Abstract
The interaction between p53 and MDM2 represents a key therapeutic target in several cancers where MDM2 overexpression suppresses p53 activity. Despite extensive research, the discovery of potent and selective MDM2 inhibitors remains challenging, underscoring the need for computational strategies specifically designed for this target. In this study, we developed a machine learning\x{2013}based approach to improve structure-based virtual screening (SBVS) for identifying MDM2 inhibitors at the p53 binding site. The models were developed, trained, and tested using experimentally validated bioactivity data from ChEMBL and PubChem, complemented with challenging decoy compounds to enhance predictive accuracy. Protein\x{2013}ligand interactions were encoded using Protein\x{2013}Ligand Extended Connectivity (PLEC) and Grid fingerprints, and multiple machine learning algorithms were evaluated. Among the implemented models, PLEC\x{2013}Random Forest and PLEC\x{2013}Support Vector Machine achieved the highest predictive performance, outperforming commonly used structure-based scoring functions, including Smina, CNN-Score, and SCORCH. Overall, these ML-based scoring functions enhance the in-silico identification of MDM2 inhibitors and provide a practical framework to guide future experimental validation and drug repurposing strategies for cancers driven by MDM2 overexpression.
1 Introduction
The tumor suppressor protein p53 is a crucial transcription factor involved in regulating the cell cycle and apoptosis. It binds to damaged DNA and initiates programmed cell death (apoptosis), ensuring that cells with DNA alterations do not continue to proliferate (; ). Under normal conditions p53 is expressed at low levels due to negative regulation by the murine double minute two protein (MDM2) (Hou et al., 2019). Disruption of p53-MDM2 regulation is strongly associated with the development and progression of various types of cancers (Wang et al., 2024; Yao et al., 2024; Peuget et al., 2024; Li et al., 2020).
For instance, in cancers like colorectal cancer (CRC), the p53 activation pathway often fails to function correctly (Liebl and Hofmann, 2021; Lai et al., 2022) frequently due to the overexpression of MDM2 in the tumor, which leads to a significant reduction in p53 levels (Hou et al., 2019). Consequently, damaged DNA cannot bind to the tumor suppressor, cellular apoptosis is not activated, and abnormal cells replicate uncontrollably. Given the critical role of the p53-MDM2 interaction in cancer pathogenesis, (Liebl and Hofmann, 2021; Traweek et al., 2022), inhibiting MDM2 to restore p53 function could represent an effective approach to induce apoptosis in tumor cells (; Sedzro et al., 2022; Twarda-Clapa, 2024) offering a potential pathway for targeted cancer treatment.
Recent developments in small-molecule inhibitors targeting the p53-MDM2 interaction have focused on four key scaffolds (Figure 1A): cis-imidazoline (Vassilev et al., 2004), oxindole (Wang et al., 2014), piperidin-2-one (Sun et al., 2014), and dihydroisoquinolinone (Holzer et al., 2015). These compounds mimic the interaction of the p53 binding domain (residues 17–125) by targeting the hydrophobic pocket of MDM2, where key residues Phe19, Trp23, and Leu26 of p53 typically bind (Figure 1B). These residues establish π-stacking and hydrophobic interactions with MDM2 (Wang et al., 2024; ; ).
FIGURE 1
Nutlin-3a (also known as RG7112), based on the cis-imidazoline scaffold, is a well-characterized example of this class. It contains three bulky functional groups positioned analogously to key p53 residues (Zhu et al., 2019). Its phenyl and heterocyclic rings mimic the π–π stacking interactions of Phe19 and Trp23, while an aliphatic isopropyl group mimics Leu26 by fitting snugly into a complementary hydrophobic pocket (Thakur et al., 2020; Wang and Chen, 2022; ). This structural similarity allows Nutlin-3a to effectively compete with p53 for binding to the hydrophobic pocket of MDM2, thereby inhibiting their interaction (Sriraman et al., 2016). Although Nutlin-3a effectively disrupts the p53–MDM2 interaction and demonstrates good selectivity, its clinical application is limited due to poor bioavailability and hematological toxicities, such as thrombocytopenia (Siu et al., 2014; Zhu et al., 2019).
Similarly, oxindole (MI-77301/SAR405838) shows high affinity for MDM2 and strong tumor regression ability, but it has been in clinical trials since 2015, facing challenges such as the emergence of new p53 mutations (Watters et al., 2015). Piperidin-2-one (AMG-232) is the most potent reported inhibitor; however, its clinical development has been slow, with no conclusive results on long-term safety (Zhu et al., 2019). At the same time, dihydroisoquinolinone (NVP-CGM097/MK-8242) remains in early clinical stages with limited publicly available data (Holzer et al., 2015). These challenges highlight the need to discover and design new inhibitors with improved safety and efficacy profiles.
Recent studies have reported a variety of new synthetic molecules with potential anticancer activity across diverse cancer types (Kryukova et al., 2020; Kinzhalov et al., 2017; Kinzhalov et al., 2016; Osipyan et al., 2018; Melekhova et al., 2017; Guranova et al., 2019; ; Grudova et al., 2022; Osmanov et al., 2022; ). This continuous progress highlights the importance of identifying effective inhibitors, and computational tools can help accelerate this search, particularly during the early stages of drug discovery (Fu and Chen, 2025; Zhang et al., 2025; ; ; Reutskaya et al., 2018; ; Kulish et al., 2016; Khrustalev et al., 2021; Il’in et al., 2019; Mikherdov et al., 2018; ; Sysoeva et al., 2024). Among these, structure-based virtual screening (SBVS) has proven especially useful for identifying active molecules. Its effectiveness is highest when the three-dimensional structure of the protein and its binding site are known, and when large chemical libraries of ligands are available (Zhu et al., 2019).
A wealth of information is available for the p53-MDM2 interaction. For MDM2 inhibition, databases such as ChEMBL include 4,585 ligand structures reported as MDM2 inhibitors together with their biological activities (Zdrazil et al., 2024), while PubChem contains 561,201 reported bioactivities related to MDM2 inhibition (Kim et al., 2025). Additionally, the availability of MDM2 three-dimensional structure in the PDB, such as PDB ID: 4HG7 (), further supports this process. Therefore, the available data on MDM2 makes SBVS a powerful approach for discovering new inhibitors targeting the p53-MDM2 interaction (Oyedele et al., 2022; Sedzro et al., 2022; Lee et al., 2025).
Molecular docking is an essential tool in SBVS, capable of predicting the best pose of a ligand within a target binding site and its relative affinity using scoring functions (Lionta et al., 2014). However, classical scoring functions, such as those based on force fields (Huang et al., 2006), knowledge-based (Gohlke et al., 2000) approaches, and empirical methods (e.g., Vina (Morris et al., 2009), Glide (Friesner et al., 2004), PMF (Muegge, 2006)), often rely on linear regression and limited structural features, which can lead to inaccuracies and unreliable predictions (Guvench and MacKerell Jr, 2009; Li et al., 2019; Guedes et al., 2018). Recently, machine learning (ML) has emerged as a promising and significantly more robust solution (Li et al., 2019; Meli et al., 2022; ; Niazi, 2023; Wu et al., 2024; McDermott et al., 2022). ML can learn non-linear relationships between the structures of protein-ligand complex and their reported binding affinities (; Wang et al., 2019). Additionally, experimental data is increasingly being used to infer intermolecular interactions, further improving the accuracy of predictions especially when applied to large datasets (Fresnais and Ballester, 2021; ). Recent studies on targets like PDL1 (Gómez-Sacristán et al., 2025), PARP1 (), CDK2 (Shahab et al., 2023) and RosettaGenFF-VS (Zhou et al., 2024) have demonstrated that ML models, trained with extensive datasets, significantly enhance affinity prediction accuracy, yielding metrics Area Under the Curve (AUC) between 0.8 and 0.9, Enrichment Factor in the top 1% (EF1%) >12, and Normal Enrichment Factor in the top 1% (NEF1%) (0.7–1) (Gómez-Sacristán et al., 2025; ; Zhou et al., 2024; Jiménez et al., 2018; He et al., 2025; Morales et al., 2024). This demonstrates the potential of ML models as a promising and effective approach for discovering specific inhibitors, such as those targeting the p53-MDM2 interaction.
In this study, we develop a machine learning (ML)-based structure-based virtual screening (SBVS) approach to identify potent MDM2 inhibitors at the p53 binding site. Using Protein Ligand Extended Connectivity (PLEC) (Wójcikowski et al., 2019) and Grid (Wu et al., 2018) descriptors, we represent protein-ligand interactions and apply 5 ML algorithms support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGB), artificial neural network (ANN), and deep neural network (DNN), with hyperparameters set through systematic search and cross-validation. We compare our models against standard scoring functions like Smina (Morris et al., 2009), CNN-Score (Ragoza et al., 2017) and SCORCH-generic (McGibbon et al., 2023) to benchmark their performance and ensure reliability. We discuss how the combination of advanced descriptors (PLEC and Grid) with optimized ML models significantly improves the efficiency of virtual screening for specific targets like MDM2 and its inhibition at the p53 binding site. This approach enhances the identification of promising inhibitors, reduces the need for exhaustive experimental screening, and accelerates the early stages of drug discovery.
2 Materials and methods
2.1 Data set
The MDM2 inhibitors dataset was built from experimental data gathered from the ChEMBL (version 29) (Zdrazil et al., 2024) and PubChem (version 29) (Kim et al., 2019) databases, containing 4,545 and 561,201 inhibitors with available bioactivity data, respectively. The data was preprocessed following a protocol published in 2023 (; Tran-Nguyen et al., 2023), considering SMILES, ID, modulator type, and activity concentrations (IC50, Kd, Ki), which were log-transformed into potency values using pIC50 = −log (IC50 × 10−9). We included potency determinations along with the minimum, maximum, median, and mean values for each metric, categorizing data with relationships: ‘<’ (active), ‘=’ (certain data), and ‘>’ (inactive). Filtering criteria were applied: (i) compounds with activity concentrations ≤10 μM were classified as active and those ≥10 μM as inactive, with this range established based on experimental data from ChEMBL, (ii) only inhibitors were selected, (iii) duplicate molecules were identified and removed based on Morgan fingerprints; 2,048 bits; radius 2 with a Tanimoto similarity of 1, and compounds lacking bioactivity values were also excluded. All structures were preprocessed using the RDKit Standardizer, including normalization of common functional groups, charge neutralization, and removal of counterions. Finally, (iv) for molecules with multiple biological activity data points, the mean and standard deviation were calculated, and only those with a standard deviation less than two were retained (Oyedele et al., 2022; ; Tran-Nguyen et al., 2023). The final dataset contained 2,687 molecules (2,184 actives and 503 inactives). Given the low number of inactives, we generated decoys using DeepCoy with an initial active-to-decoy ratio of 1:50. The generated decoys were filtered based on 27 physicochemical properties, structural diversity, and redundancy. Although 50 decoys per molecule were initially created, the filtering process produced a refined and balanced dataset with an approximate active-to-decoy ratio of 1:2. This rigorous selection ensures high-quality decoys and minimizes dataset bias, as highlighted in previous studies (Imrie et al., 2021; Segler et al., 2018). As a result, 4,629 decoys were included in the dataset.
To evaluate the molecular physicochemical properties and detect possible outliers, we used Mordred (Moriwaki et al., 2018; Tran-Nguyen et al., 2023). Using the processed data (actives, inactives, and decoys), we randomly split the dataset into training (70%) and test (30%) sets (Tran-Nguyen et al., 2023). The training set contained 1,677 actives and 3,885 inactives, while the test set included 507 actives and 1,247 inactives. To make sure the training and test sets are different in structure and better for virtual screening (VS), we took out molecules from the test set that had a Tanimoto similarity of 0.70 or higher (Morgan fingerprints; 2,048 bits; radius 2) to any molecule in the training set.
2.2 Molecular descriptors
This dataset was encoded using 3D molecular descriptors of protein-ligand interactions, including actives, inactives, and decoys. The ligands were docked into the binding site of MDM2, which interacts with p53. After docking, the MDM2-ligand interactions were characterized using two descriptors separately: PLEC and Grid.
