MMPatho

MMPatho predicts pathogenicity of missense mutations (MMs) by integrating Ensembl VEP v104 and dbNSFP v4.1a annotations, Ensembl API-derived ENSP protein sequences, protein language model embeddings (ESM-1b, ProtTrans-T5, ProtT5-XL-U50), and XGBoost with SHAP to classify gain-of-function (GOF) and loss-of-function (LOF) variants.

Key Features:

• Benchmark dataset: Large-scale nonredundant missense mutation (MM) benchmark derived from Ensembl with a focused blind test set targeting pathogenic GOF and LOF MMs.
• Feature extraction: Extracts variant-level, amino acid-level, individual outputs, and genome-level features using Ensembl VEP v104 and dbNSFP v4.1a.
• Protein sequence processing: Generates protein sequences via ENSP identifiers using the Ensembl API and encodes sequences for analysis.
• Evolutionary embeddings: Extracts embeddings from mutant sites using protein language models ESM-1b, ProtTrans-T5, and ProtT5-XL-U50 to capture evolutionary and structural information.
• ConsMM component: Uses individual outputs with an XGBoost classifier and SHAP explanation analysis to interpret feature contributions.
• EvoIndMM component: Incorporates evolutionary information from ESM-1b and ProtT5-XL-U50 embeddings to enhance predictive capability and assess mutation impacts.
• Interpretability and scoring: Produces SHAP-based explanations, reliability index scores, and extensive variant annotations.

Scientific Applications:

• Pathogenicity classification: Predicts pathogenicity of missense mutations, including classification of GOF and LOF variants.
• Variant annotation: Supplies extensive variant-level and genome-level annotations to support genetic research.
• Benchmarking and model validation: Provides a nonredundant benchmark and blind test set to support validation and development of computational models in genomics.

Methodology:

Extracts features with Ensembl VEP v104 and dbNSFP v4.1a, generates ENSP protein sequences via the Ensembl API and encodes them, obtains mutant-site embeddings from ESM-1b, ProtTrans-T5, and ProtT5-XL-U50, and trains XGBoost-based ConsMM with SHAP and an EvoIndMM incorporating embeddings; models were compared in experiments on a blind test set excluding overlapping variants and proteins from training data, reporting AUROC values of 0.9836 and 0.9854 and AUPR values of 0.9852 and 0.9902.