Systematic promoter perturbation reveals limitations of AlphaGenome gene expression predictions

Abstract

Sequence-to-function (S2F) models promise to facilitate the interpretation of noncoding variants by predicting their effects on gene regulation directly from genomic sequence. However, whether these models recapitulate fundamental principles of transcriptional biology remains unclear. Here, we tested whether AlphaGenome recognizes the expected consequences of promoter disruption on predicted gene expression. We systematically perturbed annotated promoter elements in silico for 2,111 promoter-gene pairs using dinucleotide shuffling, GC-adjusted random base substitution, and N-masking. We then compared AlphaGenome-predicted CAGE and RNA-sequencing signals across 20 tissues for reference and promoter-disrupted sequences. While CAGE and RNA-sequencing predictions were strongly correlated for reference sequences, promoter disruption markedly reduced predicted CAGE activity, with N masking producing reductions of approximately three orders of magnitude. By contrast, the predicted RNA-sequencing signals were substantially less affected, resulting in an uncoupling of two molecular readouts of gene expression. Exon-level analyses showed that AlphaGenome predicted reduced RNA-sequencing signal when exonic sequences are within the disrupted region, but frequently failed to propagate this effect to downstream exons. This divergence between the CAGE and RNA-sequencing readouts was apparent from the second exon onward and was more pronounced in longer genes. Although alternative transcription initiation and shared downstream exons may explain individual cases, it is unlikely to account for the systematic discordance between CAGE and RNA-sequencing predictions. These findings indicate that AlphaGenome can recognize the local loss of promoter activity without consistently capturing its consequences across the corresponding transcript. More broadly, our results show that predictions from individual S2F output tracks do not necessarily reflect a coherent model of transcription. Systematic biological benchmarking and experimental validation therefore remain essential before such predictions can be used to prioritize or classify noncoding variants in research or diagnostic settings.