Improved protein structure prediction by deep learning irrespective of co-evolution information

Abstract

We describe our latest study of the deep convolutional residual neural networks (ResNet) for protein structure prediction, including deeper and wider ResNets, the efficacy of different input features, and improved 3D model building methods. Our ResNet can predict correct folds (TMscore>0.5) for 26 out of 32 CASP13 FM (template-free-modeling) targets and L/5 long-range contacts for these targets with precision over 80%, a significant improvement over the CASP13 results. Although co-evolution analysis plays an important role in the most successful structure prediction methods, we show that when co-evolution is not used, our ResNet can still predict correct folds for 18 of the 32 CASP13 FM targets including several large ones. This marks a significant improvement over the top co-evolution-based, non-deep learning methods at CASP13, and other non-coevolution-based deep learning models, such as the popular recurrent geometric network (RGN). With only primary sequence, our ResNet can also predict correct folds for all 21 human-designed proteins we tested. In contrast, RGN predicts correct folds for only 3 human-designed proteins and zero CASP13 FM target. In addition, we find that ResNet may fare better for the human-designed proteins when trained without co-evolution information than with co-evolution. These results suggest that ResNet does not simply denoise co-evolution signals, but instead is able to learn important sequence-structure relationship from experimental structures. This has important implications on protein design and engineering especially when evolutionary information is not available.