Predicting the Disease Risk of Protein Mutation Sequences With Pre-training Model

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Missense mutations in tumor suppressor genes (e.g., BRCA1, PTEN) can cause protein dysfunction and increase disease risk.
• Accurate identification of disease-associated missense mutations is crucial for risk assessment and potential therapeutic strategies.

Purpose of the Study:

• To develop a hybrid framework, BertVS, for predicting the disease risk associated with protein missense mutations.
• To leverage deep learning and biochemical properties for enhanced variant classification.

Main Methods:

• Pre-training BERT models on the Pfam database to capture protein sequence context.
• Integrating amino acid hydrophilic properties with sequence representations.
• Concatenating learned representations and classifying missense mutations using a classifier.

Main Results:

• The pre-trained BERT model achieved 0.984 accuracy on masked language model prediction.
• BertVS demonstrated superior performance on BRCA1 and PTEN datasets, achieving 0.920 AUROC and 0.915 AUPR.
• The framework effectively classified known clinical variants in the ClinVar dataset.

Conclusions:

• BertVS successfully learns functional information from protein sequences.
• The proposed method effectively predicts disease risk for missense variants, including those with uncertain clinical significance.
• BertVS offers a promising tool for variant interpretation in cancer genomics.