Excited to share my first contribution here at Illumina! We developed PromoterAI, a deep neural network that accurately identifies non-coding promoter variants that disrupt gene expression.🧵 (1/)

Acknowledging that all benchmarks and related comments are based on expression effects of promoter variants. Should’ve clarified that upfront — my apologies!

Huge thanks to Illumina, Kyle Farh, Nicole Ersaro, Gherman Novakovsky, and the entire team behind PromoterAI for making this happen. Full paper: www.science.org/doi/10.1126/...
@science.org (10/)

The PromoterAI source code is available at github.com/Illumina/Pro.... Precomputed scores for all promoter SNVs are freely available for academic and non-commercial research. (9/)

In the @genomicsengland.bsky.social cohort, variants prioritized by PromoterAI are enriched in clinically relevant genes. These variants account for 6% of the genetic burden and, when combined with SpliceAI and PrimateAI-3D, match the genetic burden of protein-truncating variants. (8/)

PromoterAI effectively identifies pathogenic ClinVar variants that disrupt diverse regulatory motifs — often conserved and supported by ChIP-seq evidence. (7/)

In the @ukbiobank.bsky.social cohort, PromoterAI predictions correlate strongly with protein levels and quantitative traits, suggesting that promoter variants contribute meaningfully to phenotypic variation. (6/)

PromoterAI’s internal representations reveal three promoter categories: ubiquitously active, bivalent chromatin, and enhancer-like. The enhancer-like category, enriched for TATA boxes, may represent enhancers co-opted as promoters. (5/)

PromoterAI achieves the best performance across diverse benchmarks spanning RNA, proteins, QTLs, and MPRA. Basenji2 < Enformer < Borzoi @drkbio.bsky.social. ChromBPNet @anshulkundaje.bsky.social (trained on accessibility) does well in MPRA but not in tissues. Evo2 lags far behind. (4/)

Fine-tuning was done using twin networks that contrasted observed expression between ref and alt alleles, enabling the model to attribute differences to the variant rather than unrelated confounders. This was key to generalizing across unseen genes and datasets. (3/)

We first trained PromoterAI to predict epigenetic and expression profiles at nucleotide resolution, then fine-tuned it on rare promoter variants associated with aberrant expression. (2/)

We're thrilled to introduce PromoterAI — a tool for accurately identifying promoter variants that impact gene expression. 🧵 (1/)