; VAEs, Variational Autoencoders; WHO, World Health Organization.
References
1
AbbasiM.VasconcelosV.VicenteE. M.SantosA. L.ArraisJ. P. (2025). “ A novel deep learning framework for predicting antimicrobial peptide activity using protBert and neural networks”, in: Pattern Recognition and Image Analysis: 12th Iberian Conference, IbPRIA 2025, Proceedings, Part II (Cham, Switzerland: Springer).
2
AbdelbakyI.ElhakeemM.TayaraH.BadrE.Abdul SalamM. (2024). Enhanced prediction of hemolytic activity in antimicrobial peptides using deep learning-based sequence analysis. BMC Bioinf.25, 368. doi: 10.1186/s12859-024-05983-4
3
AbramsonJ.AdlerJ.DungerJ.EvansR.GreenT.PritzelA.et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature630, 493–500. doi: 10.1038/s41586-024-07487-w
4
Agüero-ChapinG.AntunesA.Marrero-PonceY. (2023). A 2022 update on computational approaches to the discovery and design of antimicrobial peptides12, 1011. doi: 10.3390/antibiotics12061011
5
AlamN.GoldsteinO.XiaB.PorterK. A.KozakovD.Schueler-FurmanO. (2017). High-resolution global peptide-protein docking using fragments-based PIPER-FlexPepDock. Pcb13, e1005905. doi: 10.1371/journal.pcbi.1005905
6
AlcaideE.GaoZ.KeG.LiY.ZhangL.ZhengH.et al. (2026). “ Uni-Mol Docking V2: towards realistic and accurate binding pose prediction”, in: Artificial Neural Networks and Machine Learning – ICANN 2025: International Workshops and Special Sessions (Cham: Springer Nature Switzerland).
7
AlhossaryA.HandokoS. D.MuY.KwohC.-K. (2015). Fast, accurate, and reliable molecular docking with QuickVina 2. Bioinformatics31, 2214–2216. doi: 10.1093/bioinformatics/btv082
8
AliF.KhalidM.AlsiniR.YafozA.AlkhalifahT.DuM.-Z. (2026). A generative explainable model for antimicrobial peptide prediction using bidirectional temporal convolutional neural network. Sci. Rep.16, 13801. doi: 10.1038/s41598-026-43370-6
9
AllenW. J.BaliusT. E.MukherjeeS.BrozellS. R.MoustakasD. T.LangP. T.et al. (2015). DOCK 6: impact of new features and current docking performance. Japa36, 1132–1156. doi: 10.1002/jcc.23905
10
AnsariM.WhiteA. D. (2023). Serverless prediction of peptide properties with recurrent neural networks. JJoCI Modeling63, 2546–2553. doi: 10.1101/2022.05.18.492545
11
AnsariM.WhiteA. D. (2024). Learning peptide properties with positive examples only. JDD3, 977–986. doi: 10.1039/d3dd00218g
12
Badaczewska-DawidA. E.KmiecikS.KolińskiM. (2021). Docking of peptides to GPCRs using a combination of CABS-dock with FlexPepDock refinement. BiB22, bbaa109. doi: 10.1093/bib/bbaa109
13
BaeD.KimM.SeoJ.NamH. (2025). AI-guided discovery and optimization of antimicrobial peptides through species-aware language model. Brief Bioinform.26, bbaf343. doi: 10.1093/bib/bbaf343
14
BajiyaN.ChoudhuryS.DhallA.RaghavaG. P. J. A. (2024). AntiBP3: a method for predicting antibacterial peptides against Gram-positive/negative/variable bacteria13, 168. doi: 10.3390/antibiotics13020168
15
BaliusT. E.TanY. S.ChakrabartiM. (2024). DOCK 6: incorporating hierarchical traversal through precomputed ligand conformations to enable large-scale docking. Japa45, 47–63. doi: 10.26434/chemrxiv-2023-vkw87
16
BanerjeeP.EckertA. O.SchreyA. K.PreissnerR. (2018). ProTox-II: A webserver for the prediction of toxicity of chemicals. Nucleic Acids Res.46, W257–WW63. doi: 10.1093/nar/gky318
17
BhadraP.YanJ.LiJ.FongS.SiuS. W. I. (2018). AmPEP: sequence-based prediction of antimicrobial peptides using distribution patterns and random forest. Sci. Rep.8, 1697. doi: 10.1038/s41598-018-19752-w
18
BhatnagarP.KhandelwalY.MishraS.GS. K.DuttaA.MitraD.et al. (2024). Predicting antibacterial activity, efficacy, and hemotoxicity of peptides using an explainable machine learning framework. Process Biochem.145, 163–174. doi: 10.1016/j.procbio.2024.06.027
19
BlaskovichM. A. T.CooperM. A. (2025). Antibiotics re-booted—time to kick back against drug resistance. NPJ Antimicrobials Resistance3, 47. doi: 10.1038/s44259-025-00096-1
20
BournezC.RioolM.de BoerL.CordfunkeR. A.de BestL.van LeeuwenR. (2023). CalcAMP: accurate prediction of antimicrobial activity. Antibiotics12, 725. doi: 10.3390/antibiotics12040725
21
BuccheriR.RescifinaA. (2025). High-throughput, high-quality: benchmarking GNINA and AutoDock Vina for precision virtual screening workflow. Molecules (Basel Switzerland)30, 3361. doi: 10.3390/molecules30163361
22
BuchfinkB.ReuterK.DrostH.-G. (2021). Sensitive protein alignments at tree-of-life scale using DIAMOND. Nat. Methods18, 366–368. doi: 10.1038/s41592-021-01101-x
23
BuehlerM. J. (2023). A generative pretrained autoregressive transformer graph neural network applied to the analysis and discovery of novel proteins. J. Appl. Phys.134, 081101. doi: 10.1063/5.0157367
24
BurdukiewiczM.SidorczukK.RafaczD.PietluchF.ChilimoniukJ.RödigerS.et al. (2020). Proteomic screening for prediction and design of antimicrobial peptides with AmpGram21, 4310. doi: 10.3390/ijms21124310
25
CaiJ.YanJ.UnC.WangY.Campbell-ValoisF.-X.SiuS. W. I. (2025). BERT-AmPEP60: A BERT-based transfer learning approach to predict the minimum inhibitory concentrations of antimicrobial peptides for Escherichia coli and Staphylococcus aureus. J. Chem. Inf. Model.65, 3186–3202. doi: 10.1021/acs.jcim.4c01749
26
CamachoC.CoulourisG.AvagyanV.MaN.PapadopoulosJ.BealerK.et al. (2009). BLAST+: architecture and applications. BMC Bioinf.10, 421. doi: 10.1186/1471-2105-10-421
27
CaoQ.GeC.WangX.HarveyP. J.ZhangZ.MaY.et al. (2023). Designing antimicrobial peptides using deep learning and molecular dynamics simulations. Briefings Bioinf.24, bbad058. doi: 10.1093/bib/bbad058
28
CapecchiA.CaiX.PersonneH.KöhlerT.van DeldenC.ReymondJ.-L. (2021). Machine learning designs non-hemolytic antimicrobial peptides. JCs12, 9221–9232. doi: 10.1039/d1sc01713f
29
ChenB.QinH.LuoG. (2025). Ai-aided drug development in antimicrobial resistance: a systematic review. Int. J. Antimicrob. Agents58, 105563.
30
ChenJ.JiangY.GongK.YangY. (2026). CFlowAMP: Property controllable de novo AMP design via integrating ESM-2 with conditional flow matching. J. Mol. Graphics Modell.146, 109401. doi: 10.1016/j.jmgm.2026.109401
31
ChenN.YuJ.ZheL.WangF.LiX.WongK. C. (2024). TP-LMMSG: peptide prediction via graph neural network. Briefings Bioinf.25, bbae308. doi: 10.1093/bib/bbae308
32
ChenY.JiangX.ZhaoW. (2024). “ AMPpred-DLFF: prediction of AMPs based on deep learning and multi-view features fusion”, in: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). Piscataway, NJ: IEEE.
33
ChiuY.-P.YaoL.TangY.ChungC.-R.PangY.ChiangY.-C.et al. (2026). ANIA: An inception-attention network for predicting minimum inhibitory concentration of antimicrobial peptides. Briefings Bioinf.27, bbag023. doi: 10.1093/bib/bbag023
34
ChungC. R.ChienC. Y.TangY.WuL. C.HsuJ. B. K.LuJ. J.et al. (2024). An ensemble deep learning model for predicting minimum inhibitory concentrations of antimicrobial peptides against pathogenic bacteria. iScience27, 110718. doi: 10.1016/j.isci.2024.110718
35
ColeT. J.BrewerM. S. J. P. (2019). TOXIFY: A deep learning approach to classify animal venom proteins7. doi: 10.7287/peerj.preprints.27498v1
36
CorsoG.StärkH.JingB.BarzilayR.JaakkolaT. (2022). Diffdock: diffusion steps, twists, and turns for molecular docking. Japa.
