Pharmacokinetic-Pharmacodynamic Trade-offs in SARS-CoV-2 Main Protease Inhibitors Unveiled through Machine Learning and Molecular Dynamics Simulations
Anacleto Silva de SouzaiD, Vitor Martins de Freitas AmorimiD, Eduardo Pereira SoaresiD, Robson Francisco de SouzaiD, Cristiane Rodrigues GuzzoiD
Abstract
The SARS-CoV-2 main protease (Mpro) is a validated therapeutic target for inhibiting viral replication. Despite the screening of over 55,000 compounds, few candidates have advanced clinically, underscoring the difficulty in optimizing both target affinity and drug-like properties. Thus, developing effective Mpro inhibitors requires balancing high-affinity binding with favorable pharmacokinetic (PK) properties, such as solubility and permeability. To address this challenge, we integrated machine learning (ML) and molecular dynamics (MD) simulations to investigate the trade-offs between pharmacodynamic (PD) and PK properties in Mpro inhibitor design. We developed ML models to classify Mpro inhibitors based on experimental IC50 data, combining molecular descriptors with structural insights from MD simulations. Our Support Vector Machine (SVM) model achieved strong performance (training accuracy = 0.84, ROC AUC = 0.91; test accuracy = 0.79, ROC AUC = 0.86), while our logistic regression model (training accuracy = 0.78, ROC AUC = 0.85; test accuracy = 0.76, ROC AUC = 0.83) identified key molecular features influencing activity, including quantitative estimation of drug-likeness (QED), Log P and molecular weight (ExactMolWt). Notably, PK descriptors often exhibited opposing trends to binding affinity: hydrophilic features enhanced binding affinity but compromised PK properties, whereas hydrogen bonding, hydrophobic and{pi} -{pi} interactions in subsites S2 and S3/S4 are fundamental for binding affinity. Our findings highlight the need for a balanced approach in Mpro inhibitor design, strategically targeting these subsites may balance PD and PK properties. This study provides a computational framework for rational Mpro inhibitor discovery, combining ML and MD to investigate the complex interplay between enzyme inhibition and drug-likeness. These insights may guide in hit-to-lead optimization of the novel next-generation Mpro inhibitors of the SARS-CoV-2 with preclinical and clinical potential.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 16
- FWCI
- 8.24
- Field percentile
- top 2%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (15 May 2025) · CC BY 4.0
SVM and logistic regression classifiers of Mpro inhibitor potency reach test ROC AUC 0.86 and 0.83; potency favours hydrophilic compounds, trading off against drug-likeness.
The paper notes that few of the over 55,000 compounds tested on SARS-CoV-2 Mpro have advanced clinically, and that predictive models for guiding pharmacokinetic optimization are lacking. It pairs IC50-based classifiers with molecular dynamics contact analysis to relate descriptors to potency and subsite use.
Key results
ShownThe best SVM (Model 9, RBF kernel, fluorescence-assay data, 36 descriptors) separated active from inactive Mpro inhibitors on held-out compounds, classifying inactives more reliably than actives (26 actives missed, 40 inactives falsely called active).
Training accuracy 0.84 and ROC AUC 0.91; test accuracy 0.79, precision 0.75, ROC AUC 0.86. Test confusion matrix: 79 true positives, 26 false negatives, 174 true negatives, 40 false positives. Chosen from 1584 SVM models. · Abstract; Support Vector Machine (SVM) models
ShownThe IC50 measurement method strongly affected model quality. SVMs trained on the mixed FRET, fluorescence and SPR dataset discriminated poorly, and FRET-only SVMs had high recall but lower precision.
Mixed-assay SVMs had ROC AUC between 0.20 and 0.28. FRET-based SVMs had recall close to 1 and precision of 0.72 to 0.73. Mixed-assay logistic regression models (condition 31) did better, with AUC 0.76 to 0.77. · Support Vector Machine (SVM) models; Logistic Regression (LR) Models
SuggestedThe interpretable logistic regression model (Model 10, 85 descriptors) generalized consistently. Its standardized coefficients link lower lipophilicity and higher molecular weight to higher activity probability, the opposite of typical drug-likeness preferences.
Training accuracy about 0.78 and AUC about 0.85; test accuracy about 0.76 and AUC 0.83. Coefficients: MolLogP about -0.6, LabuteASA about -0.77, ExactMolWt about 0.37, QED 0.37, TPSA -0.000156 (near zero). · Abstract; Logistic Regression (LR) Models
ShownSix Mpro inhibitors in preclinical or clinical development (boceprevir, ensitrelvir, lufotrelvir, nirmatrelvir, ritonavir, simnotrelvir) were used to build a descriptor profile with 95% confidence intervals. It serves as a reference for hit selection.
