PromoterAI is a deep learning model from Illumina that predicts the impact of single nucleotide variants in gene promoter regions. It scores all possible substitutions within 500 bp of annotated transcription start sites (TSS), covering approximately 39.5 million genomic positions across all protein-coding genes.
Scores range from -1 to 1. Positive scores indicate predicted disruption of promoter function, negative scores indicate the variant is predicted to be tolerated. The model was trained using primate conservation and promoter sequence features, similar in approach to the related PrimateAI-3D model for coding variants.
This track is a composite with four bigWig subtracks, one for each possible alternate allele (A, C, G, T). When zoomed in, the exact score for each possible mutation is shown on mouseover. When zoomed out, the display shows an average across the visible window; this average is indicated by a "~" prefix in the mouseover.
A fifth subtrack ("PromoterAI overlaps") shows positions where overlapping transcripts produce different scores for the same variant. At these positions, the bigWig shows the score with the largest absolute value, while the overlap track shows all per-transcript scores. About 3.8% of positions have overlapping transcripts with differing scores. The track shows the list of transcripts and scores for these positions. Of these, for more than 60% of these positions, the difference is smaller than 0.01, which is why we added a filter, active per default, that hides all annotations in this track where the difference is smaller than this cutoff. The filter can be switched off on the track configuration page.
Due to the data license, this track is not available for bulk download from UCSC. The source data can be downloaded from the PromoterAI GitHub page.
The PromoterAI hg38 TSS-500 file was downloaded. The file contains pre-computed scores for all possible single nucleotide substitutions within 500 bp of annotated TSS positions. For positions covered by multiple transcripts, the score with the largest absolute value was used for the bigWig tracks. Positions where transcripts produced different scores (4.45M of 118.6M unique variants, 3.8%) were additionally written to a bigBed overlap track with per-transcript detail. A conversion script is available from our Github.
Thanks to Illumina for making PromoterAI predictions publicly available.
Gao H, Hamp T, Ede J, Schraiber JG, McRae J, Singer-Berk M, Yang Y, Dietrich ASD, Fiziev PP, Kuderna LFK et al. The landscape of tolerated genetic variation in humans and primates. Science. 2023 Jun 2;380(6648):eabn8197. PMID: 37262156; PMC: PMC10187174