2.2.1 Molecular docking
We used the protein with Protein Data Bank (PDB) ID 4HG7 (chain A) () as the receptor. This crystal structure has a 1.6 Å resolution and contains a co-crystallized scaffold, a small molecule non-covalently bound to the MDM2 binding site of p53 (Zhang et al., 2022). We used Chimera (Pettersen et al., 2004) and its Dock Prep tool to prepare the receptor structure and the co-crystallized ligand. Solvent molecules and the co-crystallized small molecule sulfate ion were removed, missing residues were completed, and partial charges for histidines in the receptor were assigned using the AM1-BCC method (Jakalian et al., 2002).
For the molecules in the dataset, we used Open Babel (O’Boyle et al., 2011) to generate 3D conformations using the MMFF94 force field. Once the structures were prepared, we performed molecular docking using Smina (Morris et al., 2009; Oyedele et al., 2022). The search grid was defined using the co-crystallized ligand as a reference. The grid size was set to 30 Å in each axis, providing enough space for all ligands to rotate freely. We selected one pose for each molecule, meaning the one with the lowest (most negative) Smina docking score. Smina uses a stochastic sampling method to create different poses, but each docking run repeats the optimization eight times by default. This means that the best pose for each molecule is usually stable, even if the docking is repeated (Torres et al., 2019; Masters et al., 2020). Therefore, we refer to this pose as the most representative one for each molecule in the dataset.
2.2.2 Orthogonal validation of docking poses
To confirm that the docking poses were chemically and physically consistent, and not influenced by artifacts from Smina, we applied orthogonal validation using a reliable control set, following the method of Bender et al (Nat. Protoc. 2021) (). First, we grouped the actives from our dataset into clusters by 2D similarity, which was calculated using ECFP4 fingerprints with a Tanimoto cutoff of 0.35. From each cluster, we kept only the most potent compound. Finally, this process gave us a control set of 30 active ligands that were structurally diverse and showed the highest experimental potencies.
These ligands were analyzed with Protein–Ligand Interaction Profiler (PLIP) to detect hydrophobic and aromatic contacts (Schake et al., 2025). First, we grouped the actives from our dataset into clusters by 2D similarity, which was calculated using ECFP4 fingerprints with a Tanimoto cutoff of 0.35. From each cluster, we kept only the most potent compound. This yielded a structurally diverse control set of 30 active ligands with the highest experimental potencies. Each ligand–MDM2 complex was then analyzed with the Protein–Ligand Interaction Profiler (PLIP) to identify hydrophobic and aromatic contacts. PLIP reported the interacting atoms, residues, and distances, confirming the expected key contacts across representative ligands. Furthermore, to evaluate whether Smina produced physically realistic docking poses, we used PoseBusters () which applies 18 criteria to assess the physical plausibility of docking poses, including intra- and intermolecular consistency. PoseBusters outputs a boolean descriptor (PB_valid) indicating the overall pose validity.
2.2.3 Featurization strategies
Protein–ligand extended connectivity (PLEC) fingerprints were generated using the PLEC function from the ODDT Python package (Wójcikowski et al., 2015), following the method described in previous studies (Tran-Nguyen et al., 2023). The fingerprint size was set to 4092, with ligand and protein fragment depths of 1 and 5, respectively. Additionally, 3D Grid-based (Grid) features were computed using the RDKitGridFeaturizer function from the DeepChem Python package (), based on the approach described in previous studies (Gómez-Sacristán et al., 2025). The extracted features included extended connectivity fingerprints (ECFP) (Rogers and Hahn, 2010), spatial ligand-based interaction fingerprints (SPLIF) (), hydrogen bonds, and salt bridges. The selected parameters were ecfp_power = 9, splif_power = 9, and voxel_width = 16.0, resulting in a total of 2052 features. To improve efficiency, feature extraction was performed in parallel using 20 cores, and the resulting features were saved as.npy files for further analysis.
2.3 Develop a specific machine learning scoring function to identify MDM2 inhibitors
To build models for classifying MDM2 inhibitors, we used different ML algorithms. These include random forest (RF) (Mach, 2001), support vector machine (SVM) (Vapnik, 1995), extreme gradient boosting (XGB) (), artificial neural network (ANN) (Jain et al., 1996), and deep neural network (DNN) (). The models were implemented using scikit-learn (Kramer and Kramer, 2016), XGBoost (), and Keras ().
Our method builds upon the ML framework proposed by Tran-Nguyen et al (2023) (Tran-Nguyen et al., 2023), particularly their strategy for developing ML-based scoring functions in virtual screening. However, we extended this approach by implementing more extensive cross-validation procedures to enhance predictive performance and robustness.
The RF model uses 200 decision trees combined with bootstrap aggregation for better accuracy. The SVM model was implemented with a radial basis function (RBF) kernel and hyperparameter settings (C = 1.0, gamma = “scale”, kernel = “rbf”). For XGB, we used gradient boosting to refine decision trees and reduce errors. The ANN model consisted of a single hidden layer with 100 neurons and ReLU activation and was trained using the Adam optimizer for up to 9000 iterations. The DNN model has three fully connected hidden layers and includes batch normalization and dropout regularization to avoid overfitting.
All ML models were trained and tested on the processed dataset and virtual screening was performed on the test molecules to compare performance and select the best model. The dataset, the code for all evaluated ML models, and the cross-validation procedures are available on GitHub (https://github.com/MarciaC18/Identification-of-MDM2-P53-Inhibitors-Using-Machine-Learning-Guided-Screening). The repository also includes scripts for hyperparameter optimization; however, in this study the models were run with default settings, detailed in Supplementary Table S1 (Supplementary Material).
The statistical analysis is presented in 2.5 Section Statistical Analysis.
2.4 Model evaluation by measuring virtual screening performance
The performance of each model was assessed using the enrichment factor in the top 1% (EF1%) (Truchon and Bayly, 2007), the normal enrichment factor in the top 1% (NEF1%) (Liu et al., 2018), Precision-recall curve (PRC) and its respectively Area Under the Precision-Recall Curve (AUC-PRC) (Truchon and Bayly, 2007). With these metrics, we evaluated the model’s ability to correctly classify active inhibitors on the test set. The values were computed with Python scripts developing by us and available at https://github.com/MarciaC18/Identification-of-MDM2-P53-Inhibitors-Using-Machine-Learning-Guided-Screening. Besides, the precision-recall curve was computed using the precision_recall_curve function from the sklearn.metrics module, based on the predicted probabilities and true labels (Gulli and Pal, 2017).
2.5 Statistical analysis
To assess potential overfitting, we applied five-fold stratified cross-validation within the training set. For each fold, the receiver operating characteristic area under the curve (ROC-AUC) was calculated using the roc_auc_score function from sklearn.metrics (Pedregosa et al., 2011), and the mean and SEM across folds were reported. This allowed us to verify that the model performed consistently across folds and was not overfitting the training data.
To evaluate the generalization capabilities of the models, the ROC-AUC was also calculated on the independent test set for each of the 10 independent runs. The final performance was reported as the mean ± SEM of these ROC-AUC values, providing a robust estimate of model performance on unseen data. We developed the Python script, which can be found in this GitHub repository: https://github.com/MarciaC18/Identification-of-MDM2-P53-Inhibitors-Using-Machine-Learning-Guided-Screening.
To assess whether there were statistically significant differences in ROC-AUC performance between models with similar values, we followed the procedure below. First, we collected the ROC-AUC values from the five cross-validation folds for both models. Second, we tested if the values were normally distributed using the Shapiro-Wilk test. If both distributions followed a normal distribution (p > 0.05), we used a paired t-test to compare their means. If not, we used the non-parametric Wilcoxon signed-rank test. A p-value below 0.05 was considered indicative of a statistically significant difference. This analysis helped determine whether the observed differences in performance were consistent and not attributable to random variation.
2.6 Comparison with traditional docking methods
The performance of the ML models was compared to that of traditional molecular docking methods such as Smina (Morris et al., 2009), SCORCH-generic (McGibbon et al., 2023), and CNN-Score (Ragoza et al., 2017). Both approaches were evaluated based on their ability to identify potential inhibitors from the docked poses of the test set.
2.7 Evaluation of the diversity in the chemical structures identified by the best PLEC-based models
To evaluate the chemical diversity of the molecules identified in the top 1%, we focused on the models that achieved the highest performance according to NEF1% and EF1%. First, we collected the SMILES and ChEMBL IDs of the compounds predicted in the top 1% for each PLEC-ML model. Second, we converted the SMILES into molecular structures using RDKit and calculated Morgan fingerprints for each compound. Third, we computed the Tanimoto similarity between all pairs of fingerprints to create a similarity matrix. Fourth, using this matrix, we performed hierarchical clustering to group molecules with similarity higher than 0.7. Finally, we analyzed the clusters and calculated the average Tanimoto similarity within each group to measure how similar the molecules are in each cluster. This process helped us understand the structural diversity of the selected compounds.
3 Results and discussion
3.1 Building of data set
The development of ML models for drug discovery begins with constructing an appropriate dataset (Walters and Murcko, 2020; ). The selection, processing and representation of data play a crucial role in the model’s performance in identifying potential drug candidates. Accordingly, we developed a specific dataset for the identification of potential MDM2 inhibitors at the p53 binding site. Our methodology involved identifying the protein crystal structure, curating experimental bioactivity data, preprocessing molecular structures, including protein-ligand complexes, and generating decoys to compare and enhance ML model performance for our specific target, as summarized in Figure 2.
FIGURE 2
Since the p53 binding site on MDM2 has been extensively studied through experimental data, we collected structural information from existing databases. Specifically, we retrieved the crystal structure of MDM2 from UniProt (ID: Q00987) and the Protein Data Bank (PDB) (Kouranov et al., 2006). Then, human MDM2 protein structures were identified through a search of the Uniprot database, allowing the retrieval of corresponding crystal structures from the PDB. As a result, we selected a crystal structure that fulfilled the following key criteria: it was related to human MDM2, included a co-crystallized scaffold, had high resolution, and was from a recent study. The final selection was PDB ID 4HG7 (; Kouranov et al., 2006), which includes a Nutlin-3a inhibitor bound to the p53 binding site on MDM2.
Additionally, UniProt information allowed us to identify relevant MDM2 ligands by specifying the known binding site for p53. Using this reference, we searched the ChEMBL and PubChem databases for compounds annotated as inhibitors of MDM2 at the p53 binding site (Zdrazil et al., 2024; Kim et al., 2025; ; Oyedele et al., 2022). Specifically, we selected only those bioactivity assays that explicitly referred to the p53-MDM2 interaction or to inhibition at the p53 binding site, based on their assay names. In ChEMBL, we found 4,545 experimental structures with bioactivity data, including IC50 (from biochemical or cellular assays), Ki, and Kd (from biophysical assays). Similarly, PubChem contained 561,201 bioactivity data entries related to ligands (IC50, Ki, and Kd).
Once the relevant data was collected, we preprocessed it to ensure its usability. The key selection criteria included: molecule ID, bioactivity, and SMILES, as well as confirmation that the molecules were inhibitors (see Methods section). To classify compounds as active or inactive, we applied a commonly used threshold based on the literature: compounds were considered active if their inhibitory concentration was ≤10 μM, while those above 10 µM were classified as inactive (Zhao et al., 2015; Zhang et al., 2021; Wang et al., 2016; ). After preprocessing, the dataset included 2,184 active compounds and 503 inactive compounds.
To further enhance model performance, we expanded the inactive dataset by generating decoys, following recommendations from previous studies as well as from virtual screening studies specific to the p53-MDM2 interaction (Li B.-H. et al., 2021; Oyedele et al., 2022; Liu et al., 2025). In this context, decoys have mainly been used to validate pharmacophore models, which have shown promising results in identifying MDM2 inhibitors. Although several compounds have been proposed through docking-based virtual screening targeting the p53-MDM2 interaction, these approaches do not explicitly report the use of decoys or provide performance benchmarks for ML models based on decoy datasets. Therefore, in our work, we followed standard practices in ML that include decoy generation, as demonstrated in recent studies on other targets such as PDL1 (Gómez-Sacristán et al., 2025), PARP1 (), and in general purpose ML-based virtual screening frameworks (Tran-Nguyen et al., 2023).
