Open Source Bayesian Models. 2. Mining a "Big Dataset" To Create and Validate Models with ChEMBL.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• A reference implementation of Bayesian model building using extended connectivity (ECFP) and functional class (FCFP) fingerprints was previously described.
• A large-scale validation is necessary to confirm the generalizability of this technique across diverse drug discovery datasets.

Purpose of the Study:

• To validate the generalized effectiveness of Laplacian-corrected naïve Bayesian model building with ECFP/FCFP fingerprints.
• To assess the impact of fingerprint folding and training/testing set partitioning on model performance.
• To provide publicly accessible datasets and model data for further research.

Main Methods:

• Utilized the ChEMBL database (version 20), splitting it into over 2000 target- and activity-specific datasets.
• Developed an automated algorithm for active/inactive threshold determination for two-state Bayesian classification.
• Applied ECFP and FCFP fingerprints (maximum diameter 6) for model building and validation.

Main Results:

• The Bayesian model implementation proved effective for the majority of drug discovery datasets.
• Quantified the impact of fingerprint folding on receiver operator curve cross-validation metrics.
• Evaluated the influence of training/testing set partitioning on recall rates.

Conclusions:

• The validated Bayesian model offers a robust approach for drug discovery applications.
• Novel visualization methods attribute molecular regions to activity contributions, aiding lead optimization.
• The publicly available data facilitates broader research into drug target identification and off-target effect prediction.