Hugging Face's logo

Datasets:
liubuing
/
BioCandidateRanker-training-library

Tasks:
Other
Tags:
biology
chemistry
enzymes
enzyme-kinetics
kcat
homology-split
License:
Dataset card Files
xet
 Community
BioCandidateRanker-training-library / README.md
liubuing's picture
liubuing
Add training library: frozen homology-cold splits, audits, evaluation artifacts, and v10 release archive
92cfea3 verified
preview code
|
Raw History Blame Contribute Delete
8.51 kB
metadata 
license: other
license_name: mixed-mit-cc0-cc-by-4.0
task_categories:
  - other
tags:
  - biology
  - chemistry
  - enzymes
  - enzyme-kinetics
  - kcat
  - homology-split
  - mmseqs2
  - protein-language-model
  - enzyme-engineering
size_category: 10K<n<100K
pretty_name: >-
  BioCandidateRanker Training Library (frozen splits, audits, and release
  archive)

BioCandidateRanker Training Library

中文说明见下方 中文说明 一节。

This repository is the organized, license-compliant training library of the BioCandidateRanker project — a host-aware multimodal enzyme candidate scoring research system whose first trained task predicts log10(kcat) from protein sequence, substrate molecular graph, organism/EC/reaction context, and optional FBA features.

It packages everything needed to reproduce the exact training setup: the frozen homology-cold split assignments, ingestion audit manifests, source identity hashes, external evaluation selections, and the official software release archive.

What is deliberately NOT here: the UniKP/DLKcat measurement rows themselves (protein sequences + substrates + kcat values). The upstream compilation has no redistribution grant on record (license: unknown, record-level provenance corpus-level only), so the project's fail-closed license governance (configs/data_license_inventory.json in the archive) marks it redistribution: not_verified and blocks any measurement row from shipping in a public archive. See Reconstructing the full training corpus to rebuild the corpus from the original public source and verify it bit-for-bit.

Contents 

data/
  homology_split.parquet          16,838 rows: source_row, sequence_id, cluster, split
  homology_split.json             original frozen split manifest (redacted paths)
  unikp_audit.json                ingestion audit of the development corpus (row counts,
                                  duplicates, conflicts, source SHA256; redacted path)
  unikp-km-training-blocker-audit.json      why governed Km training is blocked
  absolute-activity-training-blocker-audit.json
  full-homology-baseline/test_metrics.json  frozen baseline holdout metrics
  enzengdb-v1/                    EnzEngDB v1 selection, homology-cold audits,
                                  rank split, zero-shot transfer result, 3-seed summary
  lunzer-imdh/                    Lunzer IMDH selection, homology audit, frozen
                                  BioCandidateRanker and DLKcat evaluations
  temporal/                       temporal absolute-kinetics pool readiness audits,
                                  accepted-component FASTA and exact row mapping
  uncertainty-calibration-v2/grouped-cv.json
configs/
  unikp_source_manifest.json      source identity and adapter statistics
  data_license_inventory.json     the full license governance file governing this repo
  temporal_absolute_kinetics_protocol.json
release/
  software-implementation-v10-redacted.zip        official release archive (408 files)
  software-implementation-v10-redacted-manifest.json   per-file SHA256 manifest
RECONSTRUCTION.md                 step-by-step corpus reconstruction guide

The frozen homology-cold split

The development corpus is split protein-homology-cold with MMseqs2 linclust (30% min identity, 80% coverage, coverage mode 0, seed 42), producing 2,204 clusters. Whole clusters are assigned to a partition, so no protein family leaks between train and evaluation:

split rows
train 13,470
validation 1,684
test 1,684

The split is keyed by source_row (the 0-based row index in the original Kcat_combination_0918_wildtype_mutant.json), so it can be applied to any byte-identical copy of the source file:

from datasets import load_dataset
import json

split = load_dataset("<repo-id>", data_files="data/homology_split.parquet")["train"]
# row: {"source_row": 0, "sequence_id": "seq_...", "cluster": "seq_...", "split": "train"}

Note: the split manifest contains cluster/row/split assignments only — no sequences and no kcat values — which is exactly why it is redistributable while the corpus itself is not.

