OpenADMET PXR Induction Challenge Dataset: sample notebook
A worked notebook for OpenADMET PXR Induction Challenge Dataset, with 4 code cells and the output each one printed. On Scigantic it opens in a hosted JupyterLab with the dataset already mounted, so nothing is downloaded first.
- Provider
- OpenADMET
- License
- apache-2.0
- Format
- CSV (in sampled files)
- Stored at
- Hugging Face (https://huggingface.co/datasets/openadmet/pxr-challenge-train-test)
- Access
- Free and public. No account with the data provider is needed.
The notebook
OpenADMET PXR Induction Challenge — Getting Started
This archive is the training/test data behind OpenADMET's blind challenge for predicting human Pregnane-X Receptor (PXR) induction — the largest public PXR activity dataset (11,000+ compounds screened). PXR induction matters because it drives CYP3A4 expression: a compound that induces PXR can silently speed up how fast the body clears co-administered drugs.
The dataset ships as 8 configs measuring different things at different scales. This notebook loads the actual modeling target and shows the one thing that trips people up: the primary screen looks like more potency data, but it isn't.
from datasets import load_dataset
import pandas as pd
ds = load_dataset("openadmet/pxr-challenge-train-test", "default")
train = ds["train"].to_pandas()
test = ds["test"].to_pandas()
print(f"train: {len(train)} rows, test: {len(test)} rows (SMILES + Molecule Name only)")
train.head()
Output:
train: 4139 rows, test: 513 rows (SMILES + Molecule Name only)
Molecule Name SMILES \
0 OADMET-0006089 CC1C2CCCCC2CN1C(=O)C1=CC(F)=CC2=C1CNCC2
1 OADMET-0006088 CN(C(=O)C1=CC=CC=C1SCC(=O)N1CCC2=CC=CC=C21)C1C...
2 OADMET-0006087 O=C(C1CCCC1)N1CC(S(=O)(=O)NC2CC2C2=C(F)C=CC=C2...
3 OADMET-0006086 CC1CN(C(=O)C2=CC=CC(C3=C(F)C=CC=C3F)=C2)CCN1C1...
4 OADMET-0006085 CC1(C2=NOC(C3=NN(CC4=CC=C(F)C=C4)C(=O)CC3)=N2)...
OCNT Batch pEC50 Emax_estimate (log2FC vs. baseline) \
0 OCNT-2318682-AA-004 5.27 1.67
1 OCNT-2318669-AA-004 5.08 1.70
2 OCNT-2318665-AA-004 5.12 1.88
3 OCNT-2318651-AA-004 5.23 2.20
4 OCNT-2318648-AA-004 5.28 1.82
...
The modeling target: potency + efficacy
pEC50 is induction potency (-log10(EC50 in M)); Emax_estimate (log2FC vs. baseline) is induction efficacy (how strong the response gets at saturation). Both ship with standard error and 95% confidence intervals — use them to down-weight noisy measurements instead of treating every row as equally certain.
import matplotlib.pyplot as plt
fig, axes = plt.subplots(1, 2, figsize=(11, 4))
train["pEC50"].dropna().hist(bins=30, ax=axes[0], color="steelblue")
axes[0].set_xlabel("pEC50"); axes[0].set_title("Induction potency")
train["Emax_estimate (log2FC vs. baseline)"].dropna().hist(bins=30, ax=axes[1], color="darkorange")
axes[1].set_xlabel("Emax (log2FC vs. baseline)"); axes[1].set_title("Induction efficacy")
plt.tight_layout()
Don't pool the primary screen in as more pEC50 data
single_concentration (21,003 rows) is a single-dose screen reported as log2 fold-change, not a dose-response pEC50 — a different readout on a different scale. It's the primary screen that decided which compounds got promoted to the dose-response assay above, so it's worth looking at, just not worth merging in as another potency column.
screen = load_dataset("openadmet/pxr-challenge-train-test", "single_concentration", split="train").to_pandas()
# Note: "Molecule Name" is mostly empty here — this config joins on OCNT_ID instead.
screen[["OCNT_ID", "log2_fc_estimate", "fdr_bh"]].head()
Output:
OCNT_ID log2_fc_estimate fdr_bh 0 OCNT-2312888 -0.060085 0.665857 1 OCNT-2310099 0.544938 0.000133 2 OCNT-2310557 -0.015018 0.912907 3 OCNT-2317350 0.678915 0.000033 4 OCNT-2314761 0.637109 0.000015
What the strongest inducers look like
from rdkit import Chem
from rdkit.Chem import Draw
def draw(smiles_list, legends, per_row=4):
mols, labels = [], []
for s, l in zip(smiles_list, legends):
m = Chem.MolFromSmiles(s)
if m is not None:
mols.append(m)
labels.append(l)
return Draw.MolsToGridImage(mols, molsPerRow=per_row, subImgSize=(220, 180), legends=labels)
top = train.dropna(subset=["pEC50"]).nlargest(8, "pEC50")
draw(top["SMILES"], [f"pEC50={v:.1f}" for v in top["pEC50"]])
Where to go next
- The blind challenge closed July 1, 2026 — phase_1_unblinded + phase_2_unblinded (253 + 260 rows) together reveal the 513 test labels.
- structure holds 78 fragment-sized molecules with paired X-ray crystal structures.
- counter_assay is the PXR-null specificity control, not a second potency readout — use it to flag non-specific hits, not to add training rows.
- Open this archive's schema card (the AI panel on the archive page) for the full column reference and every measured quirk.
Run it
Open this notebook to run it, see the dataset page for the full description, or browse the full catalog.