# OpenADMET CYP Inhibition Blind Challenge Dataset

## Facts

- Provider: OpenADMET
- License: apache-2.0
- Size: 5 CSV files, 6.7 MB total
- Records: default: 4,905 train + 750 test · tdi: 6,145 · single_concentration: 17,504 (4,376 compounds x 4 enzymes) · emax: 6,146
- Format: CSV (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/openadmet/cyp-challenge-train-test
- Access: public, readable without an account
- Category: Drug Discovery
- Source: https://openadmet.ghost.io/announcing-openadmets-cyp-inhibition-blind-challenge/
- Page: https://scigantic.com/archive/15b4d4e6-d24e-5344-9789-9501cded9c3c
- Keywords: Drug Discovery, Hugging Face

## About

Training and blinded test data for the OpenADMET CYP Inhibition Blind Challenge (submissions open August 17 2026, close November 3 2026), covering direct and time-dependent inhibition of the four major drug-metabolizing cytochrome P450 isoforms: CYP1A2, CYP2C9, CYP2D6 and CYP3A4. CYPs clear most marketed small molecules, so unexpected inhibition causes drug-drug interactions, and mechanism-based (time-dependent) inhibition causes interactions that outlast the parent compound. Five CSV files across four HuggingFace configs. The default config holds the direct-inhibition dose-response training set (4,905 compounds with pIC50 plus 95% interval and standard error per isoform) and the 750-compound blinded test set, which ships SMILES and Molecule_Name only. The tdi config holds 6,145 compounds with is_TDI classification labels for CYP2D6 and CYP3A4 plus paired direct-arm and pre-incubation-arm pIC50s. The single_concentration config holds the primary screen that decided which compounds were promoted to dose-response: 17,504 measurements over 4,376 compounds against all four enzymes at one concentration (49.5 uM), reported as log2 fold-change in probe turnover with FDR. The emax config holds 6,146 compounds with maximal-effect values and, unlike the tdi config, is_TDI labels for all four isoforms. The training label matrix is deliberately sparse: compounds went to a full dose-response curve only for the isoform where they hit the primary screen, so 3,596 of 4,905 compounds carry exactly one of the four pIC50 labels and only 41 carry all four, while the test set was assayed densely against all four. The test set is a purchased analog expansion of confirmed screening hits, which makes it chemically adjacent to the training set without being a memorization exercise: 76% of test compounds carry a Murcko scaffold absent from training, yet 96% have a training neighbor at ECFP4 Tanimoto 0.5 or better. Two tracks: regression on the four pIC50 endpoints scored by macro-averaged soft-threshold RAE, where a prediction landing inside the fitted curve's credible interval incurs zero error, and binary TDI classification on CYP3A4 and CYP2D6 scored by Matthews correlation coefficient. Data produced by OpenADMET with Octant Bio; Apache-2.0. The two official tutorial notebooks from OpenADMET/CYP-Challenge-Tutorial ship as this archive's starters, rewritten to run on the pre-baked image with no installs.

## Load the data

```python
import pandas as pd

# The HF repo is FUSE-mounted read-only; reading the CSV in place avoids the datasets cache.
ROOT = "/mnt/archive"
train = pd.read_csv(f"{ROOT}/cyp-challenge-TRAIN_inhibition.csv")   # 4,905 x 18
test = pd.read_csv(f"{ROOT}/cyp-challenge-TEST-BLINDED.csv")        # 750 x 2, SMILES only

CYPS = ["CYP1A2", "CYP2C9", "CYP2D6", "CYP3A4"]
labels = [f"{c}_pIC50_direct_inhibition" for c in CYPS]

# The sparsity is the first thing to look at: most compounds have ONE of the four.
print(train[labels].notna().sum())                              # 1412 / 1285 / 1493 / 2335
print(train[labels].notna().sum(axis=1).value_counts().sort_index())  # 1:3596  2:1039  3:229  4:41
```

Written for a Scigantic notebook, where /mnt/archive (SCIGANTIC_MOUNT_PATH) is the mounted dataset.

