How Effective is Contact-assisted Protein Threading?

Abstract

Protein threading is a powerful approach for predicting protein three-dimensional structure particularly when direct homologous relationships with known structures cannot be easily detected. However, remote homology detection via threading remains challenging, in part, due to the limitations of threading scoring function for selecting optimal structural templates. In light of the recent advancements in residue-residue contact prediction technologies powered by sequence co-evolution and deep learning, we have developed a new threading method by integrating residue-residue contact information with various sequential and structural features to improve threading scoring function for better template selection [1]. Our contact-assisted threading attains a statistically significant boost in threading performance with the incorporation of true contacts as well as contacts predicted by RaptorX, a state-of-the-art contact predictor based on ultra-deep learning model, compared to a baseline contact-free threading acting as control on a large benchmark set comprising of 500 targets (Test500). Furthermore, contact-assisted threading on a dataset of 150 targets (PSICOV150) using contacts predicted by MetaPSICOV, a classical neural network-based meta approach, greatly outperforms cutting-edge contact-assisted ab initio folding method CONFOLD2 that utilizes the same MetaPSICOV predicted contacts - indicating that contact-assisted threading can be advantageous over contact-driven ab initio folding. A systematic exploration of the mutual association between the quality of predicted contacts and the resulting threading performance reveals that contact-driven threading with low-quality contacts predicted from pure co-evolutionary analysis (mfDCA and PSICOV) is not as useful as incorporating high-quality contacts from hybrid approaches that combine sequence co-evolution and machine learning (RaptorX and MetaPSICOV) in that high-quality contacts lead to improved threading performance while low-quality contacts deteriorate it. On the recent Critical Assessment of protein Structure Prediction (CASP13) targets, our method outperforms contact-assisted threading methods EigenThreader and map_align based on average TM-score of top ranked models as well as success rate of identifying correct overall folds, thereby attaining state-of-the-art performance. To further investigate the relative contributions of comparable-quality contact maps on threading performance, we consider the top five officially ranked contact predictors from CASP13 to perform contact-assisted threading on CASP13 targets. The results show that the top-ranked contact predictor, TripletRes, delivers the highest threading performance boost by attaining the best TM-score that is comparable to other high performing predictors including RaptorX, but statistically significantly better than fifth-ranked contact predictor DeepMetaPSICOV. Overall, these results demonstrate that incorporating residue-residue contact information is highly effective for improved protein threading particularly in the presence of high-quality contact maps. With rapidly expanding sequence databases and steady progress in deep learning technologies, contact prediction is likely to mature further -making our contact-assisted threading method an evolving new direction for improved remote homology protein modeling.