Means across the six compounds: ExactMolWt 561.56 Da, QED 0.33, TPSA 146.4 Å^2, MolLogP 2.03. QED values ranged from 0.50 (nirmatrelvir) and 0.46 (simnotrelvir) down to 0.11 (ritonavir). · Logistic Regression (LR) Models
SuggestedIn 150 ns simulations of four inhibitors, 3AN's quinoline and benzothiazole groups and inhibitor 21's imidazole contacted mainly S2 and S3/S4, while the pyrazole of 37 and imidazole of 40 were solvent-exposed and contacted mainly S1'.
For 3AN, the quinoline group had 1355 contacts with S3/S4 vs 643 with S2; the benzothiazole group had 1799 with S2 vs 618 with S3/S4. IC50 was 0.09 µM (3AN) and 0.06 µM (21) vs 0.33 µM (37) and 0.17 µM (40). · Molecular docking and molecular dynamics…; Molecular docking and molecular dynamics…
Methods, data and reproducibility
The authors curated Mpro IC50 data, computed RDKit 2D and 3D descriptors, and trained SVM and logistic regression classifiers with active defined as pIC50 of at least 5. They also ran docking-based molecular dynamics for four inhibitors, with correlation, free-energy landscape and subsite contact analysis.
- Curated Mpro inhibitor records from the ASAP platform plus 17 manually extracted papers (initial uncleaned set) · SARS-CoV-2 (virus) · n = 55,419 molecules
- Dataset 1: IC50 by FRET assay · SARS-CoV-2 (virus) · n = 353 molecules
- Dataset 2: IC50 by fluorescence assay (used for SVM Model 9 and LR Model 10) · SARS-CoV-2 (virus) · n = 1595 molecules
- Dataset 3: combined FRET, fluorescence and SPR data · SARS-CoV-2 (virus) · n = 1943 molecules
- 150 ns MD simulations of Mpro (AlphaFold3 monomer) docked with inhibitors 3AN, 37, 21 and 40 · SARS-CoV-2 (virus) · n = 4 inhibitor complexes
- Methods & tools
- RDKit, scikit-learn, Dimorphite-DL, UCSF Chimera, AutoDock Vina, AlphaFold3, GROMACS 2022, LigParGen, OPLS-AA, PyMOL, Open Babel, ASAP
- Data & accessions
- https://github.com/anacletosouza/Mpro_AI-ML_pipeline/
- Software & environment
- GROMACS version 2022 and PyMOL v2.5.0 are stated; the other Python libraries are named without versions.
- Assay conditions & compute
- HPC cluster with four nodes with high-memory CPUs and Nvidia GPUs (L40S and A30); run times are not given.
- Reproducibility gaps
- Model parameters, descriptor tables and some scripts are in supplementary files; library versions are not stated.
Limitations (4)
- Class imbalance (training set 475 active vs 801 inactive) is linked to the SVM missing more actives than inactives. · stated by the authors
- Pooling IC50 values from different assay types (FRET, fluorescence, SPR) degraded SVM performance (ROC AUC 0.20-0.28), which the authors attribute to the mixture of measurement methods. · stated by the authors
- Pharmacokinetic behaviour is inferred only from computed descriptors and their model coefficients; no PK properties were measured or predicted directly, and there is no external validation set. · noted in this summary
- Interaction conclusions rest on four randomly selected inhibitors; replicate simulations are not described, so contact counts may not generalize to other scaffolds. · noted in this summary
Who should read:Read this if you design or triage SARS-CoV-2 Mpro inhibitors and want descriptor-based activity classifiers plus subsite interaction analysis, including the effect of assay type on QSAR training data.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- A blueprint for high affinity SARS-CoV-2 Mpro inhibitors from activity-based compound library screening guided by analysis of protein dynamics
J. Gossen, S. Albani, … G. Rossetti·bioRxiv·Dec 14, 2020
- Large library docking for novel SARS-CoV-2 main protease non-covalent and covalent inhibitors
E. A. Fink, C. Bardine, … C. S. Craik·bioRxiv·Jul 5, 2022
- Molecular dynamics and in silico mutagenesis on the reversible inhibitor-bound SARS-CoV-2 Main Protease complexes reveal the role of lateral pocket in enhancing the ligand affinity
Y. L. Weng, S. R. Naik, … A. Ganesan·bioRxiv·Oct 31, 2020
- Discovery of SARS-CoV-2 Mpro Peptide Inhibitors from Modelling Substrate and Ligand Binding
H. T. Henry Chan, Marc A. Moesser, … Garrett M. Morris·Chemical SciencePeer-reviewed·Jan 1, 2021
- Structure-based identification of naphthoquinones and derivatives as novel inhibitors of main protease Mpro and papain-like protease PLpro of SARS-CoV-2
L. H. Santos, T. Kronenberger, … R. Ferreira·bioRxiv·Jan 5, 2022
- High Throughput Virtual Screening and Validation of a SARS-CoV-2 Main Protease Non-Covalent Inhibitor
A. Clyde, S. Galanie, … R. Stevens·bioRxiv·Mar 27, 2021