ENCODE BPNet Atlas β€” float32 ONNX with native controls

Conversion finished: 11,695 accepted / 11,695 source checkpoints. All 2,339 experiments were attempted. The release has 0 exclusions, recorded in 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 BPNet authors. GenomeSpy supplies format conversion and numerical validation; this is not an upstream-endorsed release. No retraining, quantization or fine-tuning. These are the original BPNet Atlas models with native control inputs preserved, including a sequence-only inference route using zeros as permitted by the publisher.

Select a model

Use source-inventory.json for the original repository, full source commit, H5 path, byte size/SHA-256 and corresponding ONNX path. Experiment/fold identities remain separate even when upstream weights have identical hashes. Original cards, configs and full source file metadata are preserved under source/<source-repository-name>/. Reported biosample, target, assay, QC and training information remain publisher claims. Other source files remain accessible at the pinned upstream revisions.

Collection: BPNet ONNX.

Load one selected fold on demand. Each float32 ONNX is about 465 kB; the complete atlas is about 5.4 GB. Conversion fidelity does not establish biological accuracy, fold independence or count calibration across experiments.

Scientific contract

Tensor Shape and meaning
sequence float32 [batch,4,2114], A/C/G/T channels; N is all zero
profile_controls float32 [batch,2,1000], native per-position matched WCE/Input profile controls; zero permitted
count_controls float32 [batch,2], native independent strand count-control values; zero permitted
profile_logits float32 [batch,2,1000], raw plus/minus strand logits
log_counts float32 [batch,1], raw natural-log shared total read count

For DNA at zero-based half-open hg38 interval [s,s+2114), predictions occupy [s+557,s+1557). Preserve the submitted genomic coordinates. There is no reverse-complement averaging or new strand normalization in the graph.

Both strands share ONE softmax across all 2,000 logits and ONE exp(log_counts) count total. Normalize stably after flattening both strand/position axes, multiply by the total, and reshape to [batch,2,1000]. Do not independently normalize strands or use expm1.

Inputs are the native independent controls with axes transposed only for ONNX convolution layout. The publisher labels count-control inputs as log2 totals in its source instructions; this release preserves their values and the original graph's natural reduce_logsumexp operation without adding a base conversion. Follow the selected model's original control preprocessing. Do not derive count controls from profile sums. Two zero count-control values reduce to log(2) internally, before the trained final projection. Supplying zeros does not remove that projection or its bias. Synthetic nonzero controls used for numerical testing are not measured experimental controls or biological validation.

Browser inference

Use matching onnxruntime-web@1.31.1 module and WASM assets over HTTPS or localhost. WebGPU requires an actual supported adapter.

import * as ort from 'onnxruntime-web/webgpu';
const session = await ort.InferenceSession.create(modelUrl, {executionProviders: ['webgpu']});
const feeds = {
  sequence: new ort.Tensor('float32', acgtOneHot, [1,4,2114]),
  profile_controls: new ort.Tensor('float32', new Float32Array(2000), [1,2,1000]),
  count_controls: new ort.Tensor('float32', new Float32Array(2), [1,2]),
};
const outputs = await session.run(feeds);
// Keep raw heads or jointly decode both strands as documented above.
for (const value of Object.values(feeds)) value.dispose();
for (const value of Object.values(outputs)) value.dispose();
await session.release();

Representative real GATA2/K562 folds 0 and 1 and CTCF/activated-CD8 fold 0 were tested headlessly with Apple M5 / Metal 3, Chromium 153.0.8010.12 and ORT Web 1.31.1. Both zero and independent synthetic nonzero controls pass original TensorFlow comparisons for raw heads and decoded counts at batches 1/3/6. The full-control trials produced 3,828 actual GPU compute dispatches per model over 66 calls. Warm full-control synthetic-nonzero inference medians were approximately 2.35–2.70 ms for batch 1 and 6.55–6.70 ms for batch 6; exact per-case timings and hashes are retained in the validation reports. Local session creation excludes internet download latency. WASM single-thread fallback was also tested, approximately 8.4–8.75 ms for batch 1 and 50.6–53.2 ms for batch 6 across the trial variants.

These are representative tests on one device/browser, not atlas-wide WebGPU qualification. Some graph operations may execute on CPU; no all-nodes-on-GPU assertion is made. GPU timestamp callbacks returned no samples; timings measure complete session.run, including transfers. CPU parity alone is not WebGPU evidence. Forward-only ONNX inference does not provide gradients for sequence optimization or attribution. No Sequence Designer UI/backend integration or installed-PWA qualification is part of this release.

Per-checkpoint acceptance and implementation

Every accepted checkpoint passes full ONNX checking and original TensorFlow 2.18.1 / tf-keras 2.18.0 CPU float32 comparisons with both mapped PyTorch and ONNX Runtime CPU at rtol=2e-4, atol=1e-5. Cases cover GATA2 and HBG1 loci, reverse complement, an SNV, internal ambiguous bases and all-N DNA; dynamic batches 1,3,6; zero and independent synthetic nonzero controls. Checks include both raw heads and jointly decoded expected counts. Source SHA-256 and size are verified before reading weights; only accepted exports survive temporary conversion. The manifest records per-checkpoint errors and exact converter/runtime/input-fixture provenance. Resume verifies actual artifact bytes and matching provenance.

