Efficient exploration of peptide libraries using active learning with AlphaFold-based screening.

Area of Science:

• Computational biology
• Structural biology
• Drug discovery

Background:

• AlphaFold2 (AF) enables screening of peptide-binding epitopes for Bromodomain and Extraterminal (BET) protein domains.
• Exhaustive AF screening is computationally intensive, limiting its application to large datasets like viral proteomes.

Purpose of the Study:

• To develop an efficient computational strategy for exploring peptide sequence space to identify binding epitopes.
• To reduce the number of required AlphaFold2 calculations for identifying functional peptide-protein interactions.

Main Methods:

• Implementation of an active learning strategy using Thompson sampling (TS) for peptide sequence space exploration.
• Utilizing binary labels for peptide-protein binding predictions.
• Testing the TS approach on a peptide library derived from BRD3 pull-down experiments.

Main Results:

• Thompson sampling (TS) identified 50% of all binders using only 15% of the queries needed for exhaustive sampling, a 3.3-fold improvement over random sampling.
• TS consistently identified known binding epitopes with significantly fewer computational queries.
• The method demonstrated high efficiency in exploring peptide sequence space for binding epitope discovery.

Conclusions:

• Active learning with Thompson sampling offers a computationally efficient method for identifying peptide binders targeting protein domains like BET.
• This approach is transferable to various protein-peptide binding systems and other peptide-property prediction tasks.
• The TS strategy significantly reduces the experimental and computational burden in epitope discovery and drug development pipelines.