Active Learning for Protein Structure Prediction

Abstract

Accurate protein structure prediction is challenging especially for families and types that are not well-represented in current training data. Active learning selects candidates for labeling with the aim of most rapidly improving model performance. In a general, many labeling strategies have been proposed. However, with protein structure prediction, most of these strategies dont apply or are difficult due to the high-dimensional, variable-dimensional regression target and the inherent complexity of the models involved. We applied a novel active learning strategy, DEW-DROP, to protein structure prediction on two different protein datasets: VHH-only antibodies (Nanobodies), and Mycobacterial proteins. We introduce a domain-specific fine-tuned Equifold model for VHH structures and apply DEWDROP to generate ensembles of predictions using Monte Carlo dropout. Using the statistics of these we select batches with high information content for labeling. We show that DEWDROP (1) improves model training efficiency through batch optimization outper-forming baselines, and (2) selects data with relevant high information content.