HERMES: Holographic Equivariant neuRal network model for Mutational Effect and Stability prediction.

Accurately predicting how amino acid substitutions alter protein function is a central challenge in biology, with applications from interpreting disease variants to engineering vaccines and therapeutic proteins. We introduce HERMES, a family of fast, structure-based models that predict mutational effects from the local three-dimensional atomic environment around each residue. Pre-trained on the masked amino-acid prediction task, HERMES shows strong zero-shot performance for predicting changes in thermodynamic stability and protein-protein binding affinity. We find that this pre-training induces a bias toward substitutions with similar size to the wild-type. To address this, we develop an amortized fine-tuning strategy that incorporates packing flexibility, substantially reducing size-based bias while preserving sensitivity to mutational effects. We demonstrate that HERMES can then be fine-tuned on experimental measurements without adding parameters or relying on costly data augmentation, achieving performance competitive with state-of-the-art stability predictors. Finally, we show that HERMES identifies antigen-stabilizing mutations across multiple viral envelope proteins, enabling computationally efficient, structure-guided vaccine design. Together, these results establish HERMES as a practical and accurate framework for structure-based mutational effect prediction.