Cherimoya BATv1 - float32 ONNX

ONNX conversions of the published BATv1 atlas for predicting human transcription factor / DNA-associated protein binding (ChIP-seq). Original models and metadata are by programmable-genomics / the Cherimoya authors. GenomeSpy provides serialization and numerical validation; this is not an upstream-endorsed release.

Qualified subset: 17,143 of 17,145 published source checkpoints, representing all 3,429 experiments. Two folds are excluded because native/ONNX CPU parity fails at the unchanged bound: ENCSR641XWH fold 3 and ENCSR681NOM fold 1. The other four folds of each experiment remain independently identified. accepted-manifest.json enumerates only accepted artifacts and pins both exclusions; no complete-atlas manifest is claimed. See unqualified-checkpoints.json and validation-summary.json.

Source revision: c631ebd86b46d78c2e239ee06f3fce5ff15f4ad9. The exact paginated source inventory and publisher manifest are retained under source/. The full per-artifact records are consolidated in accepted-manifest.json: original repository, revision, path, SHA-256, converted SHA-256, tensor contract, CPU numerical errors and conversion software. Its normalized conversions map preserves the complete producer provenance. Individual receipts remain local for resumable conversion; they are omitted here to stay below the Hub repository file limit. Download one experiment/fold as needed; no atlas-wide download is required.

Known WebGPU limitation: ENCSR297CGF fold 2 passes the full native/ONNX CPU gate, but its all-N decoded counts in batch 6 with nonzero controls exceed the same tolerance on the tested Apple M5 / Metal runtime (maximum tolerance ratio 1.81). Raw heads remain within tolerance. It is not WebGPU-qualified at this bound; see webgpu-validation.json. The separately identified representative BAT checkpoint passes all cases.

Additional WebGPU limitation: ENCSR681NOM fold 2 passes the CPU gate but fails profile-logit and decoded-count bounds with nonzero controls on the tested runtime (maximum tolerance ratios about 1.80 and 2.18). This is an ordinary GATA2 input case, not only all-N DNA. It is also not WebGPU-qualified at the stated bound.

Scientific contract

No quantization, retraining or fine-tuning is applied. Inputs and raw profile/count heads remain float32. This release covers forward inference and perturbation scoring; it does not provide gradients for DeepLIFT, PISA or Ledidi.

Tensor Contract
sequence float32 [batch, 4, 2114], A/C/G/T channel order; N is all zero
controls float32 [batch, 2, 2114], required; plus then minus strand counts
profile_logits float32 [batch, 2, 1000], raw logits
log_counts float32 [batch, 1], raw log1p total count

GRCh38/hg38. Input/output intervals are zero-based, half-open: input [s, s+2114) produces output [s+557, s+1557). Use the actual submitted coordinates. Training and biological annotations, publisher metrics, and chromosome-fold details are preserved verbatim under source/; numerical conversion parity is not biological validation or cross-experiment count calibration.

Decode one example by flattening all profile channels and all 1000 positions together, applying one numerically stable softmax, multiplying by max(expm1(log_counts), 0), then reshaping. BAT has plus/minus channels sharing one jointly normalized profile and one total count; separate strand softmaxes incorrectly double the total. Keep raw heads if downstream calculations need them.

BAT controls are required, same-locus plus/minus signal over the full input window. Zero controls permit sequence-only prediction and do not establish bias correction or reproduce measured-control inference. CAT has no control input.

Browser inference

Use onnxruntime-web@1.31.1 and its matching WASM/module assets, served over HTTPS or localhost. Import the WebGPU entry point and explicitly select WebGPU:

import * as ort from 'onnxruntime-web/webgpu';
const session = await ort.InferenceSession.create(modelUrl, {
  executionProviders: ['webgpu'],
});
const feeds = {sequence: new ort.Tensor('float32', sequence, [1, 4, 2114])};
// BAT only: feeds.controls = new ort.Tensor('float32', controls, [1, 2, 2114]);
const outputs = await session.run(feeds);
// Preserve outputs.profile_logits and outputs.log_counts; decode as described above.
for (const tensor of Object.values(feeds)) tensor.dispose();
for (const tensor of Object.values(outputs)) tensor.dispose();
await session.release();

Read accepted-manifest.json/receipts before selecting weights. A browser needs a working WebGPU adapter; provider selection does not promise every graph node runs on GPU. No CPU/WASM-only run is labeled WebGPU evidence.

Validation and limits

Every accepted export passes full ONNX checking and native PyTorch versus ONNX Runtime CPU comparison of raw heads and decoded counts at rtol=1e-4, atol=5e-5. Cases cover GATA2 and HBG1 DNA, reverse complement, an SNV, internal Ns and all-N DNA, batches 1/3/6, and BAT zero/nonzero controls. Native computation is the pinned Cherimoya CPU float32 implementation. Long normalization reductions are split into equal chunks to reduce accumulation error; tolerances are unchanged. There is no custom ONNX operator.

Real ENCSR000AHD fold 0 was tested headlessly on Apple M5 / Metal WebGPU in Chromium 153 with ORT Web 1.31.1: all raw-head and decoded-count comparisons passed. Actual compute dispatches were observed. See webgpu-validation.json for exact file identities, GPU/runtime, errors, session creation, cold and warm timings, and provider-placement limits. This is representative browser qualification, not a claim that every atlas checkpoint or browser/GPU was tested on WebGPU. CPU parity alone does not establish browser compatibility.

Some graph nodes are assigned to CPU, including ORT shape work; disabling CPU fallback rejects the session. GPU timestamp profiling produced no callbacks in this harness; reported performance is wall-clock session.run, including transfers and potentially affected by concurrent conversion/logging. Initial shader compilation is slower than warm inference. No installed-PWA, Safari, Firefox or mobile acceptance is claimed.

Attribution, licenses and provenance

Model weights and retained publisher metadata are under the original CC BY 4.0, reproduced in LICENSE and source/LICENSE. Attribution: programmable-genomics, the Cherimoya authors and ENCODE contributors; see the unmodified source/README.md and metadata for scientific citations and training provenance. The conversion changes serialization and reduction grouping while preserving validated scientific input/output semantics.

Cherimoya software is separately MIT, pinned to 73ce51841b6af672bcd423b969ce6f06dd40fc9d; its license is retained in CHERIMOYA-SOFTWARE-LICENSE. Conversion code and exact software versions/hashes are retained for reproducibility under conversion/ and receipts. Model weight licensing does not relicense runtime software. ONNX Runtime and browser runtime packages remain under their own licenses.

Original model: programmable-genomics/BATv1. Grouped in the Cherimoya ONNX collection.

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. ๐Ÿ™‹ Ask for provider support

Model tree for genome-spy/BATv1-ONNX

Quantized
(1)
this model

Collection including genome-spy/BATv1-ONNX