The source Keras topology is checked before reconstruction. Its sole serialized count-control Lambda is replaced by a reviewed explicit reduce_logsumexp; no serialized Python is executed. The existing reviewed BPNet Atlas layer/weight loader supplies PyTorch weights for unfused opset-17 serialization. Learned profile/count projections and biases are preserved.

Numerical stabilization: a rejected BHLHE40/HepG2 fold-0 count head exposed float32 long-sum rounding on all-N DNA. The converter can retry an equivalent centered average, anchor + mean(features-anchor), without changing weights, cases or tolerances. Only a candidate passing every original TensorFlow/PyTorch/ONNX test is accepted. Its receipt records centered_mean; both converter snapshots are retained for exact provenance. The centered artifact also passed headless WebGPU at every same batch/control case, with 4,092 actual compute dispatches over 66 calls; warm synthetic-nonzero batch-1/batch-6 medians were 2.70/7.25 ms. See webgpu-centered-validation.json. This remains representative qualification on one device.

TensorFlow-order averaging: NR3C1/A549 fold 4 required an explicit four-accumulator spatial mean matching the order in TensorFlow 2.18.1's bundled Eigen code. This preserves the same mathematical mean and weights while reproducing native float32 rounding. tensorflow_four_lane_mean identifies this variant. Its exact 534,834-byte artifact passes every unchanged TensorFlow/PyTorch/ONNX CPU gate and real WebGPU case; 39,996 actual dispatches over 66 calls, with warm synthetic-nonzero medians 11.60 ms (batch 1) / 12.75 ms (batch 6). See tensorflow-mean-provenance.json and webgpu-tensorflow-mean-validation.json. This additional serialization is used only when the ordinary/centered candidates fail; no tolerance is relaxed.

The trial found an incorrect fused profile-bias mapping in bpnet-lite 1.0.0 BasePairNet.from_bpnet, also present in the inspected upstream revision 6236e4283000e852a5ad1ee8d4d2bc933fb7535a. The fused bias should include the learned final profile mixing matrix. This release avoids that fusion and preserves native independent count controls; it does not use the unmodified bpnet-lite loader for conversion. trial-results.md and reports retain the evidence and exact source identities. No installed third-party package or original H5 weight was patched.

License, citation and provenance

Every pinned source card declares MIT. Original cards and their attribution requests are retained unmodified. LICENSE and BPNET-SOFTWARE-LICENSE retain the upstream BPNet MIT text; the inspected bpnet-lite software license is retained separately. software-license-provenance.json pins the license sources. Model-card declarations, original software terms and converter/runtime licenses remain distinct. Conversion asserts no additional rights or biological guarantees.

Please cite ENCODE, the exact experiment/annotation and references in the selected original card; Avsec et al., Base-resolution models of transcription-factor binding reveal soft motif syntax, Nature Genetics (2021), https://doi.org/10.1038/s41588-021-00782-6; and the atlas reference, Deshpande et al., A unified lexicon of predictive DNA sequence motifs from ENCODE transcription factor binding and chromatin accessibility assays (2025), as linked by the source cards.

The conversion/ directory retains converter sources, shared helpers, pinned requirements, source inventory and browser/reference harnesses. SHA256SUMS binds release files. accepted-manifest.json is authoritative for accepted outputs; files not yet enumerated there should not be selected. Local weights, environments and generated reports remain outside Git.

Complete release accounting

All 11,695 pinned folds are accepted with zero exclusions. release-statistics.json reconciles every source identity, artifact hash, unchanged numerical gate and actual converter snapshot. release-browser-identities.json binds all five executed browser-qualified artifacts to exact release hashes; qualification remains representative. Worker cooling idleness is recorded in execution-verification.json.

Direct hosted browser verification

All five qualified checkpoints were fetched without credentials directly from their public Hugging Face URLs at complete weights revision 3bc53e821537885b03dc1a99b60970a9e4b4e000. Browser SHA-256 values matched the accepted artifacts, proving cross-origin download access. The exact hosted bytes passed every original TensorFlow raw-head/decoded-count comparison on real Apple M5 / Metal 3 WebGPU at batches 1/3/6 and both control modes. Warm synthetic-nonzero batch-1 medians were 2.35–2.40 ms for ordinary/centered graphs and 11.65 ms for the TensorFlow-order graph; batch-6 medians were 6.50–6.60 ms and 12.65 ms respectively. webgpu-hosted-validation.json records downloads, hashes, runtime, actual GPU dispatches, errors and timings. These five artifacts cover all three serialization variants; the other atlas checkpoints retain CPU qualification.

Hosted release verification

At complete revision 11537bbcca2c9e90acabb6146539d32cee0b4dac, all 11,695 model LFS SHA-256 values and 7,078 metadata Git/LFS hashes matched the release records. Public Git LFS download actions were verified without credentials for all 11,695 unique model objects; five representative payloads were downloaded anonymously and matched their sizes/hashes. Collection membership was verified. publication-verification.json records this complete snapshot. The following verification-record commit adds this evidence and changes no weights.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including genome-spy/BPNet-ONNX