37
DarukaL.CzikkelyM. S.SziliP.FarkasZ.BaloghD.GrézalG.et al. (2025). ESKAPE pathogens rapidly develop resistance against antibiotics in development in vitro. Nat. Microbiol.10, 313–331. doi: 10.1038/s41564-026-02267-w
38
DavisI. W.BakerD. J. (2009). RosettaLigand docking with full ligand and receptor flexibility. Jomb385, 381–392. doi: 10.1016/j.jmb.2008.11.010
39
de AlbernazD. T. F.AllendS. O.NetoA. C. P. S.SentaD. D. O. D.PintoL. D. S.KremerF. S.et al. (2025). Novel antimicrobial peptides against Pseudomonas aeruginosa: in silico design and experimental validation. J. Appl. Microbiol.136, lxaf287. doi: 10.1093/jambio/lxaf287
40
DeanS. N.AlvarezJ. A. E.ZabetakisD.WalperS. A.MalanoskiA. P. (2021). PepVAE: Variational autoencoder framework for antimicrobial peptide generation and activity prediction. Front. Microbiol.12. doi: 10.3389/fmicb.2021.725727
41
DeanS. N.WalperS. A. (2020). Variational autoencoder for generation of antimicrobial peptides. ACS Omega5, 20746–20754. doi: 10.1021/acsomega.0c00442
42
DeeW. L. (2022). LMPred: predicting antimicrobial peptides using pre-trained language models. Bioinf. Adv.2, vbac021. doi: 10.1093/bioadv/vbac021
43
DeLucaS.KharK.MeilerJ. (2015). Fully flexible docking of medium sized ligand libraries with RosettaLigand. JPo10, e0132508. doi: 10.1371/journal.pone.0132508
44
de VriesS. J.ReyJ.SchindlerC. E.ZachariasM.TufferyP. (2017). The pepATTRACT web server for blind, large-scale peptide–protein docking. NAR45, W361–W364. doi: 10.1093/nar/gkx335
45
de VriesS. J.van DijkM.BonvinA. M. J. J. (2010). The HADDOCK web server for data-driven biomolecular docking. Np5, 883–897. doi: 10.1038/nprot.2010.32
46
DonaghyA. C.MorganS. J.KaufmanG. E.MorgenrothD. C. (2020). Team approach to prosthetic prescription decision-making. Curr. Phys. Med. Rehabil. Rep.8, 386–395. doi: 10.1007/s40141-020-00289-x
47
DongR.CaoQ.SongC. J. A. S. (2025a). Painting peptides with antimicrobial potency through deep reinforcement learning. 12. doi: 10.1002/advs.202506332
48
DongR.LiuR.LiuZ.LiuY.ZhaoG.LiH.et al. (2025b). Exploring the repository of de novo-designed bifunctional antimicrobial peptides through deep learning. eLife13, RP97330. doi: 10.7554/elife.97330.3
49
DongruiZ.MiyamotoM.YokooH.DemizuY. (2024). Innovative peptide architectures: advancements in foldamers and stapled peptides for drug discovery. Expert Opin. Drug Discov.19, 699–723. doi: 10.1080/17460441.2024.2350568
50
DuZ.DingX.XuY.LiY. (2023). UniDL4BioPep: a universal deep learning architecture for binary classification in peptide bioactivity. Briefings Bioinf.24, bbad135. doi: 10.1093/bib/bbad135
51
DuanY.Santos-JúniorC. D.SchmidtT. S.FullamA.de AlmeidaB. L. S.ZhuC.et al. (2024). A catalog of small proteins from the global microbiome. Nat. Commun.15, 7563. doi: 10.1038/s41467-024-51894-6
52
Duque-SalazarG.Mendez-OtalvaroE.Ceballos-ArroyoA. M.OrduzS. (2020). Design of antimicrobial and cytolytic peptides by computational analysis of bacterial, algal, and invertebrate proteomes. Amino Acids52, 1403–1412. doi: 10.1007/s00726-020-02900-w
53
EbrahimikondoriH.SutherlandD.YanaiA.RichterA.SalehiA.LiC.et al. (2024). Structure-aware deep learning model for peptide toxicity prediction. Protein Science: A. Publ. Protein Soc.33, e5076. doi: 10.1002/pro.5076
54
FangY.XuF.WeiL.JiangY.ChenJ.WeiL.et al. (2023). AFP-MFL: accurate identification of antifungal peptides using multi-view feature learning. Briefings Bioinf.24, bbac606. doi: 10.1093/bib/bbac606
55
FarhaM. A.TuM. M.BrownE. D. (2025). Important challenges to finding new leads for new antibiotics. Curr. Opin. Microbiol.83, 102562. doi: 10.1016/j.mib.2024.102562
56
FengJ.SunM.LiuC.ZhangW.XuC.WangJ.et al. (2024). SAMP: Identifying antimicrobial peptides by an ensemble learning model based on proportionalized split amino acid composition23, 879–890. doi: 10.1093/bfgp/elae046
57
FernandesF. C.CardosoM. H.Gil-LeyA.LuchiL. V.da SilvaM. G. L.MacedoM. L. R.et al. (2023). Geometric deep learning as a potential tool for antimicrobial peptide prediction. Front. Bioinf.3, 1216362. doi: 10.3389/fbinf.2023.1216362
58
Fernández-DíazR.Cossio-PérezR.AgoniC.LamH. T.LopezV.ShieldsD. C. (2024). AutoPeptideML: a study on how to build more trustworthy peptide bioactivity predictors. JB40, btae555. doi: 10.1093/bioinformatics/btae555
59
FinnR. D.ClementsJ.EddyS. R. (2011). HMMER web server: Interactive sequence similarity searching. Nucleic Acids Res.39, W29–W37. doi: 10.1093/nar/gkr367
60
FriesnerR. A.BanksJ. L.MurphyR. B.HalgrenT. A.KlicicJ. J.MainzD. T.et al. (2004). Glide: a new approach for rapid, accurate docking and scoring. 1. Method and assessment of docking accuracy. J. Med. Chem.47, 1739–1749. doi: 10.1021/jm0306430
61
FuL.NiuB.ZhuZ.WuS.LiW. (2012). CD-HIT: Accelerated for clustering the next-generation sequencing data. Bioinformatics28, 3150–3152. doi: 10.1093/bioinformatics/bts565
62
GabereM. N.NobleW. S. J. B. (2017). Empirical comparison of web-based antimicrobial peptide prediction tools33, 1921–1929. doi: 10.1093/bioinformatics/btx081
63
GacesaR.BarlowD. J.LongP. F. J. P. C. S. (2016). Machine learning can differentiate venom toxins from other proteins having non-toxic physiological functions2. doi: 10.7717/peerj-cs.90
64
GargateN.LawsM.RahmanK. M. (2025). Current economic and regulatory challenges in developing antibiotics for Gram-negative bacteria. NPJ Antimicrobials Resistance3, 50. doi: 10.1038/s44259-025-00123-1
65
GawdeU.ChakrabortyS.WaghuF. H.BaraiR. S.KhanderkarA.IndraguruR. (2022). CAMPR4: database of natural and synthetic antimicrobial peptides. Nucleic Acids Res.51, D377–DD83. doi: 10.1093/nar/gkac933
66
GeorgoulisE.ZervouM. A.PantazisY. J. S. R. (2025). Transfer learning on protein language models improves antimicrobial peptide classification15, 37456. doi: 10.1038/s41598-025-21223-y
67
GertzE. M.YuY. K.AgarwalaR.SchäfferA. A.AltschulS. F. (2006). Composition-based statistics and translated nucleotide searches: Improving the TBLASTN module of BLAST. BMC Biol.4, 41. doi: 10.1186/1741-7007-4-41
68
GhorbaniM.PrasadS.BrooksB. R.KlaudaJ. B. (2022). Deep attention-based variational autoencoder for antimicrobial peptide discovery. doi: 10.1101/2022.07.08.499340
69
Gonzalez-de la RosaT.Herreros-IsidroA.Marquez-ParadasE.Barrera-ChamorroL.LeonM. J.Montserrat-de la PazS. (2025). Olive leaf protein hydrolysate as a novel source of antimicrobial peptides: Peptidomic characterization and in silico evaluation. SJM30, 3382. doi: 10.3390/molecules30163382
70
GuanC.FernandesF. C.FrancoO. L.de la Fuente-NunezC. (2025). Leveraging large language models for peptide antibiotic design. Cell Rep. Phys. Sci.6, 102359. doi: 10.1016/j.xcrp.2024.102359
71
GuptaS.KapoorP.ChaudharyK.GautamA.KumarR.Consortium OSDDet al. (2013). In silico approach for predicting toxicity of peptides and proteins8. doi: 10.1371/journal.pone.0073957
72
HanC.LuJ.QiuJ. (2025). iAMPM: A deep learning approach for improved recognition of antimicrobial peptides. Artif. Intell. Technol. Res.2.
