Systematic evaluation of predictors for binding free energy changes upon mutations in protein complexes

Area of Science:

• Computational Biology
• Biophysics
• Protein Engineering

Background:

• Accurate prediction of binding free energy changes (ΔΔG) from mutations is critical for disease mechanism studies and therapeutic antibody design.
• Missense mutations disrupting molecular interactions account for ~60% of pathogenic mutations.
• Existing ΔΔG predictors show benchmark performance but lack generalization and adaptability for complex mutation scenarios.

Purpose of the Study:

• To systematically evaluate the performance and limitations of mainstream ΔΔG predictors.
• To assess predictor generalization, metric alignment, and adaptability across diverse mutation types and protein microenvironments.
• To provide insights for improving ΔΔG prediction methods.

Main Methods:

• Assessed eight leading ΔΔG predictors (physical energy function-based and machine learning-based).
• Constructed an independent evaluation dataset.
• Employed multi-dimensional metrics including regression accuracy and classification capability.
• Analyzed performance variations across mutation types, stability categories, and local protein structures.

Main Results:

• >60% of evaluated predictors (5/8) exhibit a bias towards overestimating mutational instability.
• Predictors show limited recall (<0.1) for identifying stabilizing mutations (ΔΔG < -0.5 kcal/mol).
• Predictive performance is significantly influenced by the local protein structure at the mutation site.

Conclusions:

• Current ΔΔG predictors have limitations in generalization and adaptability, particularly for complex mutation scenarios.
• There is a need for balanced, standardized datasets and advanced algorithms (e.g., local-global fusion) for future breakthroughs.
• Findings offer a reference for optimizing and practically applying ΔΔG prediction tools in molecular biology and drug design.