Causal Prediction of TP53 Variant Pathogenicity Using a Perturbation-Informed Protein Language Model

Area of Science:

• Genomics
• Bioinformatics
• Cancer Research

Background:

• Predicting variant functional impact is vital for disease understanding, especially for cancer genes like TP53.
• Current variant effect prediction (VEP) models struggle with missense mutations due to their non-gene-specific nature.

Purpose of the Study:

• To develop CaVepP53, a TP53-specific VEP model.
• To improve the accuracy and interpretability of missense mutation classification for TP53.

Main Methods:

• Fine-tuned a TP53-specific predictor using perturbation-based experimental variants.
• Integrated protein language models with experimental functional data.
• Quantified pathogenicity using Euclidean distances and logistic transformation for confidence scores.

Main Results:

• CaVepP53 significantly outperformed general VEP models (AlphaMissense, PrimateAI-3D) in accuracy, precision, and F1-score.
• Experimental validation confirmed CaVepP53's robustness, identifying novel functional variants.
• The framework was extended to other cancer-related genes (VHL, ATM, BRCA1, RAD51C, BAP1).

Conclusions:

• CaVepP53 provides accurate and interpretable VEP for TP53, overcoming limitations of existing models.
• The gene-specific VEP framework shows potential for precision medicine applications.
• This approach enhances understanding of cancer-related gene mutations.