Fig. 1: A multimodal mutation effect predictor for protein engineering tasks.

a ProMEP combines the sequence context and the structure context of a protein to accurately predict mutation effects in a zero-shot manner. It takes an arbitrary WT protein as input and uses the pre-trained multimodal deep representation learning model to calculate semantic-rich representations for each amino acid of a protein. Specifically, for arbitrary mutations, ProMEP first extracts both sequence embeddings and structure embeddings from the WT protein. These embeddings are then aligned and fed into the pre-trained transformer encoder to generate protein representations at residual resolution. With the sequence decoder, fine-grained protein representations are eventually decomposed into the conditional probabilities on each amino acid under the contexts of both sequence and structure. Effects of an arbitrary mutation can be interpreted as the difference in predicted log-likelihood between the mutated sequence and the WT sequence. A customized protein point cloud is adopted to introduce protein structure context at atomic resolution. b 3D translations and rotations of the input protein structure will not affect the structure context of a protein. ProMEP applies a rotation- and translation-equivariant structure embedding module to guarantee such invariance. c ProMEP can be used to guide protein engineering without the requirements for labeled datasets or a holistic understanding of the protein structure and molecular function. It enables the user to recognize beneficial (multiple) mutants by efficiently traversing the protein fitness landscape.