Black boxes and petri dishes: AI models for antimicrobial pharmacokinetics, drug properties, and resistance

Abstract

The convergence of artificial intelligence (AI) and antimicrobial pharmacology is creating unprecedented opportunities to accelerate antimicrobial discovery, optimize therapeutic strategies, and reduce the global burden of infectious diseases. A critical determinant of the successful integration of AI into antimicrobial pharmacology is the accurate and interpretable prediction of pharmacokinetic (PK) and pharmacodynamic (PD) properties, including measures such as the minimum inhibitory concentration (MIC), to identify the PK/PD indices and dosing regimens that maximize antimicrobial efficacy while ensuring safety. We review recent developments in AI models for antimicrobial PK/PD and critically examine their performance, interpretability, generalizability, and readiness for clinical translation. We provide a practical framework to help researchers, microbiologists, pharmacologists, and clinicians evaluate the strengths, limitations, and appropriate applications of these emerging approaches. We anticipate that rigorous development, transparent validation, and thoughtful implementation of AI-driven PK/PD models will accelerate antimicrobial discovery, optimize precision dosing, and ultimately improve patient outcomes.

Introduction

Pharmacokinetics (PK), or the study of drug movement through the body, is a critical determinant in the success, and failure, of drug development and clinical practice (Dowden and Munro, 2019; Sun et al., 2022). To capture the processes that underly PK, including absorption, distribution, metabolism and excretion (ADME), scientists and clinicians have turned to artificial intelligence (AI) to predict the outcomes of these processes on their experimental therapeutics and approved drugs (Cheng et al., 2012; Pires et al., 2015; Schyman et al., 2017; Dong et al., 2018; Yang et al., 2018; Yang et al., 2019; Venkatraman, 2021; Xiong et al., 2021; Tian et al., 2022; Zhang et al., 2022; Fu et al., 2024; Gu et al., 2024; Myung et al., 2024; Swanson et al., 2024; Yi et al., 2024; Seal et al., 2025). Additionally important in antimicrobial research is the application of AI to predict the minimum inhibitory concentration (MIC) that is incorporated into the pharmacodynamics (PD) component of PK/PD indices (Drusano, 2004; Landersdorfer and Nation, 2021). Some recent examples where AI has led to great leaps and strides include the prediction of protein structures (Jumper et al., 2021; Bryant et al., 2022; Abramson et al., 2024), the detection of Staphylococcus aureus bloodstream infections (Nigo et al., 2024) and sepsis (Adams et al., 2022; Liesenfeld et al., 2025; Watson et al., 2026), and the discovery of novel antibiotics (Imai et al., 2019; Stokes et al., 2020; Liu et al., 2023; Catacutan et al., 2025; Wong et al., 2025; Shehadeh et al., 2026). As AI continues to transform antimicrobial research (Anahtar et al., 2021; Kim et al., 2022; Codde et al., 2025; Howard et al., 2026; Sardar et al., 2026), there is a growing opportunity to equip scientists and clinicians with the knowledge and tools needed to evaluate AI models with the same rigor applied to scientific evidence and clinical trials.

Artificial intelligence (AI), a term first coined by John McCarthy in a 1955 proposal for the Dartmouth Workshop, describes “making a machine behave in ways that would be called intelligent if a human were so behaving” (McCarthy et al., 2006; Howard, 2019), Thus, AI is defined as “the science and engineering of making intelligent machines” (McCarthy, 2007). Today, two key methodological approaches within AI include machine learning (Jordan and Mitchell, 2015) and its subset, deep learning (LeCun et al., 2015). However, the phrase “all models are wrong, but some are useful” attributed to the renowned statistician George E.P. Box (Box and Draper, 1987), reminds us to take a critical stance on these models before accepting the results of their implementation.

Global initiatives to improve the standardization and benchmarking of AI models of ADMET prediction have emerged. For the standardized prediction of ADMET properties across many classes of therapeutics, the Therapeutic Data Commons (TDC) (https://tdcommons.ai/) (Huang et al., 2021) provides a benchmark of curated training data for the development of prediction models. More specifically to anti-infective therapeutics, the CAMO-Net GDR initiative (https://camonet.org/) (Mutegeki et al., 2025) aims to provide a secure trusted research environment of antimicrobial PK/PD. To further enhance the efforts of antimicrobial PK/PD modeling, a critical review of ADME and related models should include the following questions: What was the source and quality of the training data? How was the data cleaned? Why was that algorithm or architecture selected? In what areas did the model perform well versus not so well? Was there a prospective in vitro or in vivo experimental validation performed? In this context, a critical evaluation of AI models for antimicrobial PK/PD is an underrepresented gap in the literature, and to address this gap, we apply the above questions to critically evaluate the development, interpretability and validation of AI models for the prediction of PK/PD and MIC of novel and established antimicrobial agents.

Critical evaluation of AI models of PK and ADMET

Anti-infective drug development in the United States is comparatively less costly than many other therapeutic areas, with a mean expected capitalized cost of ~$378.7 million per agent (95% CI, $244.6–$556.1 million), inclusive of the cost of failures and the cost of capital, compared with a mean expected capitalized cost of ~$879.3 million across all drug classes (Sertkaya et al., 2024). This remains an incredibly hefty price tag given that only approximately 10% of new molecular entities entering Phase I clinical development ultimately achieve regulatory approval following successful progression through subsequent phases of clinical testing (Hay et al., 2014; Wong et al., 2019). Notably, this ~10% overall approval probability reflects new molecular entities across all therapeutic areas; infectious-disease agents have historically demonstrated higher-than-average phase-transition and approval probabilities (Phase I–to–approval ~25%) (Wong et al., 2019), likely reflecting well-validated microbial targets and objective, short-term efficacy endpoints (e.g., microbiologic or virologic clearance) that reduce late-stage efficacy failures, together with frequent use of expedited regulatory pathways (Darrow et al., 2020; DiMasi et al., 2020). Recent literature suggests an increased appreciation for the structure-tissue exposure/selectivity-activity relationship (STAR) (Sun et al., 2022) might lead to better therapeutics, and not simply more potent molecules in a dish. Additionally, the Food and Drug Administration (FDA) released a roadmap to reduce animal testing and increase the use of in vitro and in silico computational modeling in drug development (Ingber, 2022; Sadybekov and Katritch, 2023; Hoeg et al., 2026; Roadmap to Reducing Animal Testing in Preclinical Safety Studies, 2026).

Since 2010, the number of publications in PubMed that include either “artificial intelligence”, “machine learning” or “deep learning” in the title or abstract, and (conditional) “pharmacokinetics” (or related pharmacological parameters) have exponentially increased, especially in the last few years (Figure 1). While these numbers capture all papers on the topic and are not necessarily the development of an AI model for pharmacokinetics, here we critically review the notable AI models with multi-parameter prediction capability – ADMET-AI (Swanson et al., 2024), Deep-PK (Myung et al., 2024), pkCSM (Pires et al., 2015), ADMETlab (Dong et al., 2018; Xiong et al., 2021; Fu et al., 2024), FP-ADMET (Venkatraman, 2021), OptADMET (Yi et al., 2024), ADMETboost (Tian et al., 2022), Helix ADMET (Zhang et al., 2022), ADMETopt (Yang et al., 2018), admetSAR (Cheng et al., 2012; Yang et al., 2019; Gu et al., 2024), vNN-ADMET (Schyman et al., 2017), and PKSmart (Seal et al., 2025) (Table 1). The critical analysis of these leading AI models for PK and ADMET reveals a few examples of strengths and limitations that may be informative in translatability and real-world applications. Specifically, these models were trained on diverse datasets of variable size and data cleaning approaches, with differences in model architecture, performance metrics, and external validation.

Figure 1

Training data and data quality

A practical first question to consider is the source and quality of training data. Many platforms, such as ADMET-AI (Swanson et al., 2024), ADMETlab (Dong et al., 2018; Xiong et al., 2021; Fu et al., 2024), and admetSAR (Cheng et al., 2012; Yang et al., 2019; Gu et al., 2024) (ADMETopt (Yang et al., 2018) optimization module), rely on aggregated public datasets (such as PubChem (Kim et al., 2016; Kim et al., 2025) or ChEMBL (Gaulton et al., 2012; Zdrazil et al., 2024)) or partially proprietary datasets. However, the provenance and curation of these data are not always reported in sufficient detail. For example, plasma protein binding (PPB) may be reported in a dataset, but the species from which it was derived (humans, mice, rats, or an in vitro system), the experimental conditions, and the assay methodology are often unclear. Similarly, large chemical databases frequently contain inconsistent annotations, duplicate compounds, variable assay methodologies, and heterogeneous physiological conditions, all of which can influence model development. Careful data curation is therefore essential. At the same time, overly aggressive filtering or inappropriate preprocessing may introduce systematic bias and reduce model generalizability. Although some approaches, such as vNN-ADMET (Schyman et al., 2017), seek to mitigate these challenges by incorporating variable nearest-neighbor applicability domains, transparent reporting of data curation procedures – including compound standardization, duplicate handling, outlier treatment, normalization, and quality-control criteria – remains inconsistent across platforms. Greater transparency in these processes would facilitate more meaningful comparisons among models and improve confidence in their reported performance.

Model architecture and selection

Another important consideration is the choice of modeling architecture. Current platforms span a broad range of approaches, including graph-based signatures in pkCSM (Pires et al., 2015), graph neural networks in Deep-PK (Myung et al., 2024), ensemble and boosting methods in ADMETboost (Tian et al., 2022), and fingerprint-based or hybrid frameworks in FP-ADMET (Venkatraman, 2021) and admetSAR (Cheng et al., 2012; Yang et al., 2019; Gu et al., 2024) (ADMETopt (Yang et al., 2018)). Each offers distinct advantages depending on the prediction task and the available training data. Graph-based representations and graph neural networks are well suited to capturing molecular topology and structural relationships, whereas fingerprint-based methods often provide computational efficiency and robustness, and may offer greater interpretability. Ensemble and boosting approaches can further improve predictive performance by integrating complementary models and iteratively correcting prediction errors. Ultimately, no single architecture is universally optimal across all PK/PD or ADMET endpoints, and model performance frequently depends on the underlying dataset, endpoint definition, and chemical space represented during training. Greater transparency regarding the rationale for model selection, together with systematic benchmarking across diverse datasets and clinically relevant endpoints, would help investigators determine which approaches are best suited for specific pharmacological applications while promoting reproducibility and broader clinical translation.

Comparative model performance and generalizability

The comparison of multiple newer AI models versus classical statistical approaches, such as nonlinear mixed effects (NLME) modeling (NONMEM), was very nicely illustrated by Mao et al. (2025) While previous studies focus on the comparison of one AI model versus NONMEM, these investigators understood that, and evaluated the predictive performance of several AI models. Specifically, they compared five machine learning models (linear regression, LASSO, XGBoost, Light Gradient Boosting Machine, and Random Forest), three deep learning models (neural networks, gated recurrent units, and long short-term memory), and one neural ordinary differential equation model on simulated and real clinical pharmacokinetic datasets. They showed AI models can outperform traditional NLME approaches such as NONMEM in predictive accuracy, however they also note the sensitivity of the models to data characteristics, such as the different phases of clinical trials. Overall, they concluded that the AI-based approaches are promising complementary tools for pharmacokinetic modeling but are not yet sufficient to replace mechanistic frameworks in settings that require robust extrapolation and clinical interpretability (Mao et al., 2025).

Direct comparison of ADMET prediction platforms remains challenging because models are frequently trained on different datasets, evaluated on distinct endpoints, and reported using heterogeneous performance metrics. Although newer deep learning and transfer-learning approaches often outperform traditional machine learning methods on benchmark datasets, the magnitude of improvement is frequently endpoint-dependent and may not always translate into improved generalizability on external datasets. Nevertheless, several general trends emerge. Models such as ADMET-AI (Swanson et al., 2024), ADMETLab (Dong et al., 2018; Xiong et al., 2021; Fu et al., 2024), HelixADMET (Zhang et al., 2022), and Deep-PK (Myung et al., 2024) tend to demonstrate strong performance across multiple ADMET endpoints due to their use of large-scale datasets, multitask learning frameworks, graph neural networks, and transfer-learning strategies. For example, ADMET-AI (Swanson et al., 2024) reported median AUROC values approaching 0.84 across benchmark classification tasks, while Deep-PK (Myung et al., 2024) achieved R² values up to approximately 0.72 for selected pharmacokinetic endpoints. Similarly, ADMETLab (Dong et al., 2018; Xiong et al., 2021; Fu et al., 2024) and HelixADMET (Zhang et al., 2022) reported AUROC values frequently exceeding 0.80 across several ADMET classification tasks (Table 1).

Performance differences across models become particularly apparent when evaluating specific pharmacokinetic endpoints. While physicochemical properties such as lipophilicity, solubility, and permeability are generally associated with stronger predictive performance because they are closely tied to molecular structure and supported by relatively large in vitro datasets, the prediction of biological half-life (t½) remains substantially more challenging. Thus, the biological half-life of the drug (defined as the time required for a 50% reduction in drug concentration in human plasma) reflects the combined influence of absorption, distribution, metabolism, excretion, plasma protein binding, tissue penetration, transporter activity, and physiological variability.

