Fitness translocation: improving variant effect prediction with biologically-grounded data augmentation

Abstract

Data scarcity limits the characterization of protein fitness landscapes and the development of accurate variant effect prediction models. To address this challenge, we introduce fitness translocation, a data augmentation strategy that generates synthetic variants for a target protein by leveraging variant fitness data previously measured in homologous proteins. Using embeddings from protein language models, the method computes the difference between each homolog variant and its wild type and applies these offsets to the target wild-type embedding to create synthetic variants in embedding space. We illustrate the utility of fitness translocation in the context of variant effect prediction on three protein families: IGPS, GFP, and SARS-CoV-2 spike proteins, across different models and training data sizes. Fitness translocation consistently improves predictive performance, particularly under limited training data, and is effective even when augmenting with remote homologs sharing as little as 35% sequence identity. These results illustrate how biologically grounded data augmentation can expand and diversify protein fitness landscapes, supporting more data-efficient protein engineering. The code is available at https://github.com/adrienmialland/ProtFitTrans.