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

Genetic variants can impact the structure of the corresponding protein, which can have detrimental effects on protein function. While the effect of protein-truncating variants is often easier to evaluate, most genetic variants that affect the protein-coding region of the human genome are missense variants. These variants are mostly single nucleotide variants, which result in the exchange of a single amino acid. The effect on protein function of these variants can be challenging to deduce. To aid the interpretation of missense variants, a variety of bioinformatic algorithms have been developed, yet current algorithms rarely directly use the protein structure as a feature to consider. We developed a machine learning workflow that utilizes the protein-language-model ESMFold to predict the protein structure of missense variants, which is subsequently embedded using graph autoencoders. The generated embeddings are used in a classifier model, which predicts pathogenicity. We provide evidence that graph embeddings can be used for pathogenicity prediction and that they can be used to enhance the widely applied CADD score. Additionally, we explored different levels of abstraction of the graph embeddings and their influence on the classifier. Finally, we compare the utility of graph embeddings from different protein-folding models.