skybluee/longbench-chatnt-style

DNALongBench ChatNT-Style v1

This repository contains a derived, ChatNT-style serialization of two long-range binary classification tasks from DNALongBench: enhancer-target gene prediction (ETGP) and eQTL prediction (eQTLP). It is intended to make the tasks convenient for DNA + text multimodal model evaluation and supervised fine-tuning.

It is not an official DNALongBench release and does not include the original source tables or reference genomes. Please cite and comply with the original DNALongBench data sources and licenses when using this derivative.

Contents

etgp_v1/
  train.jsonl.gz
  valid.jsonl.gz
  test.jsonl.gz
  manifest.json
eqtlp_v1/
  train.jsonl.gz
  valid.jsonl.gz
  test.jsonl.gz
  manifest.json

All JSONL files are gzip-compressed. Each sample has a fixed 450,000 bp DNA context.

Task	Train	Validation	Test	Positive labels in test
ETGP	2,066	266	270	10
eQTLP	20,364	6,554	4,297	190

The official DNALongBench splits are preserved. Both tasks are strongly imbalanced, so AUROC should be accompanied by AUPRC and class-aware metrics.

Tasks

ETGP: enhancer-target gene prediction

Given a 450 kb genomic context and metadata for an enhancer candidate, target gene, and K562 cell type, predict whether the enhancer regulates the gene (Yes or No).

eQTLP: eQTL prediction

Given matched 450 kb reference and alternate allele contexts plus tissue and gene metadata, predict whether the variant is an eQTL (Yes or No).

Data format

Each line is a standalone JSON object. The conversation format follows ChatNT-style multimodal references: text references a named DNA span, while DNA is stored separately in dna_sequences.

{
  "id": "dnalongbench_etgp_k562_test_0000022",
  "messages": [
    {
      "role": "user",
      "content": "Task: enhancer-target gene prediction. Does the tested enhancer regulate BAX in K562, based on the long genomic sequence <DNA_1>? Answer Yes or No."
    },
    {"role": "assistant", "content": "No"}
  ],
  "dna_sequences": [
    {"name": "DNA_1", "role": "genomic_context", "sequence": "ACGT..."}
  ],
  "target": "No",
  "meta": {
    "task_name": "etgp",
    "split": "test",
    "label": 0,
    "sequence_length": 450000
  }
}

For eQTLP, dna_sequences contains DNA_REF and DNA_ALT rather than a single DNA_1 sequence. Metadata contain provenance fields such as chromosome, gene ID, tissue/cell type, genomic distance, alleles, and masking statistics where available. The user message never contains the target label.

Sequence construction

The conversion mirrors the official DNALongBench EPI/eQTL dataset loaders:

• ETGP uses a gene TSS window (+/- 3 kb) and enhancer-region window (+/- 500 bp), then takes the genomic interval spanning both.
• eQTLP follows the same long-context construction and creates reference/alternate sequence pairs from the two alleles.
• Intermediate blacklist regions are masked with N; sequences shorter than 450 kb are padded with N, while longer contexts are truncated to 450 kb.
• Cross-chromosome and over-distance eQTL records are excluded consistently with the official loading logic. The resulting eQTLP conversion skipped 67 records.
• When genomic orientation is reversed relative to gene direction, the sequence is reverse complemented.

The 450 kb context is a benchmark design choice: it makes input length fixed for fair long-context comparison and allows models to use distal regulatory evidence. It is not a claim that every base is functional.

Loading example

import gzip
import json

path = "etgp_v1/train.jsonl.gz"
with gzip.open(path, "rt", encoding="utf-8") as handle:
    sample = json.loads(next(handle))

print(sample["messages"][0]["content"])
print(sample["target"])
print(len(sample["dna_sequences"][0]["sequence"]))

Recommended evaluation

Use the provided validation/test splits without resampling them. Report at least AUROC and AUPRC; additionally report MCC, accuracy, and threshold-selection protocol. Because text fields may carry useful biological priors, DNA + text results should be compared with the following controls:

1. DNA-only.
2. Text/metadata-only.
3. DNA + text.
4. DNA + shuffled metadata.

Provenance and citation

Source benchmark: DNALongBench repository and its associated paper, DNALongBench: Benchmarking long-context genomic sequence models (Nature Communications, 2025).

Please cite the DNALongBench paper and the original underlying datasets. Cite this repository as a derived conversion, not as the source of the biological labels or reference sequence.

Limitations

• This is a task-format conversion, not a newly curated biological dataset.
• The natural-language instructions are templated from task metadata; they should not be interpreted as free-form annotations.
• Large compressed files contain raw DNA strings. Use streaming readers rather than loading all samples into memory.
• Public redistribution is subject to the terms of the original benchmark, reference genome, and source datasets. Verify compliance before redistribution or publication.