Given that the combination of generated decoys and ML models can potentially introduce dataset biases affecting the reliability of virtual screening results, (Tran-Nguyen et al., 2020; Gowthaman et al., 2015; Segler et al., 2018), we aimed to minimize such biases by using DeepCoy, a generative method that creates decoys with physicochemical properties similar to those of active compounds while maintaining structural dissimilarity (see Methods section) (Imrie et al., 2021). This approach helps reduce artificial enrichment and analog bias, thereby decreasing the risk of including false negatives (Gómez-Sacristán et al., 2025). Consequently, the generated decoys were structurally dissimilar yet physicochemically comparable to known active compounds and were treated as inactive. A total of 4,629 decoys were obtained, resulting in a final dataset of 7,316 compounds.
To ensure the dataset’s relevance for drug discovery, we analyzed its physicochemical properties. Since the compounds must exhibit physicochemical characteristics representative of potential drugs, we calculated and compared their properties against Lipinski’s rule-of-five descriptors (MW, ALogP, PSA, nHBDon, and nHBAcc). The results, presented in Supplementary Material: Supplementary Figure S1, confirmed that all compounds fell within standard physicochemical ranges, making the dataset suitable for ML model development.
For model training and evaluation, the data set was randomly divided into a training set and a test set. The training set comprised 70% of the molecules in the data set, including 1,677 active compounds (bioactivity concentration ≤10 µM) and 3,885 inactive compounds. The remaining 30% formed the test set, which contained 507 active and 1,247 inactive compounds. To ensure the test set contained molecules structurally dissimilar to the training set and avoid analogue bias, a similarity filter was applied (see Methods). This makes the test set more challenging and a better benchmark for evaluating the model’s generalization capability.
Moreover, this class imbalance was maintained because previous studies suggest that models trained with a large number of negative samples tend to be more reliable (Gómez-Sacristán et al., 2025; Kurczab et al., 2014). A higher proportion of inactive compounds helps alleviate overfitting to false positives, provides a more realistic representation of the chemical space, and improves model precision (; Gómez-Sacristán et al., 2025; Kurczab et al., 2014). This approach also allows the model to better distinguish between active and inactive compounds.
Finally, it is important to highlight that both the training and test set were docked and characterized using chemical descriptors, including structural information captured in a target-ligand complex or a ligand structure in 3D space. Specifically, we extracted a single binding pose for each molecule in MDM2 at the p53 binding site, based on the best docking score. This ensures that ML models can learn from 3D interactions and associate them with activity values, improving their predictive power.
3.2 Orthogonal validation of docking poses
Docking provides practical binding poses and descriptors for ML, but docking results may be vulnerable to sampling artifacts and scoring-function bias, (Prieto et al., 2024; ; Fischer et al., 2021), which can propagate into downstream ML models, reducing their reliability. For this reason, we used orthogonal validation steps to make sure that the docking poses were chemically and physically plausible, and not dominated by artifacts.
First of all, we curated a control set of 30 representative active compounds. Following Bender et al (Nat. Protoc. 2021) (), redundant molecules were clustered by 2D similarity (ECFP4 Tanimoto cutoff = 0.35), and the most potent molecule from each cluster was retained (see Methods). This ensured chemical diversity, avoided overrepresentation of specific scaffolds, and provided a robust control set for downstream validation.
With this control set, we implemented automated structural sanity checks based on the chemistry of the p53–MDM2 interaction. The binding site of p53 in MDM2 is formed by 14 residues (Leu54, Leu57, Ile61, Met62, Tyr67, Gln72, Val75, Phe86, Phe91, Val93, His96, Ile99, Tyr100, Ile101) (; Schon et al., 2002). Therefore, p53 binding to MDM2 is mainly stabilized by hydrophobic forces, particularly involving residues Leu54, His96, and Tyr100 (; ; Kannan et al., 2019; ; Wei et al., 2025). Based on this knowledge, we used the Protein–Ligand Interaction Profiler (PLIP) to extract hydrophobic and aromatic contacts from MDM2–inhibitor complexes and confirm their presence across representative ligands (see Methods) (Schake et al., 2025). Results are shown in Supplementary Table S2 and Figure 3.
FIGURE 3
Figure 3 shows that PLIP analysis confirmed that the docking poses reproduced more than three key contacts within the p53-binding site of MDM2. Supplementary Table S2 highlights interactions of Leu54, His96, and Tyr100 with the control inhibitors involving hydrophobic contacts, π–π stacking, and halogen bonds. Similar interactions have been observed for the p53 peptide and Nutlin-3 (example, PDB 4HG7) (Walter et al., 2018; Li H. et al., 2021; Wei et al., 2025; ). The finding that all ligands occupied the expected hydrophobic pocket and consistently formed at least three relevant hydrophobic interactions, comparable to those of p53, supports their chemical plausibility.
Using PoseBusters (), we confirmed that Smina generated at least one physically valid docking pose. PoseBusters applies 18 criteria to assess the physical and chemical plausibility of poses, summarized by the boolean descriptor PB_valid. The results, reported in Supplementary Table S3, show that all control compounds passed the 18 PoseBusters checks. In particular, we made sure that no steric clashes were present and that the aromatic rings kept their flat shape, which is important for stable hydrophobic interactions. These findings confirm that the generated poses are physically plausible.
Together, the mechanistic interaction analysis and PoseBusters validation comparison provide strong and converging evidence that the generated docking poses are both structurally and physically reliable. These complementary validation steps substantially strengthen confidence that the ML models are capturing genuine biological features rather than artifacts of the docking process.
3.3 Developing a specific machine learning scoring function to identify MDM2 inhibitors
3.3.1 Selection of the best descriptor type
The use of ML scoring functions has significantly improved the performance of SBVS for specific targets (Scantlebury et al., 2023; Zhang et al., 2023; Junaid et al., 2024; Gómez-Sacristán et al., 2025). This is because ML models can extract and learn target-specific 3D molecular features and correlate them to the experimental activity of these molecules (Gómez-Sacristán et al., 2025; Tran-Nguyen et al., 2023). As a result, ML-based scoring functions for virtual screening can achieve higher accuracy compared to general scoring functions, which rely on recognizing generic target features and applying them to evaluate molecular interactions. Furthermore, previous studies have demonstrated that developing ML models using a dataset enriched with inactive compounds enhances performance (Gómez-Sacristán et al., 2025; Kurczab et al., 2014; Réau et al., 2018). Therefore, to develop an MDM2-specific ML model, and considering the aforementioned insights, we constructed and evaluated 5 ML algorithms using a dataset enriched with inactive compounds and 3D interaction features. The results are summarized in Supplementary Material: Supplementary Table S4 and Figure 4.
FIGURE 4
The performance of ML models was evaluated using the PRC and AUC (Keilwagen et al., 2014). These metrics help measure how well the model can separate active and inactive compounds, which is important for handling imbalanced datasets (Tran-Nguyen et al., 2023; Richardson et al., 2024). A good model has an AUC-PRC close to 1, meaning it can distinguish well between the two classes. In drug discovery, AUC-PRC values between 0.6 and 0.8 are considered acceptable (Gómez-Sacristán et al., 2025; ; Zhu et al., 2025; Tahir ul Qamar et al., 2022). For example, Zhou et al., in 2024 showed that ML models for virtual screening achieved a similar level of performance (Zhou et al., 2024). In summary, we evaluated the performance of ML models for discovering MDM2 inhibitors targeting the p53 binding site and compared their results to those of previous virtual screening studies to examine whether they achieved similar or better results. A more detailed comparison with traditional methods is presented in the section of Comparison with Generic Virtual Screening Models.
To achieve this, we tested five supervised learning algorithms using a classification approach to explore their predictive capabilities. The models were trained and tested to classify each compound as either active or inactive. Although the selected algorithms such as SVM, RF, XGB, ANN, and DNN can also be used for regression tasks, in this study we only applied them in classification mode. These models were chosen because they have been successfully applied in various drug discovery studies (Patel et al., 2020; Zhou et al., 2024; ; ) and are also described in the protocol we followed, which is based on the approach proposed by Tran-Nguyen et al (2023) in their guide on ML-based scoring functions for virtual screening (Tran-Nguyen et al., 2023).
To train these ML models with representative data, a numerical representation of molecular structures is required. Therefore, we encoded the 3D interaction information of the MDM2-L protein-ligand complex at the p53 binding site using the chemical descriptors Grid and PLEC. These descriptors were applied to both active and inactive compounds from the previously constructed database. Grid and PLEC numerically encode the 3D interaction information between the protein and ligands derived from docking poses, and have been successfully used in prior virtual screening studies (; Gómez-Sacristán et al., 2025). Notably, previous studies have reported that ML model performance improves when descriptors incorporate 3D protein-ligand interaction information.
To evaluate the effectiveness of these descriptors in our models, we examined the performance of ML algorithms using Grid-ML and PLEC-ML representations. Figure 4 shows the combined prediction results of the Grid-ML and PLEC-ML models across 10 test runs (Keilwagen et al., 2014). For the Grid dataset (Figure 4A), DNN achieved the highest AUC-PRC (0.353), followed by SVM (0.309). However, the results show that all AUC-PRC values were relatively low, indicating that Grid descriptors do not provide good performance for machine learning models.
In contrast, the performance improved significantly when using the PLEC dataset, as shown in Figure 4B. The SVM achieved the highest AUC-PRC (0.767), followed closely by RF with 0.766 and XGB with 0.756. These results suggest that PLEC descriptors provide more useful features for classification. However, in this case, the DNN had the lowest performance, with an AUC-PRC of 0.525. This indicates that more complex models are not always the best choice, as their effectiveness depends on the type of descriptor and specific target. Therefore, according to these results, simpler or more traditional models such as PLEC-SVM, PLEC-RF, or PLEC-XGB may achieve higher performance in correctly classifying true actives (inhibitors), demonstrating greater discriminatory power between the active and inactive compounds for MDM2, as evidenced by their superior performance in the PRC, which is a well-established metric for evaluating classifier performance in imbalanced datasets (; Saito and Rehmsmeier, 2015).
3.3.2 Virtual screening performance: EF1% and NEF1% metrics
Within our Grid-ML and PLEC-ML frameworks, specific ML models such as PLEC-RF, and PLEC-SVM showed a remarkable ability to correctly classify active compounds (inhibitors), demonstrating a strong discriminatory power between active and inactive compounds for MDM2. To evaluate the predictive ability of our models in selecting the most active compounds, we used the EF1% metric calculated on the test set, which measures a model’s ability to retrieve active compounds within the top 1% of its predictions (Truchon and Bayly, 2007). Specifically, in SBVS, a high EF1% indicates that the model efficiently identifies the most promising compounds and rank them at the top of the list based on affinity or activity probability (Zhou et al., 2024; Li H. et al., 2021; Mahdizadeh and Eriksson, 2025). This is crucial for prioritizing candidates for experimental testing.
The results of our models are shown in Figure 5. For the PLEC-ML models, the mean EF1% values ranged between 2.742 and 3.450, respectively, for PLEC-RF (3.450 ± 0.00), PLEC-XGB (3.290 ± 0.03), PLEC-SVM (3.450 ± 0.00), PLEC-ANN (3.290 ± 0.05), and PLEC-DNN (2.742 ± 0.18). These results indicate that the PLEC-ML models identified between 2.742 and 3.450 times more active compounds than random selection. On the other hand, the mean EF1% values obtained for the Grid-ML models ranged from 0.471 to 2.030, respectively, for Grid-RF (1.420 ± 0.00), Grid-XGB (0.810 ± 0.03), Grid-SVM (2.030 ± 1.40), Grid-ANN (0.471 ± 0.06), and Grid-DNN (0.772 ± 0.16). Thus, the Grid-ML models only identified between 0.772 and 2.030 times more active compounds than random selection. Therefore, it is remarkable that all PLEC-ML models outperformed all Grid-ML models. This is consistent with the previously presented results.
