Unsupervised prioritization of gene-regulatory variants with isolation forests

eQIF, an isolation-forest method integrating Enformer outputs, produces a single gene-regulatory score per common variant and predicts eQTLs strongly.

• Why it matters: LD and non-coding architecture make causal variant identification difficult, and existing supervised or deep-learning prioritizers can yield many scores per variant.
• What they did: The approach aggregates Enformer predictions into one unsupervised gene-regulatory score per variant using isolation forests, and it was evaluated for eQTL and gene-regulatory GWAS variant prediction.
• The result: Top-scoring variants fall in evolutionarily conserved TFBS footprints, and their enrichment across overrepresented TFBS motifs matches motif information content; scores are precomputed for 9,431,969 common variants.

AI summary of the abstract · may contain errors · not medical advice · read the paper.