(9/10) Our BPReveal package provides tools to engineer sequences with desired properties. For example, we designed mutations to alter a nucleosome’s presence in vivo, and our design was corroborated experimentally.

(8/10) These models detect barrier elements in nucleosome occupancy and chromatin organization. Around barriers, motif effects are asymmetric, and the most asymmetric regions align with known domain boundaries.

(7/10) Overcoming enzymatic sequence biases using PISA reveals the nucleosome motif grammar. Upon correcting the AT-rich sequences of a BPNet model trained to predict MNase-seq nucleosome occupancy, resulting attribution scores clearly identify motifs.

(6/10) PISA can detect enzymatic biases in various sequencing data. Sequence-to-function models learn both enzymatic bias of the sequencing experiment and the underlying biology. Here, a BPNet MNase-seq model trained in yeast exhibits expected preference for AT-rich sequences.

(5/10) PISA reveals complex motif effects on histone modification ChIP-seq data. In a H3K27ac BPNet model predicting activity in early embryo fly, a pioneering Zelda motif produces a dual response in H3K27ac profile: a central depletion, flanked by an increase in activity.

(4/10) PISA can reveal “hidden” motifs that possess disparate effects on the output window. In a ChromBPNet model predicting accessibility in early embryo fly, a CA-rich “Cackle” motif is revealed to possess a positive and negative contributions depending on the output locus.

(3/10) PISA can visualize the influence range of sequences, including TF motifs. In a mESC BPNet model predicting Nanog binding, PISA can distinguish the nucleosome-range effects of pioneering motif Oct4-Sox2 versus the binding-range effects of the Nanog motif itself.

(2/10) Our new interpretation tool called PISA (pairwise influence by sequence attribution) overcomes this limitation and quantifies how each individual base impacts the predicted readout at each genomic coordinate.

(1/10) When interpreting sequence-to-function models, current attribution methods typically summarize an input base’s effect on the entire output. But what happens if a single nucleotide causes one effect at one output position, but a different effect elsewhere?

(7/7) Our Model: All promoters use TFIID to load TBP, but TATA promoters additionally allow direct TBP binding to the TATA box. Such dual initiation likely enables faster TBP re-loading and larger transcriptional bursts at TATA promoters. For more details, check out our work!