For a cross-comparison of performance in the prediction of biological half-life, a critical determinant in dosing rationale, Deep-PK (Myung et al., 2024) achieved a mean Matthews correlation coefficient of 0.32 (Matthews correlation coefficient ranges from -1 to +1, with +1 indicating perfect classification and 0 indicating no better performance than random classification) across three-fold cross-validation, with a between-fold standard deviation of 0.056, indicating modest but relatively consistent classification performance across the internal validation folds. This is compared to the ADMETlab 2.0 (Xiong et al., 2021) reported MCC of 0.478 and AUROC of 0.801 (accuracy [ACC]: 0.727, specificity: 0.658, sensitivity: 0.827); ADMETlab 3.0 (Fu et al., 2024) upgraded the biological half-life from a classification to a regression task, with their directed message passing neural network (DMPNN) with traditional molecular descriptors (DMPNN-Des) achieving an average predictive performance (R²) of 0.653 ± 0.070 and 0.748 ± 0.052 for the held-out and validation test sets, respectively. Furthermore, the development of ADMET-AI (Swanson et al., 2024) (individual Chemprop-RDKit) reported an average Spearman correlation of 0.239 ± 0.019 for the five-model ensemble on the scaffold-split TDC Half_Life_Obach leaderboard test set of 667 compounds with a half-life range of 0.065 – 1,200 hours. This compares to the earlier ADMETboost (Tian et al., 2022) reported a Spearman correlation of 0.396 ± 0.027 on the TDC half-life dataset. This contrasts the reported performance of HelixADMET (Zhang et al., 2022) on the classification in the prediction of biological half-life of 0.736, as opposed to the continuous half-life regression prediction used by comparator models.

However, these values should not be directly ranked because the platforms used different endpoint formulations, datasets, validation procedures, and performance measures. Nevertheless, their generally modest and variable performance indicates that half-life remains one of the more difficult pharmacokinetic endpoints to predict from molecular structure alone. Each reported metric (MCC, AUROC, and Spearman correlation) is drawn from its originating publication and reflects distinct endpoint formulations, datasets, and validation procedures; these values are therefore presented for context rather than direct ranking (Fu et al., 2024; Swanson et al., 2024).

Experimental validation and translational considerations

The evaluation of model performance also highlights an uneven landscape across these models based on their strengths and weaknesses. Platforms such as ADMETlab (Dong et al., 2018; Xiong et al., 2021; Fu et al., 2024) and ADMET-AI (Swanson et al., 2024) often demonstrate strong predictive capability for well-characterized physicochemical endpoints (lipophilicity, solubility, permeability), where large and relatively consistent datasets are available. Conversely, predictions of more complex biological processes – such as cytochrome P450 inhibition, transporter interactions, or metabolic clearance – tend to be more variable across models including Deep-PK (Myung et al., 2024), Helix ADMET (Zhang et al., 2022), and ADMETboost (Tian et al., 2022), This disparity likely reflects both data limitations and the inherent biological complexity of these endpoints, which involve multi-scale interactions not easily captured by structure-based features alone. While tools like vNN-ADMET (Schyman et al., 2017) attempt to define applicability domains to improve confidence in predictions, many platforms still heavily rely on internal cross-validation, with relatively limited external benchmarking on independent datasets. This raises the concern of model overfitting and inflated performance metrics.

How frequently was a prospective in vitro or in vivo experimental evaluation performed? A critical limitation across the field is the general lack of prospective, experimental validation accompanying many of these computational models. While tools such as pkCSM (Pires et al., 2015), ADMETlab (Dong et al., 2018; Xiong et al., 2021; Fu et al., 2024), and OptADMET (Yi et al., 2024) are widely used for virtual screening, a relatively limited number of studies report the systematic in vitro or in vivo validation of their predictions. When validation is included, it is often limited to a small subset of compounds or specific endpoints, reducing confidence in the generalizability of these models across diverse chemical spaces. For newer platforms, such as Helix ADMET (Zhang et al., 2022) and Deep-PK (Myung et al., 2024) experimental corroboration remains particularly sparse in the literature. The absence of robust experimental benchmarking not only limits the translational utility of these models but also future opportunities for iterative model refinement based on empirical feedback.

Model interpretability and explainable artificial intelligence

A recurring challenge in AI-driven antimicrobial PK/PD modeling is balancing predictive performance with interpretability. Simpler models, including linear regression, Elastic Net, decision trees, and nearest-neighbor approaches, provide transparent relationships between input variables and model predictions, facilitating biological interpretation, hypothesis generation, and mechanistic insight. In contrast, more complex architectures such as deep neural networks, graph neural networks, and transformer-based models are capable of learning highly nonlinear relationships and may achieve superior predictive performance, particularly when large and complex datasets are available.

Beyond predictive performance, interpretability is not merely a technical consideration but a prerequisite for biological discovery and clinical translation (Figure 2). In explainable AI, understanding the prediction rationale could provide investigators with hypothesis-generating ideas of otherwise overlooked biological mechanisms. For example, interpretable models may reveal specific genomic determinants associated with antimicrobial resistance, highlight features influencing MIC predictions, or identify physicochemical properties linked to clearance or tissue penetration, or the pharmacogenomics of therapeutic underexposure. Such insights can guide hypothesis generation, antimicrobial design, experimental validation, and clinical decision-making. Consequently, the most appropriate model is not necessarily the most complex or highest-performing architecture, but rather the one that provides the most aid in human decision-making. As AI models continue to move toward real-world implementation, improving interpretability will be essential for fostering trust, supporting regulatory acceptance, and ensuring that model predictions can be critically evaluated by microbiologists, pharmacologists, and clinicians (Warrier et al., 2026).

Figure 2

Explainable AI approaches are increasingly being developed to improve transparency within complex models. Methods such as SHAP (Shapley Additive Explanations), feature attribution analyses, and saliency mapping (Lundberg and Lee, 2017) attempt to quantify the contribution of individual variables to a model prediction. In antimicrobial PK/PD applications, these approaches may identify molecular descriptors influencing ADMET predictions, genomic features associated with antimicrobial resistance, or physicochemical properties contributing to tissue penetration and drug exposure. Similarly, attention visualization methods can highlight sequence motifs, molecular substructures, or genomic regions that disproportionately influence model outputs, thereby providing investigators with biologically meaningful hypotheses for further study (Lundberg and Lee, 2017).

Other explainability approaches focus on evaluating the reliability and robustness of model predictions. For example, applicability-domain analyses estimate whether a compound, peptide, or microbial isolate falls within the region of chemical or biological space represented by the training data, whereas counterfactual explanations identify the minimal changes required to alter a model prediction (Molnar, 2020). These methods help investigators distinguish between predictions supported by strong evidence and those associated with greater uncertainty. While explainable AI does not completely eliminate the “black box” nature of complex deep learning systems, it can improve transparency, increase confidence in model outputs, facilitate experimental validation, and promote alignment between computational predictions and established biological and pharmacological knowledge. As a result, explainable AI represents an important strategy for reconciling the trade-off between predictive performance and interpretability.

Finally, the issues around interpretability and translatability remain a significant challenge (Figure 2). Many deep-learning based systems, including Deep-PK (Myung et al., 2024) and aspects of ADMET-AI (Swanson et al., 2024) function as “black boxes”, offering limited mechanistic insight into the features driving the predictions. However, Deep-PK (Myung et al., 2024) does incorporate molecular interpretation features, and ADMET-AI (Swanson et al., 2024) uses Chemprop (Heid et al., 2024; Graff et al., 2026)-based models that are not entirely opaque. In contrast, simpler frameworks such as vNN-ADMET (Schyman et al., 2017) or descriptor-based approaches in FP-ADMET (Venkatraman, 2021) may offer greater transparency but sometimes at the cost of predictive performance. This trade-off between accuracy and interpretability remains an unresolved tension across the discipline. Although some platforms, such as ADMETlab (Dong et al., 2018; Xiong et al., 2021; Fu et al., 2024), have begun incorporating visualization tools and fragment-based explanations, these efforts are not yet standardized or consistently informative for guiding medicinal chemistry decisions.

AI models for the prediction of antimicrobial susceptibility and resistance

A renewed interest in antimicrobial PK/PD is evident from the journal podcasts of professional societies, most notably – Communicable E51: We will make you love PK/PD, part 1 (https://communicable.transistor.fm/episodes), the many episodes of Dosing Consults by the Breakpoints podcast (https://breakpoints-sidp.org/), and the recent call for a global data repository (GDR) in a secure Trusted Research Environment by the CAMO-Net GDR initiative (https://camonet.org/) (Mutegeki et al., 2025). Classically, there exist three mainstay antimicrobial PK/PD profiles: the percentage of time in a single dosing period the unbound, free (f) drug concentration is above the minimum inhibitory concentration (MIC) (f Time>MIC), the max free drug concentration as a ratio to the MIC (f Cmax/MIC), and the mathematical product of both time- and concentration-dependent exposure, calculated as the area under the curve (AUC) of the free drug concentration as a ratio to MIC (f AUC/MIC) (Drusano, 2004). From the above, one can easily observe an outsized emphasis placed on the denominator, MIC. Thus, herein we critically evaluate recent publications that predict MIC across different species and antimicrobial agents.

Quantitative prediction of MIC

Several AI models that leverage microbiologic and genomic data to predict the MIC of pathogens were identified and provided for discussion (Table 2). Substantial methodological differences exist across the reported literature with respect to data generation, data preprocessing, rationale of model selection, evaluation strategies, and experimental validation. For example, data quality and preprocessing remain a fundamental challenge for MIC measurements, which are often noisy and aggregated from variable assay conditions, strains, and inoculum sizes. These challenges are further compounded when training data draws from large, global data repositories of genotype-phenotype relationships, such as CARD (Jia et al., 2017; Alcock et al., 2020; Alcock et al., 2023), AMRFinderPlus (Feldgarden et al., 2019; Feldgarden et al., 2021; Feldgarden et al., 2022), MEGARes (Lakin et al., 2017), ResFinder (Bortolaia et al., 2020), PointFinder (Zankari et al., 2017), WHO GLASS (Tornimbene et al., 2018), EARS-NET, ATLAS (Gagliotti et al., 2011; Ayobami et al., 2020; El Mammery et al., 2023), Pathogenwatch (Argimon et al., 2021a; Argimon et al., 2021b; Sanchez-Buso et al., 2021), and BV-BRC (Davis et al., 2020), to name a few. However, this approach provides limited discussion on data harmonization across disparate sources, inter-laboratory variability, and breakpoint inconsistencies. To account for this limitation of a pure database approach, isolate collections (or biobanks) provide the unique opportunity to conduct prospective, parallel in vitro and in silico validations of AI model building. A few noteworthy studies that did this include Rehman et al., Humphries et al., Pham et al., Vergauwe et al., Gao et al., and Gerada et al. (Rehman et al., 2021; Humphries et al., 2023; Gao et al., 2024; Pham et al., 2024; Vergauwe et al., 2025; Gerada et al., 2026) While these AI models require real-world evidence of impact upon implementation, they provide a refreshing perspective on the value of AI models developed with parallel in vitro and in silico approaches.

Relatively few studies that develop AI models for MIC prediction provide detailed descriptions of how data inconsistencies – such as duplicates, conflicts, controls, and differing units or thresholds – are resolved. To address these issues, the study by Batisti Biffignandi et al. positively stands out because they explicitly address left- and right-censored MIC values (e.g. < 0.125 or ≥ 256 µg/mL), varying concentration ranges across different antimicrobial agents, and the implications of representing MICs as either continuous or categorical outcomes (Batisti Biffignandi et al., 2024). The investigators explicitly tested the impact of multiple data representations (continuous versus categorical) as a core modeling decision tied to data structure, recommending regression when the number of observed antibiotic concentration levels is large and classification when fewer levels are available. Additionally, by removing the inequality symbols (< or ≥ etc.) and converting values to log2 scale data, they were able to harmonize heterogeneous lab outputs into consistent modeling space while preserving ordering (Batisti Biffignandi et al., 2024). By simulating semi-quantitative traits with varying genetic architectures and heritability levels, they specifically evaluated how data representation choices influenced model performance.

Similarly, the CRyPTIC Consortium leveraged a highly standardized microtiter plate assay across more than 15,000 Mycobacterium tuberculosis isolates, providing one of the most rigorously curated datasets on genotype-to-phenotype relationships (Consortium, C. R, 2024). These investigators quantified data inconsistencies across a large dataset rather than hiding them. Specifically, they used multiple reading methods – human laboratory readings, automated image analysis, and citizen-science scoring (BashTheBug) – to assess measurement concordance across their dataset. By this approach, they were able to filter data and even observe that in some cases – data inconsistency can reflect true biological variation rather than solely technical artifacts. Using a multivariable linear mixed model, they incorporated population structure and measurement variation into their model rather than “cleaning away” the inconsistencies. By doing so, they identified 492 unique MIC-elevating variants and 91 mutations linked to hypersensitivity across 13 drugs, demonstrating that formally accounting for heterogeneity can reveal genetic determinants of both resistance and susceptibility that might otherwise be obscured. Taken together, these two studies underscore a deeper paradigm shift: rather than treating inconsistency as noise to be sanitized, MIC modeling can be strengthened by formally encoding measurement uncertainty, experimental heterogeneity, and data ambiguity as integral features of the learning process – thereby producing models that better reflect the underlying biological complexity and are more resilient and generalizable (Consortium, C. R, 2024).

