appliedscientific/cardiosafe-benchmark

CardioSafe ion-channel benchmark

Curated multi-task ion-channel labels + Tanimoto-controlled splits + supplementary materials for CardioSafe: multi-task prediction of cardiac ion channel activity with reverse-leak audited benchmarking (Jovanović et al., 2026, bioRxiv).

The canonical source is the CardioSafe-benchmark GitHub repository. This HF Datasets repo mirrors data/ from that deposit and adds pre-joined per-fold Parquet shards so you can write:

from datasets import load_dataset

# v1.1 (audit-clean) test fold on the tan70 Tanimoto cutoff:
ds = load_dataset("appliedscientific/cardiosafe-benchmark",
                  "tan70_v1_1", split="test")
print(ds.column_names)
# ['row_idx', 'smiles', 'inchikey', 'herg_pchembl', 'herg_blocker_10um',
#  'herg_blocker_1um', 'nav15_pchembl', 'nav15_blocker', 'cav12_pchembl',
#  'cav12_blocker', 'iks_blocker']

Each row is one curated compound. Label columns are NaN-sparse — only those compounds with primary-screen evidence for the relevant channel have non-null values.

Configs (pre-joined splits)

Config	Split strategy	Version	Compounds (train / val / test)
tan70_v1_0	Tanimoto ≥ 0.70 cutoff	v1.0 preprint	241,792 / 46,326 / 46,326
tan60_v1_0	Tanimoto ≥ 0.60 cutoff	v1.0 preprint	306,665 / 13,889 / 13,889
tan70_v1_1	Tanimoto ≥ 0.70 cutoff	v1.1 retrain	241,790 / 46,328 / 46,326
tan60_v1_1	Tanimoto ≥ 0.60 cutoff	v1.1 retrain	306,662 / 13,892 / 13,889

v1.1 differs from v1.0 by 2 force-routed analogs in the cardiac-cliff cluster (terfenadine/fexofenadine/HMT). The test fold is identical across v1.0 and v1.1 — paper Table 2/3 metrics are unchanged. See supplementary/note_s3_v1_1_audit_correction.md.

Label schema

Column	Head	Type
herg_pchembl	regression — hERG pIC50	float
herg_blocker_10um	hERG blocker @ 10 µM	binary {0, 1}
herg_blocker_1um	hERG blocker @ 1 µM	binary {0, 1}
nav15_pchembl	regression — Nav1.5 pIC50	float
nav15_blocker	Nav1.5 blocker @ 10 µM	binary {0, 1}
cav12_pchembl	regression — Cav1.2 pIC50	float
cav12_blocker	Cav1.2 blocker @ 10 µM	binary {0, 1}
iks_blocker	IKs blocker @ 10 µM (exploratory)	binary {0, 1}

IKs has no regression head (n = 115 labelled compounds; treated as exploratory).

Per-channel label counts (primary binary head, all folds combined):

Channel	n labelled	n blockers
hERG (10 µM)	331,127	11,881
Nav1.5 (10 µM)	3,160	1,240
Cav1.2 (10 µM)	1,138	548
IKs (10 µM)	115	30

Raw canonical files

raw/ contains the source-of-truth files exactly as published in the GitHub repo, for power users doing custom merges:

raw/
├── compounds.parquet              # 334,444 × (row_idx, smiles, inchikey)
├── labels.parquet                 # 334,444 × 8 sparse labels
├── splits/
│   ├── tan70.parquet              # v1.0 — paper preprint splits
│   ├── tan60.parquet
│   ├── tan70_v1_1.parquet         # v1.1 — audit-clean retrain splits
│   └── tan60_v1_1.parquet
├── labels_MANIFEST.json           # curation provenance (sources, voting policy)
├── splits_CHANGELOG.md            # v1.0 → v1.1 diff
└── README.md                      # full data-deposit notes from GitHub

All splits share the same row_idx keying — join on it for arbitrary slicing.

Comparators

comparators/ ships the CToxPred2 and CardioGenAI predictions on the v1.0 tan70 test fold, the inputs to the reverse-leak audit and the head-to-head comparison in paper Tables 3 / 3b / S2 / S3 / Figure 4.

Supplementary materials

supplementary/ ships verbatim:

• Notes S1, S2, S3 — Y-randomization mechanism; activity-cliff curation provenance + bibliography + per-pair composition + filtered cliff manifests; audit-driven v1.1 correction + retrain metrics + the cardiac-cliff case study
• Tables S0, S1, S2, S3, S5, S6, S7, S8, S9 — descriptor spec, per-head confusion matrices, comparator panels pre/post de-leak, tan60 drug panel, failure-mode SMILES, AD per-bin metrics, L1000 threshold sweep, curation sensitivity (no S4 — paper supplementary numbering jumps S3 → S5)
• Figure S1 — hERG reliability curves across applicability-domain bins (PDF + PNG + JSON of the underlying decile data)

Source data

• Labels are derived from ChEMBL 36 (source dump SHA-256 b25820eef0f0481ad7712bdf4bac3b45f354e3cbacb76be1fdbf4205d6b48fb9, available from https://www.ebi.ac.uk/chembl/) and the hERG Central primary screen, under a pharmacology-aware curation policy that retains censored values and inhibition-percentage votes. Full provenance lives in raw/labels_MANIFEST.json.
• Splits are Tanimoto-controlled across train / val / test on Morgan-r2-2048 fingerprints, with terfenadine and fexofenadine force-routed to val so the canonical cardiac activity cliff is available as a held-out case study. v1.1 additionally force-routes the hydroxymethyl-terfenadine analogs flagged by an exhaustive O(n_train × n_other) Tanimoto leakage audit (scripts/audit_tanimoto_leak.py in the GitHub repo).

What is not here

• L1000 raw signatures used to train the expression encoder — available from GEO under GSE92742 (Phase I) and GSE70138 (Phase II).
• Continually-updated production ensemble — served at platform.appliedscientific.ai/cardiosafe. This deposit is the frozen paper snapshot.
• Model weights — at appliedscientific/cardiosafe (5-seed v1.0 + v1.1 ensembles + L1000 encoder, CC-BY-NC-4.0).
• Interactive demo — at appliedscientific/cardiosafe (Space).

License

CC-BY-4.0. Use with attribution; commercial use allowed under the license terms. See the full LICENSE-DATA in the GitHub repo for the exact text.

Note: the model weights are CC-BY-NC-4.0 (non-commercial), not the data. They live at a separate HF repo — appliedscientific/cardiosafe.

Citation

@article{cardiosafe2026,
  title   = {CardioSafe: multi-task prediction of cardiac ion channel
             activity with reverse-leak audited benchmarking},
  author  = {Jovanović, Mihailo and Weidener, Lukas and Brkić, Marko and
             Ulgac, Emre and Meduri, Aakaash},
  year    = {2026},
  journal = {bioRxiv},
  doi     = {10.64898/2026.05.06.723181},
  url     = {https://www.biorxiv.org/content/10.64898/2026.05.06.723181v1}
}