Variant effect prediction using structure-informed protein language models

The workhorse molecule of life, proteins play a central role in cellular functions. Their variations in humans and pathogens often lead to genetic diseases and therapeutic resistance. The ability to decipher the association between protein variations and resulting effects would facilitate disease prognostics and biologics design (e.g. antibodies). Although multiplexed assays of variant effects (MAVE) are generating data ranging from protein stability to cell viability, their speed and applicability are dwarfed by the amount of variants and effects to characterize. Therefore, there is a critical need to develop high-throughput and accurate computational tools with explainability to speed up variant effect evaluation and provide clues for mechanism analysis.