Utilizing protein structure graph embeddings to predict the pathogenicity of missense variants

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Protein Structure Prediction

Background:

• Genetic variants, particularly missense variants, can alter protein structure and function, posing challenges for pathogenicity prediction.
• Existing bioinformatic tools for variant interpretation often do not directly incorporate protein structural information.
• Accurate prediction of missense variant pathogenicity is crucial for understanding genetic diseases.

Purpose of the Study:

• To develop a machine learning workflow for predicting the pathogenicity of missense variants using protein structure.
• To evaluate the utility of graph embeddings derived from predicted protein structures for pathogenicity prediction.
• To enhance existing pathogenicity prediction scores, such as CADD, using structural information.

Main Methods:

• Utilized the ESMFold protein language model to predict protein structures for missense variants.
• Employed graph autoencoders to generate embeddings from the predicted protein structures.
• Trained a classifier model using these graph embeddings to predict variant pathogenicity and compared performance with existing methods.

Main Results:

• Demonstrated that graph embeddings from predicted protein structures are effective for pathogenicity prediction.
• Showed that incorporating these graph embeddings can enhance the performance of the CADD score.
• Explored the impact of different graph embedding abstraction levels and compared embeddings from various protein-folding models.

Conclusions:

• The developed machine learning workflow provides a novel approach to pathogenicity prediction by leveraging predicted protein structures.
• Graph embeddings offer a valuable feature for improving the accuracy of missense variant interpretation.
• This method has the potential to refine our understanding of genetic variant effects on protein function and human health.