Prediction of antimicrobial resistance phenotypes

While minimum inhibitory concentration (MIC) prediction and antimicrobial resistance (AMR) prediction are closely related, they represent distinct modeling tasks with different biological and clinical objectives (Li et al., 2024). MIC prediction seeks to estimate a quantitative measure of antimicrobial susceptibility, whereas AMR prediction typically focuses on classifying susceptibility phenotypes (e.g., resistant versus susceptible) or identifying genomic determinants associated with resistance. Consequently, AMR prediction models frequently leverage whole-genome sequencing, resistance gene databases, mutation profiles, and other genomic features to infer resistance mechanisms and predict antimicrobial susceptibility. As genomic surveillance programs continue to expand, AI-driven AMR prediction offers the potential for rapid resistance profiling, identification of emerging resistance determinants, and improved antimicrobial stewardship (Olatunji et al., 2024).

The comparison of multiple algorithms is an approach that provides direct insight into the importance of model accuracy versus mechanistic explainability. The study by Ren et al. positively stands out for their comparison of both machine learning (logistic regression, support vector machines, Random Forest) and deep learning (convolutional neural networks) for the prediction of E. coli resistance from whole genome sequencing (WGS) data (Ren et al., 2022). The investigators derived resistance labels from phenotypic antimicrobial susceptibility testing and trained models using label encoding, one-hot encoding, and frequency matrix chaos game representation (FCGR). From this approach, their AI models achieved high predictive performance with AUCs up to 0.96 without reliance on prior knowledge of resistance, indicating genomic features can effectively predict resistance and identify mutations associated with AMR. While this study was limited to E. coli and treated resistance as a simplified binary phenotype (resistant [R] versus susceptible [S]), this study positively demonstrates the feasibility of WGS-based AMR prediction with the potential for data-driven approaches to mechanistic discovery. Since this work, WGS-based machine-learning models have been extended to additional pathogens—for example, achieving AUC >0.9 and categorical agreement of ~0.96 in Klebsiella pneumoniae (Jia et al., 2026), while systematic benchmarking and review studies confirm that predictive performance is species- and antibiotic-dependent (Ardila et al., 2024; Hu et al., 2024), with notably weaker performance for certain β-lactams such as amoxicillin/clavulanate (Hu et al., 2024; Wan et al., 2026), a discordance attributable to its quantitative rather than binary resistance phenotype (Davies et al., 2020).

Other investigations, such as the genome-wide growth prediction analysis in Escherichia coli by Benkwitz-Bedford et al., illustrate the impact of strain and condition-dependent variability on unexplained variance in genomic predictive models (Benkwitz-Bedford et al., 2021). Although this study predicted growth parameters under subinhibitory antimicrobial concentrations rather than MICs directly, its finding that “missing heritability” limits prediction accuracy is broadly relevant to MIC modeling. Even with improved feature engineering, as demonstrated by the k-mer-based prediction framework developed by Gao et al. for Acinetobacter baumannii, data heterogeneity across the field remains a limiting factor for robust MIC prediction (Gao et al., 2024). In contrast, studies that relied on large surveillance databases with paired genomic data often provide limited discussion of laboratory confirmation, duplicate removal, and discrepancy resolution. These methodological gaps in harmonization across contributing labs can lead to systematic biases in the models (Benkwitz-Bedford et al., 2021; Gao et al., 2024; Yu et al., 2025).

The challenge of data preprocessing is even more pronounced in antimicrobial peptide (AMP) studies (Yan et al., 2023; Chung et al., 2024; Yao et al., 2024; Bajiya et al., 2025; Cai et al., 2025; Chiu et al., 2026). A few notable studies utilize experimentally derived datasets of peptide MICs, but place relatively less emphasis on the potential effects of assay variability, differing peptide synthesis methods, lab-specific protocols, or reporting standards. For example, the peptide regression framework introduced by Bajiya et al. (2025), the transfer-learning-based BERT-AmPEP60 model by Cai et al. (2025), and the hybrid Inception- Transformer architecture (“ANIA”) described by Chiu et al. (2026), each integrate large, aggregated peptide datasets but provide limited discussion of cross-study assay variability despite achieving strong predictive performance. Finally, AMPActiPred by Yao et al., a technically sophisticated and biologically meaningful three-stage framework, provides limited transparency in negative sample construction, redundancy reduction, and cross-dataset harmonization (Yao et al., 2024). Given that AMP datasets are typically aggregated from multiple experimental sources, the insufficient attention to data origin and quality may limit reproducibility and inflate apparent predictive accuracy (Wilkinson et al., 2016; Wilkinson et al., 2019).

The choice of model algorithms or architecture also varies considerably across studies. Genome-based investigations frequently employ ensemble learning methods such as Random Forest, XGBoost, and CatBoost because these algorithms perform well with high-dimensional genomic data and complex nonlinear relationships. For example, Nguyen et al. used XGBoost to predict MICs in f using 5,278 genomes (Nguyen et al., 2019), Yasir et al. employed machine learning approaches to model Neisseria gonorrhoeae susceptibility (Yasir et al., 2022), and Yang et al. leveraged pan-genome features to predict MIC values across multiple bacterial species (Yang et al., 2023). More generally, algorithm selection is typically guided by benchmarking against alternative models and predictive performance.

While this approach is entirely appropriate, model architecture selection should ideally be guided not only by empirical predictive performance but also by the biological and pharmacological mechanisms underlying the endpoint being predicted. For example, graph-based approaches may be particularly suitable for properties that arise directly from molecular structure, whereas sequence-based architectures may be better aligned with antimicrobial peptide activity because biologically relevant information is encoded within amino acid sequences. In the context of MIC prediction, model architecture should ideally account for the biological determinants of susceptibility, including pathogen genetics, resistance mechanisms, microbial phenotypes, and assay-specific variability, rather than relying solely on structural or sequence-derived features. Consequently, the most appropriate architecture is often the one whose training data and underlying assumptions most closely reflect the biological and pharmacological determinants of the endpoint of interest, rather than simply the model with the highest reported predictive performance. A similar pattern is observed in studies such as Gao et al., which compared Random Forest, support vector machine, and XGBoost for MIC prediction in Acinetobacter baumannii (Gao et al., 2024). As AI models continue to evolve, closer integration of model architecture with the underlying biological, microbiological, and pharmacological mechanisms may improve both interpretability and confidence in their application to antimicrobial PK/PD.

Rather than the binary variable of susceptibility based on MIC breakpoints, other studies in E. coli, such as that of Pataki et al., also leverage machine learning and whole genome sequencing to predict MICs (Pataki et al., 2020). By treating MICs as a continuous value, the investigators achieved 65% predictive accuracy within a two-fold dilution and 93% within four-fold dilutions of the MIC from WGS data using 704 E. coli genomes from multiple countries. The investigators identified four key predictors—mutations in gyrA residues Ser83 and Asp87, a mutation in parC residue Ser80, and the presence of the qnrS1 gene—all of which have been experimentally validated. In their validation, they used leave-one-country-out (LOCO) validation to prevent the model from learning country-specific biases, providing greater external generalization. Finally, an additional strength of their approach was the ability of the model to identify known causal resistance mechanisms of fluoroquinolone resistance, such as gyrA, parC, and qnrS1. The main strength of this study is the simplified nature of the prediction, as well as the use of an international dataset from multiple countries, providing a strong degree of confidence in the generalizability and explainability of the model (Pataki et al., 2020).

Other studies prioritize the interpretability of their predictions over maximizing predictive performance, employing methods such as Elastic Net and linear mixed models that allow more direct examination of feature contributions (Batisti Biffignandi et al., 2024). Batisti Biffignandi et al. employed Elastic Net, Random Forest, and linear mixed models to balance predictive accuracy with mechanistic insight (Batisti Biffignandi et al., 2024). Earlier work, such as the multivariate regression analysis by Demczuk et al., similarly demonstrates how interpretable statistical frameworks can directly link specific resistance mutations to quantitative MIC variation, achieving predicted-to-phenotypic MIC correlations of 90–98% within one doubling dilution across six antibiotics in Neisseria gonorrhoeae, reinforcing their value for mechanistic understanding (Demczuk et al., 2020). Similarly, Gerada et al. emphasized transparent feature attribution and biological interpretability when predicting MICs from genomic information (Gerada et al., 2026). These studies highlight an important tension within the field: highly interpretable models may offer greater mechanistic value for antimicrobial development and stewardship, whereas more complex models often provide superior predictive accuracy at the expense of biological transparency.

Model performance has generally been strongest when predictions are restricted to specific species-antibiotic combinations. Nguyen et al. achieved high MIC prediction accuracy in nontyphoidal Salmonella (Nguyen et al., 2019), while Humphries et al. reported agreement between predicted and observed cefepime susceptibility in Escherichia coli (Humphries et al., 2023). Likewise, the Acinetobacter baumannii k-mer study demonstrated categorical agreement exceeding 95% for several antibiotics (Gao et al., 2024), and Demczuk et al. showed that known resistance determinants explain substantial variation in Neisseria gonorrhoeae (Demczuk et al., 2020). Comparable trends are seen in organism-specific studies such as Gao et al. (2024), where performance improved within a single species, and in Benkwitz-Bedford et al., where unexplained variance in predictive models was attributed primarily to missing heritability — genetic effects too rare or weak to be captured — rather than to dataset inconsistency alone (Benkwitz-Bedford et al., 2021). Together, these studies illustrate that machine learning can effectively capture susceptibility patterns within relatively homogeneous datasets.

AI models for the prediction of antimicrobial peptide activity and MIC

Antimicrobial peptides (AMPs) represent a distinct class of anti-infective agents that differ substantially from conventional small-molecule therapeutics in their structure, physicochemical properties, mechanisms of action, and data representation. Consequently, AI models developed for AMP activity and MIC prediction have largely followed a separate methodological trajectory, relying on amino acid sequence information, protein language models, convolutional neural networks, and transformer architectures rather than molecular descriptors, fingerprints, or graph-based representations commonly used for small molecules. These biological and computational differences warrant separate consideration when evaluating AI approaches for antimicrobial susceptibility and MIC prediction.

A major distinction between AMP and small-molecule modeling lies in the nature of the underlying training datasets. While small-molecule models frequently leverage large repositories such as ChEMBL, PubChem, and DrugBank, AMP-focused studies typically rely on curated peptide databases assembled from experimental studies and literature reports. Common resources include APD (Wang et al., 2026), CAMP (Waghu et al., 2016), DRAMP (Ma et al., 2025), and DBAASP (Pirtskhalava et al., 2021), which provide peptide sequences, antimicrobial activity annotations, physicochemical properties, and, in some cases, quantitative MIC measurements. However, AMP datasets are often substantially smaller and more heterogeneous than traditional small-molecule datasets, creating unique challenges for model development and evaluation (Ramazi et al., 2022; Wang et al., 2022).

Data harmonization presents a particularly important challenge for AMP datasets because peptide activity data are frequently aggregated from multiple laboratories using different peptide synthesis protocols, purification methods, microbial strains, inoculum sizes, growth media, susceptibility testing methodologies, and MIC reporting standards. Additional challenges include sequence redundancy, inconsistent sequence annotation, differing approaches to negative sample construction, and varying definitions of antimicrobial activity. Consequently, these differences can introduce systematic bias, inflate apparent model performance, or reduce generalizability across independent datasets. Transparent reporting of data curation, redundancy reduction, activity definitions, and cross-dataset harmonization procedures is therefore essential for developing robust and reproducible AMP prediction models.

These distinctive biological features have driven the development of specialized AI architectures for antimicrobial peptide prediction. For example, BERT-AmPEP60 (Cai et al., 2025) leverages transformer-based protein language models to learn contextual sequence relationships associated with peptide activity and MIC prediction, achieving Pearson correlation coefficients of approximately 0.81 and 0.75 for Escherichia coli and Staphylococcus aureus MIC prediction, respectively. Similarly, ANIA (Chiu et al., 2026) combines Chaos Game Representation with Inception modules and transformer-based attention mechanisms to capture both local sequence motifs and long-range sequence dependencies associated with antimicrobial activity, achieving Pearson correlation coefficients of approximately 0.75–0.79 (MSE 0.23–0.26) across S. aureus, E. coli, and P. aeruginosa MIC prediction tasks. Other approaches, such as MBC-Attention (Yan et al., 2023), employ multi-branch convolutional neural networks with attention mechanisms to identify biologically meaningful sequence patterns that contribute to peptide efficacy, reporting an average Pearson correlation coefficient of 0.775 (RMSE 0.533 log µM) for E. coli MIC prediction. Likewise, Chung et al. developed an ensemble deep-learning framework integrating protein language models, convolutional neural networks, and recurrent architectures that achieved improved predictive performance relative to individual component models (Chung et al., 2024). Collectively, these studies demonstrate that sequence-based deep-learning architectures are capable of capturing biologically relevant determinants of AMP activity and MIC while achieving performance metrics comparable to or exceeding many traditional machine-learning approaches.

