All-atom protein sequence design based on geometric deep learning

Abstract The development of advanced deep learning methods has revolutionized computational protein design. Although the success rate of design has been significantly increased, the overall accuracy of de novo design remains low. Many computational sequence design approaches are devoted to recover the original sequences for given protein structures by encoding the environment of the central residue without considering atomic details of side chains. This may limit the exploration of new sequences that can fold into the same structure and restrain function design that depends on interaction details. In this study, we proposed a novel deep learning frame-work, GeoSeqBuilder, to learn the relationship between protein structure and sequence based on rotational and translational invariance by extracting the information from relative locations. We utilized geometric deep learning to fetch the spatial local geometric features from protein backbones and explicitly incorporated three-body interactions to learn the inter-residue coupling information, and then determined the central residue type. Our model recovers over 50% native residue types and simultaneously gives highly accurate prediction of side-chain conformations which gives the atomic interaction details and circumvents the dependence of protein structure prediction tools. We used the likelihood confidence log P as scoring function for sequence and structure consistence evaluation which exhibits strong correlation with TM-score, and can be applied to recognize near-native structures from protein decoys pool in protein structure prediction. We have used GeoSeqBuilder to design sequences for two proteins, including thiore-doxin and a de novo hallucinated protein. All of the 15 sequences experimentally tested can be expressed as soluble monomeric proteins with high thermal stability and correct secondary structures. We further solved one crystal structure for thioredoxin and two for the hallucinated structure and all the experimentally solved structures are in good agreement with the designed models. The two designed sequences for the hallucination structure are novel without any homologous sequences within the latest released database clust30. The ability of GeoSeqBuilder to design new sequences for given protein structures with atomic details makes it applicable, not only for de novo sequence design, but also for protein-protein interaction and functional protein design.