Contact-assisted Protein Threading

Abstract

Protein threading is a popular template-based approach for predicting protein three-dimensional structure in the absence of close templates. However, identifying remote homologous templates remains challenging. Inspired by recent progress in inter-residue contact prediction driven by sequence co-evolution and deep learning, we have developed a new contact-assisted threading method to explore whether incorporating residue-residue contact information along with various sequential and structural features improves threading performance [1]. Experimental results show statistically significant threading performance boost with the incorporation of native contacts as well as ultra-deep learning based state-of-the-art RaptorX predicted contacts compared to baseline contact-free threading method on Test500 dataset. We also observe that our contact-assisted method outperforms popular threading method MUSTER on the same dataset (Test500) using both true and predicted contacts. Furthermore, benchmarking on PSICOV150 dataset reveals that our contact-assisted threading method significantly outperforms cutting-edge contact-assisted ab initio folding method CONFOLD2, both using the same classical neural network based MetaPSICOV predicted contact maps. Overall, these results demonstrate that incorporating residue-residue contact maps boost protein threading performance. Furthermore, our systematic study unravels the mutual association between the quality of contact maps and threading performance by demonstrating that contact-assisted threading using low-quality pure co-evolutionary based contacts (mfDCA and PSICOV) badly degrades threading performance as opposed to high-quality contacts (RaptorX and MetaPSICOV) that combine machine learning with co-evolutionary information, with RaptorX, the most accurate contacts, assisted threading attaining the best threading performance. On the recently concluded 13th Critical Assessment of protein Structure Prediction (CASP13) experiment, our work outperforms state-of-the-art contact-assisted threading methods EigenThreader and map_align based on mean TM-score of top ranked models as well as success rate of identifying correct folds, establishing our method as one of the best cutting-edge contact-assisted threading methods. Furthermore, to investigate the relative contribution of highly accurate state-of-the-art contact maps on threading performance, we consider the top five officially ranked contact predictors of comparable qualities from CASP13. The results demonstrate that the first-ranked contact predictor, TripletRes, attains the best threading performance by reaching a mean TM-score that is comparable to other top ranked predictors, but statistically significantly better than the fifth-ranked contact predictor, DeepMetaPSICOV. In view of the rapid developments in contact prediction technologies, this study opens a fertile line of research - how contact-assisted threading pushes the frontier of low-homology protein modeling beyond what is currently feasible.