Cherimoya Accessibility aTlas (CATv1)

This is the first release of the Cherimoya Accessibility aTlas (CATv1): a collection of over 7,500 Cherimoya models trained on DNase-seq and ATAC-seq experiments from the ENCODE Project. Cherimoya models are state-of-the-art predictors of local chromatin accessibility, mapping a DNA sequence to a base-resolution accessibility profile together with its total read count.

This release covers 1,518 ENCODE experiments (1,149 DNase-seq and 369 ATAC-seq). Every experiment is trained and evaluated across 5 folds (fold_0–fold_4) of a chromosome-held-out cross-validation. Each model predicts accessibility for a single experiment, so download only the model(s) you need.

Model description

Each Cherimoya model is a compact residual convolutional network (~0.6M parameters; 9 layers, 128 filters) in the BPNet / ChromBPNet family. It takes a one-hot DNA sequence of in_window=2114 bp and produces two outputs over the central out_window=1000 bp: a profile head (base-resolution logits, shape (N, 1, 1000)) describing the shape of the accessibility signal, and a count head (log total counts, shape (N, 1)) describing the overall magnitude. The published checkpoints are the exponential-moving-average (EMA) shadow weights selected on validation performance.

See https://github.com/jmschrei/cherimoya for more details on the Cherimoya model.

Training data

Models are trained on the human genome (GRCh38 / hg38) against the observed ENCODE accessibility signal for each experiment. Training loci are the experiment's peak regions plus GC-matched background (non-peak) negatives, with ENCODE blacklist regions excluded. The five folds partition the genome by chromosome, so for every fold the validation and test chromosomes are held out of training; reported metrics are validation-set performance, not the test set.

Repository layout

Checkpoints live under models/, one directory per experiment:

file	description
models/<encid>/cherimoya.fold_<i>.torch	checkpoint for fold i (EMA shadow weights)

Repository-level files:

file	description
performance.tsv	validation metrics for every experiment and fold
CATv1-metadata.tsv	experiment metadata, keyed by experiment_accession
manifest.csv	the full list of experiments and folds included
fit.json / evaluate.json	a representative example of the per-fold training / evaluation settings

performance.tsv has one row per experiment-fold: the first two columns are experiment_accession and fold, followed by profile_mnll profile_jsd profile_pearson profile_spearman count_pearson count_spearman count_mse. The headline metrics are profile_pearson (profile shape) and count_pearson (total accessibility).

Each experiment_accession is the source ENCODE experiment the models were trained on (https://www.encodeproject.org/experiments/<experiment_accession>/), and the paired annotation_accession points to the ENCODE ChromBPNet model annotation for that same experiment (https://www.encodeproject.org/annotations/<annotation_accession>/).

Which model should I use?

Pick the experiment that matches your assay and biosample using CATv1-metadata.tsv, then download that experiment's checkpoint(s). The five folds are cross-validation replicates of the same experiment: use a single fold (e.g. fold_0), or average the predictions of all five folds for a more robust estimate. A model only reflects the experiment it was trained on, so do not use it for a different cell type or assay.

Intended uses and limitations

Beyond profile and count prediction, the models support downstream sequence analyses such as attribution, motif marginalization, and in-silico mutagenesis. They are human (GRCh38) only, are specific to the experiment they were trained on, and report validation (not test) metrics.

Installation

pip install cherimoya    # PyPI (Python 3.10+)

Install-from-source, Docker images, and GPU/Triton notes: https://github.com/jmschrei/cherimoya

Usage

import torch
from huggingface_hub import hf_hub_download
from cherimoya import Cherimoya

path = hf_hub_download("programmable-genomics/CATv1", "models/<encid>/cherimoya.fold_0.torch")
model = Cherimoya.load(path, device="cuda", compile=False).eval()

# Input: one-hot DNA, shape (batch, 4, in_window=2114), ACGT channels.
X = torch.zeros(2, 4, 2114, device="cuda")
X[:, 0] = 1.0   # placeholder: replace with your encoded sequences

# Option A: call the model directly.
profile_logits, log_counts = model(X)          # (N,1,1000), (N,1)
profile = torch.softmax(profile_logits.float(), dim=-1)

# Option B: via tangermeme (auto-batches, handles device/dtype).
from tangermeme.predict import predict
profile_logits, log_counts = predict(model, X, device="cuda", batch_size=64)

Contact

Jacob Schreiber jacob.schreiber@umassmed.edu — Programmable Genomics Laboratory @ UMass Chan.