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

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Protein Engineering

Background:

• Protein fitness landscape characterization and variant effect prediction are hindered by data scarcity.
• Existing methods struggle with limited datasets, impacting protein engineering advancements.

Purpose of the Study:

• To introduce fitness translocation, a novel data augmentation strategy for protein variant effect prediction.
• To address data scarcity challenges in characterizing protein fitness landscapes.

Main Methods:

• Fitness translocation leverages variant fitness data from homologous proteins to generate synthetic variants for a target protein.
• Protein language model embeddings are used to compute variant differences and create synthetic variants in embedding space.
• The method involves applying fitness offsets from homolog variants to the target wild-type embedding.

Main Results:

• Fitness translocation consistently improves predictive performance across different protein families (IGPS, GFP, SARS-CoV-2 spike proteins).
• The strategy is particularly effective in low-data regimes, enhancing variant effect prediction accuracy.
• The method demonstrates efficacy even when augmenting with remote homologs (as low as 35% sequence identity).

Conclusions:

• Fitness translocation expands and diversifies protein fitness landscapes through biologically grounded data augmentation.
• This approach supports more data-efficient protein engineering by improving variant effect prediction models.
• The study highlights the potential of leveraging homologous protein data for enhanced predictive modeling.