Dynamics-informed multigraph neural networks for protein thermostability prediction and residue-level interpretation

Thermostability is crucial for protein engineering, but experimental determination of melting temperatures is costly and time-consuming. Most existing machine learning approaches for melting temperature prediction rely on sequential and structural features while largely overlooking protein dynamics, a key determinant of conformational stability. Here, we introduce three dynamics-informed graphs and evaluate their utility relative to sequential or structural representations. Specifically, we propose three dynamical graphs derived from normal mode analysis-co-directionality, coordination, and deformation graphs-and assess whether they can serve as alternatives to contact graphs in predicting melting temperatures. We then integrate sequential, structural, and dynamical information within a unified multigraph learning framework. Our results show that dynamical graphs achieve comparable predictive performance to conventional contact graphs, and that combining structural and dynamical graphs yields consistent, albeit modest, improvements over contact-only models. Furthermore, Laplacian centrality analysis on coordination graphs reveals enrichment tendencies and mechanical signals, providing interpretability. Overall, this work demonstrates the value of protein dynamics-informed multigraph representations for learning protein properties.