73
HaoY.LiuX.FuH.ShaoX.CaiW. (2024). PGAT-ABPp: harnessing protein language models and graph attention networks for antibacterial peptide identification with remarkable accuracy. Bioinformatics40, btae497. doi: 10.1093/bioinformatics/btae497
74
HasanM. M.SChaduangratN.BasithS.LeeG.ShoombuatongW.ManavalanB. (2020). HLPpred-Fuse: Improved and robust prediction of hemolytic peptide and its activity by fusing multiple feature representation. Bioinformatics36, 3350–3356. doi: 10.1093/bioinformatics/btaa160
75
HoffmanS. C.ChenthamarakshanV.WadhawanK.ChenP. Y.DasP. (2022). Optimizing molecules using efficient queries from property evaluations. Nat. Mach. Intell.4, 21–31. doi: 10.1038/s42256-021-00422-y
76
HonoratoR. V.TrelletM. E.Jiménez-GarcíaB.SchaarschmidtJ. J.GiuliniM.ReysV.et al. (2024). The HADDOCK2.4 web server for integrative modeling of biomolecular complexes. Japa19, 3219–3241. doi: 10.1038/s41596-024-01011-0
77
HuangJ.XuY.XueY.HuangY.LiX.ChenX. (2023). Identification of potent antimicrobial peptides via a machine-learning pipeline. Nat. Biomed. Eng.7, 797–810. doi: 10.1038/s41551-022-00991-2
78
JainM.BengioE.Hernandez-GarciaA.Rector-BrooksJ.DossouB. F. P.EkboteC. A.et al. (2022). “ Biological sequence design with GFlowNets”, in: Proceedings of the 39th International Conference on Machine Learning (Baltimore, MD: PMLR).
79
JainA.KiharaD. (2019). NNTox: Gene ontology-based protein toxicity prediction using neural network. Sci. Rep.9, 17923. doi: 10.1038/s41598-019-54405-6
80
JanA.HayatM.WedyanM.AlturkiR.GazzaweF.AliH. (2022). Target-AMP: Predicting antimicrobial peptides via sequential info & evolutionary profile. Comput. Biol. Med.151, 106311.
81
JesudasonT. J. T. L. I. D. (2026). Who publishes update on global antibiotic resistance26. doi: 10.1016/s1473-3099(25)00737-6
82
JinS.ZengZ.XiongX.HuangB.TangL.WangH.et al. (2025). AMPGen: an evolutionary information-reserved and diffusion-driven generative model for de novo design of antimicrobial peptides. Commun. Biol.8, 839. doi: 10.1038/s42003-025-08282-7
83
JonesP.BinnsD.ChangH.-Y.FraserM.LiW.McAnullaC.et al. (2014). InterProScan 5: Genome-scale protein function classification. Bioinformatics30, 1236–1240. doi: 10.1093/bioinformatics/btu031
84
JorgeP.LourencoA.PereiraM. O. (2012). New trends in anti-biofilm strategies: achievements & bioinformatic approaches. Biofouling28, 1033–1061. doi: 10.1080/08927014.2012.728210
85
JoyS.NairP. S.HariharanR.PillaiM. R. (2006). Detailed comparison of the protein-ligand docking efficiencies of GOLD, a commercial package and ArgusLab, a licensable freeware. Isb6, 601–605. doi: 10.3233/isb-00268
86
JumperJ.EvansR.PritzelA.GreenT.FigurnovM.RonnebergerO.et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature596, 583–589. doi: 10.1038/s41586-021-03819-2
87
KarasevD. A.MalakhovG. S.SobolevB. N. (2024). Quantitative prediction of hemolytic activity of peptides. Comput. Toxicol.32, 100335. doi: 10.1016/j.comtox.2024.100335
88
KelleyM. L.SailaR. K.LiuZ.PetriC. A.AbeykoonA. (2024). PanCleave: protease cleavage prediction to evaluate AMP stability. Nat. Methods21, 310–318.
89
KhorramfardA.PirgaziJ.Ghanbari SorkhiA.PourseifM. M. (2026). AMP-distillation: A knowledge distillation framework for accurate and efficient antimicrobial peptide prediction. Comput. Biol. Chem.124, 109129. doi: 10.1016/j.compbiolchem.2026.109129
90
KoesD. R.BaumgartnerM. P.CamachoC. J. (2013). Lessons learned in empirical scoring with smina from the CSAR 2011 benchmarking exercise. Joci Modeling53, 1893–1907. doi: 10.1021/ci300604z
91
KumarV.KumarR.AgrawalP.PatiyalS.RaghavaG. P. S. (2020). A method for predicting hemolytic potency of chemically modified peptides from its structure. Front. Pharmacol.11, 54. doi: 10.3389/fphar.2020.00054
92
KurcinskiM.JamrozM.BlaszczykM.KolinskiA.KmiecikS. (2015). CABS-dock web server for the flexible docking of peptides to proteins without prior knowledge of the binding site. Nucleic Acids Res.43, W419–W424. doi: 10.1093/nar/gkv456
93
LaiC. W.LinC. Y.TsaiM. C.ChenW. J.HsiehC. C.LinZ. J. (2025). Antimicrobial peptides engineered by GANs to combat resistant bacteria. Chem. Eng. J.519, 164905. doi: 10.1016/j.cej.2025.164905
94
LaiL.LiuY.SongB.LiK.ZengX. (2025). Deep generative models for therapeutic peptide discovery: A comprehensive review. ACM Comput. Surv.57. doi: 10.1145/3714455
95
LangP. T.BrozellS. R.MukherjeeS.PettersenE. F.MengE. C.ThomasV.et al. (2009). DOCK 6: combining techniques to model RNA–small molecule complexes. Japa15, 1219–1230. doi: 10.1261/rna.1563609
96
LawrenceT. J.CarperD. L.SpanglerM. K.CarrellA. A.RushT. A.MinterS. J. (2021). amPEPpy 1.0: a portable and accurate antimicrobial peptide prediction tool. Bioinformatics37, 2058–2064. doi: 10.1093/bioinformatics/btaa917
97
LeeH.HeoL.LeeM. S.SeokC. (2015). GalaxyPepDock: a protein-peptide docking tool based on interaction similarity and energy optimization. Nucleic Acids Res.43, W431–W435. doi: 10.1093/nar/gkv495
98
LeeH.LeeS.LeeI.NamH. (2023). AMP‐BERT: prediction of antimicrobial peptide function via BERT model. Protein Sci.32, e4529. doi: 10.1145/1006091.1006094
99
LemmonG.MeilerJ. (2012). “ Rosetta ligand docking with flexible XML protocols,” in Computational Drug Discovery and Design. Ed. BaronR. ( Springer New York, New York, NY), 143–155.
100
LiC.BirolI.WarrenR. L. (2023). AMPlify methods in bioinformatics and statistical data processing. BMC Res. Notes16, 11.
101
LiF.ChenJ.LeierA.Marquez-LagoT.LiuQ.WangY.et al. (2020a). DeepCleave: a deep learning predictor for caspase and matrix metalloprotease substrates and cleavage sites. Bioinformatics36, 1057–1065. doi: 10.1093/bioinformatics/btz721
102
LiW.HuangB.GuoM.ZengZ.CaiT.FengL.et al. (2025). Unveiling the evolution of antimicrobial peptides in gut microbes via foundation-model-powered framework. Cell Rep.44, 115773. doi: 10.1016/j.celrep.2025.115773
103
LiF.LeierA.LiuQ.WangY.XiangD.AkutsuT.et al. (2020b). Procleave: predicting protease-specific substrate cleavage sites by combining sequence and structural information. Genomics Proteomics Bioinf.18, 52–64. doi: 10.1016/j.gpb.2019.08.002
104
LiY.LiX.LiuY.YaoY.HuangG. (2022). MPMABP: CNN & Bi-LSTM-based prediction of multi-activities of bioactive peptides. Pharmaceuticals15, 707.
