Publication date: Oct 01, 2026
The COVID-19 pandemic highlighted the urgent necessity for effective therapeutic agents targeting SARS-CoV-2. The study focuses on predicting the bioactivity of small molecules against the polyprotein 1ab of SARS-CoV-2, a crucial target for antiviral therapeutics. Using data from the ChEMBL database (CHEMBL4523582), 3747 compounds were curated with IC values, following a stringent preprocessing and standardization protocol. Molecular fingerprints, including PubChem and Morgan fingerprints, were employed for feature generation, resulting in 1905 descriptors. Performed the Feature importance analysis using a Random Forest Regressor, SHAP and RFE to select the top features for model building. Machine learning models were trained and evaluated, including Linear Regression, Decision Tree, Random Forest, Partial Least Squares, LightGBM, and XGBoost. The Stacking model combining LightGBM, Random Forest, and XGBoost performed the best, using the top 80 features (SHAP) with an R^2 of 0. 811 (95% CI: [0. 751, 0. 860]), MSE of 0. 899, and MAE of 0. 604 in the external validation set. Deep learning models such as MLP, GRU, VAE, FFNN, and CNN were also explored, with the MLP showing the highest R^2 of 0. 807 (95% CI: [0. 755, 0. 851]). The results suggest that feature selection significantly enhances predictive accuracy, with XGBoost and MLP models being particularly effective. A web application (CoviPredX) was developed using PHP for user-friendly prediction of pIC values, accessed from http://covipredx. bicpu. edu. in/. The study comprehensively compares machine learning and deep learning approaches for predicting small molecule bioactivity, offering valuable insights for drug discovery efforts against SARS-CoV-2.
Semantics
| Type | Source | Name |
|---|---|---|
| disease | MESH | COVID-19 pandemic |
| disease | MESH | MAE |