Designing Interpretability Methods to Decode Deep Learning Models for Protein Fitness Prediction

About

Deep learning models have become powerful tools for predicting the functional impact of variants in proteins, with important applications in clinical genomics, drug development, and protein engineering. However, their lack of interpretability limits their adoption in domains where understanding model decisions is critical. This project aims to develop a concept-based interpretability framework tailored to protein fitness prediction. By linking model attributions to biologically meaningful concepts, such as structural sites and functional domains, the team seeks to create explanations that are both human-interpretable and biologically grounded.