Open Source Bayesian Models. 3. Composite Models for Prediction of Binned Responses.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Bayesian models using structure-derived fingerprints are effective for binary active/inactive classification in drug discovery.
• These models offer high interpretability and are suitable for automated workflows, avoiding overtraining issues common in QSAR/QSPR.
• Current methods are limited to binary classification, restricting their application to multi-state bioactivity data.

Purpose of the Study:

• To develop a novel composite Bayesian modeling approach for multi-state bioactivity prediction.
• To extend the utility of Bayesian models beyond binary classification in drug discovery research.
• To provide a chemically intuitive and robust method for analyzing compounds with multiple activity states.

Main Methods:

• A composite group of Bayesian models was created, with each model handling a specific activity range (bin).
• Compounds are analyzed by predicting their likelihood for each bin, with the highest likelihood indicating the predicted activity state.
• The method was evaluated on numerous datasets from ChEMBL v20, including ADME/Tox and bioactivity data.

Main Results:

• The composite Bayesian model successfully extends binary classification to multi-state predictions.
• The approach maintains the interpretability and speed of traditional Bayesian models.
• Evaluation on ChEMBL and ADME/Tox datasets demonstrated the method's effectiveness.

Conclusions:

• The composite Bayesian modeling approach offers a powerful extension for analyzing multi-state bioactivity data in drug discovery.
• This method enhances the applicability of Bayesian models in automated workflows and complex research scenarios.
• The approach provides chemically intuitive rankings and predictions for candidate molecules across various activity levels.