Spectral Interpretability for Molecular Potentials Learning with Graph Neural Networks

The SIMPL project (Spectral Interpretability for Molecular Interaction PotentiaLs) is designed to enhance the interpretability and efficiency of Graph Neural Networks (GNNs) used to model Molecular Interaction Potentials (MIPs). GNNs have greatly advanced the prediction of molecular interactions, which is crucial for fields like drug discovery and material science. However, the complexity of these models often makes it challenging to identify which molecular interactions are most critical, limiting their practical application. SIMPL addresses this challenge by incorporating Spectral Graph Theory and Algebraic Topology into GNNs, allowing for clearer interpretation and streamlined models, focusing on the most relevant interactions without sacrificing accuracy. Compared to traditional gradient-based methods, which can be unstable and sensitive to specific molecular conformations, SIMPL offers a robust approach. Instead of a localized insights that could vary significantly depending on the particular state of the molecule or be subject to vanishing or exploding gradients problems (especially in complex models), SIMPL provides a global perspective by assessing the relevance of interactions across the entire learning task. This ensure that the identified interactions are consistently important and reliable. This approach in practice forces a multi-body expansion of the MIP leading to a comprehensive understanding of molecular systems.While SIMPL has broad applications, it will be applied to the study of immune recognition processes, such as peptide-Major Histocompatibility Complex (MHC) binding. The framework’s ability to reliably identify critical molecular interactions could significantly enhance our understanding of these processes, leading to more informed approaches in immunotherapy and vaccine development.