UMass releases BATv1: 46k-parameter models spanning 3,429 TF ChIP-seq experiments

The Programmable Genomics lab at UMass Chan Medical School (@jmschreiber91) has released the Cherimoya Binding Atlas (BATv1): a collection of genome sequence-to-function (S2F) models comprising 17,145 Cherimoya models that cover 3,429 transcription factor binding ChIP-seq experiments from the ENCODE project, spanning 183 biosamples and 1,139 targets. @finchshenq called it the headline release of the week.

Confirmed

• BATv1 comprises 17,145 models covering 3,429 TF binding ChIP-seq experiments, 183 biosamples, and 1,139 targets, with data from ENCODE.
• The models are extremely small: a single BATv1 model has roughly 46K parameters, compared with about 614K for the CATv1 accessibility model and around 114K for BPNet.
• The authors report that in preliminary comparisons BATv1 is typically more accurate than BPNet, despite having roughly an order of magnitude fewer parameters.
• The models are open source.
• In response to concerns that Enformer often confuses transcription factor motifs with open-chromatin GC bias, the authors noted that BATv1 uses exactly the same training data and loci as the BPNet family (with only minor differences in sampling ratios), which avoids the GC bias issue.

Why it matters

• Higher accuracy with an order of magnitude fewer parameters challenges the "bigger is better" intuition for genomic S2F models, and makes large-scale deployment and reuse far easier.
• The systematic atlas covering 3,429 experiments provides a unified, batch-usable resource for transcription factor binding prediction.

2026-10-31 ~ 2026-10-31 · 5 related posts

Full story(3 episodes)→

• Episode 1: UMass releases BATv1: 46k-parameter models spanning 3,429 TF ChIP-seq experiments(2026-10-31, 5 posts)
• Episode 2: Researcher uses TF binding models to interpret variants and QC designed DNA elements(2026-10-31, 7 posts)
• Episode 3: Stanford researcher explains how genomic models correct chromatin confounders(2026-10-31, 2 posts)

• [source] BATv1: 17,145 Cherimoya models map TF binding across 3,429 ENCODE ChIP-seq experiments — jmschreiber91 · 2026-10-31
• Cherimoya Binding Atlas BATv1: 46k-param genomic models for 3,429 TF binding experiments — jmschreiber91 · 2026-10-31
• [source] BATv1: 46k-Parameter Genomics Models Beat 114k-Parameter BPNet on Accuracy — jmschreiber91 · 2026-10-31
• BATv1 genomics model trained on same data as BPNet, addressing GC-bias confounds — jmschreiber91 · 2026-10-31

1 near-duplicate retellings: finch_shenq