Two faces of drug repurposing data in machine learning-driven structure-based virtual screening supported by interaction fingerprints
Clinical Snapshot
PICO Framework
| P — Population | SARS-CoV-2 main protease (Mpro) inhibitor candidates drawn from drug repurposing libraries and rationally designed compound libraries |
| I — Intervention | Random forest machine learning models trained on protein-ligand interaction fingerprints derived from ensemble docking of three distinct compound libraries (drug repurposing datasets and Mpro-specific inhibitor datasets) |
| C — Comparator | Models trained on drug repurposing data alone versus models trained on rationally designed Mpro inhibitors alone versus hybrid (combined) datasets |
| O — Outcomes | Classification performance of random forest models in predicting Mpro inhibitor activity; applicability domain coverage; predictive accuracy metrics across different fingerprint implementations and training set compositions |
Bottom Line
This computational benchmarking study from the Chumakov Federal Scientific Center investigates how the chemical origin of training data affects machine learning models used to identify SARS-CoV-2 main protease (Mpro) inhibitors via structure-based virtual screening. The central finding is that random forest classifiers trained solely on drug repurposing datasets perform poorly at classifying Mpro inhibitors, but that blending repurposing data with libraries of rationally designed Mpro inhibitors meaningfully expands the applicability domain of the resulting models. This is a methodologically important observation for computational drug discovery: drug repurposing screens, while pragmatically attractive, may introduce chemical noise that degrades model specificity when used in isolation. The practical recommendation — to curate balanced, chemically diverse training sets that combine target-specific and repurposing data — is actionable and relevant to any virtual screening campaign. However, the absence of quantitative performance metrics, confidence intervals, and prospective experimental validation in the reported abstract limits independent assessment of result magnitude and robustness. Clinicians and translational researchers should regard this as a methodological signal rather than a validated discovery pipeline. For Australian practice, the relevance is indirect: improved Mpro virtual screening tools could eventually support next-generation antiviral development beyond current TGA-approved options such as Paxlovid.
Key Findings
P Value: Not reported
Effect Size: Not quantified in the abstract; directional improvement in applicability domain reported without magnitude
Primary Outcome: Random forest classifiers trained exclusively on drug repurposing compound libraries demonstrated low predictive performance for SARS-CoV-2 Mpro inhibitor classification; incorporation of drug repurposing data into libraries of rationally designed Mpro inhibitors substantially improved the applicability domain of the resulting models
Nnt Or Sensitivity: No sensitivity, specificity, AUC-ROC, or equivalent diagnostic-style metrics reported in the abstract; this is a model performance benchmarking study — relevant metrics would include AUC-ROC, Matthews Correlation Coefficient (MCC), enrichment factor, and applicability domain coverage, none of which are quantified in the available abstract text
Confidence Interval: Not reported
Clinical Application
The methodological recommendations (balanced dataset preparation combining drug repurposing and target-specific inhibitor data) are feasible to implement in existing computational drug discovery workflows using open-source tools (RDKit, scikit-learn, AutoDock). The ensemble docking component requires access to high-performance computing infrastructure, which is available in most academic and pharmaceutical research settings. Direct clinical applicability in Australia is limited at this stage. SARS-CoV-2 Mpro is the target of nirmatrelvir (Paxlovid, nirmatrelvir/ritonavir), which is TGA-approved and PBS-listed for high-risk COVID-19 patients in Australia. Improved virtual screening methodologies for Mpro could contribute to next-generation antiviral development relevant to future SARS-CoV-2 variants or emerging coronaviruses. RACGP guidelines on COVID-19 antiviral prescribing would be the downstream clinical framework. The Chumakov Institute (Russia) affiliation has no specific regulatory or funding relationship with Australian bodies (TGA, NHMRC, MRFF), but the findings are internationally generalisable as a methodological contribution. Australian researchers at institutions such as the Doherty Institute, WEHI, or Monash Institute of Pharmaceutical Sciences working on antiviral drug discovery would be the primary local beneficiaries. Not directly applicable to patient populations; findings apply to computational medicinal chemists and drug discovery scientists developing virtual screening pipelines for SARS-CoV-2 Mpro inhibitors or analogous viral protease targets
Abstract
Ensemble docking is a prominent source of protein-ligand complex structures, which could be conveniently analyzed with interaction fingerprints approach in a machine learning framework. Nevertheless, the chemical nature and provenance of the source data libraries used for model development may influence the predictive ability. We assessed the performance of SARS-CoV-2 main protease (Mpro) inhibitors classification by random forest models trained on different fingerprint implementations using three inhibitor libraries, comprising drug repurposing data or compounds rationally designed as Mpro inhibitors. Models trained on drug repurposing data showed a low prediction performance on their own, while infusion of these data into libraries with specific inhibitors greatly improved applicability domains of the models. Our results highlight the importance of balanced dataset preparation for machine learning applications in medicinal chemistry, especially when relied on drug repurposing experiments.
References
- 1.Fomina, A. D., & Osolodkin, D. I. (2026). Two faces of drug repurposing data in machine learning-driven structure-based virtual screening supported by interaction fingerprints. Journal of Computer-Aided Molecular Design. https://doi.org/10.1155/2015/471371
Related Research
Archives of toxicology
A review of machine learning in toxicology: current practices and reporting gaps
2 Aug 2026
Journal of chemical information and modeling
Characterizing the MSMP-CCR2 Interaction through Molecular Dynamics Simulations and Machine Learning Approaches
14 July 2026
Journal of integrative medicine
Deep learning in traditional Chinese medicine
1 July 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service