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

Accurate prediction of variant functional impact is crucial for understanding human diseases, particularly for cancer-related genes such as TP53. Advances in high-throughput mutational assays have enhanced variant effect prediction (VEP), but missense classification remains challenging due to the limitations of broad, non-gene-specific models. Here we present CaVepP53, a TP53-specific predictor fine-tuned on perturbation-based experimental variants. The model not only classifies mutations but also quantifies their pathogenicity by calculating Euclidean distances between the wild-type and mutant embeddings and deriving confidence scores through logistic transformation. Benchmarking demonstrates that CaVepP53 significantly outperforms general-purpose models, such as AlphaMissense (AM) and PrimateAI-3D, achieving higher accuracy, precision, and F1-score in predicting pathogenic mutations. Competitive growth assay validation of 22 mutations further confirms CaVepP53's robustness, including 7 functional novel variants absent in the ClinVar database. Thus, by integrating protein language models with experimentally validated functional data, our approach enables accurate, interpretable VEP for TP53, overcoming limitations of predictors trained solely on evolutionary or clinical associations. We further extended this framework to five additional cancer-related genes (VHL, ATM, BRCA1, RAD51C, and BAP1), establishing a generalizable framework for gene-specific VEP with potential applications in precision medicine.