trRosetta

Abstract

he prediction of interresidue contacts and distances from coevo-lutionary data using deep learning has considerably advancedprotein structure prediction. Here, we build on these advances bydeveloping a deep residual network for predicting interresidueorientations, in addition to distances, and a Rosetta-constrainedenergy-minimization protocol for rapidly and accurately generat-ing structure models guided by these restraints. In benchmarktests on 13th Community-Wide Experiment on the Critical Assess-ment of Techniques for Protein Structure Prediction (CASP13)-and Continuous Automated Model Evaluation (CAMEO)-derivedsets, the method outperforms all previously described structure-prediction methods. Although trained entirely on native proteins,the network consistently assigns higher probability to de novo-designed proteins, identifying the key fold-determining residuesand providing an independent quantitative measure of the “ide-ality” of a protein structure. The method promises to be useful fora broad range of protein structure prediction and design problems.