105
LiY.LiL.WangS.TangX. (2023). EQUIBIND: a geometric deep learning-based protein-ligand binding prediction method. Drug Discoveries Ther.17, 363–364. doi: 10.5582/ddt.2023.01063
106
LiT.RenX.LuoX.WangZ.LiZ.LuoX.et al. (2024). A foundation model identifies broad-spectrum antimicrobial peptides against drug-resistant bacterial infection. Nat. Commun.15, 7538. doi: 10.1038/s41467-024-51933-2
107
LiC.SutherlandD.HammondS. A.YangC.TahoF.BergmanL.et al. (2022). AMPlify: attentive deep learning model for discovery of novel antimicrobial peptides effective against WHO priority pathogens23, 77. doi: 10.3390/ph15060707
108
LiF.WangC.GuoX.AkutsuT.WebbG. I.CoinL. J. M.et al. (2023). ProsperousPlus: a one-stop and comprehensive platform for accurate protease-specific substrate cleavage prediction and machine-learning model construction. Briefings Bioinf.24. doi: 10.1093/bib/bbad372
109
LiG.WangL.LuoJ.LiangC. (2025). Antimicrobial peptide prediction based on contrastive learning and gated convolutional neural network. Sci. Rep.15, 45274. doi: 10.1038/s41598-025-29666-z
110
LiY.XuX.ZhangX.XuZ.ZhaoJ.ZhuR.et al. (2025). BroadAMP-GPT: AI-driven generation of broad-spectrum antimicrobial peptides for combating multidrug-resistant ESKAPE pathogens. Gut Microbes17, 2523811. doi: 10.1080/19490976.2025.2523811
111
LiC.YanS.DaiE. (2025). Unizyme: a unified protein cleavage site predictor enhanced with enzyme active-site knowledge. Japa. doi: 10.52202/085713-0624
112
LiX.ZhaoW.LiangX.ZhuoX.YuS.LiangG. (2025). HemPepPred: Quantitative prediction of peptide hemolytic activity based on machine learning and protein language model–derived features14, 4143. doi: 10.3390/foods14234143
113
LiC.ZouQ.JiaC.ZhengJ. (2024). AMPpred-MFA: interpretable antimicrobial peptide predictor using stacking architecture and multihead attention. J. Chem. Inf. Model.64, 2393–2404. doi: 10.1021/acs.jcim.3c01017
114
LinC.XiongS.CuiF.ZhangZ.ShiH.WeiL. (2025). Deep learning in antimicrobial peptide prediction. J. Chem. Inf. Model. doi: 10.1021/acs.jcim.5c00530
115
LinC.XiongS.LiJ.CuiF.ZhangZ.ShiH.et al. (2026). PepGraphormer: an ESM-GAT hybrid deep learning framework for antimicrobial peptide prediction. J. Cheminf.18, 15. doi: 10.1186/s13321-025-01144-8
116
LinW.XuD. (2016). Imbalanced multi-label learning for identifying antimicrobial peptides. Bioinformatics32, 3745–3752. doi: 10.1093/bioinformatics/btw560
117
LinT. T.YangL. Y.LinC. Y.WangC. T.LaiC. W.KoC. F.et al. (2023). Intelligent de novo design of novel antimicrobial peptides against antibiotic-resistant bacterial strains. Int. J. Mol. Sci.24, 6788. doi: 10.3390/ijms24076788
118
LinT. T.YangL. Y.LuI. H.ChengW. C.HsuZ. R.ChenS. H. (2021). AI4AMP: Antimicrobial peptide predictor using physicochemical encoding and deep learning. mSystems6, e00299-21. doi: 10.1128/msystems.00299-21
119
LipskyB. A.HolroydK. J.ZasloffM. (2008). Topical versus systemic antimicrobial therapy for treating mildly infected diabetic foot ulcers: a randomized, controlled, double-blinded, multicenter trial of pexiganan cream. Clin. Infect. Dis. An. Off. Publ. Infect. Dis. Soc. America47, 1537–1545. doi: 10.1086/593185
120
LiuL.WangC.ZhangM.ZhangZ.WuY.ZhangY. (2022). An efficient evaluation system accelerates α-helical antimicrobial peptide discovery and its application to global human genome mining13, 870361. doi: 10.3389/fmicb.2022.870361
121
LiuY.ZhangX.LiuY.SuY.ZengX.YenG. G. (2023). Evolutionary multi-objective optimization in searching for various antimicrobial peptides [Feature. IEEE Comput. Intell. Mag.18, 31–45. doi: 10.1109/mci.2023.3245731
122
LobanovM. Y.SlizenM. V.DovidchenkoN. V.PanfilovA. V.SurinA. A.LikhachevI. V.et al. (2024). Comparison of deep learning models with simple method to assess the problem of antimicrobial peptides prediction. Mol. Inf.43, e202200181. doi: 10.1002/minf.202200181
123
LondonN.RavehB.CohenE.FathiG.Schueler-FurmanO. (2011). Rosetta FlexPepDock web server—high resolution modeling of peptide–protein interactions. Nar39, W249–W253. doi: 10.1093/nar/gkr431
124
LuW.WuQ.ZhangJ.RaoJ.LiC.ZhengS. (2022). “ TANKBind: trigonometry-aware neural networks for drug-protein binding structure prediction”, in: Advances in Neural Information Processing Systems 35 (New Orleans, LA: Curran Associates Inc). Article 525.
125
LuoZ.GengA.WeiL.ZouQ.CuiF.ZhangZ. (2025). CPL-Diff: A diffusion model for de novo design of functional peptide sequences with fixed length. Adv. Sci.12, e2412926. doi: 10.1002/advs.202412926
126
LvH.YanK.GuoY.ZouQ.HeshamA.-L.LiuB. (2022). AMPpred-EL: an effective antimicrobial peptide prediction model based on ensemble learning146, 105577. doi: 10.1016/j.compbiomed.2022.105577
127
MaY.GuoZ.XiaB.ZhangY.LiuX.YuY. (2022). Identification of antimicrobial peptides from the human gut microbiome using deep learning. Nat. Biotechnol.40, 921–931. doi: 10.1038/s41587-022-01226-0
128
MaaschJ. R. M. A.TorresM. D. T.MeloM. C. R.de la Fuente-NunezC. (2023). Molecular de-extinction of ancient antimicrobial peptides. Cell. Host Microbe31, 1260. doi: 10.1016/j.chom.2023.07.001
129
MananA.BaekE.IlyasS.LeeD. (2025). Digital alchemy: the rise of machine and deep learning in small-molecule drug discovery26, 6807. doi: 10.3390/ijms26146807
130
MaoJ.GuanS.ChenY.ZebA.SunQ.LuR.et al. (2023). Application of a deep generative model produces novel and diverse functional peptides against microbial resistance21, 463–471. doi: 10.1016/j.csbj.2022.12.029
131
MartinsF. G.SantosH. A.SousaS. F. (2026). A review of current computational tools for peptide–protein docking. JoCC47, e70328. doi: 10.1002/jcc.70328
132
MathurD.SinghS.MehtaA.AgrawalP.RaghavaG. P. J. P. O. (2018). In silico approaches for predicting the half-life of natural and modified peptides in blood13. doi: 10.1371/journal.pone.0196829
133
MayrA.KlambauerG.UnterthinerT.HochreiterS. (2016). DeepTox: Toxicity prediction using deep learning3, 2015. doi: 10.1016/j.toxlet.2017.07.175
134
MengQ.ChenG.LinB.ZhengS.LinY.TangJ.et al. (2024). DMAMP: A deep-learning model for detecting antimicrobial peptides and their multi-activities21, 2025–2034. doi: 10.1109/tcbb.2024.3439541
135
MengH. J. P.ProteinsA. (2025). Ai-driven discovery and design of antimicrobial peptides: progress, challenges, and opportunities1–23. doi: 10.1007/s12602-025-10856-0
136
Mera-BangueroC.OrduzS.CardonaP.OrregoA.Munoz-PerezJ.Branch-BedoyaJ. W. J. A. D. A. B. D. C. (2024). AmpClass: an antimicrobial peptide predictor based on supervised machine learning96. doi: 10.1590/0001-3765202420230756
137
MitraU.SinghA. K. (2023). “ AMPpred-CNN: prediction of antimicrobial peptide by using 1D convolution neural network and composition/transition/distribution (CTD) encoding”, in: Soft Computing: Theories and Applications: Proceedings of SoCTA 2023, Volume 2 ( Springer).
138
MittalA.TiwariS.BurrellR.DubeyS. (2025). Computational modelling in AMP clinical translation. Adv. Drug Delivery Rev.205, 115655.