Earlier applications of artificial intelligence to antimicrobial peptides primarily focused on peptide identification and activity classification rather than quantitative MIC prediction. For example, DeepAMP (Veltri et al., 2018) demonstrated that deep neural networks could distinguish antimicrobial from non-antimicrobial peptides using sequence-derived features, but like many early AMP classifiers, its performance was evaluated primarily on curated datasets that may not fully represent the diversity of naturally occurring peptides. Similarly, AMPlify (Li et al., 2022) employed attention-based deep learning architectures to improve AMP discovery and identify sequence features associated with antimicrobial activity; however, its predictions remained largely centered on classification rather than quantitative assessment of antimicrobial potency. More recently, generative frameworks such as AMP-GAN (Dean and Walper, 2020) expanded the use of AI beyond prediction toward de novo peptide design, generating novel peptide sequences with favorable predicted characteristics. Nevertheless, evaluation of generated peptides frequently relies on computational filtering and limited experimental validation, making it difficult to determine how well model-generated sequences generalize to real-world biological systems. Likewise, CAMPER (Shehadeh et al., 2026) represents a notable step toward mechanism-informed peptide design by integrating artificial intelligence with biological knowledge of persister-cell physiology. However, as with many contemporary AI-driven peptide discovery platforms, broader validation across diverse pathogens, infection models, and experimental conditions remains necessary to establish generalizability and translational utility. Collectively, these studies illustrate the evolution of AMP-focused AI from sequence classification toward mechanism-guided peptide design, while simultaneously highlighting persistent challenges related to dataset quality, experimental validation, biological interpretability, and external generalization.

The shift toward deep learning is especially apparent in MIC predictions of antimicrobial peptides. For example, Cai et al. employed BERT-derived embeddings for peptide activity prediction (Cai et al., 2025), Chiu et al. introduced ANIA, combining Chaos Game Representation with Inception modules and Transformer encoders (Chiu et al., 2026), and Chung et al. developed an ensemble framework integrating protein language models, convolutional neural networks, and recurrent architectures (Chung et al., 2024). The MBC-Attention model by Yan et al. incorporated multi-branch CNN architecture with global and local attention mechanisms specifically designed to capture sequence-level determinants of antimicrobial activity (Yan et al., 2023). In parallel, models such as the Random Forest-based peptide MIC predictor described by Bajiya et al. highlight how both traditional machine learning and modern deep learning frameworks are being applied to similar AMP datasets with differing levels of architectural complexity (Bajiya et al., 2025). Although these architectures frequently outperform traditional machine learning methods, the justification for their adoption is typically empirical rather than mechanistic, making it difficult to ascertain whether improvements reflect superior biological representation or increased model complexity.

Experimental validation remains the most significant gap across literature. Prospective in vitro and in vivo experimental validation studies were not consistently performed across studies. Most relied on retrospective validation using public datasets or external genomic collections (Nguyen et al., 2019; Demczuk et al., 2020; Yasir et al., 2022; Yang et al., 2023; Batisti Biffignandi et al., 2024; Gerada et al., 2026). The same criticism applies to many AMP-focused studies (Yan et al., 2023; Chung et al., 2024; Bajiya et al., 2025; Cai et al., 2025; Chiu et al., 2026). For example, Nguyen et al., Demczuk et al., and Bajiya et al. all demonstrate strong retrospective predictive performance but rely primarily on existing datasets rather than prospective experimental validation, highlighting a persistent translational gap (Nguyen et al., 2019; Demczuk et al., 2020; Bajiya et al., 2025). Across both scenarios, whether predicting pathogen susceptibility to existing antimicrobials or predicting peptide activity, the ideal starting dataset would be the internally validated biobank of isolates with genomic sequencing and well characterized and standardized phenotypic data. Consequently, such integration of in silico computational chemistry with in vitro microbiology and in vivo pharmacology could indeed lead to successful AI models for antimicrobial discovery (Imai et al., 2019; Stokes et al., 2020; Liu et al., 2023; Catacutan et al., 2025; Wong et al., 2025; Shehadeh et al., 2026).

AI models for the prediction of physicochemical properties

Dissociation constant (pKa)

In partition coefficient (or distribution) theory (Shore et al., 1957; Hogben et al., 1959), the acid dissociation constant (pKa) is one of the most important physicochemical properties governing the behavior of small molecules and ligands in biological systems because ionization state directly influences membrane permeability, protein binding, tissue distribution, and interactions with biological targets. In the context of antimicrobial PK/PD, the pKa determines the fraction of drug ionized at physiological pH, which directly affects oral absorption, penetration into acidic infection sites (e.g., abscesses, biofilms, intracellular compartments), protein binding, and renal elimination, all of which determine whether PK/PD targets such as f T>MIC, f AUC/MIC, or f Cmax/MIC are achieved at the site of infection.

For small-molecule pKa prediction (Table 3), representative AI models include Epik 7.0 (Shelley et al., 2007; Johnston et al., 2023), which employs an ensemble of atomic graph convolutional neural networks (GCNNs), MolGpKa (Pan et al., 2021), which utilizes a graph convolutional neural network architecture, and pKaLearn (Genzling et al., 2026), which is based on graph neural networks. More recent hybrid approaches, such as QupKake (Abarbanel and Hutchison, 2024), integrate graph-based learning with quantum chemical descriptors to better capture electronic effects governing protonation equilibria. For peptides the prediction of pKa navigates the complex relationship between protein sequence, structure, and protonation behavior. Notable examples include PyPKa (Reis et al., 2024), which represent modern computational approaches for predicting residue ionization states, as well as Protein pKa Language Model (PKALM) (Xu and Onoda, 2025) and its pKALM-shifted variant (pKALMs) (Xu and Onoda, 2025) which leverage large-scale protein language model architectures to extract contextual information from amino acid sequences.

Solubility, aqueous (LogS)

Aqueous solubility is a critical physicochemical property that influences drug absorption, bioavailability, formulation development, and overall pharmacokinetic performance. For antimicrobials, aqueous solubility directly determines the maximum absorbable dose for oral agents and the feasibility of intravenous formulations, thereby constraining the drug concentrations achievable at the site of infection and the ability to meet PK/PD targets. Poor aqueous solubility remains one of the leading causes of attrition during drug discovery and development, particularly as modern screening campaigns increasingly identify highly lipophilic and structurally complex compounds.

Early computational approaches relied on molecular descriptors and quantitative structure–property relationship (QSPR) methods (Table 3), exemplified by models such as ESOL (Delaney, 2004) and AqSolPred (Sorkun et al., 2019; Sorkun et al., 2021). More recently, advances in deep learning have led to the development of Chemprop framework (Heid et al., 2024; Graff et al., 2026), which utilizes graph-based representations to capture atomic and bond-level interactions associated with solubility. Additional graph neural network and representation-learning frameworks, including MolCLR (Wang et al., 2022) and SchNet (Schutt et al., 2018; Gao et al., 2020) have further improved predictive performance by leveraging self-supervised learning, attention mechanisms, and message-passing strategies to model complex molecular features Transformer-based solubility prediction models, such as SolTranNet (Francoeur and Koes, 2021) and SolvaFormer (Broadbent et al., 2025), employ attention mechanisms to model long-range molecular dependencies relevant to aqueous solubility. In parallel, foundation molecular language models, such as MolBERT (Fabian et al., 2020) and MolFormer (Ross et al., 2022), generate contextual molecular embeddings that can be fine-tuned for downstream molecular property prediction tasks, including aqueous solubility. Ensemble and gradient-boosting approaches, including XGB-298 (Ali et al., 2025) and WaSPred (Di Stefano et al., 2024), have also demonstrated strong predictive performance by integrating diverse molecular descriptors and learning strategies. In parallel, protein-informed architectures such as ProteinMPNN (Wang et al., 2023; Dolorfino et al., 2024; Sumida et al., 2024; Sun et al., 2025) have expanded the broader application of deep learning to molecular design tasks, including the generation of sequences with favorable physicochemical properties.

Because experimental determination of peptide and protein solubility is often time-consuming, resource-intensive, and highly dependent on environmental factors, early computational approaches primarily relied on sequence-derived features and conventional machine learning algorithms to estimate protein solubility. Representative examples include PROSO (Smialowski et al., 2007; Smialowski et al., 2012), SOLpro (Magnan et al., 2009), SoluProt (Hon et al., 2021), PaRSnIP (Rawi et al., 2018), Protein-Sol (Hebditch et al., 2017; Hebditch and Warwicker, 2019a; Hebditch and Warwicker, 2019b; Hebditch and Warwicker, 2020), PON-Sol2 (Yang et al., 2021), the Solubility-Weighted Index (SWI) (Bhandari et al., 2020), and ProG-SO (Li et al., 2025), which utilize amino acid composition, physicochemical properties, hydrophobicity profiles, charge distribution, and evolutionary information to identify proteins likely to exhibit favorable solubility characteristics. These models demonstrated that substantial predictive information could be extracted from primary amino acid sequences alone and provided valuable tools for prioritizing recombinant proteins for expression and downstream development. Additional methods such as SurfSol (Fang et al., 2026)and MPSol (Zhang et al., 2025) incorporated information related to solvent accessibility, surface properties, and structural descriptors, further improving prediction accuracy by accounting for features that influence protein–solvent interactions and aggregation propensity.

For peptide solubility, deep neural network architectures such as DeepSol (Khurana et al., 2018), DeepSoluE (Wang and Zou, 2023), and DSResSol (Madani et al., 2021) employ convolutional and residual neural networks to learn nonlinear relationships between amino acid sequence patterns and solubility outcomes. Graph-based approaches, including GraphSol (Chen et al., 2021) and GATSol (Li and Ming, 2024), further enhance predictive performance by representing proteins as interaction networks and leveraging graph neural networks and attention mechanisms to model residue-level relationships that contribute to solubility behavior. More recently, advances in protein foundation models and protein language models have transformed sequence-based developability prediction. Models such as NetSolP (Thumuluri et al., 2022), PROTSOLM (Tan et al., 2024), PLM_Sol (Zhang et al., 2024), and SOuLMuSiC (Attanasio et al., 2025) leverage large-scale self-supervised learning on millions of protein sequences to generate contextual embeddings that capture evolutionary, structural, and functional information relevant to solubility. These approaches benefit from transfer learning strategies similar to those used in natural language processing and often demonstrate superior generalizability compared with earlier machine learning methods.

Lipophilicity (logP) and distribution (logD)

Lipophilicity, commonly expressed as logP or logD, is a fundamental determinant of absorption, tissue distribution, membrane permeability, and toxicity (Table 3). For antimicrobials, lipophilicity directly governs the volume of distribution, intracellular penetration, and protein binding — properties that determine free drug concentrations at the site of infection and whether PK/PD targets (e.g., f AUC/MIC, f Cmax/MIC) are achieved. Highly lipophilic antimicrobials tend to exhibit greater tissue penetration but also higher protein binding, which can reduce the pharmacologically active free fraction. Traditional experimental determination relies on shake-flask assays or chromatographic methods; AI models can potentially predict the lipophilicity of experimental therapeutics within a margin of accuracy.

ALipSol (Wu et al., 2022) demonstrated that deep learning architectures can accurately estimate lipophilicity-related properties from molecular descriptors and graph-derived features. Similarly, the multitask D-MPNN (Lenselink and Stouten, 2021) prediction of logP simultaneously learned multiple physicochemical endpoints, including logP, improving generalizability through shared molecular representations. Other examples of GNN-based modeling include those of Tang et al. (2020), Wieder et al. (2021), and Li et al. (2026) Experimental chromatographic data has also been increasingly incorporated into machine learning workflows. RTlogD (Wang et al., 2023) employed retention time information to improve logD prediction accuracy, while subsequent models (Win et al., 2023) directly integrated liquid chromatography retention data with molecular descriptors to predict both logP and logD. More recently, improved GNN-based logD7.4 predictors (Duan et al., 2023) incorporated advanced message-passing schemes and larger training datasets, achieving enhanced accuracy for pharmaceutical compounds under physiologically relevant ionization conditions.

Conclusions

The application of AI to antimicrobial PK/PD modeling and MIC prediction has demonstrated considerable potential to accelerate drug development, optimize dosing strategies, and support precision antimicrobial therapy. Across the reviewed literature of AI models, both machine and deep learning approaches frequently achieved high predictive performance, particularly when large, well-characterized datasets were available. However, the transparency and relevance of these models remain heavily dependent on the underlying data, methodological choices, and the rigor of validation procedures. Many studies emphasized predictive accuracy while providing limited discussion of data curation practices, model interpretability, external generalizability, or biological plausibility, thereby restricting confidence in the robustness of reported findings.

A recurring finding is that most studies relied on heterogeneous datasets aggregated from public databases, surveillance programs, or previously published experiments, yet relatively few provide detailed descriptions of data curation, preprocessing, de-duplication, correction of annotation errors, and standardization of susceptibility testing and reporting. Because MIC measurements are inherently influenced by laboratory methodology, inoculum effects, and inter-laboratory variability, inadequate data cleaning introduces noise that may be incorrectly learned as biologically meaningful patterns. This challenge is further compounded by concept drift, the phenomenon whereby the relationship between genomic inputs and resistance phenotypes changes over time due to the evolving trajectory of AMR, which can degrade model performance if not explicitly addressed through time-varying or continuously updated modeling approaches (Howard et al., 2026).

