#20 · Genomics

Interpretable sequence models link DNA motifs to regulatory activity across species and…

Three preprints train convolutional or related interpretable models to predict regulatory readouts from local DNA sequence, and the learned features are examined as motifs. One builds an atlas from over 200 human transcription initiation profiling experiments. CNN-CR learns a compact set of shared promoter motif families predicting expression across 116 human biosamples, with cell-type-specific interpretation. A Heliconiini study uses re-analyzed wing ATAC-seq and histone ChIP data to train an enhancer predictor (AUROC 0.909) applied across 34 genomes. The readouts and species differ, so the shared element is methodological. The Cell Reports study quantifies enhancer RNA from GTEx and reports sex-biased enhancer activity (47% across 41 tissues), but it is an expression-based analysis rather than a sequence model. The new emphasis is on shared, compact sequence representations that are read differently across contexts.