139
MonsalveD.MesaA.MiraL. M.MeraC.OrduzS.Branch-BedoyaJ. W. (2024). Antimicrobial peptides designed by computational analysis of proteomes. Antonie van Leeuwenhoek117, 55. doi: 10.1007/s10482-024-01946-0
140
MorettaA.SalviaR.ScieuzoC.Di SommaA.VogelH.PucciP.et al. (2020). A bioinformatic study of antimicrobial peptides identified in the Black Soldier Fly (BSF) Hermetia illucens (Diptera: Stratiomyidae). Sci. Rep.10, 16875. doi: 10.1038/s41598-020-74017-9
141
MorrisG. M.HueyR.LindstromW.SannerM. F.BelewR. K.GoodsellD. S.et al. (2009). AutoDock4 and AutoDockTools4: automated docking with selective receptor flexibility. Jomb30, 2785–2791. doi: 10.1002/jcc.21256
142
MurakamiY.IshidaS.DemizuY.TerayamaK. (2023). Design of antimicrobial peptides containing non-proteinogenic amino acids using multi-objective Bayesian optimisation. Digital Discov.2, 1347–1353. doi: 10.1039/d3dd00090g
143
MusaI. (2024). Evaluation of antibiotic potential of metabolic peptides using in silico approach. Egyptian J. Sheep Goats Sci.19, 1–4. doi: 10.21608/ejsgs.2024.335711.1013
144
MusinK.AsyanovaE. (2025). How machine learning helps in combating antimicrobial resistance: a review of AMP analysis and generation methods. Int. J. Pept. Res. Ther.31, 1–14. doi: 10.1007/s10989-025-10716-z
145
National Center for Biotechnology I (2026). BLAST: basic local alignment search tool. doi: 10.32388/rhq6vj
146
NevilleB. A.LawleyT. D. (2025). Mining microbiomes for microproteins. JNRM23, 146. doi: 10.1038/s41579-025-01154-1
147
NguyenQ. H.Nguyen‐VoT. H.DoT. T.NguyenB. P. (2024). An efficient hybrid deep learning architecture for predicting short antimicrobial peptides. Pharmacology24, 2300382. doi: 10.1002/pmic.202300382
148
NoorA.SyedF.AliS. (2025). Emerging paradigms in AI-based peptide therapeutics. Cell Rep. Med.6, 100141.
149
NorvellD. C.SuckowB. D.WebsterJ. B.LandryG.HendersonA. W.TwineC. P. (2021). Development and usability of AMPREDICT decision support tool. Eur. J. Vasc. Endovascular Surg.62, 304–311. doi: 10.26226/morressier.614ca2d887a68d83cb5d5cdc
150
Oliveira JúniorN. G.SouzaC. M.BucciniD. F.CardosoM. H.FrancoO. L. J. N. R. M. (2025). Antimicrobial peptides: structure, functions and translational applications23, 687–700. doi: 10.1038/s41579-025-01200-y
151
OrsiM.ReymondJ.-L. (2024). Can large language models predict antimicrobial peptide activity and toxicity? RSC Med. Chem.15, 2030–2036. doi: 10.1039/d4md00159a
152
PalmoT.JamwalV.KumariD.SinghK. J. R. M. C. (2026). Harnessing biological insights to accelerate drug discovery against ESKAPE pathogens. 174, 109682. doi: 10.1039/d5md00358j
153
PandeyA.LeninR. R.PatiyalS.AgrawalP. J. P.ProteinsA. (2025). High throughput meta-analysis of antimicrobial peptides for characterizing class specific therapeutic candidates: an in silico approach17, 1–21. doi: 10.1007/s12602-025-10596-1
154
PandeyP.SrivastavaA. (2025). sAMP-VGG16: force-field assisted image-based deep neural network prediction model for short antimicrobial peptides93, 372–383. doi: 10.1101/2023.06.04.543607
155
PandiA.AdamD.ZareA.TrinhV. T.SchaeferS. L.BurtM.et al. (2023). Cell-free biosynthesis combined with deep learning accelerates de novo development of antimicrobial peptides. Nat. Commun.14, 7197. doi: 10.1038/s41467-023-42434-9
156
PangY.YaoL.XuJ.WangZ.LeeT.-Y. (2022). Integrating transformer and imbalanced multi-label learning to identify antimicrobial peptides and their functional activities. Bioinformatics38, 5368–5374. doi: 10.1093/bioinformatics/btac711
157
PanjlaA.JoshiS.SinghG.BamfordS. E.MechlerA.VermaS. J. C. A. A. J. (2024). Applying machine learning for antibiotic development and prediction of microbial resistance19. doi: 10.1002/asia.202400102
158
PerišićO.Sevim BayrakC.GunadyM. K. J. F. I. M. B. (2025). Machine learning in computer-aided drug design. Front. Media SA12, 1568437. doi: 10.3389/fmolb.2025.1568437
159
PlissonF.Ramírez-SánchezO.Martínez-HernándezC. (2020). Machine learning-guided discovery and design of non-hemolytic peptides. Sci. Rep.10, 16581. doi: 10.1038/s41598-020-73644-6
160
QiaoZ.NieW.VahdatA.MillerT. F.AnandkumarA. (2024). State-specific protein–ligand complex structure prediction with a multiscale deep generative model. Nat. Mach. Intell.6, 195–208. doi: 10.1038/s42256-024-00792-z
161
RajakumarP.Mohammed Abdul RasheedA.BaluP. M.MurugesanK. (2006). Synthesis, characterization, and anti-bacterial efficacy of some novel cyclophane amide. Bioorg. Med. Chem.14, 7458–7467. doi: 10.1016/j.bmc.2006.07.020
162
RamaziS.MohammadiN.AllahverdiA.KhaliliE.AbdolmalekiP. J. D. (2022). A review on antimicrobial peptides databases and the computational tools2022. doi: 10.1093/database/baac011
163
Ramos-LlorensM.Bello-MadrugaR.ValleJ.AndreuD.TorrentM. (2024). PyAMPA: a high-throughput prediction and optimization tool for antimicrobial peptides9. doi: 10.1128/msystems.01358-23
164
RathoreA. S.KumarN.ChoudhuryS.MehtaN. K.RaghavaG. P. S. (2025). Prediction of hemolytic peptides and their hemolytic concentration. Commun. Biol.8, 176. doi: 10.1038/s42003-025-07615-w
165
RavehB.LondonN.ZimmermanL.Schueler-FurmanO. (2011). Rosetta FlexPepDock ab-initio: simultaneous folding, docking and refinement of peptides onto their receptors. Po6, e18934. doi: 10.1371/journal.pone.0018934
166
RenaudS.MansbachR. A. (2023). Latent spaces for antimicrobial peptide design. Digital Discov.2, 441–458. doi: 10.1039/d2dd00091a
167
RivesA.MeierJ.SercuT.GoyalS.LinZ.LiuJ.et al. (2021). Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences118. doi: 10.1073/pnas.2016239118
168
RuffoloJ. A.MadaniA. (2024). Designing proteins with language models. Nat. Biotechnol.42, 200–212. doi: 10.1038/s41587-024-02123-4
169
SadeeqM.LiY.WangC.HouF.ZuoJ.XiongP. J. F. I. C.et al. (2025). Unlocking the power of antimicrobial peptides: advances in production, optimization, and therapeutics15, 1528583. doi: 10.3389/fcimb.2025.1528583
170
SaleemN.KumarN.El-OmarE.WillcoxM.JiangX. T. (2025). Harnessing machine learning approaches for the identification, characterization, and optimization of novel antimicrobial peptides. Antibiotics (Basel Switzerland)14, 1263. doi: 10.3390/antibiotics14121263
171
SaleemN.KumarN.El-OmarE.WillcoxM.JiangX.-T. (2026). Nano-antimicrobial peptides (Nano-AMPs) to combat resistant gram-negative bacteria. Drug Delivery Trans. Res. doi: 10.1007/s13346-026-02085-x
172
SalemM.Keshavarzi ArshadiA.YuanJ. S. (2022). AMPDeep: Hemolytic activity prediction of antimicrobial peptides using transfer learning. BMC Bioinf.23, 389. doi: 10.21203/rs.3.rs-1615895/v1
173
SalimiA.LeeJ. Y. (2026). CoLPAT-AMP: A transformer-based framework for designing novel antimicrobial peptides with property awareness and partially controllable length. Expert Syst. Appl.318, 131999. doi: 10.1016/j.eswa.2026.131999
174
Santos-JúniorC. D.TorresM. D. T.DuanY.Rodríguez del RíoÁ.SchmidtT. S. B.ChongH.et al. (2024). Discovery of antimicrobial peptides in the global microbiome with machine learning. Cell.187, 3761–78.e16. doi: 10.1016/j.cell.2024.05.013
175
Santos-MartinsD.Solis-VasquezL.TillackA. F.SannerM. F.KochA.ForliS. (2021). Accelerating AutoDock4 with GPUs and gradient-based local search. J. Chem. Theory Comput.17, 1060–1073. doi: 10.26434/chemrxiv.9702389.v1
176
SchindlerC. E.de VriesS. J.ZachariasM. (2015). Fully blind peptide-protein docking with pepATTRACT. JS23, 1507–1515. doi: 10.1016/j.str.2015.05.021
177
SelvarajV.SudhakarS.SekaranS. (2025). Global trends and projections in antimicrobial resistance. Lancet405, 1906–1907. doi: 10.1016/s0140-6736(25)00810-4
178
SharmaN.NaoremL. D.JainS.RaghavaG. P. S. (2022). ToxinPred2: An improved method for predicting toxicity of proteins. Briefings Bioinf.23, bbac174. doi: 10.1093/bib/bbac174
179
SharmaR.ShrivastavaS.SinghS.KumarA.SinghA. K.SaxenaS. (2024). AI-based model for predicting MIC of antibacterial peptides. IEEE J. Biomed. Health Inf.28, 1949–1965. doi: 10.1109/JBHI.2023.3271611
180
ShengF.NoaeenM.ShakeriZ. (2026). ProDCARL: Reinforcement learning-aligned diffusion models for de novo antimicrobial peptide design. Japa.
