Systematic promoter perturbation reveals limitations of AlphaGenome gene expression predictions

AlphaGenome’s promoter disruption tests show predicted CAGE activity drops strongly (down to ~three orders of magnitude with N masking), while predicted RNA-seq signals change far less across 20 tissues.

• Why it matters: It is unclear whether sequence-to-function models capture core transcriptional consequences of promoter element loss, which is critical for interpreting noncoding variants.
• What they did: The study in silico perturbed annotated promoter elements for 2,111 promoter–gene pairs using dinucleotide shuffling, GC-adjusted substitutions, and N-masking, then compared AlphaGenome-predicted CAGE and RNA-seq signals across 20 tissues.
• The result: CAGE and RNA-seq predictions were strongly correlated for reference sequences, but promoter disruption uncoupled them; exon-level effects on RNA-seq often failed to propagate beyond the disrupted region, especially in longer genes.

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