CardioSafe-benchmark

Splits, labels, and supplementary artifacts for CardioSafe: multi-task prediction of cardiac ion channel activity with reverse-leak audited benchmarking (Jovanović et al., 2026).

CardioSafe is a three-branch multi-task neural network that predicts blocker status and pIC50 for the four CiPA cardiac ion channels — hERG, Nav1.5, Cav1.2, and (exploratory) IKs — trained on the largest publicly reported multi-channel cardiac ion channel dataset (ChEMBL 36 + hERG Central). This repository is the public deposit for the paper: the curated label matrix, Tanimoto-controlled splits, supplementary tables / notes / figure promised in the Additional files section of the manuscript, the exhaustive Tanimoto leakage audit, the reference model + training-step code, and the v1.0 / v1.1 ensemble weights with runnable inference.

Status: initial release. The repository ships the curated dataset

• Tanimoto-controlled splits + supplementary tables / notes / figure, together with the paper-faithful reference architecture, training-step reference code, the exhaustive O(n²) Tanimoto leakage audit script that drove the v1.1 correction, **and the v1.0 + v1.1 ensemble weights
• runnable inference** (inference/predict.py). The deployed CardioSafe ensemble at platform.appliedscientific.ai/cardiosafe is continually updated with CRO-validated bioassay data and may differ from this paper snapshot. Questions: lukas@appliedscientific.ai or mihailo@appliedscientific.ai.

What's in the box

Path	Contents
data/compounds/compounds.csv	Canonical SMILES + standard InChI-keys for all 334,444 curated compounds
data/labels/labels_v1.csv + MANIFEST.json	The 8-task curated label matrix and its provenance manifest
data/splits/tan70.csv, tan60.csv	v1.0 splits — Tanimoto ≥ 0.70 / 0.60 cutoffs, as used in the bioRxiv preprint
data/splits/tan70_v1_1.csv, tan60_v1_1.csv	v1.1 splits — audit-clean (cardiac-cliff cluster fully force-routed to val). Test fold identical to v1.0. See data/splits/CHANGELOG.md.
data/supplementary/	Notes S1–S3, Tables S0–S9 (no S4), Figure S1, plus the per-split cliff manifests
data/comparators/	CToxPred2 and CardioGenAI predictions on the tan70 test fold — the inputs to the reverse-leak audit and the head-to-head comparison in Tables 3 / 3b / S2 / S3 / Figure 4
scripts/audit_tanimoto_leak.py	Exhaustive O(n_train × n_other) Tanimoto leakage audit — verifies no cross-fold Morgan-r2-2048 edges at or above the cutoff. See scripts/README.md.
model/	Reference architecture (CrossAttnIonChannelPredictor, ChemBERTa adapter, L1000 encoder) + MODEL_CARD.md.
inference/	Runnable inference: predict.py (CSV-in / CSV-out CLI), featurize.py, ensemble.py. Auto-downloads weights from Releases on first call. See inference/README.md.
train/	Paper-faithful loss functions and Stage 1 / Stage 2 training-step references. No data loaders; bring your own feature caches.
Releases	v1.0 + v1.1 ensemble weights (5 seeds each, 15 MB / seed) and the L1000 expression encoder (10 MB). Licensed CC-BY-NC-4.0 (see LICENSE-WEIGHTS).

See data/README.md for the schema, fold counts, and per-file documentation.

Quickstart

import pandas as pd

compounds = pd.read_csv("data/compounds/compounds.csv")
tan70     = pd.read_csv("data/splits/tan70.csv")
labels    = pd.read_csv("data/labels/labels_v1.csv")

# hERG 10 µM test fold (n ≈ 46,120, matching paper Table 2)
test_herg_10um = (
    labels[["row_idx", "herg_blocker_10um"]]
        .dropna()
        .merge(tan70.query("fold == 'test'"), on="row_idx")
        .merge(compounds[["row_idx", "smiles"]], on="row_idx")
)
print(len(test_herg_10um))

