Lastly we investigate all human promoters by saturation mutagenesis, identifying canonical promoter TFs and linking non-coding variant effect size to coding constraint (LoEUF), bridging the gap between coding and non-coding function.

MPAC can scale to predict 514M gnomAD variant effects and we quantify the relationship between allele frequency or evolutionary conservation with predicted skew at an unprecedented level. Notably, we find that variants causing high skew are under greater constraint than those with small effects.

In COSMIC we identify known non-coding driver mutations (TERT) and by combining variant recurrence, regulatory element annotations, and cancer-associated promoters we nominate 1,892 emVars as putative non-coding drivers.

Many clinically identified non-coding variants lack clear effects, using MPAC we can predict the impact of all ClinVar non-coding variants and observe enrichments in pathogenic alleles for highly disruptive variants (emVars).

MPAC predictions distinguish causal variants from the UK Biobank, Biobank Japan and eQTLs from GTEx with experimental accuracy but without experimental overhead!

Trained on MPRA data from a large-scale study of human trait and eQTL variants (www.biorxiv.org/content/10.1...) and extending the Malinois model architecture (www.nature.com/articles/s41...) MPAC predicts variant effects with high accuracy.

Massively Parallel Reporter Assays (MPRAs) quantify the activity of 10-100s of thousands of sequences, however, it is not feasible to test all known variation. Modeling MPRA can increase scale and lead to better understanding of complex traits, somatic and germline diseases, and population genetics.