ChromBPNet Remote Tool (MCP Server)

Serves ChromBPNet (Pampari et al., Nature Methods 2025) — base-resolution, bias-corrected deep learning of chromatin accessibility from DNA sequence — as the ToolUniverse remote tools run_chrombpnet_predict and run_chrombpnet_variant_effect.

ChromBPNet predicts ATAC-seq/DNase-seq accessibility from a 2,114 bp sequence with the Tn5/DNase enzyme bias regressed out. It is the modern successor to DeepSEA/Basset for non-coding regulatory variant interpretation (GWAS/eQTL fine-mapping) and TF-motif discovery, and underlies the ENCODE accessibility model zoo. The model has two output heads: a 1,000 bp accessibility profile (shape) and a scalar log total count (magnitude).

Note on DeepSEA: the classic DeepSEA model (and HumanBase/FUMA front-ends) is browser-only with no maintained programmatic API. ChromBPNet is the maintained, installable, bias-corrected equivalent and is what this tool wraps.

Served remotely because it carries a heavy TensorFlow/Keras stack and requires a trained, cell-type-specific model (.h5), referenced per call by model_path on the server. Get models from the ChromBPNet model zoo / ENCODE, or train your own with the chrombpnet package.

Operations

• run_chrombpnet_predict — predicted accessibility (log total counts + base-resolution profile) for one sequence.

• run_chrombpnet_variant_effect — ref-vs-alt count log2 fold-change (magnitude effect) + profile Jensen-Shannon divergence (shape effect), the canonical ChromBPNet variant scores.

Models

Trained, cell-type-specific models live in the HF ENCODE ChromBPNet zoo — e.g. kundajelab/encode-chrombpnet-DNASE-ENCSR000EMK-ENCSR816AQM. Each repo has 5 folds; use the bias-corrected fold_N/model.chrombpnet_nobias.fold_N.*.h5 for variant scoring. Pull one with huggingface_hub.hf_hub_download(...) and pass the path as model_path.

Deploy

pip install -r requirements.txt             # tensorflow + tf-keras + numpy + huggingface_hub
export TF_USE_LEGACY_KERAS=1                 # ENCODE zoo models are Keras 2 (.h5)
python chrombpnet_tool.py                    # starts the MCP server on 127.0.0.1:8032

TF_USE_LEGACY_KERAS=1 + tf-keras are required on TensorFlow ≥ 2.16 (Keras 3) so load_model can read the legacy Keras-2 .h5. GPU recommended. Expose remotely only behind TOOLUNIVERSE_API_TOKEN (SMCP bind guard).

Validated end-to-end against a real ENCODE DNASE model (kundajelab/encode-chrombpnet-DNASE-ENCSR000EMK): the model’s (None, 2114, 4) input and [(None, 1000), (None, 1)] outputs match the tool’s INPUT_LEN/ OUTPUT_LEN and head ordering; predict returns a valid accessibility distribution, and variant scoring scales correctly (identical → 0; a 50 bp disruption → larger count log2FC + profile JSD than a single SNP).

Register in ToolUniverse

Tool definition: src/tooluniverse/data/remote_tools/chrombpnet_tools.json (type: RemoteTool).