FIGURE 5
Specifically, for PLEC, the PLEC-RF and PLEC-SVM models identified 3.450 times more active compounds, respectively, compared to random selection. This confirms that PLEC descriptors capture relevant structural information for classifying MDM2 inhibitors. Overall, these results support the use of classical ML approaches, such as RF and SVM, combined with PLEC descriptors to improve the selection of MDM2 inhibitors in SBVS processes.
Although EF1% indicates the ability to retrieve compounds in the top 1% of VS, this metric is not normalized. This means it does not confirm if a model is 100% efficient in classifying only actives in the top 1%. To address this, we measured the Enrichment Factor Normalized to 1% (NEF1%), which is calculated as the observed EF1% divided by the maximum possible EF1% for a given test set (Liu et al., 2018). NEF1% indicates how well a model classifies actives in the top 1%. A value of 1 or close to it means the model selects only actives in the top 1%. The results of this analysis are shown in Figure 6, with Supplementary Material: Supplementary Table S5.
FIGURE 6
In Figure 6, we can observe the normalized enrichment factor at 1% mean (NEF1%) for different models using Grid and PLEC descriptors. The results indicate that PLEC-RF (1.000 ± 0.00), PLEC-XGB (0.952 ± 0.01), PLEC-SVM (1.000 ± 0.00), PLEC-ANN (0.952 ± 0.02), and PLEC-DNN (0.792 ± 0.05) achieved the highest NEF1% values. This means that they efficiently classified actives within the top 1% of the ranked compounds. These models demonstrated strong predictive capabilities in prioritizing MDM2 inhibitors.
However, Grid-RF (0.410 ± 0.00) and Grid-SVM (0.590 ± 0.00) obtained moderate NEF1% values, while Grid-XGB (0.24 ± 0.02), Grid-ANN (0.136 ± 0.02), and Grid-DNN (0.227 ± 0.05) showed a low NEF1% value. This suggests that, although they retrieved active compounds, they were not effective in ranking them exclusively at the top, highlighting their inability to effectively classify actives at the top 1% of the predictions.
These results reinforce the effectiveness of classical ML models, such as RF and SVM, particularly when combined with PLEC descriptors. This combination proves highly suitable for identifying MDM2 inhibitors in SBVS. The higher NEF1% values for PLEC-based models confirm their effectiveness. These descriptors provide meaningful structural information, which helps prioritize active compounds better.
3.3.3 Generalization ability: cross-validation ROC-AUC
To assess the models’ consistent performance without overfitting and their generalization capability, we performed cross-validation on the training set and evaluated ROC-AUC scores on the independent test set (see Methods section). ROC-AUC shows the balance between sensitivity and specificity. In other words, it measures model’s ability to identify active compounds and distinguish them from inactive ones (Robinson et al., 2020; LeDell et al., 2015). Besides, it helps to estimate the performance across different datasets. Ideal ROC-AUC values are near to one because this indicates that can perfectly distinguish between active and inactive classes (LeDell et al., 2015; Robinson et al., 2020). Nonetheless, it is an ideal model. Some studies on drug discovery using ML models have showed that ROC-AUC >0.8 yields good results for predicting new or repurposed drugs (; Monsia and Bhattacharyya, 2025; Zhou et al., 2024; ). In addition, this metric allows us to determine whether a model has overfitting. In brief, good generalization capability is demonstrated with ROC-AUC> 0.8 in drug discovery.
As a result, we calculated and compared the ROC-AUC on the training set using five-fold cross-validation, reporting the mean and standard deviation (SD) to assess consistent performance without overfitting. Additionally, ROC-AUC was calculated on the independent test set to evaluate the generalization capability and the quality of predictions for different MDM2 inhibitors. ROC-AUC mean and SD scores for the training set models are summarized in Supplementary Figure S2, while ROC-AUC mean and SEM scores for the test set models are reported in Figure 7.
FIGURE 7
The results in Supplementary Figure S2 show the ROC-AUC from cross-validation on the training set. Values remained high across folds for PLEC-RF (0.810 ± 0.01), PLEC-XGB (0.812 ± 0.02), PLEC-SVM (0.820 ± 0.006), and PLEC-ANN (0.798 ± 0.01) indicating low variability and minimal overfitting. In contrast, PLEC-DNN achieved a ROC-AUC of 0.714 ± 0.07, which was lower than the other models and exhibited higher variability. Figure 7 presents the mean ROC-AUC scores on the independent test set illustrating each model’s generalization ability to unseen MDM2 inhibitors. The models achieved the following mean ROC-AUC scores: PLEC-RF (0.866 ± 0.0007), PLEC-XGB (0.862 ± 0.0001), PLEC-SVM (0.871 ± 3.8 × 10−6), and PLEC-ANN (0.747 ± 0.061). These results are consistent with the training set performance, confirming that the models generalize well to unseen data. This indicates that the best-performing PLEC models not only achieve high classification ability but also maintain robustness. In contrast, PLEC-DNN remained at chance level (0.500 ± 0.0001), indicating poor generalization in this dataset.
Compared to the previous models, Grid-based models showed lower ROC-AUC values than PLEC-based models on both the test and training sets. The results in the test set (Figure 7) were: Grid-RF (0.478 ± 1.00 × 10−5), Grid-XGB (0.477 ± 2.00 × 10−5), Grid-SVM (0.513 ± 2.00 × 10−6), Grid-ANN (0.505 ± 1.00 × 10−2), and Grid-DNN (0.484 ± 5.00 × 10−4). Furthermore, in Supplementary Figure S2, ROC-AUC values for the training set were near 0.5. This indicates that Grid-based models have poor performance in both the training and test sets and do not have a predictive advantage over random classification because they have low generalization capacity.
In summary, these results highlight the importance of selecting an appropriate feature representation for the specific MDM2-inhibitor system. In this case, the PLEC strategy effectively captures structural and interaction information relevant for distinguishing active and inactive MDM2 inhibitors, whereas the Grid strategy does not. Although previous SBVS performance suggested that both PLEC-SVM and PLEC-RF were strong contenders, PLEC-SVM achieved a slightly higher mean ROC-AUC during cross-validation (0.820 ± 0.01) and lower standard deviation compared to PLEC-RF (0.800 ± 0.01). However, a paired t-test (p = 0.075) showed that this difference was not statistically significant. Thus, both models exhibit strong predictive performance, with PLEC-SVM showing a slight, though not statistically confirmed, advantage. In conclusion, these results confirm that PLEC-SVM and PLEC-RF are the most reliable models, combining strong training performance with good generalization to independent test data.
3.3.4 Comparison with generic virtual screening models
Some studies have shown that reclassifying docking poses with ML improves the identification of active molecules compared to generic docking methods for specific targets (Gómez-Sacristán et al., 2025; ; Li et al., 2014; Hany et al., 2025). To evaluate this in our case, we focused on MDM2 inhibitors at the p53 binding site and measured the mean PRC and AUC scores, EF1% and NEF1% between our models and classical generic methods like Smina (Morris et al., 2009), SCORCH-generic (McGibbon et al., 2023), and CNN-generic (Ragoza et al., 2017). These methods evaluate docking poses by generating scores that approximate or correlate with the binding free energy (ΔGbinding) between the protein and ligand, rather than computing it explicitly. However, each method employs different scoring functions and algorithms to estimate binding affinity.
Smina, a derivative of AutoDock Vina 1.1.2, calculates ΔGbinding using a linear regression model (Morris et al., 2009). This approach assumes that weak interactions between proteins and ligands can be modeled as linear relationships. Smina was trained with the Community Structure-Activity Resource (CSAR)-NRC HiQ 2010 dataset, which is known for its high-quality protein-ligand complex data (Koes et al., 2013). As a result, several studies have shown that it improves the performance of scoring docking poses of active compounds against decoys across multiple biological targets. Its performance was benchmarked against ten other docking programs (MVP, DockIt, FlexX, Flo+, Fred, Dock4, Glide, Gold, MOEDock, and LigandFit) (Warren et al., 2006; ; ) using the EF1% metric. In this evaluation, Smina achieved an EF1% of 11.6, significantly outperforming the best competing programs, which had EF1% values between 5 and 7. These results highlight Smina’s superior ability to prioritize active compounds in virtual screening, on average across a range of evaluated biological targets.
SCORCH is a ML model that combines gradient-boosted decision trees (XGB), feedforward neural networks (FNN), and wide and deep neural networks (DNN). It was trained on protein-ligand datasets like PDBbind, Binding MOAD, and Iridium, along with decoys created using DeepCoy (McGibbon et al., 2023). Then, it was tested on different targets and is considered highly effective because it outperforms traditional scoring functions in virtual screening (VS) and pose prediction.
Finally, CNN-Score is a deep learning model consisting of five convolutional neural networks (CNN) with up to 20 hidden layers. It was trained on protein-ligand interaction data, including decoys from the DUD-E database, which contains 22,645 positive instances and 1,407,145 negative ones (Mysinger et al., 2012). Then, it was tested on the LIT-PCBA dataset (Tran-Nguyen et al., 2020), designed for virtual screening and ML with 149 dose-response PubChem bioassays. CNN-Score outperformed traditional scoring functions like Smina and improved pose re-scoring without overestimating detection power. We employed these high-performing models for a comparative analysis on our specific target (MDM2 at the p53 binding site), and the results are presented in Figure 8 and Supplementary Table S6.
FIGURE 8
In Figure 8, the mean PRC is shown for the best ML models that were previously evaluated. The figure includes the models based on PLEC (SVM, RF, XGB, ANN, and DNN) and the generic scoring functions SCORCH, CNN-Score, and Smina. The PLEC-ML models show a decreasing order of mean scores, starting with SVM (0.767), RF (0.766), XGB (0.756), ANN (0.665), and DNN (0.525) while generic models show lower values, such as CNN (0.131), SCORCH (0.023), and Smina (0.022).
With regard to the foregoing, the PLEC-ML models outperform the generic scoring functions in terms of AUC-PRC. Then, we can see that, although CNN has shown good results in other studies, (Ragoza et al., 2017), in the specific case of identifying inhibitors targeting the p53 binding site in MDM2, its AUC-PRC is lower than that of the generic models. This indicates that its ability to re-rank docked poses in this dataset is limited. Besides, SCORCH, which is based on decision trees enhanced with neural networks, also exhibited very low performance in our specific case, suggesting that its re-scoring capability is not optimal in this scenario. Finally, Smina, a scoring function based on linear regression, showed the worst performance in this specific case. This confirms that classical docking approaches may be less effective in distinguishing active from inactive docked poses in specific targets.
In addition, Supplementary Table S6 illustrates the model’s ability to retrieve active compounds within the top 1% of the ranking generated using the NEF1% metric. The results indicate that the PLEC-ML models have scores closest to 1 whereas the generic models have scores closest to 0. Firstly, the SCORCH generic model showed an NEF1% of 0.30, meaning that it can only prioritize the classification of actives by 30% in the top 1% ranking. Therefore, this indicates that this model has a lower enrichment capability in retrieving active compounds at the top positions compared to PLEC-ML models. Secondly, CNN showed an NEF1% of 0.30, meaning that it can only prioritize the classification of active compounds by 30% in the top 1% ranking. Therefore, this suggests that this model has a lower enrichment capability in retrieving active compounds at the top positions compared to PLEC-ML models. Third, Smina exhibited an NEF1% of 0 implying that this model cannot prioritize molecules in the top-ranking search for actives, specifically inhibitors of the p53 binding site in MDM2.
In summary, the results show that ML models trained with PLEC descriptors outperform traditional and generic scoring functions in identifying inhibitors of the p53 binding site in MDM2. In particular, the PLEC-SVM and PLEC-RF models have a high capability to prioritize/classify active molecules in the top 1%. This suggests that re-scoring with PLEC-SVM and PLEC-RF significantly improves the prediction of actives in the top-ranked classification performed by SBVS.
