Bioactivity Comparison across Multiple Machine Learning Algorithms Using over 5000 Datasets for Drug Discovery.

Area of Science:

• Computational chemistry
• Pharmacology
• Data science

Background:

• Machine learning (ML) is increasingly vital in pharmaceutical research for drug discovery.
• Previous studies have evaluated ML methods on smaller scales, often using limited datasets.
• Large-scale comparisons of diverse ML algorithms are needed to identify optimal approaches.

Purpose of the Study:

• To conduct the largest-scale comparison to date of various ML algorithms for drug discovery.
• To evaluate the performance of proprietary software Assay Central against established ML methods.
• To assess the utility of ML models in predicting compound activity and toxicity.

Main Methods:

• Extracted over 5000 datasets from the ChEMBL database.
• Applied ECFP6 fingerprints and compared Assay Central with random forest, k-nearest neighbors, support vector classification, naïve Bayesian, AdaBoosted decision trees, and deep neural networks.
• Assessed model performance using fivefold cross-validation metrics (AUC, F1, Cohen's kappa, MCC).

Main Results:

• All evaluated ML methods demonstrated comparable performance on ranked normalized scores.
• Assay Central and support vector classification showed similar top-tier performance.
• Deep neural networks did not exhibit an advantage over other methods in this extensive dataset comparison.

Conclusions:

• Assay Central and support vector classification are highly effective ML methods for drug discovery applications.
• This large-scale study provides valuable insights into the comparative performance of ML algorithms.
• Future work should explore additional databases, descriptors, and ML algorithms for further refinement.