Zero-shot prediction of mutation effects on protein function with multimodal deep representation learning

Abstract Mutations in amino acid sequences can disrupt protein structures and functions. Accurate and unsupervised prediction of mutation effects is critical in biotechnology and biomedicine, but remains a fundamental challenge. To resolve this challenge, we first presented a multimodal deep representation learning model that comprehensively learns both sequence context and structural constraints from ~160 million proteins with reliable structures. Based on the proposed model, we developed ProMEP (Protein Mutational Effect Predictor) to predict mutation effects in a zero-shot manner. ProMEP can capture multi-scale signatures of proteins at atomic-resolution and achieves state-of-the-art (SOTA) performance on mutational effects prediction. ProMEP showed an average accuracy of 91.67% during the engineering of the TnpB protein. Remarkably, ProMEP identified a TnpB variant of triple-mutations with approximately 2.4-fold editing efficiency than the wild-type. ProMEP enables efficient exploration of gigantic protein space and will greatly facilitate studies in protein structures and functions.