Synthetic MPRA perturbation supervision improves variant-effect prediction in sequence-to-function models

Fine-tuning Borzoi on synthetic MPRA variant-effect labels across 1,000 K562 genes increased Pearson correlation from 0.28 to 0.54 on CAGI5 promoter variants.

• Why it matters: S2F models trained mainly on reference sequence struggle to predict how variants change gene expression, and saturation-mutagenesis MPRA supervision is scarce.
• What they did: An ensemble of 16 models was trained on 110,359 K562 MPRA measurements, distilled into a student model using MPRA elements linked to 1,000 genes, then used to generate 600 variants per gene for synthetic labels.
• The result: Compared with pre-trained Borzoi, fine-tuning on 600,000 synthetic samples improved correlation to 0.54 on CAGI5 promoter variants and from 0.39 to 0.46 on 903 GTEx fine-mapped whole-blood eQTLs, while preserving performance across all 7,611 output tracks.

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