Predicting functional effect of missense variants using graph attention neural networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate prediction of damaging missense variants is crucial for genome interpretation.
• Existing computational methods have limitations in predicting variant pathogenicity.
• Machine learning and large-scale genomic data offer opportunities for improved predictions.

Purpose of the Study:

• To introduce gMVP, a novel method for predicting missense variant pathogenicity.
• To leverage graph attention neural networks and co-evolutionary data for enhanced prediction.
• To improve the interpretation of missense variants in clinical and research settings.

Main Methods:

• Developed gMVP, a method utilizing graph attention neural networks.
• Constructed a graph where nodes represent amino acid features and edges represent co-evolution strength.
• Integrated local protein context and distal correlated positions for information pooling.

Main Results:

• gMVP outperformed existing methods in identifying damaging variants in TP53, PTEN, BRCA1, and MSH2 using deep mutational scan data.
• gMVP achieved superior separation of de novo missense variants in neurodevelopmental disorder cases versus controls.
• The model demonstrated successful transfer learning for gain- and loss-of-function predictions in ion channels.

Conclusions:

• gMVP represents a significant advancement in predicting missense variant pathogenicity.
• The method enhances the interpretation of missense variants for clinical genetic testing.
• gMVP shows promise for improving genetic studies and understanding disease mechanisms.