To address challenges in AI-enabled drug discovery and pharmacology, benchmarking initiatives such as the Therapeutic Data Commons (TDC) (Huang et al., 2021) provide standardized evaluation frameworks for ADMET prediction models across 22 endpoints, including permeability, clearance, protein binding, and toxicity. Although these benchmarks facilitate rigorous assessment and comparison of AI models for drug property prediction, they do not currently encompass antimicrobial-specific PK/PD endpoints such as fT>MIC or fAUC/MIC. In contrast, the CAMO-Net GDR initiative (Mutegeki et al., 2025) aims to establish a global data resource specifically for antimicrobial PK/PD, addressing the urgent need for accessible, standardized, and interlinked datasets to improve the rigor, reproducibility, and translatability of pharmacokinetic and pharmacodynamic modeling across diverse populations (Tian et al., 2022; Swanson et al., 2024; Mutegeki et al., 2025). Together these collective initiatives aim to provide greater rigor, transparency, and translatability of these AI models into real-world applications.

Future studies that develop AI models for PK/PD should prioritize standardized data preprocessing protocols, transparent justification of model architecture selection, robust internal validation (e.g., cross-validation), external validation on independent datasets from distinct geographic and temporal settings, and prospective experimental verification through in vitro, in vivo, and other external validations. Adherence to FAIR (Findable, Accessible, Interoperable, Reusable) data principles, as advocated by recent consensus frameworks, will be essential to ensure that both datasets and algorithms are transparent, reproducible, and minimally biased (Wilkinson et al., 2016; Wilkinson et al., 2019). Furthermore, model evaluation should extend beyond conventional performance metrics to include cost-sensitive frameworks that explicitly account for the clinical consequences of false-positive and false-negative predictions. These evaluations should also incorporate antimicrobial stewardship outcomes, including time to effective therapy, spectrum narrowing, and total days of therapy. Greater emphasis on explainable AI, reproducibility, and mechanistic integration with established pharmacological principles will be essential to ensure that predictive performance translates into meaningful clinical utility (Kim et al., 2022; Howard et al., 2026; Sardar et al., 2026; Warrier et al., 2026). Addressing these challenges will ultimately determine whether artificial intelligence serves as a transformative tool for antimicrobial stewardship and drug development or remains primarily an academic proof of concept.

References

• 1

  AbarbanelO. D.HutchisonG. R. (2024). QupKake: Integrating machine learning and quantum chemistry for micro-pK(a) predictions. J. Chem. Theory Comput.20, 6946–6956. doi: 10.1021/acs.jctc.4c00328

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  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

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  AdamsR.HenryK. E.SridharanA.SoleimaniH.ZhanA.RawatN.et al. (2022). Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis. Nat. Med.28, 1455–1460. doi: 10.1038/s41591-022-01894-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  AlcockB. P.HuynhW.ChalilR.SmithK. W.RaphenyaA. R.WlodarskiM. A.et al. (2023). CARD 2023: expanded curation, support for machine learning, and resistome prediction at the Comprehensive Antibiotic Resistance Database. Nucleic Acids Res.51, D690–D699. doi: 10.1093/nar/gkac920

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  AlcockB. P.RaphenyaA. R.LauT. T. Y.TsangK. K.BouchardM.EdalatmandA.et al. (2020). CARD 2020: antibiotic resistome surveillance with the comprehensive antibiotic resistance database. Nucleic Acids Res.48, D517–D525. doi: 10.1093/nar/gkz935

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  AliM.VanderheidenS.GrathwolC. W.KramerK.FriederichP.JungN.et al. (2025). Advancing aqueous solubility prediction: A machine learning approach for organic compounds using a curated data set. J. Chem. Inf. Model.65, 8426–8434. doi: 10.1021/acs.jcim.4c02399

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  AnahtarM. N.YangJ. H.KanjilalS. (2021). Applications of machine learning to the problem of antimicrobial resistance: an emerging model for translational research. J. Clin. Microbiol.59, e0126020. doi: 10.1128/JCM.01260-20

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  ArdilaC. M.YadalamP. K.Gonzalez-ArroyaveD. (2024). Integrating whole genome sequencing and machine learning for predicting antimicrobial resistance in critical pathogens: a systematic review of antimicrobial susceptibility tests. PeerJ12, e18213. doi: 10.7717/peerj.18213

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  ArgimonS.DavidS.UnderwoodA.AbrudanM.WheelerN. E.KekreM.et al. (2021a). Rapid genomic characterization and global surveillance of Klebsiella using Pathogenwatch. Clin. Infect. Dis.73, S325–S335. doi: 10.1093/cid/ciab784

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  ArgimonS.YeatsC. A.GoaterR. J.AbudahabK.TaylorB.UnderwoodA.et al. (2021b). A global resource for genomic predictions of antimicrobial resistance and surveillance of Salmonella Typhi at pathogenwatch. Nat. Commun.12, 2879. doi: 10.1038/s41467-021-23091-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11

  AttanasioS.KwasigrochJ.RoomanM.PucciF. (2025). SOuLMuSiC, a novel tool for predicting the impact of mutations on protein solubility. Sci. Rep.15, 27531. doi: 10.1038/s41598-025-11326-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 12

  Food and Drug Administration. (2026). Roadmap to Reducing Animal Testing in Preclinical Safety Studies (Silver Spring, MD: Food and Drug Administration).

  • Google Scholar

• 13

  AyobamiO.WillrichN.SuwonoB.EckmannsT.MarkwartR. (2020). The epidemiology of carbapenem-non-susceptible Acinetobacter species in Europe: analysis of EARS-Net data from 2013 to 2017. Antimicrob. Resist. Infect. Control9, 89. doi: 10.1186/s13756-020-00750-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  BajiyaN.KumarN.RaghavaG. P. S. (2025). Prediction of inhibitory peptides against E.coli with desired MIC value. Sci. Rep.15, 4672. doi: 10.1038/s41598-025-86638-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  Batisti BiffignandiG.ChindelevitchL.CorbellaM.FeilE. J.SasseraD.LeesJ. A.. (2024). Optimising machine learning prediction of minimum inhibitory concentrations in Klebsiella pneumoniae. Microb. Genom.10. doi: 10.1099/mgen.0.001222

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  Benkwitz-BedfordS.PalmM.DemirtasT. Y.MustonenV.FarewellA.WarringerJ.et al. (2021). Machine learning prediction of resistance to subinhibitory antimicrobial concentrations from Escherichia coli genomes. mSystems6, e0034621. doi: 10.1128/mSystems.00346-21

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  BhandariB. K.GardnerP. P.LimC. S. (2020). Solubility-Weighted Index: fast and accurate prediction of protein solubility. Bioinformatics36, 4691–4698. doi: 10.1093/bioinformatics/btaa578

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18

  BortolaiaV.KaasR. S.RuppeE.RobertsM. C.SchwarzS.CattoirV.et al. (2020). ResFinder 4.0 for predictions of phenotypes from genotypes. J. Antimicrob. Chemother.75, 3491–3500. doi: 10.1093/jac/dkaa345

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 19

  BoxG. E.DraperN. R. (1987). Empirical Model-Building and Response Surfaces (New York, NY: John Wiley & Sons).

  • Google Scholar
  • View reference in article

• 20

  BroadbentJ.BaileyM.LiM.PaulA.De LescureL.ChauvinP.et al. (2025). Solvaformer: An SE(3)-equivariant graph transformer for small molecule solubility prediction. doi: 10.48550/arXiv.2511.09774

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  BryantP.PozzatiG.ElofssonA. (2022). Improved prediction of protein-protein interactions using AlphaFold2. Nat. Commun.13, 1265. doi: 10.1038/s41467-022-28865-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  CaiZ.PengH.SunS.HeJ.LuoF.HuangY. (2024). DeepKa web server: high-throughput protein pK(a) prediction. J. Chem. Inf. Model.64, 2933–2940. doi: 10.1021/acs.jcim.3c02013

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  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

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  CatacutanD. B.TranV.ArnoldA.AlexanderJ.CorsoG.YousefiY.et al. (2025). Discovery and artificial intelligence-guided mechanistic elucidation of a narrow-spectrum antibiotic. Nat. Microbiol.10, 2808–2822. doi: 10.1038/s41564-025-02142-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  ChenJ.ZhengS.ZhaoH.YangY. (2021). Structure-aware protein solubility prediction from sequence through graph convolutional network and predicted contact map. J. Cheminf.13, 7. doi: 10.1186/s13321-021-00488-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  ChengF.LiW.ZhouY.ShenJ.WuZ.LiuG.et al. (2012). admetSAR: a comprehensive source and free tool for assessment of chemical ADMET properties. J. Chem. Inf. Model.52, 3099–3105. doi: 10.1021/ci300367a

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  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

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 28

  ChungC. R.ChienC. Y.TangY.WuL. C.HsuJ. B.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

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 29

  CoddeC.FaucherJ. F.WoillardJ. B. (2025). Use of artificial intelligence in current fight against antimicrobial resistance. Microb. Drug Resist.32, mdr20240241. doi: 10.1089/mdr.2024.0241

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  Consortium, C. R (2024). Quantitative measurement of antibiotic resistance in Mycobacterium tuberculosis reveals genetic determinants of resistance and susceptibility in a target gene approach. Nat. Commun.15, 488. doi: 10.1038/s41467-023-44325-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 31

  DarrowJ. J.NajafzadehM.StefaniniK.KesselheimA. S. (2020). Regulatory approval characteristics of antimicrobial versus non-antimicrobial products, 1984-2018: an evaluation of Food and Drug Administration flexibilities. Lancet Infect. Dis.20, e159–e164. doi: 10.1016/S1473-3099(20)30197-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 32

  DaviesT. J.StoesserN.SheppardA. E.AbuounM.FowlerP.SwannJ.et al. (2020). Reconciling the potentially irreconcilable? Genotypic and phenotypic amoxicillin-clavulanate resistance in Escherichia coli. Antimicrob. Agents Chemother.64, e02026-19. doi: 10.1128/AAC.02026-19

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 33

  DavisJ. J.WattamA. R.AzizR. K.BrettinT.ButlerR.ButlerR. M.et al. (2020). The PATRIC Bioinformatics Resource Center: expanding data and analysis capabilities. Nucleic Acids Res.48, D606–D612. doi: 10.1093/nar/gkz943

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  DeanS. N.WalperS. A. (2020). Variational autoencoder for generation of antimicrobial peptides. ACS Omega5, 20746–20754. doi: 10.1021/acsomega.0c00442

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35

  DelaneyJ. S. (2004). ESOL: Estimating aqueous solubility directly from molecular structure. J. Chem. Inf. Comput. Sci.44, 1000–1005. doi: 10.1021/ci034243x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 36

  DemczukW.MartinI.SawatzkyP.AllenV.LefebvreB.HoangL.et al. (2020). Equations to predict antimicrobial MICs in Neisseria gonorrhoeae using molecular antimicrobial resistance determinants. Antimicrob. Agents Chemother.64, e02005-19. doi: 10.1128/AAC.02005-19

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  DiMasiJ. A.FlorezM. I.StergiopoulosS.PenaY.SmithZ.WilkinsonM.et al. (2020). Development times and approval success rates for drugs to treat infectious diseases. Clin. Pharmacol. Ther.107, 324–332. doi: 10.1002/cpt.1627

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38

  Di StefanoM.GalatiS.LonziC.GranchiC.PoliG.TuccinardiT.et al. (2024). WaSPred: A reliable AI-based water solubility predictor for small molecules. Int. J. Pharm.666, 124817. doi: 10.1016/j.ijpharm.2024.124817

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39

  DolorfinoM.SamantaR.VorobievaA. (2024). “ ProteinMPNN recovers complex sequence properties of transmembrane beta-barrels,” in Biorxiv. (Cold Spring Harbor, NY: Cold Spring Harbor Laboratory). doi: 10.1101/2024.01.16.575764

  • CrossRef
  • Google Scholar

• 40

  DongJ.WangN. N.YaoZ. J.ZhangL.ChengY.OuyangD.et al. (2018). ADMETlab: a platform for systematic ADMET evaluation based on a comprehensively collected ADMET database. J. Cheminf.10, 29. doi: 10.1186/s13321-018-0283-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 41

  DowdenH.MunroJ. (2019). Trends in clinical success rates and therapeutic focus. Nat. Rev. Drug Discov.18, 495–496. doi: 10.1038/d41573-019-00074-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42

  DrusanoG. L. (2004). Antimicrobial pharmacodynamics: critical interactions of 'bug and drug'. Nat. Rev. Microbiol.2, 289–300. doi: 10.1038/nrmicro862

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43

  DuanY. J.FuL.ZhangX. C.LongT. Z.HeY. H.LiuZ. Q.et al. (2023). Improved GNNs for log D(7.4) prediction by transferring knowledge from low-fidelity data. J. Chem. Inf. Model.63, 2345–2359. doi: 10.1021/acs.jcim.2c01564

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44

  El MammeryA.Ramirez De ArellanoE.Canada-GarciaJ. E.CercenadoE.Villar-GomaraL.Casquero-GarciaV.et al. (2023). An increase in erythromycin resistance in methicillin-susceptible Staphylococcus aureus from blood correlates with the use of macrolide/lincosamide/streptogramin antibiotics. EARS-Net Spain (2004-2020). Front. Microbiol.14, 1220286. doi: 10.3389/fmicb.2023.1220286

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45

  FabianB.EdlichT.GasparH.SeglerM.MeyersJ.FiscatoM.et al. (2020). Molecular representation learning with language models and domain-relevant auxiliary tasks. doi: 10.48550/arXiv.2011.13230