181
ShinM. K.LeeB.KimS. T.YooJ. S.SungJ.-S. (2022). Designing a novel functional peptide with dual antimicrobial and anti-inflammatory activities via in silico methods13, 821070. doi: 10.3389/fimmu.2022.821070
182
SinghV.ShrivastavaS.Kumar SinghS.KumarA.SaxenaS. (2022). StaBle-ABPpred: stacked ensemble predictor for antibacterial peptides. Briefings Bioinf.23, bbab439. doi: 10.1093/bib/bbab439
183
SongJ.LiF.LeierA.Marquez-LagoT. T.AkutsuT.HaffariG.et al. (2018). PROSPERous: high-throughput prediction of substrate cleavage sites for 90 proteases with improved accuracy. Bioinformatics34, 684–687. doi: 10.1093/bioinformatics/btx670
184
SongJ.TanH.PerryA. J.AkutsuT.WebbG. I.WhisstockJ. C.et al. (2012). PROSPER: An integrated feature-based tool for predicting protease substrate cleavage sites7. doi: 10.1371/journal.pone.0050300
185
SowersA.WangJ.HuG.XingM.LiB. (2026). Design and classify innovative antimicrobial and dual- and multi-functional peptides using generative artificial intelligence. Sci. Rep. 16, 23365. doi: 10.1038/s41598-026-49612-x
186
SpitzerR.JainA. N. (2012). Surflex-Dock: docking benchmarks and real-world application. Joc-amd26, 687–699. doi: 10.1007/s10822-011-9533-y
187
SteineggerM.SödingJ. (2017). MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. Biotechnol.35, 1026–1028. doi: 10.1038/nbt.3988
188
StokesJ. M.YangK.SwansonK.JinW.Cubillos-RuizA.DonghiaN. M.et al. (2020). A deep learning approach to antibiotic discovery180, 688–702. doi: 10.1016/j.cell.2020.04.001
189
SuX.XuJ.YinY.QuanX.ZhangH. (2019). Antimicrobial peptide identification using multi-scale convolutional network20, 730. doi: 10.1186/s12859-019-3327-y
190
SuZ.YuH.LvT.ChenQ.LuoH.ZhangH. (2025). Progress in the classification, optimization, activity, and application of antimicrobial peptides. Front. Microbiol. 16. doi: 10.3389/fmicb.2025.1582863
191
SultanM. F.ShaonM. S. H.KarimT.AliM. M.HasanM. Z.AhmedK.et al. (2024). MLAFP-XN: leveraging neural network model for development of antifungal peptide identification tool. Heliyon10, e37820. doi: 10.1016/j.heliyon.2024.e37820
192
SunS. (2025). Progress in the identification and design of novel antimicrobial peptides against pathogenic microorganisms. Probiotics Antimicrob. Proteins17, 918–936. doi: 10.1007/s12602-024-10402-4
193
SunZ.XuJ.ZhangY.ZhangY.WangZ.WangX. (2025). Multimodal geometric learning for antimicrobial peptide identification. Briefings Bioinf.26, bbaf261. doi: 10.1093/bib/bbaf261
194
SunK.ZhouR.ShiM. (2025). Using ML pipelines for AMP hit optimization. Drug Discov. Today30, 402–415.
195
SzymczakP.MozejkoM.GrzegorzekT.JurczakR.BauerM.NeubauerD.et al. (2023). Discovering highly potent antimicrobial peptides with a deep generative model, HydrAMP. Nat. Commun.14, 1453. doi: 10.1038/s41467-023-36994-z
196
SzymczakP.ZarzeckiW.WangJ.DuanY.WangJ.CoelhoL. P.et al. (2025). Ai-driven antimicrobial peptide discovery: mining and generation58, 1453–1871. doi: 10.1021/acs.accounts.0c00594
197
TangW.DaiR.YanW.ZhangW.BinY.XiaE. (2022). Identifying multifunctional bioactive peptide functions using multi-label deep learning. Briefings Bioinf.23, bbab414. doi: 10.1093/bib/bbab414
198
TimmonsP. B.HewageC. M. (2020). HAPPENN is a novel tool for hemolytic activity prediction for therapeutic peptides which employs neural networks. Sci. Rep.10, 10869. doi: 10.1038/s41598-020-67701-3
199
TorresM. D. T.de la Fuente-NunezC.RezaeiM.MyintM. (2025). Antimicrobial peptides: a renaissance in the AI era. Cell Rep. Med.6, 100124.
200
TorresM. D. T.ZengY.WanF.MausN.GardnerJ.de la Fuente-NunezC. (2024). A generative artificial intelligence approach for antibiotic optimization. doi: 10.1101/2024.11.27.625757
201
TrottiA.GardenA.WardeP.SymondsP.LangerC.RedmanR.et al. (2004). A multinational, randomized phase III trial of iseganan HCl oral solution for reducing the severity of oral mucositis in patients receiving radiotherapy for head-and-neck Malignancy. Int. J. Radiat. Oncol. Biol. Phys.58, 674–681. doi: 10.1016/s0360-3016(03)01627-4
202
TsabanT.VargaJ. K.AvrahamO.Ben-AharonZ.KhramushinA.Schueler-FurmanO. (2022). Harnessing protein folding neural networks for peptide–protein docking. Nc13, 176. doi: 10.1038/s41467-021-27838-9
203
TsaiC.-T.LinC.-W.YeG.-L.WuS.-C.YaoP.LinC.-T.et al. (2024). Accelerating antimicrobial peptide discovery for who priority pathogens through predictive and interpretable machine learning models9, 9357–9374. doi: 10.1021/acsomega.3c08676
204
TuralićA.HegedüsZ.AltieriF.MartinekT. A.ĐeđibegovićJ. J. (2025). Discordance between in silico and in vitro results on proteolytic stability of milk protein-derived DPP-4 inhibitory peptides. Efr Technol.251, 3199–3214. doi: 10.1007/s00217-025-04833-8
205
van KempenM.KimS. S.TumescheitC.MirditaM.LeeJ.GilchristC. L. M.et al. (2024). Fast and accurate protein structure search with Foldseek. Nat. Biotechnol.42, 243–246. doi: 10.1038/s41587-023-01773-0
206
Van OortC. M.FerrellJ. B.RemingtonJ. M.WshahS.LiJ. (2021). AMPGAN v2: Machine learning-guided design of antimicrobial peptides. J. Chem. Inf. Model.61, 2198–2207. doi: 10.1021/acs.jcim.0c01441
207
VasconcelosA. A.SilvaA.de SouzaC.CruzA.SilvaM. (2024). GAN-for-AMP design: Generative adversarial network for de novo antimicrobial peptide generation. iScience27, 109876.
