Fingerprint-Based Machine Learning for SARS-CoV-2 and MERS-CoV Mpro Inhibition: Highlighting the Potential of Bayesian Neural Networks
Niklas Piet Doering, Valerij Talagayev, Sijie Liu, Gerhard Wolber · ChemRxiv · 2025
Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) and Middle East respiratory syndrome coronavirus (MERS-CoV) are two important targets in current drug discovery, mainly due to the COVID-19 pandemic and the MERS-CoV outbreaks in recent years. An important target of both SARS-CoV-2 and MERS-CoV is the main protease ($M^{pro}$). The ASAP Discovery Consortium focused in the recent years on the acceleration of $M^{pro}$ inhibitors with a part of this initiative being an open blind challenge in collaboration with Valence lab using the Polaris platform, where datasets of previously undisclosed inhibitors of SARS-CoV-2 $M^{pro}$ and MERS-CoV $M^{pro}$ was shared with researchers, to allow the development of machine learning and deep learning models for the prediction of the potency. We used this opportunity to evaluate and compare traditional machine learning models consisting of a random forest (RF) and gradient boosting model (XGBoost) with a bayesian neural network (BNN) model. For this purpose, we created single task models for the predictions for each of the targets. The results obtained showed that the BNN model outperformed both traditional machine learning models for both targets, indicating that BNNs are a promising deep learning framework in low-data regimes.