3.3.5 Evaluation of the diversity in the chemical structures identified by the best PLEC-based models
A ML model designed to predict the re-scoring of docking poses and potential inhibitors must be capable of identifying diverse molecular structures. This ensures that the model does not favor specific compounds while overlooking others with inhibitory potential. This aspect is essential to improving its applicability in drug discovery (Nogueira and Koch, 2019; Tran-Nguyen et al., 2023).
Thus, we evaluated the chemical diversity of PLEC-SVM and PLEC-RF, both of which demonstrated a high generalization capacity and strong performance in classifying molecules with inhibitory potential for p53 binding site in MDM2. This analysis ensures that the model recognizes not only compounds similar to those in the training set but also molecular structures that differ from it.
We evaluated chemical diversity by calculating pairwise Tanimoto similarity using Morgan fingerprints (2048 bits, radius 2), applying a similarity threshold of 0.7. Compounds with a Tanimoto similarity equal to or greater than this threshold were grouped together, following the approach described by Tran-Nguyen et al. (2023). The classified molecules correspond to the true active compounds identified by PLEC-SVM and PLEC-RF in the top 1%. The results are shown in Figure 9.
FIGURE 9
The results show that for PLEC-SVM there are 11 different groups of molecular structures, which were clustered based on Tanimoto similarity coefficients higher than 0.7. This indicates that the PLEC-SVM model can identify distinct molecular structures as true actives, specifically inhibitors of the p53 binding site in MDM2. In contrast, PLEC-RF identified 7 different groups of molecular structures, also clustered by Tanimoto coefficients greater than 0.7. Therefore, among the groups identified by PLEC-SVM and PLEC-RF, it is evident that both models can find molecules with structural chemical diversity. This difference is expected in ML models (), as each can capture different patterns or chemical features, which can be beneficial for exploring chemical diversity in the search for inhibitors of the p53 binding site in MDM2. Therefore, the application of virtual screening by structure-based methods (SBVS) is highly promising in this context.
Moreover, it was observed that both PLEC-SVM and PLEC-RF models coincide on a subset of molecules with high similarity (7 common: CHEMBL1834286, CHEMBL1688304, CHEMBL3927813, CHEMBL3800573, CHEMBL3260848, CHEMBL1688271, CHEMBL4451740), since both models identified these molecules as representative or relevant according to their respective criteria (SVM using kernels and RF using decision trees).
We think that the PLEC-SVM and PLEC-RF models find different chemical patterns. The SVM model looks for more diverse molecules, while the RF model focuses on molecules with stronger evidence. Because of this, using both models together may help find more different and important molecules, which can improve virtual screening for MDM2 inhibitors at the p53 binding site.
4 Conclusion
In this study, we successfully developed and validated a ML-based virtual screening approach for identifying potential MDM2 inhibitors at the p53 binding site. Our findings demonstrate that using advanced descriptors (PLEC) with ML models outperformed Grid descriptors and generic models in prediction accuracy across multiple scoring metrics. The comparative analysis of descriptor types revealed that PLEC descriptors provide more informative features for classifying active and inactive molecules than Grid descriptors.
Compared to traditional scoring functions like Smina, the PLEC-ML models more effectively captured non-linear relationships between ligand structures and their binding affinity to MDM2. This highlights the potential of ML to overcome the limitations of classical methods in predicting protein-ligand interactions.
Additionally, the PLEC-SVM and PLEC-RF models exhibited high performance in classifying active MDM2 inhibitors from inactive molecules. They also effectively ranked active molecules in the top 1% of the SBVS ranking and identified structurally diverse molecules. This suggests strong potential for expanding the chemical space of inhibitors for this target.
From an applied perspective, our findings demonstrate that this ML-based SBVS approach can enhance virtual screening accuracy, reduce false positives, and prioritize candidates with higher likelihood of success in experimental studies. Targeting MDM2 to restore p53 activity represents a promising therapeutic strategy in cancers such as colorectal cancer. This approach provides a practical framework to accelerate the discovery of novel MDM2 inhibitors and could support drug repurposing and development for cancers driven by MDM2 overexpression.
In conclusion, this work demonstrates the effectiveness of ML-based SBVS in drug discovery and its successful application to the p53-MDM2 system, leveraging experimental bioactivity data to optimize in silico workflows. Future efforts will focus on applying these models to virtual screening of U.S. Food and Drug Administration-approved drugs, performing molecular dynamics simulations of MDM2–potential inhibitor complexes, and calculating binding free energies using methods such as Linear Interaction Energy (LIE) or MM-GBSA. Computational optimization of lead inhibitors, followed by their chemical synthesis and experimental testing, will support the development of potent and selective MDM2 inhibitors. Experimental validation will include in-vitro and in-vivo studies in cancer cell lines with MDM2 overexpression, such as colorectal cancer, to confirm the efficacy and selectivity of predicted potent MDM2 inhibitors.
Statements
Data availability statement
Publicly available datasets were analyzed in this study. This data can be found here: https://github.com/MarciaC18/Identification-of-MDM2-P53-Inhibitors-Using-Machine-Learning-Guided-Screening.
Author contributions
MC: Investigation, Conceptualization, Writing – original draft, Formal Analysis, Data curation, Resources, Visualization, Funding acquisition, Validation, Methodology, Writing – review and editing. GM: Resources, Supervision, Writing – review and editing, Project administration.
Funding
The author(s) declared that financial support was received for this work and/or its publication. This work was supported by the Universidad de los Andes through its Excellence Scholarship.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fddsv.2025.1731262/full#supplementary-material
References
1
AbadiM.AgarwalA.BarhamP.BrevdoE.ChenZ.CitroC.et al (2016). Tensorflow: large-scale machine learning on heterogeneous distributed systems. arXiv [Preprint]. Available online at: https://arxiv.org/abs/1603.04467 (Accessed January 15, 2025).
2
AcharyyaR. K.HuangL.AguilarA.HuB.BaiL.MetwallyH.et al (2025). MD-4251: a first-in-class oral MDM2 degrader inducing complete tumor regression with single-dose administration. J. Med. Chem.68, 13249–13267. 10.1021/acs.jmedchem.5c00809
3
AinQ. U.AleksandrovaA.RoesslerF. D.BallesterP. J. (2015). Machine‐learning scoring functions to improve structure‐based binding affinity prediction and virtual screening. Wiley Interdiscip. Rev. Comput. Mol. Sci.5, 405–424. 10.1002/wcms.1225
4
AnilB.RiedingerC.EndicottJ. A.NobleM. E. M. (2013). The structure of an MDM2–Nutlin-3a complex solved by the use of a validated MDM2 surface-entropy reduction mutant. Biol. Crystallogr.69, 1358–1366. 10.1107/S0907444913004459
5
AtatrehN.GhattasM. A.BardaweelS. K.RawashdehS.AlSorkhyM.Al (2018). Identification of new inhibitors of Mdm2–p53 interaction via pharmacophore and structure-based virtual screening. Drug Des. Devel Ther.12, 3741–3752. 10.2147/DDDT.S182444
6
BallesterP. J.MitchellJ. B. O. (2010). A machine learning approach to predicting protein–ligand binding affinity with applications to molecular docking. Bioinformatics26, 1169–1175. 10.1093/bioinformatics/btq112
7
BaykovS. V.MikherdovA. S.NovikovA. S.GeylK. K.TarasenkoM. V.GureevM. A.et al (2021). π–π noncovalent interaction involving 1, 2, 4-and 1, 3, 4-oxadiazole systems: the combined experimental, theoretical, and database study. Molecules26, 5672–5687. 10.3390/molecules26185672
8
BenderB. J.GahbauerS.LuttensA.LyuJ.WebbC. M.SteinR. M.et al (2021). A practical guide to large-scale docking. Nat. Protoc.16, 4799–4832. 10.1038/s41596-021-00597-z
9
BermanH. M.WestbrookJ.FengZ.GillilandG.BhatT. N.WeissigH.et al (2000). The protein data bank. Nucleic Acids Res.28, 235–242. 10.1093/nar/28.1.235
10
BistaM.WolfS.KhouryK.KowalskaK.HuangY.WronaE.et al (2013). Transient protein states in designing inhibitors of the MDM2-p53 interaction. Structure21, 2143–2151. 10.1016/j.str.2013.09.006
11
BobadillaJ. (2021). Machine learning y deep learning: Usando Python, Scikit y Keras. Bogotá, Colombia: Ediciones de la U.
12
BolotinD. S.Il’inM. V.NovikovA. S.BokachN. A.SuslonovV. V.KukushkinV. Y. (2017). Trinuclear (aminonitrone) Zn II complexes as key intermediates in zinc (ii)-mediated generation of 1, 2, 4-oxadiazoles from amidoximes and nitriles. New J. Chem.41, 1940–1952. 10.1039/c6nj03508f
13
BrázdaV.FojtaM. (2019). The rich world of p53 DNA binding targets: the role of DNA structure. Int. J. Mol. Sci.20, 5605–5627. 10.3390/ijms20225605
14
BrummerT.ZeiserR. (2024). The role of the MDM2/p53 axis in antitumor immune responses. Blood143, 2701–2709. 10.1182/blood.2023020731
15
BudennyyS.KazakovA.KovtunE.ZhukovL. (2023). New drugs and stock market: a machine learning framework for predicting pharma market reaction to clinical trial announcements. Sci. Rep.13, 12817–12830. 10.1038/s41598-023-39301-4
16
BuslovI. V.NovikovA. S.KhrustalevV. N.GrudovaM. V.KubasovA. S.MatsulevichZ. V.et al (2021). 2-Pyridylselenenyl versus 2-Pyridyltellurenyl halides: symmetrical chalcogen bonding in the solid State and reactivity towards nitriles. Symmetry (Basel)13, 2350–2361. 10.3390/sym13122350
17
ButtenschoenM.MorrisG. M.DeaneC. M. (2024). PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chem. Sci.15, 3130–3139. 10.1039/d3sc04185a
18
CabaK.Tran-NguyenV.-K.RahmanT.BallesterP. J. (2024). Comprehensive machine learning boosts structure-based virtual screening for PARP1 inhibitors. J. Cheminform16, 40–57. 10.1186/s13321-024-00832-1
19
ChaputL.MouawadL. (2017). Efficient conformational sampling and weak scoring in docking programs? Strategy of the wisdom of crowds. J. Cheminform9, 37. 10.1186/s13321-017-0227-x
20
ChaputL.Martinez-SanzJ.SaettelN.MouawadL. (2016). Benchmark of four popular virtual screening programs: construction of the active/decoy dataset remains a major determinant of measured performance. J. Cheminform8, 1–17. 10.1186/s13321-016-0167-x
21
ChenT.GuestrinC. (2016). “Xgboost: a scalable tree boosting system,” in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. New York, NY: Association for Computing Machinery, 785–794.
22
ChenH.EngkvistO.WangY.OlivecronaM.BlaschkeT. (2018). The rise of deep learning in drug discovery. Drug Discov. Today23, 1241–1250. 10.1016/j.drudis.2018.01.039
23
ChenT.HeT.BenestyM.KhotilovichV. (2019). Package ‘xgboost.’ R version 90, Vienna, Austria: CRAN, 40.
24
ChengX.ChenR.ZhouT.ZhangB.LiZ.GaoM.et al (2022). Leveraging the multivalent p53 peptide-MdmX interaction to guide the improvement of small molecule inhibitors. Nat. Commun.13, 1087–1098. 10.1038/s41467-022-28721-x
25
ChukwuemekaP. O.UmarH. I.IwaloyeO.OretadeO. M.OlowosokeC. B.ElabiyiM. O.et al (2022). Targeting p53-MDM2 interactions to identify small molecule inhibitors for cancer therapy: beyond “Failure to rescue.”. J. Biomol. Struct. Dyn.40, 9158–9176. 10.1080/07391102.2021.1924267
26
ConsortiumU. (2015). UniProt: a hub for protein information. Nucleic Acids Res.43, 204–212. 10.1093/nar/gku989
27
Corrêa VeríssimoG.Salgado FerreiraR.Gonçalves MaltarolloV. (2025). Ultra‐large virtual screening: definition, recent advances, and challenges in drug design. Mol. Inf.44, e202400305. 10.1002/minf.202400305
28
CrossJ. B.ThompsonD. C.RaiB. K.BaberJ. C.FanK. Y.HuY.et al (2009). Comparison of several molecular docking programs: pose prediction and virtual screening accuracy. J. Chem. Inf. Model49, 1455–1474. 10.1021/ci900056c
29
DaC.KireevD. (2014). Structural protein–ligand interaction fingerprints (SPLIF) for structure-based virtual screening: method and benchmark study. J. Chem. Inf. Model54, 2555–2561. 10.1021/ci500319f
30
DastidarS. G.LaneD. P.VermaC. S. (2009). “Modulation of p53 binding to MDM2: computational studies reveal important roles of Tyr100,” in BMC bioinformatics (London, United Kingdom: BMC Bioinformatics), 1–11.