Splits

Two Tanimoto-controlled splits are released in two versions:

version	split	train	val	test	excluded
v1.0 (preprint)	tan70 (primary)	241,792	46,326	46,326	0
v1.0 (preprint)	tan60 (stricter)	306,665	13,889	13,889	1
v1.1 (audit-clean)	tan70	241,790	46,328	46,326	0
v1.1 (audit-clean)	tan60	306,662	13,892	13,889	1

Both versions force-route terfenadine and fexofenadine to val so the canonical cardiac activity cliff is available as a held-out case study. v1.1 additionally force-routes 2 hydroxymethyl-terfenadine analogs on tan70 and 3 on tan60 in the same cluster (caught by scripts/audit_tanimoto_leak.py); test fold is identical between versions. Use v1.1 for new work; v1.0 is retained so the bioRxiv preprint numbers stay reproducible. See data/splits/CHANGELOG.md and Note S3 (data/supplementary/note_s3_v1_1_audit_correction.md).

Labels

labels_v1.csv is a NaN-sparse (334,444 × 8) matrix:

col	head	type
herg_pchembl	regression	hERG pIC50
herg_blocker_10um	binary	hERG blocker @ 10 µM
herg_blocker_1um	binary	hERG blocker @ 1 µM
nav15_pchembl	regression	Nav1.5 pIC50
nav15_blocker	binary	Nav1.5 blocker @ 10 µM
cav12_pchembl	regression	Cav1.2 pIC50
cav12_blocker	binary	Cav1.2 blocker @ 10 µM
iks_blocker	binary	IKs blocker @ 10 µM (exploratory)

Per-channel counts (primary binary head):

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

Labels are derived from ChEMBL 36 (source dump SHA-256 b25820eef0f0481ad7712bdf4bac3b45f354e3cbacb76be1fdbf4205d6b48fb9) and the hERG Central primary screen, under a pharmacology-aware curation policy that retains censored values and inhibition-percentage votes. Full provenance is in data/labels/MANIFEST.json.

Supplementary materials

Everything the manuscript's Additional files section promises is in data/supplementary/, with three formats per table (CSV / Markdown / JSON) where applicable:

• Notes S1, S2, S3 — Y-randomization mechanism; activity-cliff curation provenance + 25-source bibliography + per-pair_id composition + per-split filtered cliff manifests; audit-driven v1.1 correction with retrain metrics and terfenadine / fexofenadine case-study re-evaluation
• 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 — the supplementary numbering jumps S3 → S5 in the manuscript)
• Figure S1 — hERG reliability curves across applicability-domain bins (PDF + PNG + JSON of the underlying decile data)

What's NOT in this repository

• L1000 raw signatures used to train the expression encoder are available from GEO under accessions GSE92742 (Phase I) and GSE70138 (Phase II). The trained encoder weights are released; the raw input signatures are not redistributed.
• ChEMBL 36 source dump (SHA-256 b25820eef0f0481ad7712bdf4bac3b45f354e3cbacb76be1fdbf4205d6b48fb9) is available from https://www.ebi.ac.uk/chembl/.
• Continually-updated production ensemble is served at platform.appliedscientific.ai/cardiosafe and may differ from this paper snapshot — it is retrained on new CRO-validated bioassay data routed through ASI's partner network.
• Comparator predictions (CToxPred2, CardioGenAI) used in the reverse-leak audit are reproducible from each comparator's released code on this repo's tan70 test fold. The aggregated paired-bootstrap comparison numbers are shipped in Tables S2 / S3.

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}
}

License

• Code in this repository is under the MIT License.
• Data (everything under data/) is released under Creative Commons Attribution 4.0 International.
• Model weights distributed via the GitHub Releases of this repository are released under Creative Commons Attribution-NonCommercial 4.0 International. Academic, educational, and non-profit research use is permitted with attribution. Commercial use requires a separate license — contact the authors (lukas@appliedscientific.ai, mihailo@appliedscientific.ai).