Detection of protein structural hotspots using AI distillation and explainability: application to the DAX-1 protein

AlphaMissense is a valuable resource for discerning important functional regions within proteins, providing pathogenicity heatmaps that highlight the pathogenic risk of specific mutations along the protein sequence. However, due to protein folding and long-range interactions, the actual structural alterations with functional implications may be occurring at a distance from the mutation site. As a result, the identification of the most sensitive structural regions for protein function may be hampered by the presence of mutations that indirectly affect the critical regions from a distance. In this study, we illustrate how the use of AlphaMissense predictions to train an XGBoost regression model on structural features extracted from the structures of protein variants predicted by OmegaFold enables the definition of a new explainability metric: a residue-based importance score that highlights the most critical structural domains within a protein sequence. To verify the accuracy of our approach, we applied it to the extensively studied protein DAX-1 and successfully identified critical structural domains. Notably, as this score only requires knowledge of the protein's amino acid sequence, it is valuable in guiding experimental investigations aimed at discovering functionally crucial regions in proteins that have been poorly characterized.