  • CrossRef
  • Google Scholar

• 46

  FangY. H.WeiL.XingB. B.WeiZ. G. (2026). SurfSol: A multimodal surface-based deep learning framework for protein solubility prediction. J. Chem. Inf. Model.66, 3737–3746. doi: 10.1021/acs.jcim.6c00124

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  FeldgardenM.BroverV.HaftD. H.PrasadA. B.SlottaD. J.TolstoyI.et al. (2019). Validating the AMRFinder Tool and Resistance Gene Database by using antimicrobial resistance genotype-phenotype correlations in a collection of isolates. Antimicrob. Agents Chemother.63, e00483-19. doi: 10.1128/AAC.00483-19

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 48

  FeldgardenM.BroverV.Gonzalez-EscalonaN.FryeJ. G.HaendigesJ.HaftD. H.et al. (2021). AMRFinderPlus and the Reference Gene Catalog facilitate examination of the genomic links among antimicrobial resistance, stress response, and virulence. Sci. Rep.11, 12728. doi: 10.1038/s41598-021-91456-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49

  FeldgardenM.BroverV.FedorovB.HaftD. H.PrasadA. B.KlimkeW. (2022). Curation of the AMRFinderPlus databases: applications, functionality and impact. Microb. Genom.8. doi: 10.1099/mgen.0.000832

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 50

  FrancoeurP. G.KoesD. R. (2021). SolTranNet-A machine learning tool for fast aqueous solubility prediction. J. Chem. Inf. Model.61, 2530–2536. doi: 10.1021/acs.jcim.1c00331

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51

  FuL.ShiS.YiJ.WangN.HeY.WuZ.et al. (2024). ADMETlab 3.0: an updated comprehensive online ADMET prediction platform enhanced with broader coverage, improved performance, API functionality and decision support. Nucleic Acids Res.52, W422–W431. doi: 10.1093/nar/gkae236

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52

  GagliottiC.BalodeA.BaqueroF.DegenerJ.GrundmannH.GurD.et al. (2011). Escherichia coli and Staphylococcus aureus: bad news and good news from the European Antimicrobial Resistance Surveillance Network (EARS-Net, formerly EARSS), 2002 to 2009. Euro Surveill16. doi: 10.2807/ese.16.11.19819-en

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 53

  GaoY.LiH.ZhaoC.LiS.YinG.WangH. (2024). Machine learning and feature extraction for rapid antimicrobial resistance prediction of Acinetobacter baumannii from whole-genome sequencing data. Front. Microbiol.14, 1320312. doi: 10.3389/fmicb.2023.1320312

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 54

  GaoP.ZhangJ.SunY.YuJ. (2020). Accurate predictions of aqueous solubility of drug molecules via the multilevel graph convolutional network (MGCN) and SchNet architectures. Phys. Chem. Chem. Phys.22, 23766–23772. doi: 10.1039/d0cp03596c

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 55

  GaultonA.BellisL. J.BentoA. P.ChambersJ.DaviesM.HerseyA.et al. (2012). ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res.40, D1100–D1107. doi: 10.1093/nar/gkr777

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 56

  GenzlingJ.LuoZ.WeiserB.MoitessierN. (2026). Development of a pKa predictor (pKaLearn) by leveraging teaching experience to improve machine learning. Commun. Chem.9, 181. doi: 10.1038/s42004-026-01983-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 57

  GeradaA.ZhongY.HarperN.VelluvaA.RezaN.DubeyV.et al. (2026). Prediction of antimicrobial minimum inhibitory concentration from bacterial genomes using a scalable and interpretable machine learning approach. NPJ Antimicrob. Resist. 4 (1), 51 doi: 10.1038/s44259-026-00217-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58

  GoverdeC. A.PacesaM.GoldbachN.DornfeldL. J.BalbiP. E. M.GeorgeonS.et al. (2024). Computational design of soluble and functional membrane protein analogues. Nature631, 449–458. doi: 10.1038/s41586-024-07601-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 59

  GraffD. E.MorganN. K.BurnsJ. W.DonerA. C.LiB.LiS. C.et al. (2026). Chemprop v2: An efficient, modular machine learning package for chemical property prediction. J. Chem. Inf. Model.66, 28–33. doi: 10.1021/acs.jcim.5c02332

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 60

  GuY.YuZ.WangY.ChenL.LouC.YangC.et al. (2024). admetSAR3.0: a comprehensive platform for exploration, prediction and optimization of chemical ADMET properties. Nucleic Acids Res.52, W432–W438. doi: 10.1093/nar/gkae298

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 61

  HayM.ThomasD. W.CraigheadJ. L.EconomidesC.RosenthalJ. (2014). Clinical development success rates for investigational drugs. Nat. Biotechnol.32, 40–51. doi: 10.1038/nbt.2786

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 62

  HebditchM.Carballo-AmadorM. A.CharonisS.CurtisR.WarwickerJ. (2017). Protein-Sol: a web tool for predicting protein solubility from sequence. Bioinformatics33, 3098–3100. doi: 10.1093/bioinformatics/btx345

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 63

  HebditchM.WarwickerJ. (2019a). Web-based display of protein surface and pH-dependent properties for assessing the developability of biotherapeutics. Sci. Rep.9, 1969. doi: 10.1038/s41598-018-36950-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 64

  HebditchM.WarwickerJ. (2019b). Charge and hydrophobicity are key features in sequence-trained machine learning models for predicting the biophysical properties of clinical-stage antibodies. PeerJ7, e8199. doi: 10.7717/peerj.8199

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 65

  HebditchM.WarwickerJ. (2020). Protein-sol pKa: prediction of electrostatic frustration, with application to coronaviruses. Bioinformatics36, 5112–5114. doi: 10.1093/bioinformatics/btaa646

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 66

  HeidE.GreenmanK. P.ChungY.LiS. C.GraffD. E.VermeireF. H.et al. (2024). Chemprop: A machine learning package for chemical property prediction. J. Chem. Inf. Model.64, 9–17. doi: 10.1021/acs.jcim.3c01250

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 67

  HoegT. B.KozlowskiS.SaberH.MakaryM. A. (2026). Eliminating unnecessary animal testing. JAMA335, 2011–2014. doi: 10.1001/jama.2026.7074

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 68

  HogbenC. A.ToccoD. J.BrodieB. B.SchankerL. S. (1959). On the mechanism of intestinal absorption of drugs. J. Pharmacol. Exp. Ther.125, 275–282. doi: 10.1016/S0022-3565(25)12995-9

  • CrossRef
  • Google Scholar

• 69

  HonJ.MarusiakM.MartinekT.KunkaA.ZendulkaJ.BednarD.et al. (2021). SoluProt: prediction of soluble protein expression in Escherichia coli. Bioinformatics37, 23–28. doi: 10.1093/bioinformatics/btaa1102

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 70

  HowardJ. (2019). Artificial intelligence: implications for the future of work. Am. J. Ind. Med.62, 917–926. doi: 10.1002/ajim.23037

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 71

  HowardA.RezaN.GreenP. L.YinM.DuffyE.MwandumbaH. C.et al. (2026). Artificial intelligence and infectious diseases: tackling antimicrobial resistance, from personalised care to antibiotic discovery. Lancet Infect. Dis.26, e181–e192. doi: 10.1016/S1473-3099(25)00313-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 72

  HuK.MeyerF.DengZ. L.AsgariE.KuoT. H.MunchP. C.et al. (2024). Assessing computational predictions of antimicrobial resistance phenotypes from microbial genomes. Briefings Bioinf.25, bbae206. doi: 10.1093/bib/bbae206

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 73

  HuangK.FuT.GaoW.ZhaoY.RoohaniY.LeskovecJ.et al. (2021). Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development. (Ithaca, NY: Cornell University (arXiv)).

  • Google Scholar

• 74

  HumphriesR. M.BraginE.ParkhillJ.MoralesG.SchmitzJ. E.RhodesP. A. (2023). Machine-learning model for prediction of cefepime susceptibility in Escherichia coli from whole-genome sequencing data. J. Clin. Microbiol.61, e0143122. doi: 10.1128/jcm.01431-22

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 75

  ImaiY.MeyerK. J.IinishiA.Favre-GodalQ.GreenR.ManuseS.et al. (2019). A new antibiotic selectively kills Gram-negative pathogens. Nature576, 459–464. doi: 10.1038/s41586-019-1791-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 76

  IngberD. E. (2022). Human organs-on-chips for disease modelling, drug development and personalized medicine. Nat. Rev. Genet.23, 467–491. doi: 10.1038/s41576-022-00466-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 77

  JamalS.KhubaibM.GangwarR.GroverS.GroverA.HasnainS. E. (2020). Artificial Intelligence and Machine learning based prediction of resistant and susceptible mutations in Mycobacterium tuberculosis. Sci. Rep.10, 5487. doi: 10.1038/s41598-020-62368-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 78

  JiaB.RaphenyaA. R.AlcockB.WaglechnerN.GuoP.TsangK. K.et al. (2017). CARD 2017: expansion and model-centric curation of the comprehensive antibiotic resistance database. Nucleic Acids Res.45, D566–D573. doi: 10.1093/nar/gkw1004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 79

  JiaX.ZhangJ.ChenJ.WangJ.ChuX.YuW.et al. (2026). Machine learning-based prediction of multi-level antimicrobial resistance in Klebsiella pneumoniae using whole-genome sequencing data. Int. J. Antimicrob. Agents67, 107807. doi: 10.1016/j.ijantimicag.2026.107807

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 80

  JohnstonR. C.YaoK.KaplanZ.ChelliahM.LeswingK.SeekinsS.et al. (2023). Epik: pK(a) and protonation state prediction through machine learning. J. Chem. Theory Comput.19, 2380–2388. doi: 10.1021/acs.jctc.3c00044

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 81

  JordanM. I.MitchellT. M. (2015). Machine learning: trends, perspectives, and prospects. Science349, 255–260. doi: 10.1126/science.aaa8415

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 82

  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

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 83

  KearnesS.McCloskeyK.BerndlM.PandeV.RileyP. (2016). Molecular graph convolutions: moving beyond fingerprints. J. Comput.-Aided Mol. Des.30, 595–608. doi: 10.1007/s10822-016-9938-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 84

  KhuranaS.RawiR.KunjiK.ChuangG. Y.BensmailH.MallR. (2018). DeepSol: a deep learning framework for sequence-based protein solubility prediction. Bioinformatics34, 2605–2613. doi: 10.1093/bioinformatics/bty166

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 85

  KimS.ThiessenP. A.BoltonE. E.ChenJ.FuG.GindulyteA.et al. (2016). PubChem substance and compound databases. Nucleic Acids Res.44, D1202–D1213. doi: 10.1093/nar/gkv951

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 86

  KimJ. I.MaguireF.TsangK. K.GouliourisT.PeacockS. J.McAllisterT. A.et al. (2022). Machine learning for antimicrobial resistance prediction: current practice, limitations, and clinical perspective. Clin. Microbiol. Rev.35, e0017921. doi: 10.1128/cmr.00179-21

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 87

  KimS.ChenJ.ChengT.GindulyteA.HeJ.HeS.et al. (2025). PubChem 2025 update. Nucleic Acids Res.53, D1516–D1525. doi: 10.1093/nar/gkae1059

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 88

  LakinS. M.DeanC.NoyesN. R.DettenwangerA.RossA. S.DosterE.et al. (2017). MEGARes: an antimicrobial resistance database for high throughput sequencing. Nucleic Acids Res.45, D574–D580. doi: 10.1093/nar/gkw1009

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 89

  LandersdorferC. B.NationR. L. (2021). Limitations of antibiotic MIC-based PK-PD metrics: looking back to move forward. Front. Pharmacol.12, 770518. doi: 10.3389/fphar.2021.770518

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 90

  LeCunY.BengioY.HintonG. (2015). Deep learning. Nature521, 436–444. doi: 10.1038/nature14539

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 91

  LenselinkE. B.StoutenP. F. W. (2021). Multitask machine learning models for predicting lipophilicity (logP) in the SAMPL7 challenge. J. Comput.-Aided Mol. Des.35, 901–909. doi: 10.1007/s10822-021-00405-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 92

  LiY.CuiX.YangX.LiuG.ZhangJ. (2024). Artificial intelligence in predicting pathogenic microorganisms' antimicrobial resistance: challenges, progress, and prospects. Front. Cell. Infect. Microbiol.14, 1482186. doi: 10.3389/fcimb.2024.1482186

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 93

  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 pathogens. BMC Genomics23, 77. doi: 10.1186/s12864-022-08310-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 94

  LiJ.GuY.GuoQ.LaiR.HanW. (2026). Modeling method based on traditional machine learning and graph neural network for interpretable prediction of short peptide logD. J. Phys. Chem. B130, 4624–4635. doi: 10.1021/acs.jpcb.6c00299

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 95

  LiB.MingD. (2024). GATSol, an enhanced predictor of protein solubility through the synergy of 3D structure graph and large language modeling. BMC Bioinf.25, 204. doi: 10.1186/s12859-024-05820-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 96

  LiG.ZhangN.FanL. (2025). ProG-SOL: predicting protein solubility using protein embeddings and dual-graph convolutional networks. ACS Omega10, 3910–3916. doi: 10.1021/acsomega.4c09688

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 97

  LiesenfeldO.AroraS.AufderheideT. P.ClementsC. M.DevosE.FischerM.et al. (2025). Clinical validation of an AI-based blood testing device for diagnosis and prognosis of acute infection and sepsis. Nat. Med.31, 4044–4054. doi: 10.1038/s41591-025-03933-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 98

