SHINE: protein language model-based pathogenicity prediction for short inframe insertion and deletion variants

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate prediction of variant pathogenicity is crucial for genetic disease studies.
• Inframe indels pose interpretation challenges due to limited training data.
• Existing methods rely on manually encoded features, limiting predictive power.

Purpose of the Study:

• To develop a novel pathogenicity predictor for short inframe indels.
• To improve the interpretation of inframe indel variants in human diseases.
• To leverage deep learning and protein language models for enhanced variant prediction.

Main Methods:

• Developed SHINE (SHort Inframe iNsertion and dEletion) predictor.
• Utilized pretrained protein language models for latent representation of indels and protein context.
• Employed supervised machine learning models with curated training data from ClinVar and gnomAD.

Main Results:

• SHINE demonstrated superior prediction performance compared to existing methods.
• The predictor showed improved accuracy for both deletion and insertion variants.
• Performance was validated on two independent test datasets.

Conclusions:

• Unsupervised protein language models offer valuable insights into protein information.
• SHINE represents a significant advancement in variant interpretation for genetic analyses.
• The developed method enhances the prediction of inframe indel pathogenicity.