From single-sequence structure prediction to protein fitness landscape through a composable, epistasis-aware mutation atlas
Predicting the effects of multiple mutations requires modeling epistasis: interactions that make their combined effect differ from the sum of individual effects. Cerebra-Epistasis couples a single-sequence structure predictor with a fitness model that jointly represents additive and epistatic effects, placing these interactions in a three-dimensional structural context.
The framework encodes a wild-type protein once to construct a mutation-indexed atlas of single-mutation scores and epistatic representations. An assembly module retrieves and combines these components to predict fitness and epistasis for arbitrary-order mutants, making large combinatorial landscapes accessible without encoding every mutant separately.
Across diverse assays, the model extrapolated from low-order measurements to unseen higher-order combinations and recovered experimentally observed epistatic patterns, including interactions between spatially close residues.
Status: bioRxiv preprint, 2026. Under review at Nature Biotechnology.