CAMBUS

A naive CD4+ T cell carries roughly a quarter of a million accessible chromatin peaks, and most of them tolerate mutation without consequence. The few that do not are exactly the ones a geneticist wants when a disease association lands in non-coding sequence. Cross-species conservation is the usual stand-in for that importance, but it is measured once per genomic position and says nothing about which cell type a region matters in.

CAMBUS (Chromatin Accessibility Mutation Burden Score), developed at the RIKEN Center for Integrative Medical Sciences with collaborators at the University of Tokyo, Shizuoka General Hospital, and the University of Shizuoka, derives that signal from the sequence itself. A convolutional network called CAMBUS ML is trained to predict whether the central 200 bp of a 5 kb DNA window is accessible in each of 29 sorted human immune cell types. The trained, frozen model is then run over every possible single-nucleotide substitution near each open chromatin region (OCR), and the mean predicted change in accessibility within 2 kb of the region becomes its CAMBUS score. Regions whose predicted accessibility collapses under most substitutions are the ones whose sequence is load-bearing.

The approach descends from accessibility-from-sequence convolutional models such as Basset, and from the practice of reading a trained sequence model through in-silico saturated mutagenesis rather than through its raw predictions. Where long-range attention models like Enformer trade local resolution for hundreds of kilobases of context, CAMBUS ML stays local and spends its capacity on cell-type resolution.

#Key Features

• Mutation burden as a constraint metric: Constraint is estimated from a model's sensitivity to hypothetical mutations rather than from observed variation or cross-species alignment, so it applies to regions with no segregating variants and no conserved ortholog.
• Cell-type-resolved scores: Every OCR receives a separate score in each of the 29 immune cell types, turning constraint from a single genome-wide track into a matrix read per lineage.
• Wide input window: The 5 kb input is five times longer than the DeepSEA architecture it is built on, capturing flanking regulatory context around the 200 bp being classified.
• Frozen model, statistical read-out: The trained network is never refit downstream; a local-FDR mixture model over the CAMBUS distribution defines the CAMBUS+ set of constrained OCRs.
• Orthogonal to existing metrics: 59,605 CAMBUS+ regions — 90% of them — are not prioritized by any of seven established constraint metrics, including Gnocchi, GERP, and the phyloP tracks.

#Technical Details

CAMBUS ML is a deep convolutional network built on the DeepSEA architecture and trained with the Selene library (v0.4.7), taking a 5 kb reference sequence and emitting a sigmoid accessibility probability per cell type. Training data came from ATAC-seq of peripheral blood from 55 healthy Japanese donors, sorted into 27 immune populations plus PBMCs and neutrophils; after quality control, 1,386 samples yielded 7,939,714 OCRs covering 498 Mb (16.1%) of the genome, merged into 1,349,934 non-overlapping input sequences. A single model was trained once with a chromosome holdout — chromosomes 8 and 9 for testing, 6 and 7 for validation, the remaining autosomes for training — with no per-cell-type or per-fold retraining. Predictive accuracy is reported only in the May 2025 version of the preprint, which gave a mean AUROC of 0.87 across the 29 cell types on the held-out chromosomes, with a minimum of 0.84 for neutrophils; the current v3 drops those panels and states no accuracy figures. Running it across all substitutions within 1 kb of each OCR center produced a median of 505 million mutation effects per cell type, and thresholding by local FDR gave a median of 1,584 CAMBUS+ OCRs per cell type (0.056% of the genome), 66,043 in total. Those regions are strongly enriched in the gnomAD Gnocchi constrained genome (odds ratio 11.45, 95% CI 9.33–14.05, P = 4.7 × 10⁻⁶⁸), and remain enriched for known enhancers and super-enhancers after adjusting for those metrics.

#Applications

The intended use is prioritizing non-coding variants in human genetics. CAMBUS+ regions capture a significant share of heritability for immune-related and blood-cell traits in stratified LD score regression, and overlap fine-mapped causal variants from biobank GWAS in both European and East Asian cohorts, including rare variants. This yields 1,533 putative causal variants connected to 70 complex traits, each assigned to the immune cell type whose accessibility model flagged it — evidence that conservation scores and bulk annotations cannot supply.

#Impact

CAMBUS treats a trained accessibility model as an instrument for measuring functional constraint, not just as a predictor, and doing so surfaces regulatory elements that comparative genomics leaves out. The scope is correspondingly narrow: all 29 tracks are peripheral-blood immune cell types from a single Japanese cohort, so behavior outside the hematopoietic lineage is untested. The work remains a preprint, and no trained weights, training code, or score tables have been released — the availability statements promise the score and its calculation scripts after acceptance — so the results cannot currently be reproduced or applied by outside groups.