Fig. 4: Prioritizing functional genetic variants.
From: Nucleotide Transformer: building and evaluating robust foundation models for human genomics
a, Overview of the NT application of zero-shot predictions. b, Proportion of variant consequence terms across deciles based on the cosine similarity metric across models. The consequence terms are shown in order of severity (less severe to more severe) as estimated by Ensembl. c, Comparison of zero-shot predictions for prioritizing functional variants based on different distance metrics. d, Comparison of fine-tuned models and available methods for prioritizing functional variants based on GRASP eQTLs and meQTLs, ClinVar and HGMD annotated mutations. Model performance is measured with the ROC AUC. The AUC for the three best-performing models is shown.
