Sequential regulatory activity prediction across chromosomes with convolutional neural networks

Basenji predicts diverse epigenetic and transcriptional profiles from DNA sequence. (A) The AKT2 locus exemplifies the genome-wide accuracy of Basenji predictions; gene promoters and the strongest distal regulatory elements are easily identified, with some false-positive and -negative predictions for weaker elements. For each track, the darker version on top represents the experimental coverage, and the lighter version below represents Basenji predictions. (B) We computed the variance explained (R²) for each experiment and plot here the distributions classified by data set type. Basenji predicts punctate peak data, but broad chromatin marks remain challenging. (C) For the median accuracy DNase-seq experiment, mobilized CD34 cells, we plotted the log₂ predictions versus log₂ experiment coverage in 128-bp bins. (D) For all replicated experiments, we plotted log–log Pearson correlation between the replicate experiments versus the correlation between the experiment and its replicate's prediction (averaged across replicates). Both the mean and median Basenji prediction accuracy exceed the replicate accuracy.