## Data structure

File types in sampled files: CSV

| Field | Type |
| --- | --- |
| `Molecule_Name` | string  (OCNT-####### Octant compound id; the submission join key. Unique per row in every file except… |
| `SMILES` | string  (no salts or mixtures anywhere: 0 of 5,655 contain a "."; 43 train SMILES carry explicit charges; RDKit parses… |
| `{CYP}_pIC50_direct_inhibition` | float  -log10(IC50 in M), no pre-incubation. THE regression label, one column per isoform. Non-null: CYP1A2 1,412 ·… |
| `{CYP}_pIC50_direct_inhibition_conf_low / _conf_high` | float  95% interval on the curve fit. Width == 3.92 x _std (median), so it is a SYMMETRIC Wald interval, not an… |
| `{CYP}_pIC50_direct_inhibition_std` | float  standard error of the fitted pIC50. Median 0.13-0.24 by isoform; use it to down-weight noisy curves |
| `{CYP}_pIC50_TDI_condition (+ _conf_low/_conf_high/_std)` | float  same fit after active pre-incubation with the CYP and NADPH. tdi config only |
| `CYP2D6_is_TDI / CYP3A4_is_TDI` | bool  time-dependent-inhibition label, the classification target. Non-null 1,497 / 3,584. Positive 324 (21.6%) / 764… |
| `CYP1A2_is_TDI / CYP2C9_is_TDI` | bool  emax config ONLY, absent from the tdi config. Positive 23 of 1,414 (1.6%) / 38 of 1,285 (3.0%). Not scored by the… |
| `{CYP}_EmaxVsPosCtrl_direct_inhibition / _TDI_condition (+ _conf_low/_conf_high)` | float  plateau effect as a fraction of the positive control, ~-1.0 == full inhibition. Render as % inhibition with -100… |
| `enzyme` | string  single-concentration only: CYP1A2 \| CYP2C9 \| CYP2D6 \| CYP3A4, exactly 4,376 rows each |
| `concentration_M` | float  single-concentration only, CONSTANT 4.9505e-05 (49.5 uM) across all 17,504 rows. Carries no information |
| `log2fc_estimate / log2fc_median` | float  log2 fold-change in probe turnover vs. negative control. NEGATIVE == inhibition. The two correlate at r=0.9998… |
| `log2fc_fdr / p_value` | float  Benjamini-Hochberg FDR and raw p on the per-plate contrast. 78.2% of rows are FDR < 0.05 |
| `log2fc_std_error / cohens_d` | float  PLATE-LEVEL CONSTANTS, not per-compound statistics — exactly 36 distinct values across 17,504 rows, one per… |
| `plate_id` | string  36 plates: 9 per enzyme, no plate shared between enzymes |
| `OCNT_Batch` | string  physical batch, e.g. OCNT-0018684-AQ-003. Exactly one batch per compound, so it adds nothing over Molecule_Name |

Files:

```text
HuggingFace dataset, FUSE-mounted read-only at /mnt/archive (symlinked to ~/archive). Five flat CSVs at the repo root, exposed as four configs via load_dataset("openadmet/cyp-challenge-train-test", "<config>"). Reading the CSV off the mount with pandas is equivalent and avoids the datasets cache.
  default / train            cyp-challenge-TRAIN_inhibition.csv        4,905 rows x 18 cols. Direct-inhibition pIC50 for 4 isoforms. THE regression target.
  default / test             cyp-challenge-TEST-BLINDED.csv              750 rows x  2 cols. Molecule_Name + SMILES only. Labels withheld until the challenge closes.
  tdi / train                cyp-challenge-TRAIN_TDI.csv               6,145 rows x 36 cols. is_TDI for CYP2D6/CYP3A4 + both assay arms' pIC50 for all 4.
  single_concentration/train cyp-challenge-single-concentration-TRAIN.csv 17,504 rows x 12 cols. Primary screen, long format (compound x enzyme).
  emax / train               cyp-challenge-TRAIN_Emax.csv              6,146 rows x 30 cols. Maximal effect + is_TDI for ALL FOUR isoforms.
Compound-set nesting: single_concentration (4,376) subset-of default/train (4,905) subset-of tdi (6,145) ~= emax (6,146, one extra compound OCNT-2329004). The 750 test compounds appear in none of them.
```

Example file: `pd.read_csv("/mnt/archive/cyp-challenge-TRAIN_inhibition.csv")`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/15b4d4e6-d24e-5344-9789-9501cded9c3c
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/15b4d4e6-d24e-5344-9789-9501cded9c3c
- Over MCP: `get_schema_card` with id `15b4d4e6-d24e-5344-9789-9501cded9c3c` at https://api.scigantic.com/mcp

## Guide

- [How to open the OpenADMET CYP inhibition challenge data in a Jupyter notebook](https://scigantic.com/guides/open-openadmet-cyp-challenge-data)

## Related topics

- [Drug Discovery datasets](https://scigantic.com/topics/drug-discovery)
