DyVarMap: Integrating Conformational Dynamics and Interpretable Machine Learning for Cancer-Associated Missense...

Area of Science:

• Computational biology
• Structural bioinformatics
• Precision oncology

Background:

• Interpreting missense variants in cancer genes is difficult, with many classified as variants of uncertain significance (VUS).
• Receptor tyrosine kinases like FGFR2 have functions dependent on conformational dynamics, complicating variant analysis.
• Existing sequence-based predictors often lack mechanistic explanations for variant effects.

Purpose of the Study:

• To develop DyVarMap, an interpretable structural-learning framework for predicting the pathogenicity of cancer-associated gene variants.
• To integrate conformational dynamics into variant effect prediction for improved accuracy and mechanistic understanding.
• To provide testable hypotheses for experimental validation in precision oncology.

Main Methods:

• DyVarMap integrates AlphaFold2-based ensemble generation with physics-driven refinement and manifold learning.
• A supervised classification model uses five biophysically motivated geometric features.
• SHAP analysis provides mechanistic attributions for variant pathogenicity predictions.

Main Results:

• DyVarMap successfully classified pathogenicity for FGFR2 variants, generating diverse conformational ensembles and identifying metastable states.
• External validation on ten kinase-domain variants achieved an AUROC of 0.77 with superior calibration compared to PolyPhen-2 and AlphaMissense.
• Feature importance analysis highlighted K659-E565 salt-bridge distance and DFG motif dihedral angles as key predictors, linking predictions to known activation mechanisms.

Conclusions:

• DyVarMap effectively bridges the gap between static structure prediction and dynamics-aware functional assessment.
• The framework provides structurally coherent mechanistic explanations for variant effects, aiding in precision oncology.
• Incorporating conformational dynamics into variant effect prediction offers significant value for clinical applications and experimental validation.