An augmented transformer model trained on family specific variant data leads to improved prediction of variants of uncertain significance

Abstract Variants of uncertain significance (VUS) represent variants that lack sufficient evidence to be confidently associated to a disease, thus posing a challenge in the interpretation of genetic testing results. In this work, we present an improved gene specific approach to variant prediction that leverages a pre-trained protein language model for predicting VUS. Our deep learning model combines zero-shot log odd scores from evolutionary scale model (ESM-2) as a feature along with embeddings from ESM-2 as features for training a supervised model on variants associated with the gene. Our training set creation approach uses variant data from a gene family if the gene of interest has low or no functional data for training a gene specific predictor. We demonstrated the accuracy of our method by testing it on VUS of an enzyme Alpha-N-acetylglucosaminidase (NAGLU) whose deficiency due to mutations is known to cause a rare genetic disorder, Mucopolysaccharidosis IIIB or Sanfillipo B disease. Our model augmented with contextual information from the gene family improved prediction of VUS in the NAGLU gene and outperformed state-of-the-art pathogenicity predictors. Our results also indicate that for genes with sparse or no experimental variant impact data, the family variant data can serve as proxy training data for making accurate predictions.