  LiuG.CatacutanD. B.RathodK.SwansonK.JinW.MohammedJ. C.et al. (2023). Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii. Nat. Chem. Biol.19, 1342–1350. doi: 10.1038/s41589-023-01349-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 99

  LiuS.YangQ.ZhangL.LuoS. (2024). Accurate protein pK(a) prediction with physical organic chemistry guided 3D protein representation. J. Chem. Inf. Model.64, 4410–4418. doi: 10.1021/acs.jcim.4c00354

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 100

  LundbergS. M.LeeS.-I. (2017). A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst.30. doi: 10.48550/arXiv.1705.07874

  • CrossRef
  • Google Scholar

• 101

  MaT.LiuY.YuB.SunX.YaoH.HaoC.et al. (2025). DRAMP 4.0: An open-access data repository dedicated to the clinical translation of antimicrobial peptides. Nucleic Acids Res.53, D403–D410. doi: 10.1093/nar/gkae1046

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 102

  MadaniM.LinK.TarakanovaA. (2021). DSResSol: A sequence-based solubility predictor created with dilated squeeze excitation residual networks. Int. J. Mol. Sci.22. doi: 10.3390/ijms222413555

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 103

  MagnanC. N.RandallA.BaldiP. (2009). SOLpro: Accurate sequence-based prediction of protein solubility. Bioinformatics25, 2200–2207. doi: 10.1093/bioinformatics/btp386

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 104

  MaoB.GaoY.XuC.MachaS.ShaoS.AhamadiM. (2025). Opportunities for AI-based model-informed drug development: A comparative analysis of NONMEM and AI-based models for population pharmacokinetic prediction. AAPS J.28, 21. doi: 10.1208/s12248-025-01121-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 105

  MayrF.WiederM.WiederO.LangerT. (2022). Improving small molecule pK (a) prediction using transfer learning with graph neural networks. Front. Chem.10, 866585. doi: 10.3389/fchem.2022.866585

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 106

  McCarthyJ. (2007). What is artificial intelligence. Stanford University. Available online at: https://www-formal.stanford.edu/jmc/whatisai.pdf.

  • Google Scholar

• 107

  McCarthyJ.MinskyM. L.RochesterN.ShannonC. E. (2006). “ A proposal for the dartmouth summer research project on artificial intelligence, august 31, 1955,” in Ai Mag. (Menlo Park, CA: American Association for Artificial Intelligence (AAAI)), vol. 27. , 12–12.

  • Google Scholar

• 108

  MiaoR.LiuD.MaoL.ChenX.ZhangL.YuanZ.et al. (2024). GR-pKa: a message-passing neural network with retention mechanism for pKa prediction. Briefings Bioinf.25, bbae408. doi: 10.1093/bib/bbae408

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 109

  MolnarC. (2020). Interpretable machine learning. Interpretable Machine Learning: A Guide for Making Black Box Models Explainable. Available online at: https://christophm.github.io/interpretable-ml-book.

  • Google Scholar

• 110

  MutegekiH.HolmesA. H.JjingoD.GaliwangoR.KambuguA.ShafiqN.et al. (2025). There is urgent need for a Global Data Resource for Antimicrobial PK/PD: CAMO-Net GDR Initiative. Nat. Commun.16, 9766. doi: 10.1038/s41467-025-64707-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 111

  MyungY.de SaA. G. C.AscherD. B. (2024). Deep-PK: deep learning for small molecule pharmacokinetic and toxicity prediction. Nucleic Acids Res.52, W469–W475. doi: 10.1093/nar/gkae254

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 112

  NguyenM.LongS. W.McDermottP. F.OlsenR. J.OlsonR.StevensR. L.et al. (2019). Using machine learning to predict antimicrobial MICs and associated genomic features for nontyphoidal Salmonella. J. Clin. Microbiol.57, e01260-18. doi: 10.1128/JCM.01260-18

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 113

  NguyenQ. H.NgoH. H.Nguyen-VoT. H.DoT. T. T.RahardjaS.NguyenB. P. (2023). eMIC-AntiKP: Estimating minimum inhibitory concentrations of antibiotics towards Klebsiella pneumoniae using deep learning. Comput. Struct. Biotechnol. J.21, 751–757. doi: 10.1016/j.csbj.2022.12.041

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 114

  NigoM.RasmyL.MaoB.KannadathB. S.XieZ.ZhiD. (2024). Deep learning model for personalized prediction of positive MRSA culture using time-series electronic health records. Nat. Commun.15, 2036. doi: 10.1038/s41467-024-46211-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 115

  OlatunjiI.BardajiD. K. R.MirandaR. R.SavkaM. A.HudsonA. O. (2024). Artificial intelligence tools for the identification of antibiotic resistance genes. Front. Microbiol.15, 1437602. doi: 10.3389/fmicb.2024.1437602

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 116

  PanX.WangH.LiC.ZhangJ. Z. H.JiC. (2021). MolGpka: A web server for small molecule pK(a) prediction using a graph-convolutional neural network. J. Chem. Inf. Model.61, 3159–3165. doi: 10.1021/acs.jcim.1c00075

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 117

  PatakiB. A.MatamorosS.van der PuttenB. C. L.RemondiniD.GiampieriE.Aytan-AktugD.et al. (2020). Understanding and predicting ciprofloxacin minimum inhibitory concentration in Escherichia coli with machine learning. Sci. Rep.10, 15026. doi: 10.1038/s41598-020-71693-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 118

  PhamN. P.GingrasH.GodinC.FengJ.GroppiA.NikolskiM.et al. (2024). Holistic understanding of trimethoprim resistance in Streptococcus pneumoniae using an integrative approach of genome-wide association study, resistance reconstruction, and machine learning. mBio15, e0136024. doi: 10.1128/mbio.01360-24

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 119

  PiresD. E.BlundellT. L.AscherD. B. (2015). pkCSM: Predicting small-molecule pharmacokinetic and toxicity properties using graph-based signatures. J. Med. Chem.58, 4066–4072. doi: 10.1021/acs.jmedchem.5b00104

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 120

  PirtskhalavaM.AmstrongA. A.GrigolavaM.ChubinidzeM.AlimbarashviliE.VishnepolskyB.et al. (2021). DBAASP v3: Database of antimicrobial/cytotoxic activity and structure of peptides as a resource for development of new therapeutics. Nucleic Acids Res.49, D288–D297. doi: 10.1093/nar/gkaa991

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 121

  RamaziS.MohammadiN.AllahverdiA.KhaliliE.AbdolmalekiP. (2022). A review on antimicrobial peptides databases and the computational tools. Database2022, baac011. doi: 10.1093/database/baac011

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 122

  RawiR.MallR.KunjiK.ShenC. H.KwongP. D.ChuangG. Y. (2018). PaRSnIP: sequence-based protein solubility prediction using gradient boosting machine. Bioinformatics34, 1092–1098. doi: 10.1093/bioinformatics/btx662

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 123

  RehmanA.JeukensJ.LevesqueR. C.LamontI. L. (2021). Gene-gene interactions dictate ciprofloxacin resistance in Pseudomonas aeruginosa and facilitate prediction of resistance phenotype from genome sequence data. Antimicrob. Agents Chemother.65, e0269620. doi: 10.1128/AAC.02696-20

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 124

  ReisP.ClevertD. A.MachuqueiroM. (2024). PypKa server: Online pKa predictions and biomolecular structure preparation with precomputed data from PDB and AlphaFold DB. Nucleic Acids Res.52, W294–W298. doi: 10.1093/nar/gkae255

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 125

  ReisP.BertoliniM.MontanariF.RocchiaW.MachuqueiroM.ClevertD. A. (2022). A fast and interpretable deep learning approach for accurate electrostatics-driven pK(a) predictions in proteins. J. Chem. Theory Comput.18, 5068–5078. doi: 10.1021/acs.jctc.2c00308

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 126

  RenY.ChakrabortyT.DoijadS.FalgenhauerL.FalgenhauerJ.GoesmannA.et al. (2022). Prediction of antimicrobial resistance based on whole-genome sequencing and machine learning. Bioinformatics38, 325–334. doi: 10.1093/bioinformatics/btab681

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 127

  RongY.BianY.XuT.XieW.WeiY.HuangW.et al. (2020). Self-supervised graph transformer on large-scale molecular data. doi: 10.48550/arXiv.2007.02835

  • CrossRef
  • Google Scholar

• 128

  RossJ.BelgodereB.ChenthamarakshanV.PadhiI.MrouehY.DasP. (2022). Large-scale chemical language representations capture molecular structure and properties. Nat. Mach. Intell.4, 1256–1264. doi: 10.1038/s42256-022-00580-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 129

  RyuB.JeonW.KimD. (2024). Integrating genomic and molecular data to predict antimicrobial minimum inhibitory concentration in Klebsiella pneumoniae. Sci. Rep.14, 25951. doi: 10.1038/s41598-024-75973-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 130

  SadybekovA. V.KatritchV. (2023). Computational approaches streamlining drug discovery. Nature616, 673–685. doi: 10.1038/s41586-023-05905-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 131

  Sanchez-BusoL.YeatsC. A.TaylorB.GoaterR. J.UnderwoodA.AbudahabK.et al. (2021). A community-driven resource for genomic epidemiology and antimicrobial resistance prediction of Neisseria gonorrhoeae at Pathogenwatch. Genome Med.13, 61. doi: 10.1186/s13073-021-00858-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 132

  SardarS.DashS.RoychowdhuryP.SolankiJ.SikdarP.ThangappanJ.et al. (2026). Artificial intelligence for antimicrobial resistance: advancing reproducibility, interpretability, and clinical deployment. Briefings Bioinf.27, bbag269. doi: 10.1093/bib/bbag269

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 133

  SchapinN.MajewskiM.Torrens-FontanalsM.De FabritiisG. (2024). PlayMolecule pKAce: Small Molecule Protonation through Equivariant Neural Networks. doi: 10.48550/arXiv.2407.11103

  • CrossRef
  • Google Scholar

• 134

  SchuttK. T.SaucedaH. E.KindermansP. J.TkatchenkoA.MullerK. R. (2018). SchNet - A deep learning architecture for molecules and materials. J. Chem. Phys.148, 241722. doi: 10.1063/1.5019779

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 135

  SchymanP.LiuR.DesaiV.WallqvistA. (2017). vNN web server for ADMET predictions. Front. Pharmacol.8, 889. doi: 10.3389/fphar.2017.00889

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 136

  SealS.TrapotsiM. A.MahaleM.SubramanianV.GreeneN.SpjuthO.et al. (2025). PKSmart: an open-source computational model to predict intravenous pharmacokinetics of small molecules. J. Cheminf.17, 147. doi: 10.1186/s13321-025-01066-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 137

  SertkayaA.BelecheT.JessupA.SommersB. D. (2024). Costs of drug development and research and development intensity in the US, 2000-2018. JAMA Netw. Open7, e2415445. doi: 10.1001/jamanetworkopen.2024.15445

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 138

  ShehadehF.MishraB.Ferrer-EspadaR.BasuA.FelixL.DellisC.et al. (2026). CAMPER: mechanistic artificial intelligence for designing peptides that target MRSA persisters. Nat. Commun.17, 3689. doi: 10.1038/s41467-026-70348-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 139

  ShelleyJ. C.CholletiA.FryeL. L.GreenwoodJ. R.TimlinM. R.UchimayaM. (2007). Epik: A software program for pK(a) prediction and protonation state generation for drug-like molecules. J. Comput.-Aided Mol. Des.21, 681–691. doi: 10.1007/s10822-007-9133-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 140

  ShenM.KortzakD.AmbrozakS.BhatnagarS.BuchananI.LiuR.et al (2025a). Protein electrostatic properties are fine-tuned through evolution. Res. Sq. doi: 10.21203/rs.3.rs-6471091/v1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 141

  ShenM.et al. (2025b). KaMLs for predicting protein pK(a) values and ionization states: are trees all you need? J. Chem. Theory Comput.21, 1446–1458. doi: 10.1021/acs.jctc.4c01602

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 142

  ShoreP. A.BrodieB. B.HogbenC. A. (1957). The gastric secretion of drugs: A pH partition hypothesis. J. Pharmacol. Exp. Ther.119, 361–369. doi: 10.1016/S0022-3565(25)11878-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 143

  SmialowskiP.DooseG.TorklerP.KaufmannS.FrishmanD. (2012). PROSO II--a new method for protein solubility prediction. FEBS J.279, 2192–2200. doi: 10.1111/j.1742-4658.2012.08603.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 144

  SmialowskiP.Martin-GalianoA. J.MikolajkaA.GirschickT.HolakT. A.FrishmanD.. (2007). Protein solubility: Sequence based prediction and experimental verification. Bioinformatics23, 2536–2542. doi: 10.1093/bioinformatics/btl623

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 145

  SorkunM. C.KhetanA.ErS. (2019). AqSolDB, a curated reference set of aqueous solubility and 2D descriptors for a diverse set of compounds. Sci. Data6, 143. doi: 10.1038/s41597-019-0151-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 146

  SorkunM. C.KoelmanJ.ErS. (2021). Pushing the limits of solubility prediction via quality-oriented data selection. iScience24, 101961. doi: 10.1016/j.isci.2020.101961

