bpAI-TAC

Sequence-to-activity (S2A) models learn the cis-regulatory grammar of the genome by predicting measured chromatin activity directly from DNA sequence. Most are trained against a single scalar per regulatory region — the total accessibility of a peak — which discards the fine structure that assays like ATAC-seq actually produce. bpAI-TAC asks whether supervising on that discarded structure makes the model better at the coarse task it was built for. It predicts, at single-base resolution, the distribution of Tn5 transposase insertion sites across an open chromatin region and the region's total accessibility, jointly, across 90 mouse immune cell types.

Higher output resolution turns out to improve the lower-resolution predictions. Adding the base-pair profile objective consistently sharpens predictions of differential accessibility between cell types, and attribution analysis shows the profile-trained model recovers regulatory motifs the accessibility-only model misses. Profile data therefore acts less as a separate prediction target than as a supervisory signal constraining what the network learns about transcription factor binding.

bpAI-TAC was developed by Nuria Alina Chandra, Yan Hu, Jason D. Buenrostro, Sara Mostafavi, and Alexander Sasse at the University of Washington, the Broad Institute, Harvard University, and Heidelberg University. It extends AI-TAC, the same groups' earlier multi-task model of immune chromatin accessibility, and was posted as a preprint in January 2025 before publication in Bioinformatics Advances in 2026. Where genome-scale S2A models such as Enformer, Basenji2, and Borzoi trade resolution for receptive field, bpAI-TAC takes the opposite position within a single peak.

#Key Features

• Base-pair resolution profile head: Outputs a softmax distribution over Tn5 insertion positions in a 250 bp central window for each of 90 cell types, capturing where within a peak a transcription factor footprint sits rather than only how open the peak is.
• Joint profile and accessibility prediction: A second head predicts total accessibility per cell type; the two are multiplied to give per-base insertion counts trained under a Poisson negative log-likelihood loss.
• Multi-task across immune lineages: Training one model over all 90 cell types consistently outperforms an ensemble of single-task models, exploiting shared regulatory grammar across related lineages.
• Explicit Tn5 bias handling: Insertion profiles can be corrected using logits from a separately trained enzyme bias model, with four bias strategies benchmarked.
• Attribution-based motif discovery: Systematic comparison of sequence attributions against the accessibility-only baseline isolates motifs whose contributions emerge only when profile data is modeled.

#Technical Details

bpAI-TAC is a convolutional network that takes a 998 bp one-hot encoded sequence and passes it through an initial width-25 convolution followed by nine dilated residual blocks of 300 filters each (width 3, dilation doubling from 2). The profile head applies a single width-25 convolution with 90 output filters and a softmax over the 250 bp center; the accessibility head applies three rounds of width-5 max pooling with width-3 convolutions before a fully connected layer emitting 90 cell-type values. Training uses Adam at a learning rate of 1e-3 with 1,000 linear warmup steps and a batch size of 20.

Training data comes from the Immunological Genome Project ATAC-seq atlas (Yoshida et al. 2019, GEO GSE100738): 518,845 mm10 open chromatin regions across 90 sorted immune populations sequenced at 2.5–40 million fragments each, with chromosomes 12 and 15 held out for validation and 11 and 16 for testing. On held-out regions bpAI-TAC reaches a mean Pearson correlation of 0.70 for accessibility, rising to 0.804 for the best cell type. On cell-type-specific accessible regions it achieves 0.62 versus 0.51 for the accessibility-only baseline, and mean correlation for differential accessibility across 4,005 cell-type pairs improves from 0.46 to 0.58 (0.61 to 0.69 across 45 lineage pairs). Ablations training at 5, 10, and 20 bp output resolution all underperform the single-base model. Across 1,082 high-confidence regions, 23 motif clusters are significantly enriched in bpAI-TAC attributions relative to the baseline (Fisher exact test, Benjamini-Hochberg FDR < 0.05).

#Applications

bpAI-TAC serves immunologists and regulatory genomicists working with mouse immune chromatin. Its per-base output localizes putative transcription factor binding sites within open chromatin regions, and its attribution maps support de novo motif discovery and in-silico scoring of how sequence changes alter accessibility in specific lineages — useful for prioritizing noncoding variants and dissecting the regulatory programs that distinguish immune cell states. The resolution and multi-task findings also serve as design guidance for anyone building an S2A model on a profiling assay.

#Impact

The contribution is methodological rather than a scale record. By showing that a higher-resolution auxiliary target improves coarse-grained predictions and surfaces motifs otherwise invisible to attribution analysis, bpAI-TAC argues that S2A model quality is limited as much by output resolution as by receptive field or parameter count. The limits are clear: it is trained on one mouse immune atlas, covers a 998 bp local context that cannot capture distal enhancers, and is distributed as research code with checkpoints in the repository rather than as a packaged foundation model. The code carries an MIT badge in its README but no license file.