Enhancing protein structure predictions: DeepSHAP as a tool for understanding AlphaFold2
Sabbir Ahmed Sibli, Vlasios Panagiotis Panagiotou, Christos Makris
Abstract
The AlphaFold2 introduces an entirely new era in the field of computational biology, achieving phenomenal results in protein structure predictions. Despite its remarkable performance, the model’s prediction pipeline is highly complex, and the Deep Learning model itself is a black box, leaving researchers uncertain about how specific predictions are made. Explainable AI (XAI) offers a promising avenue for improving the transparency and trustworthiness of such models, making their predictions understandable and trustworthy. This research presents how DeepSHAP, an advanced XAI method, can be applied to explain AlphaFold2’s predictions through the OpenFold and ColabFold frameworks. DeepSHAP is designed to integrate Shapley Additive exPlanations (SHAP) values and Deep Learning Important Features (DeepLIFT) with a complex deep learning model to interpret the contribution of individual input features for a specific prediction. This study depicts that certain amino acids play key roles in the formation of distinct protein structures. Specifically, by calculating SHAP values and integrating them with DeepLIFT, it investigates several protein structures and identifies key amino acids that are most influential in AlphaFold2’s prediction process. The results demonstrate the promising potential of this approach to enhance interpretability and provide insights for the transparency of AI-based protein structure prediction models.