Abstract
Predicting how residue variations affect protein stability is crucial for rational protein design and for assessing the impact of disease-related mutations. Recent advances in protein language models have revolutionized computational protein analysis, enabling more accurate predictions of mutational effects. However, balancing predictive accuracy with the fundamental laws of thermodynamics remains a challenge for sequence-based models. Here we show JanusDDG, a physics-informed neural network that leverages embeddings from protein language models and a bidirectional cross-attention transformer architecture to predict stability changes for both single and multiple residue mutations. By adopting a physics-informed paradigm, the model is explicitly constrained to satisfy fundamental thermodynamic principles, such as antisymmetry and transitivity, while maintaining high predictive performance. Instead of conventional self-attention, JanusDDG employs a cross-interleaved attention mechanism that computes the relationship between wild-type and mutant embeddings to capture mutation-induced perturbations while preserving essential contextual information. Our results demonstrate that JanusDDG achieves state-of-the-art performance in predicting stability changes from sequence alone, matching or exceeding the accuracy of structure-based methods for both single and multiple mutations.