31
DavisJ.GoadrichM. (2006). “The relationship between Precision-Recall and ROC curves,” in Proceedings of the 23rd international conference on Machine learning. Pittsburgh, PA: International Machine Learning Society, 233–240.
32
DebnathP.DebnathP.RoyM.SierońL.ManiukiewiczW.AktarT.et al (2022). Novel organotin (IV) complexes of 2-[4-hydroxy-3-((2-hydroxyethylimino) methyl) phenylazo] benzoic acid: synthesis, structure, noncovalent interactions and in vitro antibacterial activity. Cryst. (Basel)12, 1582–1599. 10.3390/cryst12111582
33
DingQ.YaoR.BaiY.DaL.WangY.XiangR.et al (2025). Explainable artificial intelligence in the field of drug research. Drug Des. Devel Ther.19, 4501–4516. 10.2147/DDDT.S525171
34
DukhnovskyE. A.NovikovA. S.KubasovA. S.BorisovA. V.SikaonaN. D.KirichukA. A.et al (2024). Halogen bond-assisted supramolecular dimerization of pyridinium-fused 1, 2, 4-Selenadiazoles via four-center Se2N2 chalcogen bonding. Int. J. Mol. Sci.25, 3972–3988. 10.3390/ijms25073972
35
Estrada‐OrtizN.NeochoritisC. G.DömlingA. (2016). How to design a successful p53–MDM2/X interaction inhibitor: a thorough overview based on crystal structures. ChemMedChem11, 757–772. 10.1002/cmdc.201500487
36
FanY.ZhangY.HuaY.WangY.ZhuL.ZhaoJ.et al (2019). Investigation of machine intelligence in compound cell activity classification. Mol. Pharm.16, 4472–4484. 10.1021/acs.molpharmaceut.9b00558
37
FangY.LiaoG.YuB. (2020). Small-molecule MDM2/X inhibitors and PROTAC degraders for cancer therapy: advances and perspectives. Acta Pharm. Sin. B10, 1253–1278. 10.1016/j.apsb.2020.01.003
38
FeinbergE. N.SurD.WuZ.HusicB. E.MaiH.LiY.et al (2018). PotentialNet for molecular property prediction. ACS Cent. Sci.4, 1520–1530. 10.1021/acscentsci.8b00507
39
FerreiraF. J. N.CarneiroA. S. (2025). AI-Driven drug discovery: a comprehensive review. ACS Omega23, 23889–23903. 10.1021/acsomega.5c00549
40
FischerA.SmieskoM.SellnerM.LillM. A. (2021). Decision making in structure-based drug discovery: visual inspection of docking results. J. Med. Chem.64, 2489–2500. 10.1021/acs.jmedchem.0c02227
41
FresnaisL.BallesterP. J. (2021). The impact of compound library size on the performance of scoring functions for structure-based virtual screening. Brief. Bioinform22, bbaa095. 10.1093/bib/bbaa095
42
FriesnerR. A.BanksJ. L.MurphyR. B.HalgrenT. A.KlicicJ. J.MainzD. T.et al (2004). Glide: a new approach for rapid, accurate docking and scoring. 1. Method and assessment of docking accuracy. J. Med. Chem.47, 1739–1749. 10.1021/jm0306430
43
FuC.ChenQ. (2025). The future of pharmaceuticals: artificial intelligence in drug discovery and development. J. Pharm. Anal.15, 101248-101269. 10.1016/j.jpha.2025.101248
44
GohlkeH.HendlichM.KlebeG. (2000). Knowledge-based scoring function to predict protein-ligand interactions. J. Mol. Biol.295, 337–356. 10.1006/jmbi.1999.3371
45
Gómez-SacristánP.SimeonS.Tran-NguyenV.-K.PatilS.BallesterP. J. (2025). Inactive-enriched machine-learning models exploiting patent data improve structure-based virtual screening for PDL1 dimerizers. J. Adv. Res.67, 185–196. 10.1016/j.jare.2024.01.024
46
GowthamanR.LyskovS.KaranicolasJ. (2015). DARC 2.0: improved docking and virtual screening at protein interaction sites. PLoS One10, e0131612. 10.1371/journal.pone.0131612
47
GrudovaM. V.KubasovA. S.KhrustalevV. N.NovikovA. S.KritchenkovA. S.NenajdenkoV. G.et al (2022). Exploring supramolecular assembly space of cationic 1, 2, 4-selenodiazoles: effect of the substituent at the carbon atom and anions. Molecules27, 1029–1042. 10.3390/molecules27031029
48
GuedesI. A.PereiraF. S. S.DardenneL. E. (2018). Empirical scoring functions for structure-based virtual screening: applications, critical aspects, and challenges. Front. Pharmacol.9, 1089. 10.3389/fphar.2018.01089
49
GulliA.PalS. (2017). Deep learning with Keras. Birmingham, United Kingdom: Packt Publishing.
50
GuranovaN. I.Dar’inD.KantinG.NovikovA. S.BakulinaO.KrasavinM. (2019). Rh (II)-catalyzed spirocyclization of α-diazo homophthalimides with cyclic ethers. J. Org. Chem.84, 4534–4542. 10.1021/acs.joc.9b00245
51
GuvenchO.MacKerell JrA. D. (2009). Computational evaluation of protein–small molecule binding. Curr. Opin. Struct. Biol.19, 56–61. 10.1016/j.sbi.2008.11.009
52
HanyM. S.AhmedN. S.BoecklerF. M.IbrahimT. M. (2025). Benchmarking the structure-based virtual screening performance of wild-type and resistant Pf DHFR using docking and machine learning re-scoring. Drug Des. Devel. Ther.19, 7045–7058. 10.2147/DDDT.S537065
53
HeY.DiaoS.HouS.LiT.MengW.ZhangJ. (2025). Identification of novel potential hypoxia-inducible factor-1α inhibitors through machine learning and computational simulations. Front. Chem.13, 1585882. 10.3389/fchem.2025.1585882
54
HolzerP.MasuyaK.FuretP.KallenJ.Valat-StachyraT.FerrettiS.et al (2015). Discovery of a dihydroisoquinolinone derivative (NVP-CGM097): a highly potent and selective MDM2 inhibitor undergoing phase 1 clinical trials in p53wt tumors. J. Med. Chem.58, 6348–6358. 10.1021/acs.jmedchem.5b00810
55
HouH.SunD.ZhangX. (2019). The role of MDM2 amplification and overexpression in therapeutic resistance of malignant tumors. Cancer Cell Int.19, 216–224. 10.1186/s12935-019-0937-4
56
HuangN.KalyanaramanC.BernackiK.JacobsonM. P. (2006). Molecular mechanics methods for predicting protein–ligand binding. Phys. Chem. Chem. Phys.8, 5166–5177. 10.1039/b608269f
57
Il’inM. V.SysoevaA. A.BolotinD. S.NovikovA. S.SuslonovV. V.RogachevaE. V.et al (2019). Aminonitrones as highly reactive bifunctional synthons. An expedient one-pot route to 5-amino-1, 2, 4-triazoles and 5-amino-1, 2, 4-oxadiazoles–potential antimicrobials targeting multi-drug resistant bacteria. New J. Chem.43, 17358–17366. 10.1039/c9nj04529e
58
ImrieF.BradleyA. R.DeaneC. M. (2021). Generating property-matched decoy molecules using deep learning. Bioinformatics37, 2134–2141. 10.1093/bioinformatics/btab080
59
JainA. K.MaoJ.MohiuddinK. M. (1996). Artificial neural networks: a tutorial. Comput. Long. Beach Calif.29 (3), 31–44. 10.1109/2.485891
60
JakalianA.JackD. B.BaylyC. I. (2002). Fast, efficient generation of high‐quality atomic charges. AM1‐BCC model: II. Parameterization and validation. J. Comput. Chem.23, 1623–1641. 10.1002/jcc.10128
61
JiménezJ.SkalicM.Martinez-RosellG.De FabritiisG. (2018). K deep: protein–ligand absolute binding affinity prediction via 3d-convolutional neural networks. J. Chem. Inf. Model58, 287–296. 10.1021/acs.jcim.7b00650
62
JunaidM.WangB.LiW. (2024). Data-augmented machine learning scoring functions for virtual screening of YTHDF1 m6A reader protein. Comput. Biol. Med.183, 109268. 10.1016/j.compbiomed.2024.109268
63
KannanS.PartridgeA. W.LaneD. P.VermaC. S. (2019). The dual interactions of p53 with MDM2 and p300: implications for the design of MDM2 inhibitors. Int. J. Mol. Sci.20, 5996–6017. 10.3390/ijms20235996
64
KeilwagenJ.GrosseI.GrauJ. (2014). Area under precision-recall curves for weighted and unweighted data. PLoS One9, e92209. 10.1371/journal.pone.0092209
65
KhrustalevV. N.GrishinaM. M.MatsulevichZ. V.LukiyanovaJ. M.BorisovaG. N.OsmanovV. K.et al (2021). Novel cationic 1, 2, 4-selenadiazoles: synthesis via addition of 2-pyridylselenyl halides to unactivated nitriles, structures and four-center Se⋯ N contacts. Dalton Trans.50, 10689–10691. 10.1039/d1dt01322j
66
KimS.ChenJ.ChengT.GindulyteA.HeJ.HeS.et al (2019). PubChem 2019 update: improved access to chemical data. Nucleic Acids Res.47, 1102–1109. 10.1093/nar/gky1033
67
KimS.ChenJ.ChengT.GindulyteA.HeJ.HeS.et al (2025). PubChem 2025 update. Nucleic Acids Res.53, D1516–D1525. 10.1093/nar/gkae1059
68
KinzhalovM. A.NovikovA. S.LuzyaninK. V.HaukkaM.PombeiroA. J. L.KukushkinV. Y. (2016). Pd II-mediated integration of isocyanides and azide ions might proceed via formal 1, 3-dipolar cycloaddition between RNC ligands and uncomplexed azide. New J. Chem.40, 521–527. 10.1039/c5nj02564h
69
KinzhalovM. A.LegkodukhA. S.AnisimovaT. B.NovikovA. S.SuslonovV. V.LuzyaninK. V.et al (2017). Tetrazol-5-ylidene gold (III) complexes from sequential [2+ 3] cycloaddition of azide to metal-bound isocyanides and N4 alkylation. Organometallics36, 3974–3980. 10.1021/acs.organomet.7b00591
70
KoesD. R.BaumgartnerM. P.CamachoC. J. (2013). Lessons learned in empirical scoring with smina from the CSAR 2011 benchmarking exercise. J. Chem. Inf. Model53, 1893–1904. 10.1021/ci300604z
71
KouranovA.XieL.de la CruzJ.ChenL.WestbrookJ.BourneP. E.et al (2006). The RCSB PDB information portal for structural genomics. Nucleic Acids Res.34, 302–305. 10.1093/nar/gkj120
72
KramerO.KramerO. (2016). “Scikit-learn,” in Machine learning for evolution strategies (Cham, Switzerland: Springer International Publishing), 45–53.
