Predicting anti-SARS-CoV-2 activities of chemical compounds using machine learning models.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Infectious disease therapeutics

Background:

• Accelerating the discovery of novel Coronavirus Disease 2019 (COVID-19) therapeutics is critical.
• Existing methods for predicting anti-SARS-CoV-2 activity require optimization.
• Machine learning (ML) offers a promising approach to enhance drug discovery pipelines.

Purpose of the Study:

• To develop and evaluate ML models for predicting the anti-SARS-CoV-2 activities of screening compounds.
• To identify optimal ML algorithms and molecular descriptors for accurate activity prediction.
• To create a user-friendly web server for predicting anti-SARS-CoV-2 activity of arbitrary compounds.

Main Methods:

• Explored 6 ML algorithms (including k-nearest neighbors - KNN) with 15 molecular descriptors.
• Utilized data from 9 screening assays in the COVID-19 OpenData Portal.
• Developed descriptor-based (COVID-19-CP) and graph-based (Attentive FP) models.
• Compared model performance against REDIAL-2020 and developed a consensus prediction model.

Main Results:

• The KNN model with GAFF+RDKit descriptor demonstrated the best performance (average accuracy 0.68, AUC 0.74).
• Descriptor-based KNN models outperformed other ML algorithms and descriptors.
• Attentive FP models showed comparable performance to COVID-19-CP and outperformed REDIAL-2020.
• Consensus prediction significantly boosted accuracy compared to individual models.

Conclusions:

• Developed accurate ML models (COVID-19-CP and Attentive FP) for predicting anti-SARS-CoV-2 activity.
• The KNN model with GAFF+RDKit descriptor is highly effective for COVID-19 drug discovery.
• A web server is available for predicting compound activity, aiding therapeutic development.
• Consensus prediction strategies can enhance the success rate of identifying COVID-19 drug candidates.