Reconstructing the full training corpus

  1. Obtain Kcat_combination_0918_wildtype_mutant.json from the original UniKP repository (the file is also the wildtype+mutant kcat compilation distributed with DLKcat).

  2. Verify the frozen source identity:

    field value
    SHA256 13643b0b36374f8d3f64d8b014882cf1b3b58946eeaae2b9dcd59e8b2c2d6719
    size 12,132,719 bytes
    rows 17,010

    The audit on this machine confirmed a byte-identical copy (2026-10-24).

  3. Run the project's audit and training CLI (included in release/*.zip) as described in RECONSTRUCTION.md.

Full step-by-step instructions: RECONSTRUCTION.md.

External evaluation data included

component source license shipped content
EnzEngDB v1 Zenodo DOI 10.5281/zenodo.17310823 CC-BY-4.0 selection manifests, homology-cold audit, rank split, evaluation results (the 30 MB source archive is not bundled — fetch it from the DOI)
Lunzer IMDH landscape Dryad DOI 10.5061/dryad.7nd70 CC0-1.0 project-generated selection, homology audit, and frozen evaluations (source .xls not bundled)
Temporal absolute-kinetics pool 18+ CC0 / CC-BY-4.0 deposits CC0 / CC-BY-4.0 per source readiness audits, accepted-component FASTA, row mapping (curation pool below its gates — never scored, not a benchmark)

Release archive

release/software-implementation-v10-redacted.zip is the official deterministic release archive (software-implementation-v10): every Git-tracked source file plus 38 frozen evidence artifacts, built by scripts/build_release_archive.py with license-ownership verification and absolute-path redaction. Its embedded and sidecar manifests carry per-file SHA256 hashes; the archive SHA256 is 34e42f5b211bd590b785abe3ff59ac2cfaaca801328ddef68c9b303044b394d8.

The archive includes the trained baseline checkpoint (artifacts/full-homology-baseline/best.pt), which the license inventory explicitly permits as a derived artifact (weights are not a redistribution of measurements).

License and attribution

  • This package's own content (splits, audits, configs, code in the archive): MIT, © 2026 BioCandidateRanker authors.
  • EnzEngDB v1: CC-BY-4.0 — cite the Zenodo deposit DOI 10.5281/zenodo.17310823.
  • Lunzer IMDH landscape: CC0-1.0 — Dryad DOI 10.5061/dryad.7nd70.
  • UniKP/DLKcat corpus: not redistributed here. If you reconstruct it, respect the upstream terms and cite the UniKP and DLKcat papers.

Citation

If you use this split or the reconstruction protocol, please cite the upstream sources and this dataset:

@software{biocandidateranker2026,
  title  = {BioCandidateRanker training library: frozen homology-cold splits and audits},
  author = {BioCandidateRanker authors},
  year   = {2026},
  note   = {software-implementation-v10 release archive}
}

中文说明

本仓库是 BioCandidateRanker 项目的训练库合规整理包,目标是让任何人都能精确重建训练设置:

  • 包含:冻结的 MMseqs2 同源冷划分(2,204 个簇,16,838 行 → train 13,470 / validation 1,684 / test 1,684,按原始文件行号索引)、摄入审计 manifest、源文件 SHA256 身份、EnzEngDB(CC-BY-4.0)与 Lunzer IMDH(CC0)的外部评估选择与结果、时间池就绪审计、以及官方 software-implementation-v10 全量发布档案(含基线模型权重 best.pt)。
  • 不包含:UniKP/DLKcat 训练语料的测量行本体(序列、底物、kcat 值)。项目许可治理文件 configs/data_license_inventory.json 将其标记为 not_verified 并明确禁止任何测量行进入公开档案——上游(BRENDA 编译)无逐条记录的再分发授权。请通过 RECONSTRUCTION.md 从原始公开来源获取并按 SHA256 校验。
  • 训练语料原始出处:UniKP (Luo-SynBioLab)、UniKP 论文 (Yu et al., Nat Commun)、DLKcat 补充数据 (Zenodo)。