73
KryukovaM. A.SapeginA. V.NovikovA. S.KrasavinM.IvanovD. M. (2020). New crystal forms for biologically active compounds. Part 2: anastrozole as N-substituted 1, 2, 4-Triazole in halogen bonding and Lp-π interactions with 1, 4-Diiodotetrafluorobenzene. Crystals10, 2073–4352. 10.3390/cryst10050371
74
KulishK. I.NovikovA. S.TolstoyP. M.BolotinD. S.BokachN. A.ZolotarevA. A.et al (2016). Solid state and dynamic solution structures of O-carbamidine amidoximes gives further insight into the mechanism of zinc (II)-mediated generation of 1, 2, 4-oxadiazoles. J. Mol. Struct.1111, 142–150. 10.1016/j.molstruc.2016.01.038
75
KurczabR.SmuszS.BojarskiA. J. (2014). The influence of negative training set size on machine learning-based virtual screening. J. Cheminform6, 1–9. 10.1186/1758-2946-6-32
76
LaiC. W.XieC.RaufmanJ.-P.XieG. (2022). Targeting post-translational regulation of p53 in colorectal cancer by exploiting vulnerabilities in the p53-MDM2 axis. Cancers (Basel)14, 219–235. 10.3390/cancers14010219
77
LeDellE.PetersenM.van der LaanM. (2015). Computationally efficient confidence intervals for cross-validated area under the ROC curve estimates. Electron J. Stat.9, 1583–1607. 10.1214/15-EJS1035
78
LeeY.LeeY. Y.ParkJ.MaksakovaA.SeoD.KimJ.et al (2025). Illudin S inhibits p53-Mdm2 interaction for anticancer efficacy in colorectal cancer. Biomed. and Pharmacother.182, 117795–117805. 10.1016/j.biopha.2024.117795
79
LiB.-H.GeJ.-Q.WangY.-L.WangL.-J.ZhangQ.BianC. (2021). Ligand‐Based and docking‐based virtual screening of MDM2 inhibitors as potent anticancer agents. Comput. Math. Methods Med.2021, 3195957–3195966. 10.1155/2021/3195957
80
LiH.LeungK.-S.WongM.-H.BallesterP. J. (2014). Substituting random forest for multiple linear regression improves binding affinity prediction of scoring functions: cyscore as a case study. BMC Bioinforma.15, 1–12. 10.1186/1471-2105-15-291
81
LiH.PengJ.SidorovP.LeungY.LeungK.-S.WongM.-H.et al (2019). Classical scoring functions for docking are unable to exploit large volumes of structural and interaction data. Bioinformatics35, 3989–3995. 10.1093/bioinformatics/btz183
82
LiW.PengX.LangJ.XuC. (2020). Targeting mouse double minute 2: current concepts in DNA damage repair and therapeutic approaches in cancer. Front. Pharmacol.11, 631–649. 10.3389/fphar.2020.00631
83
LiH.SzeK.LuG.BallesterP. J. (2021). Machine‐learning scoring functions for structure‐based virtual screening. Wiley Interdiscip. Rev. Comput. Mol. Sci.11, e1478. 10.1002/wcms.1478
84
LieblM. C.HofmannT. G. (2021). The role of p53 signaling in colorectal cancer. Cancers (Basel)13, 2125–2134. 10.3390/cancers13092125
85
LiontaE.SpyrouG.K VassilatisD.CourniaZ. (2014). Structure-based virtual screening for drug discovery: principles, applications and recent advances. Curr. Top. Med. Chem.14, 1923–1938. 10.2174/1568026614666140929124445
86
LiuS.AlnammiM.EricksenS. S.VoterA. F.AnanievG. E.KeckJ. L.et al (2018). Practical model selection for prospective virtual screening. J. Chem. Inf. Model59, 282–293. 10.1021/acs.jcim.8b00363
87
LiuH.ZhangH.DingS.XueY.DuX.HanZ.et al (2025). Combining the NanaPPI toolbox and AI-driven virtual inhibitor screening for the p53-MDM2 interaction. Anal. Chem.97, 29–15697. 10.1021/acs.analchem.5c01354
88
MachL. B. (2001). Random forests. Mach. Learn. 45, 5–32.
89
MahdizadehS. J.ErikssonL. A. (2025). iScore: a ML-based scoring function for de novo drug discovery. J. Chem. Inf. Model65 (6), 2759–2772. 10.1021/acs.jcim.4c02192
90
MastersL.EagonS.HeyingM. (2020). Evaluation of consensus scoring methods for AutoDock Vina, smina and idock. J. Mol. Graph Model96, 107532. 10.1016/j.jmgm.2020.107532
91
McDermottJ.SturtevantD.KathadU.VarmaS.ZhouJ.KulkarniA.et al (2022). Artificial intelligence platform, RADR®, aids in the discovery of DNA damaging agent for the ultra-rare cancer atypical teratoid rhabdoid tumors. Front. Drug Discov.2, 1033395. 10.3389/fddsv.2022.1033395
92
McGibbonM.Money-KyrleS.BlayV.HoustonD. R. (2023). SCORCH: improving structure-based virtual screening with machine learning classifiers, data augmentation, and uncertainty estimation. J. Adv. Res.46, 135–147. 10.1016/j.jare.2022.07.001
93
MelekhovaA. A.SmirnovA. S.NovikovA. S.PanikorovskiiT. L.BokachN. A.KukushkinV. Y. (2017). Copper (I)-catalyzed 1, 3-dipolar cycloaddition of ketonitrones to dialkylcyanamides: a step toward sustainable generation of 2, 3-dihydro-1, 2, 4-oxadiazoles. ACS Omega2, 1380–1391. 10.1021/acsomega.7b00130
94
MeliR.MorrisG. M.BigginP. C. (2022). Scoring functions for protein-ligand binding affinity prediction using structure-based deep learning: a review. Front. Bioinformatics2, 885983–886013. 10.3389/fbinf.2022.885983
95
MikherdovA. S.NovikovA. S.KinzhalovM. A.ZolotarevA. A.BoyarskiyV. P. (2018). Intra-/intermolecular bifurcated chalcogen bonding in crystal structure of thiazole/thiadiazole derived binuclear (diaminocarbene) PdII complexes. Cryst. (Basel)8, 112–127. 10.3390/cryst8030112
96
MonsiaR.BhattacharyyaS. (2025). Efficient and explainable virtual screening of molecules through fingerprint-generating networks integrated with artificial neural networks. ACS Omega10, 4896–4911. 10.1021/acsomega.4c10289
97
MoralesJ. F.RuizM. E.StratfordR. E.TaleviA. (2024). Application of machine learning to predict unbound drug bioavailability in the brain. Front. Drug Discov.4, 1360732–1360746. 10.3389/fddsv.2024.1360732
98
MoriwakiH.TianY.-S.KawashitaN.TakagiT. (2018). Mordred: a molecular descriptor calculator. J. Cheminform10, 4. 10.1186/s13321-018-0258-y
99
MorrisG. M.HueyR.LindstromW.SannerM. F.BelewR. K.GoodsellD. S.et al (2009). AutoDock4 and AutoDockTools4: automated docking with selective receptor flexibility. J. Comput. Chem.30, 2785–2791. 10.1002/jcc.21256
100
MueggeI. (2006). PMF scoring revisited. J. Med. Chem.49, 5895–5902. 10.1021/jm050038s
101
MysingerM. M.CarchiaM.IrwinJ. J.ShoichetB. K. (2012). Directory of useful decoys, enhanced (DUD-E): better ligands and decoys for better benchmarking. J. Med. Chem.55, 6582–6594. 10.1021/jm300687e
102
NiaziS. K. (2023). The coming of age of AI/ML in drug discovery, development, clinical testing, and manufacturing: the FDA perspectives. Drug Des. Devel. Ther.17, 2691–2725. 10.2147/DDDT.S424991
103
NogueiraM. S.KochO. (2019). The development of target-specific machine learning models as scoring functions for docking-based target prediction. J. Chem. Inf. Model59, 1238–1252. 10.1021/acs.jcim.8b00773
104
OsipyanA.SapeginA.NovikovA. S.KrasavinM. (2018). Rare medium-sized rings prepared via hydrolytic imidazoline ring expansion (HIRE). J. Org. Chem.83, 9707–9717. 10.1021/acs.joc.8b01210
105
OsmanovV. K.ChipinskyE. V.KhrustalevV. N.NovikovA. S.AskerovR. K.ChizhovA. O.et al (2022). Facile access to 2-selenoxo-1, 2, 3, 4-tetrahydro-4-quinazolinone scaffolds and corresponding diselenides via cyclization between methyl anthranilate and isoselenocyanates: synthesis and structural features. Molecules27, 5799–5819. 10.3390/molecules27185799
106
OyedeleA.-Q. K.AdelusiT. I.OgunlanaA. T.AdeyemiR. O.AtandaO. E.BabalolaM. O.et al (2022). Integrated virtual screening and molecular dynamics simulation revealed promising drug candidates of p53-MDM2 interaction. J. Mol. Model28, 142–155. 10.1007/s00894-022-05131-w
107
O’BoyleN. M.BanckM.JamesC. A.MorleyC.VandermeerschT.HutchisonG. R. (2011). Open Babel: an open chemical toolbox. J. Cheminform3, 1–14. 10.1186/1758-2946-3-33
108
PatelL.ShuklaT.HuangX.UsseryD. W.WangS. (2020). Machine learning methods in drug discovery. Molecules25, 5277. 10.3390/molecules25225277
109
PedregosaF.VaroquauxG.GramfortA.MichelV.ThirionB.GriselO.et al (2011). Scikit-learn: machine learning in python. J. Mach. Learn. Res.12, 2825–2830.
110
PettersenE. F.GoddardT. D.HuangC. C.CouchG. S.GreenblattD. M.MengE. C.et al (2004). UCSF Chimera—A visualization system for exploratory research and analysis. J. Comput. Chem.25, 1605–1612. 10.1002/jcc.20084
111
PeugetS.ZhouX.SelivanovaG. (2024). Translating p53-based therapies for cancer into the clinic. Nat. Rev. Cancer24, 192–215. 10.1038/s41568-023-00658-3
112
PrietoJ. M.SchottlenderG.ClementeC. M.BetanzosR.Fernández Do PortoD.MartiM. A. (2024). “Docking and bias docking,” in Structure-Based Drug Design (Cham, Switzerland: Springer), 127–148.
