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

Area of Science:

• Biophysics
• Computational Biology
• Protein Engineering

Background:

• Protein thermostability is vital for engineering applications, but experimental melting temperature determination is resource-intensive.
• Current machine learning models for melting temperature prediction often neglect protein dynamics, a critical factor in conformational stability.

Purpose of the Study:

• To introduce and evaluate dynamics-informed graphs for predicting protein melting temperatures.
• To compare the performance of dynamical graphs against traditional sequential and structural representations.
• To integrate diverse data types into a unified multigraph learning framework.

Main Methods:

• Development of three dynamical graphs (co-directionality, coordination, deformation) using normal mode analysis.
• Assessment of dynamical graphs as alternatives to contact graphs for melting temperature prediction.
• Implementation of a multigraph learning framework combining sequential, structural, and dynamical features.
• Application of Laplacian centrality analysis on coordination graphs for interpretability.

Main Results:

• Dynamical graphs demonstrate predictive performance comparable to conventional contact graphs.
• Integrating structural and dynamical graphs provides modest but consistent improvements over contact-only models.
• Laplacian centrality analysis on coordination graphs reveals key mechanical signals and enrichment tendencies.

Conclusions:

• Protein dynamics-informed multigraph representations offer valuable insights for predicting protein properties like thermostability.
• Dynamical graphs serve as effective alternatives or complements to structural graphs in machine learning models.
• The study highlights the potential of incorporating dynamic information for enhanced protein engineering applications.