Bias-corrected ENCODE ChromBPNet atlas β float32 ONNX
Conversion finished: 7,560 accepted / 7,560 source checkpoints. All 1,512 experiments were attempted. 0 checkpoints are explicitly excluded after failing the unchanged numerical/topology acceptance gate; see excluded-checkpoints.json. accepted-manifest.json lists the exact accepted artifact identities, hashes, validation errors and converter provenance.
Original models, training and biological annotations are by the Kundaje Lab / ENCODE ChromBPNet authors. GenomeSpy supplies format conversion and numerical validation; this is not an upstream-endorsed release. No retraining, quantization or fine-tuning. This repository contains only the nobias accessibility branch that the publisher recommends for biological discovery; it omits the full ChromBPNet and bias-only model weights.
Select a model
Use source-inventory.json to resolve assay, experiment/fold, original repository, full source commit, exact H5 path and source size/SHA-256. Source repository names are preserved in models/<source-repository-name>/fold_<i>/<original-basename>.onnx. Unmodified original cards, configs and complete file metadata are retained in source/<source-repository-name>/; original training logs and other source objects remain available at their pinned upstream revisions indexed there. Reported biological/QC information remains the publisher's claim. Conversion parity does not establish biological accuracy or comparability of absolute counts across experiments.
Download one selected checkpoint as needed. A float32 file is approximately 25.5 MB; the entire source scope would produce about 193 GB of ONNX. Do not fetch the whole atlas for a browser session. See the ChromBPNet ONNX collection.
Scientific contract
| Tensor | Shape and meaning |
|---|---|
sequence |
float32 [batch, 4, 2114], A/C/G/T channel order; N becomes all zero |
profile_logits |
float32 [batch, 1, 1000], raw unstranded profile logits |
log_counts |
float32 [batch, 1], raw log1p total count |
Input [s, s+2114) yields output [s+557, s+1557), zero-based half-open GRCh38/hg38 coordinates. Preserve the actual submitted coordinates. There are no control inputs for this bias-corrected branch. This is an unstranded accessibility prediction, without separate plus/minus outputs.
Decode each example with one numerically stable softmax over all 1,000 logits, multiplied by max(expm1(log_counts), 0). Keep the raw heads if required downstream. The zero clamp preserves the qualified Sequence Designer expected-read decoder; original publisher pseudocode is retained verbatim in source cards. No per-sequence rescaling, thresholding or calibration is introduced into the exported raw heads.
Browser inference
Use the matching onnxruntime-web@1.31.1 WebGPU module/WASM assets, served over HTTPS or localhost:
import * as ort from 'onnxruntime-web/webgpu';
const session = await ort.InferenceSession.create(modelUrl, {executionProviders: ['webgpu']});
const input = new ort.Tensor('float32', acgtOneHot, [1, 4, 2114]);
const outputs = await session.run({sequence: input});
// Read outputs.profile_logits and outputs.log_counts; decode as documented above.
input.dispose();
for (const tensor of Object.values(outputs)) tensor.dispose();
await session.release();
An actual WebGPU adapter is required. Representative real-model inference was tested headlessly on Apple M5 / Metal 3 using Chromium 153 and ORT Web 1.31.1. A stabilized B-cell fold-2 export passed original TensorFlow comparisons of raw heads and decoded signals at batches 1, 3 and 6, including all-N DNA. It dispatched 1,704 actual GPU compute calls over 12 inferences; warm batch-1 inference was approximately 31 ms, batch 6 approximately 170 ms. See webgpu-bcell-validation.json and browser-validation-notes.json for exact file hashes, errors, timings and limitations. The earlier unmodified-serialization trial of K562/activated-CD8/memory-CD8 is retained separately and does not qualify every rewritten model.
Exact bulk-artifact qualification: accepted B-cell ATAC fold 2, K562 ATAC fold 0 and brain DNase fold 0 were independently checked on WebGPU against original TensorFlow outputs. All raw heads and decoded expected counts pass the unchanged bound at batches 1/3/6, with 1,704 actual compute dispatches per model. Warm batch-1 times were 29.9β31.9 ms and batch-6 times 171.1β173.4 ms while conversion was running concurrently. webgpu-release-validation.json binds the exact source and exported hashes and records every error/timing. This is representative qualification, not an atlas-wide WebGPU claim.
A real 64-channel variant (heart-left-ventricle ATAC ENCSR851EBF, fold 0) also passes original TensorFlow comparisons in WebGPU across every same case/batch. It produced 3,000 actual dispatches over 12 calls; warm batch-1 inference was 33.4 ms and batch 6 was 205.2 ms. See webgpu-chunk64-validation.json for its exact source/artifact identities and errors.
Some graph operations execute on CPU (including ORT shape handling). No all-nodes-on-GPU assertion is made. GPU timestamp callbacks returned no samples; timings measure complete session.run, including transfers. Only representative files are browser-qualified; CPU numerical parity is not evidence that every checkpoint/GPU/browser meets the same WebGPU tolerance. No gradients or gradient-based design/attribution methods are provided by these forward-only ONNX files. No app backend or UI integration is part of this release.
Per-checkpoint acceptance
Every accepted artifact passes full ONNX checking and original TensorFlow 2.18.1 / tf-keras 2.18.0 CPU float32 comparisons for both raw heads and decoded expected counts at rtol=2e-4, atol=1e-5. The original reviewed Keras JSON graph is reconstructed and its H5 weights loaded without executing serialized Python. The app's pinned bpnet-lite loader supplies the equivalent PyTorch graph used for opset-17 export. Dynamic batch inputs are checked at batches 1, 3 and 6 on GATA2, HBG1, reverse complement, SNV, internal Ns and all-N DNA. The manifest records every accepted file's errors, source/output hashes, chosen convolution channel chunk and exact converter/runtime provenance.
Wide convolutions use identical trained weights split into equal input-channel groups, balanced additions, and one trained bias addition. This mathematically equivalent serialization reduces float32 cancellation error in wide channel sums. The converter tries channel chunks of 128, 64, 32 and 16; a candidate is retained only if every unchanged test passes. No tolerance or input case is removed. Unsupported graph topology or failed numerical checks exclude a model rather than silently accepting it. The earlier CPU-versus-PyTorch near-zero-logit discrepancy is documented in the trial record; acceptance here uses the original TensorFlow implementation directly.
License, citations and provenance
All pinned source cards declare MIT. They also retain the ENCODE data-use policy and attribution requests; original cards/configs remain unmodified. LICENSE and CHROMBPNET-SOFTWARE-LICENSE retain the Kundaje Lab MIT text; bpnet-lite software has its own retained MIT text. Model terms, original software terms and converter/runtime licenses remain distinct. No new rights or biological guarantees are asserted by conversion.
Please cite the ENCODE Project Consortium; Pampari et al., ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants, bioRxiv; and the atlas reference and exact experiment/annotation listed in the selected original model card. The source cards link Yun et al., a unified lexicon of predictive DNA sequence motifs.
The conversion/ directory retains converter sources, shared helpers, pinned requirements, the source-inventory script and representative browser/reference harnesses. Source files, manifests and accepted artifact hashes provide exact identities; local temporary H5 downloads are removed after conversion. SHA256SUMS binds published release files. Files arriving before a matching accepted manifest entry remain unenumerated and should not be selected until the manifest is updated.