Beyond additivity: zero-shot methods cannot predict impact of epistasis on protein properties and function

Abstract Accurate prediction of properties and function of mutated proteins is crucial for both research and industrial applications. Experimental assessment of mutations relies on biochemical techniques, which, while accurate, are costly and labour-intensive. As an alternative, computational methods have emerged as a scalable and cost-effective solution. A key challenge for predicting functional consequences of mutations is epistasis, a phenomenon where the effect of one mutation is influenced by others. We evaluated the ability of 95 zero-shot models to predict the impact of epistasis on proteins using datasets from ProteinGym. Our results demonstrate that while the current models perform well for single mutations and non-epistatic combinations of mutations, they fail to predict the effect of strongly epistatic combinations of mutations. This exposes deficiencies of the state-of-the-art models and the need for focusing on capturing complex mutational interactions, which is essential for advancing both evolutionary studies and protein design.