ROPIUS0: A deep learning-based protocol for protein structure prediction and model selection and its performance in CASP14

Abstract

Protein structure prediction has recently been revolutionized when AlphaFold2 [1] predicted protein structures with near-experimental accuracy in the latest CASP14 season of critical assessment of methods of protein structure prediction (CASP). Among numerous implications, this breakthrough has led to a rapidly growing number of high-quality structural models [2]. We present a protocol ROPIUS0 for protein structure prediction and model selection and discuss its benefits in the new era of structure prediction. At the core of the ROPIUS0 protocol is the deep learning module developed for the selection of protein structural models. It is shown that the direct use of predicted inter-residue distances may be sufficient to discriminate between correct and incorrect protein folds, considering only a small fraction of predicted distances. We extensively tested the protocol: In the latest CASP14 prediction season, a ROPIUS0 variant based on model selection ranked 13th in the category of tertiary structure prediction. Its performance is on par with top-performing automated prediction servers when tested on the CASP13 dataset, and it performs similarly on a CAMEO dataset. The results suggest ways to improve searching for structurally similar and homologous proteins without considerably increasing speed. Our new open-source threading tool based on comparing a subset of inter-residue distances demonstrates the effectiveness and application of the deep learning module of the ROPIUS0 protocol.