🧬 BEND Zero-shot prediction of disease variants from ClinVar

Predicting variant effects is a binary problem, where single-bp mutations are classified as either having an effect or not. Each variant is a genomic position with a mutation $x ∈ {A, C, G, T}$ and a label $y ∈ {0, 1}$ indicating whether it is pathogenic or benign. The adjacent 512 bp serve as context.

As this is a zero-shot task, we used the cosine distance in embedding space between a variant nucleotide and its reference nucleotide as the prediction score in BEND.