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 147

  StokesJ. M.YangK.SwansonK.JinW.Cubillos-RuizA.DonghiaN. M.et al. (2020). A deep learning approach to antibiotic discovery. Cell180, 688–702, e613. doi: 10.1016/j.cell.2020.01.021

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 148

  SumidaK. H.Nunez-FrancoR.KalvetI.PellockS. J.WickyB. I. M.MillesL. F.et al. (2024). Improving protein expression, stability, and function with ProteinMPNN. J. Am. Chem. Soc146, 2054–2061. doi: 10.1021/jacs.3c10941

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 149

  SunD.GaoW.HuH.ZhouS. (2022). Why 90% of clinical drug development fails and how to improve it? Acta Pharm. Sin. B.12, 3049–3062. doi: 10.1016/j.apsb.2022.02.002

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 150

  SunC. Q.LiZ. M.JiY.SchwanebergU.LiZ. L. (2025). CMDmpnn: Combining comparative molecular dynamics and ProteinMPNN to rapidly expand enzyme substrate spectrum. J. Chem. Inf. Model.65, 2741–2747. doi: 10.1021/acs.jcim.5c00117

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 151

  SwansonK.WaltherP.LeitzJ.MukherjeeS.WuJ. C.ShivnaraineR. V.et al. (2024). ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics40, btae416. doi: 10.1093/bioinformatics/btae416

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 152

  TanR.YuA.LiuZ.LiuZ.JiangR.WangX.et al. (2021). Prediction of minimal inhibitory concentration of meropenem against Klebsiella pneumoniae using metagenomic data. Front. Microbiol.12, 712886. doi: 10.3389/fmicb.2021.712886

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 153

  TanY.ZhengJ.HongL.ZhouB. (2024). ProtSolM: protein solubility prediction with multi-modal features. doi: 10.48550/arXiv.2406.19744

  • CrossRef
  • Google Scholar

• 154

  TangB.KramerS. T.FangM.QiuY.WuZ.XuD. (2020). A self-attention based message passing neural network for predicting molecular lipophilicity and aqueous solubility. J. Cheminf.12, 15. doi: 10.1186/s13321-020-0414-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 155

  ThumuluriV.MartinyH. M.Almagro ArmenterosJ. J.SalomonJ.NielsenH.JohansenA. R. (2022). NetSolP: predicting protein solubility in Escherichia coli using language models. Bioinformatics38, 941–946. doi: 10.1093/bioinformatics/btab801

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 156

  TianH.KetkarR.TaoP. (2022). ADMETboost: a web server for accurate ADMET prediction. J. Mol. Model.28, 408. doi: 10.1007/s00894-022-05373-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 157

  TornimbeneB.EreminS.EscherM.GriskevicieneJ.ManglaniS.Pessoa-SilvaC. L. (2018). WHO Global Antimicrobial Resistance Surveillance System early implementation 2016-17. Lancet Infect. Dis.18, 241–242. doi: 10.1016/S1473-3099(18)30060-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 158

  UrquhartR. J.van TeijlingenA.TuttleT. (2024). ANI neural network potentials for small molecule pK(a) prediction. Phys. Chem. Chem. Phys.26, 23934–23943. doi: 10.1039/d4cp01982b

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 159

  VeltriD.KamathU.ShehuA. (2018). Deep learning improves antimicrobial peptide recognition. Bioinformatics34, 2740–2747. doi: 10.1093/bioinformatics/bty179

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 160

  VenkatramanV. (2021). FP-ADMET: a compendium of fingerprint-based ADMET prediction models. J. Cheminf.13, 75. doi: 10.1186/s13321-021-00557-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 161

  VergauweF.De WaeleG.SassA.HighmoreC.HanrahanN.CookY.et al. (2025). Harnessing machine learning to predict antibiotic susceptibility in Pseudomonas aeruginosa biofilms. NPJ Biofilms Microbiomes11, 205. doi: 10.1038/s41522-025-00833-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 162

  WaghuF. H.BaraiR. S.GurungP.Idicula-ThomasS. (2016). CAMPR3: A database on sequences, structures and signatures of antimicrobial peptides. Nucleic Acids Res.44, D1094–D1097. doi: 10.1093/nar/gkv1051

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 163

  WanF.TongW.WuW.LiuJ.HuangH.ZhangP.et al. (2026). Utilizing whole genome sequencing data for machine learning driven prediction of antibiotic resistance in Escherichia coli. Front. Microbiol.17, 1842717. doi: 10.3389/fmicb.2026.1842717

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 164

  WangT.JinX.LuX.MinX.GeS.LiS. (2023). Empirical validation of ProteinMPNN's efficiency in enhancing protein fitness. Front. Genet.14, 1347667. doi: 10.3389/fgene.2023.1347667

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 165

  WangY.XiongJ.XiaoF.ZhangW.ChengK.RaoJ.et al. (2023). LogD7.4 prediction enhanced by transferring knowledge from chromatographic retention time, microscopic pKa and logP. J. Cheminf.15, 76. doi: 10.1186/s13321-023-00754-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 166

  WangC. C.HungY. T.ChouC. Y.HsuanS. L.ChenZ. W.ChangP. Y.et al. (2023). Using random forest to predict antimicrobial minimum inhibitory concentrations of nontyphoidal Salmonella in Taiwan. Vet. Res.54, 11. doi: 10.1186/s13567-023-01141-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 167

  WangC. X.FengY. J.XuZ. J.XuX. H.PengB. G. (2025). MolAttnNet: A predictive model for organic drug solubility based on graph convolutional networks and transformer-attention. IEEE Access13, 139605–139615. doi: 10.1109/Access.2025.3596145

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 168

  WangG.SchmidtC.LiX.WangZ. (2026). APD6: The antimicrobial peptide database is expanded to promote research and development by deploying an unprecedented information pipeline. Nucleic Acids Res.54, D363–D374. doi: 10.1093/nar/gkaf860

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 169

  WangG.VaismanI. I.van HoekM. L. (2022). Machine learning prediction of antimicrobial peptides. Methods Mol. Biol.2405, 1–37. doi: 10.1007/978-1-0716-1855-4_1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 170

  WangY. Y.WangJ. R.CaoZ. L.FarimaniA. B. (2022). Molecular contrastive learning of representations via graph neural networks. Nat. Mach. Intell.4, 279–287. doi: 10.1038/s42256-022-00447-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 171

  WangC.ZouQ. (2023). Prediction of protein solubility based on sequence physicochemical patterns and distributed representation information with DeepSoluE. BMC Biol.21, 12. doi: 10.1186/s12915-023-01510-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 172

  WarrierS. P.NarasimhanV.MeijerE.GutlinY.NolteO.VeeraraghavanB.et al. (2026). Uncloaking the black-box: the need for explainable artificial intelligence in clinical microbiology and infectious diseases applications. Front. Public Health14, 1776922. doi: 10.3389/fpubh.2026.1776922

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 173

  WatsonG. L.StaplesG.CarverR.BhargavaA.Lopez-EspinaC.SchmalzL.et al. (2026). Interpretability of an FDA-authorized AI/ML sepsis diagnostic tool improved by SHAP values. JAMIA Open9, ooag020. doi: 10.1093/jamiaopen/ooag020

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 174

  WiederO.KuenemannM.WiederM.SeidelT.MeyerC.BryantS. D.et al. (2021). Improved lipophilicity and aqueous solubility prediction with composite graph neural networks. Molecules26, 6185. doi: 10.3390/molecules26206185

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 175

  WilkinsonM. D.DumontierM.AalbersbergI. J.AppletonG.AxtonM.BaakA.et al. (2016). The FAIR guiding principles for scientific data management and stewardship. Sci. Data3, 160018. doi: 10.1038/sdata.2016.18

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 176

  WilkinsonM. D.DumontierM.Jan AalbersbergI.AppletonG.AxtonM.BaakA.et al. (2019). Addendum: The FAIR guiding principles for scientific data management and stewardship. Sci. Data6, 6. doi: 10.1038/s41597-019-0009-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 177

  WinZ. M.CheongA. M. Y.HopkinsW. S. (2023). Using machine learning to predict partition coefficient (Log P) and distribution coefficient (Log D) with molecular descriptors and liquid chromatography retention time. J. Chem. Inf. Model.63, 1906–1913. doi: 10.1021/acs.jcim.2c01373

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 178

  WongF.OmoriS.LiA.KrishnanA.LachR. S.RufoJ.et al. (2025). An explainable deep learning platform for molecular discovery. Nat. Protoc.20, 1020–1056. doi: 10.1038/s41596-024-01084-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 179

  WongC. H.SiahK. W.LoA. W. (2019). Estimation of clinical trial success rates and related parameters. Biostatistics20, 273–286. doi: 10.1093/biostatistics/kxx069

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 180

  WozniakS.JansonG.FeigM. (2025). Accurate predictions of molecular properties of proteins via graph neural networks and transfer learning. J. Chem. Theory Comput.21, 4830–4845. doi: 10.1021/acs.jctc.4c01682

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 181

  WuJ.WangJ.WuZ.ZhangS.DengY.KangY.et al. (2022). ALipSol: an attention-driven mixture-of-experts model for lipophilicity and solubility prediction. J. Chem. Inf. Model.62, 5975–5987. doi: 10.1021/acs.jcim.2c01290

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 182

  XiongG.WuZ.YiJ.FuL.YangZ.HsiehC.et al. (2021). ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties. Nucleic Acids Res.49, W5–W14. doi: 10.1093/nar/gkab255

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 183

  XuS.OnodaA. (2025). Accurate and rapid prediction of protein pK(a): Protein language models reveal the sequence-pK(a) relationship. J. Chem. Theory Comput.21, 3752–3764. doi: 10.1021/acs.jctc.4c01288

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 184

  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: 10.1128/msystems.00345-23

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 185

  YangH.SunL.WangZ.LiW.LiuG.TangY.. (2018). ADMETopt: a web server for ADMET optimization in drug design via scaffold hopping. J. Chem. Inf. Model.58, 2051–2056. doi: 10.1021/acs.jcim.8b00532

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 186

  YangH.LouC.SunL.LiJ.CaiY.WangZ.et al. (2019). admetSAR 2.0: web-service for prediction and optimization of chemical ADMET properties. Bioinformatics35, 1067–1069. doi: 10.1093/bioinformatics/bty707

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 187

  YangM. R.SuS. F.WuY. W. (2023). Using bacterial pan-genome-based feature selection approach to improve the prediction of minimum inhibitory concentration (MIC). Front. Genet.14, 1054032. doi: 10.3389/fgene.2023.1054032

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 188

  YangY.ZengL.VihinenM. (2021). PON-Sol2: prediction of effects of variants on protein solubility. Int. J. Mol. Sci.22, 8027. doi: 10.3390/ijms22158027

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 189

  YaoL.GuanJ.XieP.ChungC. R.DengJ.HuangY.et al. (2024). AMPActiPred: A three-stage framework for predicting antibacterial peptides and activity levels with deep forest. Protein Sci.33, e5006. doi: 10.1002/pro.5006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 190

  YasirM.KarimA. M.MalikS. K.BajafferA. A.AzharE. I. (2022). Prediction of antimicrobial minimal inhibitory concentrations for Neisseria gonorrhoeae using machine learning models. Saudi J. Biol. Sci.29, 3687–3693. doi: 10.1016/j.sjbs.2022.02.047

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 191

  YiJ.ShiS.FuL.YangZ.NieP.LuA.et al. (2024). OptADMET: a web-based tool for substructure modifications to improve ADMET properties of lead compounds. Nat. Protoc.19, 1105–1121. doi: 10.1038/s41596-023-00942-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 192

  YuY.WheelerN. E.BarquistL. (2025). Biased sampling driven by bacterial population structure confounds machine learning prediction of antimicrobial resistance. PloS Biol.23, e3003539. doi: 10.1371/journal.pbio.3003539

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 193

  ZankariE.AllesoeR.JoensenK. G.CavacoL. M.LundO.AarestrupF. M. (2017). PointFinder: a novel web tool for WGS-based detection of antimicrobial resistance associated with chromosomal point mutations in bacterial pathogens. J. Antimicrob. Chemother.72, 2764–2768. doi: 10.1093/jac/dkx217

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 194

  ZdrazilB.FelixE.HunterF.MannersE. J.BlackshawJ.CorbettS.et al. (2024). The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Res.52, D1180–D1192. doi: 10.1093/nar/gkad1004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 195

  ZhangS.YanZ.HuangY.LiuL.HeD.WangW.et al. (2022). HelixADMET: a robust and endpoint extensible ADMET system incorporating self-supervised knowledge transfer. Bioinformatics38, 3444–3453. doi: 10.1093/bioinformatics/btac342

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 196

  ZhangX.HuX.ZhangT.YangL.LiuC.XuN.et al. (2024). PLM_Sol: predicting protein solubility by benchmarking multiple protein language models with the updated Escherichia coli protein solubility dataset. Briefings Bioinf.25, bbae404. doi: 10.1093/bib/bbae404

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 197

  ZhangY.ChenP.DingK.LiuH.WangS.SongQ.. (2025). MPSol: A multimodal prompt learning framework for protein solubility prediction. IEEE J. Biomed. Health Inform.29, 8807–8814. doi: 10.1109/JBHI.2025.3608273

  • Pubmed Abstract
  • CrossRef
  • Google Scholar