208
VeldsmanW. P.ZhangQ.ZhaoQ.ZhangL.YuanjieZ. (2025). The antimicrobial peptide pipeline: a bacteria-centric AMP predictor. Curr. Gene Ther. 25(5), 786–792. doi: 10.2174/0115665232343790250120071445
209
VeltriD.KamathU.ShehuA. J. (2018). Deep learning improves antimicrobial peptide recognition34, 2740–2747. doi: 10.1093/bioinformatics/bty179
210
VerdonkM. L.ColeJ. C.HartshornM. J.MurrayC. W.TaylorR. D. (2003). Improved protein–ligand docking using GOLD. PS Function Bioinf.52, 609–623. doi: 10.1002/prot.10465
211
Villegas-MorcilloA.MakrodimitrisS.van HamR. C. H. J.GomezA. M.SanchezV.ReindersM. J. T. (2020). Unsupervised protein embeddings outperform hand-crafted sequence and structure features at predicting molecular function. Bioinformatics37, 162–170. doi: 10.1093/bioinformatics/btaa701
212
VoitsitskyiT.KoleievI.StratiichukR.KotO.KyrylenkoR.SavchenkoI.et al. (2026). ArtiDock: accurate machine learning approach to protein–ligand docking optimized for high-throughput virtual screening. J. Chem. Inf. Model.66, 2117–2128. doi: 10.1021/acs.jcim.5c02777
213
WaghuF. H.BaraiR. S.GurungP.Idicula-ThomasS. (2016). CAMPR3: database of antimicrobial peptides. Nucleic Acids Res.44, D1094–DD100. doi: 10.1093/nar/gkv1051
214
WaghuF. H.GopiL.BaraiR. S.RamtekeP.NizamiB.Idicula-ThomasS. J. (2014). CAMP: Collection of sequences and structures of antimicrobial peptides42, D1154–D11D8. doi: 10.1093/nar/gkt1157
215
WaghuF. H.Idicula-ThomasS. (2020). Collection of antimicrobial peptides database and its derivatives: applications and beyond. Protein Sci. A. Publ. Protein Soc.29, 36–42. doi: 10.1002/pro.3714
216
WanF.TorresM. D.PengJ.de la Fuente-NunezC. (2024). Deep-learning-enabled antibiotic discovery through molecular de-extinction. Nat. Biomed. Eng.8, 854–871. doi: 10.1038/s41551-024-01201-x
217
WangG. J. P. S. (2023). The antimicrobial peptide database is 20 years old: recent developments and future directions32. doi: 10.1002/pro.4778
218
WangS. C. (2024). E-CLEAP: ensemble learning model for AMP identification. PloS One19, e0300125. doi: 10.1371/journal.pone.0300125
219
WangJ.FengJ.KangY.PanP.GeJ.WangY. (2025). Discovery of antimicrobial peptides with notable antibacterial potency by an LLM-based foundation model11. doi: 10.1126/sciadv.ads8932
220
WangL.FuX.YeX.SakuraiT.ZengX.LiuY. (2025a). PKAN: leveraging Kolmogorov-Arnold networks and multi-modal learning for peptide prediction with advanced language models. IEEE J. Biomed. Health Inf. doi: 10.1109/jbhi.2025.3561846
221
WangC.GarlickS.ZlohM. (2021). Deep learning for novel antimicrobial peptide design11, 471. doi: 10.3390/biom11030471
222
WangY.GongH.LiX.LiL.ZhaoY.BaoP.et al. (2024a). De novo multi-mechanism antimicrobial peptide design via multimodal deep learning. doi: 10.1101/2024.01.02.573846
223
WangY.LiangV.YinN.LiuS.SegalE. (2026). SGAC: a graph neural network framework for imbalanced and structure-aware AMP classification. Briefings Bioinf.27. doi: 10.1093/bib/bbag038
224
WangB.LinP.ZhongY.TanX.ShenY.HuangY. (2025). Explainable deep learning, and virtual evolution identifies anti-microbial peptides with activity against multidrug-resistant human pathogens. Nat. Microbiol.10, 332–347. doi: 10.1038/s41564-024-01907-3
225
WangL.LiuY.FuX.YeX.ShiJ.YenG. G.et al. (2025b). HMAMP: Designing highly potent antimicrobial peptides using a hypervolume-driven multiobjective deep generative model. J. Med. Chem.68, 8346–8360. doi: 10.1021/acs.jmedchem.4c03073
226
WangY.LiuX.HuangF.XiongZ.ZhangW.et al. (2024b). “ A multi-modal contrastive diffusion model for therapeutic peptide generation”, in: Proceedings of the AAAI Conference on Artificial Intelligence. Washington, DC: AAAI Press.
227
WangZ.LiuT.XuG.GaoH.ZhangR.ZhangH.et al. (2025). MPCutter: predicting protease-specific substrate cleavage sites using a protein language model. Genomics Proteomics Bioinf. doi: 10.1093/gpbjnl/qzaf130
228
WangG.SchmidtC.LiX.WangZ. J. N. A. R. (2026). APD6: the antimicrobial peptide database is expanded to promote research and development by deploying an unprecedented information pipeline54, D363–DD74. doi: 10.1093/nar/gkaf860
229
WangX. F.TangJ. Y.SunJ.DorjeS.SunT. Q.PengB.et al. (2024). ProT-Diff: A modularized and efficient strategy for de novo generation of antimicrobial peptide sequences by integrating protein language and diffusion models. Adv. Sci.11, e2406305. doi: 10.1002/advs.202406305
230
WangR.WangT.ZhuoL.WeiJ.FuX.ZouQ.et al. (2024). Diff-AMP: tailored-designed antimicrobial peptide framework with all-in-one generation, identification, prediction, and optimization. Briefings Bioinf.25. doi: 10.1093/bib/bbae078
231
WangD.WenZ.YeF.LiL.ZhouH. (2023). “ Accelerating antimicrobial peptide discovery with latent structure”, in: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Long Beach, CA: Association for Computing Machinery), 2243–2255.
232
WangX.WuZ.WangR.GaoX. (2024). UniproLcad: accurate identification of antimicrobial peptide by fusing multiple pre-trained protein language models16, 464. doi: 10.3390/sym16040464
233
WangY.YeX.FengY.WangH.LinX.MaX.et al. (2025c). BBATProt: a framework predicting biological function with enhanced feature extraction via interpretable deep learning26. doi: 10.1093/bib/bbaf593
234
WangY.ZhaoL.LiZ.XiY.PanY.ZhaoG.et al. (2025d). A generative artificial intelligence approach for the discovery of antimicrobial peptides against multidrug-resistant bacteria. Nat. Microbiol.10, 2997–3012. doi: 10.1038/s41564-025-02114-4
235
WeiL.YeX.XueY.SakuraiT.WeiL. (2021). ATSE: A peptide toxicity predictor by exploiting structural and evolutionary information based on graph neural network and attention mechanism. Briefings Bioinf.22. doi: 10.1093/bib/bbab041
236
WeiL.ZhouC.SuR.ZouQ. (2019). PEPred-Suite: Improved prediction of therapeutic peptides. Bioinformatics35, 4272–4280. doi: 10.1093/bioinformatics/btz246
237
WengG.GaoJ.WangZ.WangE.HuX.YaoX.et al. (2020). Comprehensive evaluation of fourteen docking programs on protein–peptide complexes. Japa16, 3959–3969. doi: 10.1021/acs.jctc.9b01208
238
WinT. S.MalikA. A.PrachayasittikulV.JES. W.NantasenamatC.ShoombuatongW. (2017). HemoPred: A web server for predicting the hemolytic activity of peptides. Future Med. Chem.9, 275–291. doi: 10.4155/fmc-2016-0188
239
WuZ.GuoX.SunY.SuX.ZhaoJ. (2024). TF-BAPred: a universal bioactive peptide predictor integrating multiple feature representations. Mathematics12, 3618. doi: 10.3390/math12223618
240
WuX.WangZ.LiX.FanY.HeG.WanY.et al. (2014). In vitro and <i>in vivo</i> activities of antimicrobial peptides developed using an amino acid-based activity prediction method58, 5342–5349.