113
RagozaM.HochuliJ.IdroboE.SunseriJ.KoesD. R. (2017). Protein–ligand scoring with convolutional neural networks. J. Chem. Inf. Model57, 942–957. 10.1021/acs.jcim.6b00740
114
RéauM.LangenfeldF.ZaguryJ.-F.LagardeN.MontesM. (2018). Decoys selection in benchmarking datasets: overview and perspectives. Front. Pharmacol.9, 11. 10.3389/fphar.2018.00011
115
ReutskayaE.OsipyanA.SapeginA.NovikovA. S.KrasavinM. (2018). Rethinking hydrolytic imidazoline ring expansion: a common approach to the preparation of medium-sized rings via side-chain insertion into [1.4] oxa-and [1.4] thiazepinone scaffolds. J. Org. Chem.84, 1693–1705. 10.1021/acs.joc.8b02805
116
RichardsonE.TrevizaniR.GreenbaumJ. A.CarterH.NielsenM.PetersB. (2024). The receiver operating characteristic curve accurately assesses imbalanced datasets. Patterns5 (5), 100994. 10.1016/j.patter.2024.100994
117
RobinsonM. C.GlenR. C.LeeA. A. (2020). Validating the validation: reanalyzing a large-scale comparison of deep learning and machine learning models for bioactivity prediction. J. Comput. Aided Mol. Des.34, 717–730. 10.1007/s10822-019-00274-0
118
RogersD.HahnM. (2010). Extended-connectivity fingerprints. J. Chem. Inf. Model50, 742–754. 10.1021/ci100050t
119
SaitoT.RehmsmeierM. (2015). The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS One10, e0118432. 10.1371/journal.pone.0118432
120
ScantleburyJ.VostL.CarberyA.HadfieldT. E.TurnbullO. M.BrownN.et al (2023). A small step toward generalizability: training a machine learning scoring function for structure-based virtual screening. J. Chem. Inf. Model63, 2960–2974. 10.1021/acs.jcim.3c00322
121
SchakeP.BolzS. N.LinnemannK.SchroederM. (2025). PLIP 2025: introducing protein–protein interactions to the protein–ligand interaction profiler. Nucleic Acids Res.53, W463–W465. 10.1093/nar/gkaf361
122
SchonO.FriedlerA.BycroftM.FreundS. M. V.FershtA. R. (2002). Molecular mechanism of the interaction between MDM2 and p53. J. Mol. Biol.323, 491–501. 10.1016/s0022-2836(02)00852-5
123
SedzroD. M.IdrisM. O.DurojayeO. A.YekeenA. A.FadahunsiA. A.AlakanseS. O. (2022). Identifying potential p53-MDM2 interaction antagonists: an integrated approach of pharmacophore-based virtual screening, interaction fingerprinting, MD simulation and DFT studies. ChemistrySelect7, e202202380. 10.1002/slct.202202380
124
SeglerM. H. S.KogejT.TyrchanC.WallerM. P. (2018). Generating focused molecule libraries for drug discovery with recurrent neural networks. ACS Cent. Sci.4, 120–131. 10.1021/acscentsci.7b00512
125
ShahabM.ZhengG.KhanA.WeiD.NovikovA. S. (2023). Machine learning-based virtual screening and molecular simulation approaches identified novel potential inhibitors for cancer therapy. Biomedicines11, 2251–2270. 10.3390/biomedicines11082251
126
SiuL. L.ItalianoA.MillerW. H.BlayJ.-Y.GietemaJ. A.BangY.-J.et al (2014). Phase 1 dose escalation, food effect, and biomarker study of RG7388, a more potent second-generation MDM2 antagonist, in patients (pts) with solid tumors. J. Clin. Oncol.32, 2535. 10.1200/jco.2014.32.15_suppl.2535
127
SriramanA.RadovanovicM.WienkenM.NajafovaZ.LiY.DobbelsteinM. (2016). Cooperation of Nutlin-3a and a Wip1 inhibitor to induce p53 activity. Oncotarget7 (22), 31623–31638. 10.18632/oncotarget.9302
128
SunD.LiZ.RewY.GribbleM.BartbergerM. D.BeckH. P.et al (2014). Discovery of AMG 232, a potent, selective, and orally bioavailable MDM2–p53 inhibitor in clinical development. J. Med. Chem.57, 1454–1472. 10.1021/jm401753e
129
SysoevaA. A.NovikovA. S.SuslonovV. V.BolotinD. S.Il’inM. V. (2024). 2, 1, 3-Benzoselenadiazole-containing zinc (II) halide complexes: chalcogen bonding in the solid state and catalytic activity in the Schiff condensation. Inorganica Chim. Acta561, 121867. 10.1016/j.ica.2023.121867
130
Tahir ul QamarM.ZhuX.-T.ChenL.-L.AlhussainL.AlshiekheidM. A.TheyabA.et al (2022). Target-specific machine learning scoring function improved structure-based virtual screening performance for SARS-CoV-2 drugs development. Int. J. Mol. Sci.23, 11003–11017. 10.3390/ijms231911003
131
ThakurA.SinghA.KaurN.OjhaR.NepaliK. (2020). Steering the antitumor drug discovery campaign towards structurally diverse indolines. Bioorg Chem.94, 103436. 10.1016/j.bioorg.2019.103436
132
TorresP. H. M.SoderoA. C. R.JofilyP.Silva-JrF. P. (2019). Key topics in molecular docking for drug design. Int. J. Mol. Sci.20, 4574. 10.3390/ijms20184574
133
Tran-NguyenV.-K.JacquemardC.RognanD. (2020). LIT-PCBA: an unbiased data set for machine learning and virtual screening. J. Chem. Inf. Model60, 4263–4273. 10.1021/acs.jcim.0c00155
134
Tran-NguyenV.-K.JunaidM.SimeonS.BallesterP. J. (2023). A practical guide to machine-learning scoring for structure-based virtual screening. Nat. Protoc.18, 3460–3511. 10.1038/s41596-023-00885-w
135
TraweekR. S.CopeB. M.RolandC. L.KeungE. Z.NassifE. F.ErstadD. J. (2022). Targeting the MDM2-p53 pathway in dedifferentiated liposarcoma. Front. Oncol.12, 1006959–1006978. 10.3389/fonc.2022.1006959
136
TruchonJ.-F.BaylyC. I. (2007). Evaluating virtual screening methods: good and bad metrics for the “early recognition” problem. J. Chem. Inf. Model47, 488–508. 10.1021/ci600426e
137
Twarda-ClapaA. (2024). An update patent review of MDM2–p53 interaction inhibitors (2019–2023). Expert Opin. Ther. Pat.34, 1177–1198. 10.1080/13543776.2024.2419836
138
VapnikV.VapnikV. (1995). Support-vector networks. Mach. Learn20, 273–297. 10.1007/bf00994018
139
VassilevL. T.VuB. T.GravesB.CarvajalD.PodlaskiF.FilipovicZ.et al (2004). In vivo activation of the p53 pathway by small-molecule antagonists of MDM2. Science (1979)303, 844–848. 10.1126/science.1092472
140
WalterR. F. H.WernerR.WessollyM.MairingerE.BorchertS.SchmellerJ.et al (2018). Inhibition of MDM2 via Nutlin‐3A: a potential therapeutic approach for pleural Mesotheliomas with MDM2‐Induced inactivation of Wild‐Type P53. J. Oncol.2018, 1986982. 10.1155/2018/1986982
141
WaltersW. P.MurckoM. (2020). Assessing the impact of generative AI on medicinal chemistry. Nat. Biotechnol.38, 143–145. 10.1038/s41587-020-0418-2
142
WangS.ChenF.-E. (2022). Small-molecule MDM2 inhibitors in clinical trials for cancer therapy. Eur. J. Med. Chem.236, 114334. 10.1016/j.ejmech.2022.114334
143
WangS.SunW.ZhaoY.McEachernD.MeauxI.BarrièreC.et al (2014). SAR405838: an optimized inhibitor of MDM2–p53 interaction that induces complete and durable tumor regression. Cancer Res.74, 5855–5865. 10.1158/0008-5472.CAN-14-0799
144
WangZ.SongT.FengY.GuoZ.FanY.XuW.et al (2016). Bcl-2/MDM2 dual inhibitors based on universal pyramid-like α-helical mimetics. J. Med. Chem.59, 3152–3162. 10.1021/acs.jmedchem.5b01913
145
WangD.CuiC.DingX.XiongZ.ZhengM.LuoX.et al (2019). Improving the virtual screening ability of target-specific scoring functions using deep learning methods. Front. Pharmacol.10, 924–935. 10.3389/fphar.2019.00924
146
WangW.AlbadariN.DuY.FowlerJ. F.SangH. T.XianW.et al (2024). MDM2 inhibitors for cancer therapy: the past, present, and future. Pharmacol. Rev.76, 414–453. 10.1124/pharmrev.123.001026
147
WarrenG. L.AndrewsC. W.CapelliA.-M.ClarkeB.LaLondeJ.LambertM. H.et al (2006). A critical assessment of docking programs and scoring functions. J. Med. Chem.49, 5912–5931. 10.1021/jm050362n
148
WattersJ. W.DicksonM. A.SchwartzG. K.Le CesneA.BahledaR.WagnerA. J.et al (2015). TP53 mutations emerge in circulating cell-free DNA obtained from patients undergoing treatment with the HDM2 antagonist SAR405838. Nat. Commun.7, 12609. 10.1038/ncomms12609
149
WeiQ.LiC.TangY.BaiJ.LiW.LiuJ.et al (2025). Mechanistic role of the Mdm2/MdmX lid domain in regulating their interactions with p53. Biomolecules15, 642. 10.3390/biom15050642
150
WójcikowskiM.ZielenkiewiczP.SiedleckiP. (2015). Open Drug Discovery Toolkit (ODDT): a new open-source player in the drug discovery field. J. Cheminform7, 1–6. 10.1186/s13321-015-0078-2
151
WójcikowskiM.KukiełkaM.Stepniewska-DziubinskaM. M.SiedleckiP. (2019). Development of a protein–ligand extended connectivity (PLEC) fingerprint and its application for binding affinity predictions. Bioinformatics35, 1334–1341. 10.1093/bioinformatics/bty757
152
WuZ.RamsundarB.FeinbergE. N.GomesJ.GeniesseC.PappuA. S.et al (2018). MoleculeNet: a benchmark for molecular machine learning. Chem. Sci.9, 513–530. 10.1039/c7sc02664a
153
WuY.MaL.LiX.YangJ.RaoX.HuY.et al (2024). The role of artificial intelligence in drug screening, drug design, and clinical trials. Front. Pharmacol.15, 1459954–1459966. 10.3389/fphar.2024.1459954
154
YaoY.ZhangQ.LiZ.ZhangH. (2024). MDM2: current research status and prospects of tumor treatment. Cancer Cell Int.24, 1–13. 10.1186/s12935-024-03356-8
155
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. 10.1093/nar/gkad1004
156
ZhangS.LouJ.LiY.ZhouF.YanZ.LyuX.et al (2021). Recent progress and clinical development of inhibitors that block MDM4/p53 protein–protein interactions. J. Med. Chem.64, 10621–10640. 10.1021/acs.jmedchem.1c00940
157
ZhangX.ZhouC.YangY.LiuH.WangS.DingX.et al (2022). The discovery of potential MDM2 inhibitors: a combination of pharmacophore modeling, virtual screening, molecular docking studies, and in vitro/in vivo biological evaluation. ChemMedChem17, e202100517. 10.1002/cmdc.202100517
158
ZhangX.ShenC.JiangD.ZhangJ.YeQ.XuL.et al (2023). TB-IECS: an accurate machine learning-based scoring function for virtual screening. J. Cheminform15, 63. 10.1186/s13321-023-00731-x
159
ZhangS.LiuK.LiuY.HuX.GuX. (2025). The role and application of bioinformatics techniques and tools in drug discovery. Front. Pharmacol.16, 1547131–1547142. 10.3389/fphar.2025.1547131
160
ZhaoY.AguilarA.BernardD.WangS. (2015). Small-molecule inhibitors of the MDM2–p53 protein–protein interaction (MDM2 Inhibitors) in clinical trials for cancer treatment: miniperspective. J. Med. Chem.58, 1038–1052. 10.1021/jm501092z
161
ZhouG.RusnacD.-V.ParkH.CanzaniD.NguyenH. M.StewartL.et al (2024). An artificial intelligence accelerated virtual screening platform for drug discovery. Nat. Commun.15, 7761–7775. 10.1038/s41467-024-52061-7
162
ZhuY.WangH.ThuraisamyA. (2019). MDM2/P53 inhibitors as sensitizing agents for cancer chemotherapy. London: Academic Press, Elsevier.
163
ZhuH.AsiaeeA.AzinfarL.LiJ.LiangH.IrajizadE.et al (2025). AUC-PR is a more informative metric for assessing the biological relevance of in silico cellular perturbation prediction models. bioRxiv, 2023–2025. 10.1101/2025.03.06.641935
Summary
Keywords
drug discovery, machine learning, MDM2, molecular docking, Protein–ligand extended connectivity fingerprints, structure-based virtual screening, support vector machine
Citation
Castillo Tarazona MY and Miscione GP (2026) Enhancing structure-based virtual screening of MDM2–p53 inhibitors: a benchmark of machine learning vs. traditional docking scoring functions. Front. Drug Discov. 5:1731262. doi: 10.3389/fddsv.2025.1731262
Received
23 October 2025
Revised
13 December 2025
Accepted
16 December 2025
Published
16 January 2026
Volume
5 - 2025
Edited by
Rodolpho C. Braga, InsilicAll, Brazil
Reviewed by
Cleber C. Melo-Filho, University of North Carolina at Chapel Hill, United States
Alexander Novikov, ITMO University, Russia
Updates
Copyright
© 2026 Castillo Tarazona and Miscione.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Gian Pietro Miscione, gp.miscione57@uniandes.edu.co
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.