Benchmarking Active Learning Protocols for Ligand-Binding Affinity Prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Active learning (AL) is crucial for identifying potent drug candidates from large molecular libraries.
• Understanding AL parameter influence is key to designing effective drug discovery protocols.

Purpose of the Study:

• To systematically evaluate machine learning models, sample selection, and batch sizes in active learning for drug discovery.
• To assess the impact of data set characteristics and noise on AL performance.

Main Methods:

• Utilized four affinity datasets (TYK2, USP7, D2R, Mpro) to test Gaussian process (GP) and Chemprop models.
• Evaluated performance using metrics like R2, Spearman rank, RMSE, Recall, and F1 score.
• Assessed the effect of initial and subsequent batch sizes and artificial data noise.

Main Results:

• GP model outperformed Chemprop on sparse datasets; performance was comparable on larger datasets.
• Larger initial batch sizes improved top binder identification (Recall) and overall correlation.
• Smaller batch sizes (20-30 compounds) were optimal for subsequent AL cycles.
• Moderate artificial noise did not hinder, but excessive noise negatively impacted, performance.

Conclusions:

• The choice of AL model and parameters significantly impacts drug discovery outcomes.
• Optimizing batch size and considering data sparsity are critical for robust active learning protocols.
• AL remains effective even with moderate data noise, but excessive noise compromises its utility.