241
WuK.XuG.TianY.LiG.YiZ.TangX. (2025). Synthesis and evaluation of aquatic antimicrobial peptides derived from marine metagenomes using a high-throughput screening approach. JMD23, 178. doi: 10.3390/md23040178
242
XiaoL.LiJ.ChenY.QiuC.WangQ.LiaoX.et al. (2026). PepGen-FB: Multi-functional antimicrobial peptide generation via curriculum learning-guided multi-objective feedback optimization. Artif. Intell. Chem.4, 100119. doi: 10.1016/j.aichem.2026.100119
243
XiaoX.ShaoY.-T.ChengX.StamatovicB. J. (2021). iAMP-CA2L: a new CNN-BiLSTM-SVM classifier based on cellular automata image for identifying antimicrobial peptides and their functional types22. doi: 10.1093/bib/bbab209
244
XiaoX.WangP.LinW. Z.JiaJ. H.ChouK. C. (2013). iAMP-2L: Two-level classifier for identifying antimicrobial peptides and functional types. Anal. Biochem.436, 168–177. doi: 10.1016/j.ab.2013.01.019
245
XieZ.EvangelopoulosX.OmarO. H.TroisiA.CooperA. I.ChenL. (2024). Fine-tuning GPT-3 for machine learning electronic and functional properties of organic molecules. Chem. Sci.15, 500–510. doi: 10.1039/d3sc04610a
246
XieP.YaoL.GuanJ.ChungC. R.ZhaoZ.LongF.et al. (2025). ConsAMPHemo: A computational framework for predicting hemolysis of antimicrobial peptides based on machine learning approaches. Protein Science: A. Publ. Protein Soc.34, e70087. doi: 10.1002/pro.70087
247
XuJ.LiF.LiC.GuoX.LandersdorferC.ShenH.-H.et al. (2023). iAMPCN: a deep-learning approach for identifying antimicrobial peptides and their functional activities24. doi: 10.1093/bib/bbad240
248
YanJ.BhadraP.LiA.SethiyaP.QinL.TaiH. K.et al. (2020). Deep-AmPEP30: improve short antimicrobial peptides prediction with deep learning20, 882–894. doi: 10.1016/j.omtn.2020.05.006
249
YanJ.ChenZ.CaiJ.XianW.WeiX.QinY.et al. (2025). Video-driven artificial intelligence for predictive modelling of antimicrobial peptide generation: Literature review on advances and challenges15, 7363. doi: 10.3390/app15137363
250
YanW.TangW.WangL.BinY.XiaJ. (2022). PrMFTP: multi-functional therapeutic peptides prediction based on multi-head self-attention mechanism and class weight optimization. JPcb18, e1010511. doi: 10.1371/journal.pcbi.1010511
251
YanJ.ZhangB.ZhouM.Campbell-ValoisF. X.SiuS. W. I. (2023). A deep learning method for predicting the minimum inhibitory concentration of antimicrobial peptides against Escherichia coli using Multi-Branch-CNN and Attention. mSystems8, e0034523. doi: 10.20944/preprints202303.0048.v1
252
YangD.LiY.LiC.ZhangQ.HuangJ.LiX.et al. (2026). Biochemical-knowledge-driven machine learning pipeline for generating potent antimicrobial peptides. Briefings Bioinf.27. doi: 10.1093/bib/bbag115
253
YangS.XuP. (2024). HemoDL: Hemolytic peptides prediction by double ensemble engines from rich sequence-derived and transformer-enhanced information. Anal. Biochem.690, 115523. doi: 10.1016/j.ab.2024.115523
254
YangS.YangZ.NiX. (2023). AMPFinder: a computational model to identify antimicrobial peptides and their functions based on sequence-derived information. Anal. Biochem.673, 115196. doi: 10.1016/j.ab.2023.115196
255
YaoL.GuanJ.XieP.ChungC. R.DengJ.HuangY.et al. (2024). AMPActiPred: A three‐stage framework for predicting antibacterial peptides and activity levels with deep forest33. doi: 10.1002/pro.5006
256
YuQ.DongZ.FanX.ZongL.LiY. (2021). HMD-AMP: Protein language-powered hierarchical multi-label deep forest for annotating antimicrobial peptides. doi: 10.1101/2021.11.10.468157
257
YuH.WangR.QiaoJ.WeiL. (2024). Multi-CGAN: Deep generative model-based multiproperty antimicrobial peptide design. J. Chem. Inf. Model.64, 316–326. doi: 10.1021/acs.jcim.3c01881
258
YuY.ZhangZ.GuM.YanW.HanJ.LiR.et al. (2025). Rapid response antimicrobial peptide design strategy driven by meta-learning for emerging drug-resistant pathogens. Japa68, 8530–8542. doi: 10.1021/acs.jmedchem.5c00188
259
YuQ.ZhangZ.LiuG.LiW.TangY. (2024). ToxGIN: An in silico prediction model for peptide toxicity via graph isomorphism networks integrating peptide sequence and structure information. Briefings Bioinf.25. doi: 10.1093/bib/bbae583
260
ZengZ.XuR.GuoJ.LuoX. (2024). “ Binary discriminator facilitates GPT-based protein design,” in Biorxiv, Cold Spring Harbor, NY: Cold Spring Harbor Laboratory (bioRxiv). 2023.11.20.567789.
261
ZervouM. A.DoutsiE.PantazisY.TsakalidesP. (2024). De novo antimicrobial peptide design with feedback generative adversarial networks. Int. J. Mol. Sci.25, 5506. doi: 10.3390/ijms25105506
262
ZervouM. A.DoutsiE.PantazisY.TsakalidesP. (2025). Classifier-driven generative adversarial networks for enhanced antimicrobial peptide design. Briefings Bioinf.26. doi: 10.1093/bib/bbaf500
263
ZhangY.CaiH.ShiC.ZhongB.TangJ. (2022). E3bind: an end-to-end equivariant network for protein-ligand docking. Japa.
264
ZhangT.JinQ.JiJ. (2025). Antimicrobial peptides and their mimetics: promising candidates of next-generation therapeutic agents combating multidrug-resistant bacteria9, 2400461. doi: 10.1002/adbi.202400461
265
ZhangK.TengD.MaoR.YangN.HaoY.WangJ. (2023). Thinking on the construction of antimicrobial peptide databases: powerful tools for the molecular design and screening24, 3134. doi: 10.3390/ijms24043134
266
ZhangS.ZhaoY.LiangY. (2023). AMP-EF: an ensemble framework of extreme gradient boosting and bidirectional long short-term memory network for identifying antimicrobial peptides. Match-Communications Math. Comput. Chem.91, 109–131. doi: 10.46793/match.91-1.109z
267
ZhangK.ZhuY.ZhangW.SunY.ZhangZ.WangY.et al. (2025). LigDockTailor: ligand-specific docking tool matching using multidimensional descriptors. Japa119, 108554. doi: 10.1016/j.compbiolchem.2025.108554
268
ZhaoW.HouK.ShenY.HuX. (2025). A conditional denoising VAE-based framework for antimicrobial peptides generation with preserving desirable properties. Bioinformatics41. doi: 10.1093/bioinformatics/btaf069
269
ZhaoW.HouK.TangC.ShenY.LiuJ.HuX. (2026). A novel generative framework for designing pathogen-targeted antimicrobial peptides with programmable physicochemical properties. PloS Comput. Biol.21, e1013833. doi: 10.1371/journal.pcbi.1013833
270
ZhaoY.LiQ.SunM.SuY.SuX.JiangL.et al. (2025). EBAMP: an efficient de novo broad-spectrum antimicrobial peptide discovery framework. Cell Rep.44, 116215. doi: 10.1016/j.celrep.2025.116215
271
ZhaoF.QiuJ.XiangD.JiaoP.CaoY.XuQ.et al. (2024). deepAMPNet: a novel antimicrobial peptide predictor employing AlphaFold2 predicted structures and a bi-directional long short-term memory protein language model12. doi: 10.7717/peerj.17729
272
ZhaoL.WangY.JiangJ.PanH.WangL.MaS.et al. (2025). Evaluation of LLM-generated peptide as foundation template for discovery of effective encrypted AMPs against clinical superbugs13. doi: 10.1128/spectrum.01504-25
273
ZhaoY.ZhangS.LiangY. (2024). HemoFuse: Multi-feature fusion based on multi-head cross-attention for identification of hemolytic peptides. Sci. Rep.14, 22518. doi: 10.1038/s41598-024-74326-3
274
ZhaoM.ZhangY.WangM.MaL. Z. (2024). dsAMP and dsAMPGAN: deep learning networks for antimicrobial peptides recognition and generation. Antibiotics13. doi: 10.3390/antibiotics13100948
275
ZhongG.LiuH.DengL. J. I. S. C. L. S. (2024). Ensemble machine learning and predicted properties promote antimicrobial peptide identification16, 951–965. doi: 10.1007/s12539-024-00640-z
276
ZhouS.SunZ.LiuL.XiY.ZhouL.YangZ.et al. (2026). MAC-AMP: A closed-loop multi-agent collaboration system for multi-objective antimicrobial peptide design. 16, 1745427.
277
ZhouP.JinB.LiH.HuangS. Y. (2018). HPEPDOCK: a web server for blind peptide-protein docking based on a hierarchical algorithm. Nucleic Acids Res.46, W443–W450. doi: 10.1093/nar/gky357
278
ZhuangS.TannerJ.WuY.HuynhD.LiuW.CadetX.et al. (2024). Non-hemolytic peptide classification using a quantum support vector machine. Quantum Inf. Process.23, 379. doi: 10.1007/s11128-024-04540-5
Summary
Keywords
antimicrobial peptides (AMPs), artificial intelligence (AI), deep learning (DL), generative-AI, machine learning (ML), molecular docking (MD), multi-drug resistance (MDR)
Citation
Saleem N, El-Omar E, Kumar N, Willcox M and Jiang X-T (2026) Lab to codes: multimodel advanced AI-driven pipelines to drive antimicrobial peptide design for the post-antibiotic era. Front. Cell. Infect. Microbiol. 16:1894311. doi: 10.3389/fcimb.2026.1894311
Received
29 May 2026
Revised
12 August 2026
Accepted
19 August 2026
Published
29 September 2026
Volume
16 - 2026
Edited by
Arun Gonzales Decano, City University of London, United Kingdom
Reviewed by
Mohd Sadeeq, Shandong University of Technology, China
Zubia Rashid, Sheikh Khalifa Medical City Department of Pediatrics, United Arab Emirates
Updates
Copyright
© 2026 Saleem, El-Omar, Kumar, Willcox and Jiang.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Naveed Saleem, n.saleem@unsw.edu.au; Mark Willcox, m.willcox@unsw.edu.au; Xiao-Tao Jiang, xiaotao.jiang@unsw.edu.au
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.