Enhancing missense variant pathogenicity prediction with protein language models using VariPred

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Genetics

Background:

• Predicting the pathogenicity of genetic variants is crucial for understanding disease mechanisms and clinical impact.
• Traditional methods rely on hand-crafted features, often requiring complex data preprocessing like structural or evolutionary analyses.
• The advent of deep learning and large protein language models offers new avenues for variant pathogenicity prediction.

Purpose of the Study:

• To introduce VariPred, a novel framework for predicting genetic variant pathogenicity.
• To leverage pre-trained protein language models for an end-to-end variant impact prediction.
• To demonstrate that VariPred outperforms existing state-of-the-art methods using only protein sequence data.

Main Methods:

• Developed VariPred, an end-to-end deep learning model utilizing a pre-trained protein language model (ESM-1b).
• Input requirement is limited to the protein sequence, eliminating the need for structural or multiple sequence alignment features.
• Evaluated VariPred's performance on six established variant impact prediction benchmarks.

Main Results:

• VariPred demonstrated comparable or superior performance against established predictors like 3Cnet, Polyphen-2, REVEL, MetaLR, FATHMM, and ESM variant.
• The model achieved robust classification accuracy across multiple benchmarks.
• The simplified input requirement (protein sequence only) streamlines the prediction process.

Conclusions:

• VariPred offers a powerful and efficient new tool for predicting variant pathogenicity.
• The framework highlights the potential of protein language models in genomic variant interpretation.
• This sequence-based approach simplifies pathogenicity prediction, making it more accessible for researchers.