openadmet/ahr-induction

OpenADMET AHR Induction

Experimental measurements of aryl hydrocarbon receptor (AHR) induction, produced by OpenADMET from Octant Bio cell-based reporter assays. AHR is a ligand-activated bHLH-PAS transcription factor; agonists drive expression of CYP1A1/CYP1A2/CYP1B1 and related genes, which makes AHR activation relevant to drug-drug interactions and to off-target toxicity.

The dataset covers three assay tiers from the same screening campaign: a single-concentration primary screen across roughly 11k compounds, dose-response curves for the follow-up set, and a counter-screen on the same follow-up compounds.

• Source snapshot: Databricks Delta Sharing, openadmet.open_admet.2026_08_26_* (released 2026-08-26)
• Produced by: OpenADMET
• Generated by: notebooks/ahr.ipynb in the data-curation repository

Files

File	Rows	Compounds	Assay tier
ahr_single_concentration.csv	34,024	11,360	Primary screen, single concentration
ahr_drc.csv	1,001	1,001	Dose-response follow-up
ahr_counter_screen.csv	974	974	Counter-screen, dose-response

Compounds are identified by OCNT_ID, with OCNT Batch giving the specific batch tested. Every compound in ahr_drc.csv also appears in the single-concentration file, and every counter-screen compound appears in ahr_drc.csv.

Column names and order match the released PXR challenge data (pxr-challenge_TRAIN.csv, _counter-assay_TRAIN.csv, _single_concentration_TRAIN.csv), so the two releases can be read by the same code. PXR's Molecule Name is absent because AHR compounds have no OADMET IDs, and PXR's Split / source columns are omitted as artefacts of its train/test workflow. A small number of extra columns are appended; they are marked in the tables below.

Assay descriptions

Single concentration (ahr_single_concentration.csv)

The primary screen. Each compound was tested at a single concentration per run, across runs OPA_AHR-011, OPA_AHR-012, OPA_AHR-013 (assay AY-138, cell line c6800). Pooling the runs gives up to three concentrations per compound: 0.83 uM, 8.25 uM, 33 uM. These are measured concentrations, so they sit slightly below the nominal 1 / 10 / 30 uM of the three runs.

The readout is log2 fold change in reporter signal against the negative control (contrast == 'NEGATIVE', the filter the data generators define for these runs), with a standard error, confidence interval, p-value, Benjamini-Hochberg FDR and Cohen's d. There is no potency estimate at this tier -- a compound is simply active or not at the dose tested.

Dose-response curves (ahr_drc.csv)

Full concentration-response follow-up on compounds selected from the primary screen. Runs OPA_AHR-014, OPA_AHR-015, OPA_AHR-016 (assay AY-138 -- "AHR - AHR assay | 1536", cell line c6800), 9 doses spanning roughly 5 nM to 50 uM.

Each row is one compound with its fitted curve parameters: pEC50 (potency, -log10 M), Emax (maximal effect, as log2 fold change and relative to the positive control), slope, and maximum observed activity. Every parameter carries an estimate, standard error and 95% confidence interval.

Counter screen (ahr_counter_screen.csv)

Runs OPA_ACA-041 (assay AY-194 -- "ACA - AHR counter assay", cell line c7432). Same dose-response format and the same fitted parameters as ahr_drc.csv, run on the same compounds but in the counter-assay cell line.

A counter-screen distinguishes genuine AHR activation from signal that appears in the reporter readout for other reasons (for example cytotoxicity or direct effects on the reporter itself). Compare a compound's counter-screen response against its ahr_drc.csv response: a similar response in both suggests the primary signal is not AHR-specific.

Loading the data

With datasets (HuggingFace)

from datasets import load_dataset

drc = load_dataset("OpenADMET/ahr-induction", "drc", split="train")
counter = load_dataset("OpenADMET/ahr-induction", "counter_screen", split="train")
single = load_dataset("OpenADMET/ahr-induction", "single_concentration", split="train")

drc.to_pandas().head()

With pandas

import pandas as pd

drc = pd.read_csv("ahr_drc.csv")
counter = pd.read_csv("ahr_counter_screen.csv")
single = pd.read_csv("ahr_single_concentration.csv")

Straight from the Hub without cloning:

import pandas as pd

BASE = "https://huggingface.co/datasets/OpenADMET/ahr-induction/resolve/main"
drc = pd.read_csv(f"{BASE}/ahr_drc.csv")

Column reference

Names follow the PXR challenge release. Rows marked + are additions beyond PXR parity.

Dose-response and counter screen

Column	Description
OCNT_ID	Compound identifier
SMILES	Structure as supplied by the vendor
OCNT Batch	Batch ID of the material tested
pEC50	Potency, -log10(M)
Emax_estimate (log2FC vs. baseline)	Maximal effect, log2 fold change vs. baseline
Emax.vs.pos.ctrl_estimate (dimensionless)	Maximal effect relative to the positive control
pEC50_std.error (-log10(molarity))	Standard error on pEC50
Emax_std.error (log2FC vs. baseline)	Standard error on Emax
Emax.vs.pos.ctrl_std.error (dimensionless)	Standard error on Emax vs. positive control
pEC50_ci.lower / ci.upper (-log10(molarity))	95% CI on pEC50
Emax_ci.lower / ci.upper (log2FC vs. baseline)	95% CI on Emax
Emax.vs.pos.ctrl_ci.lower / ci.upper (dimensionless)	95% CI on Emax vs. positive control
+ activity_status	Octant's active/inactive call (yes / no). An activity call, not a QC verdict
+ pEC50_below_tested_range	Fitted pEC50 lies below the lowest dose tested. See Inactive compounds

Slope and maximum-activity parameters are fitted by the instrument but not released, matching PXR, which kept only pEC50 and Emax.

Single concentration

Column	Description
OCNT_ID	Compound identifier
OCNT Batch	Batch ID of the material tested
SMILES	Structure as supplied by the vendor
plate_id	Assay plate barcode
compound_class	Always LIBRARY; controls are filtered out
concentration_M	Concentration tested, molar
log2_fc_estimate, log2_fc_stderr	Reporter response vs. negative control, and its standard error
t_statistic	log2_fc_estimate / log2_fc_stderr
p_value, fdr_bh, neg_log10_fdr	Significance: raw, Benjamini-Hochberg corrected, and -log10(fdr_bh)
median_log2_fc	Median response across replicates
n_replicates	Replicate wells behind this measurement
cohens_d	Standardised effect size
experiment_name	Run the measurement came from
+ log2_fc_ci_lower, log2_fc_ci_upper	95% CI on the response. Reported by the source but absent from PXR

Processing notes

Read these before modelling.

Inactive compounds have extrapolated pEC50s

Compounds with activity_status == "no" still carry a fitted pEC50, but those values lie below the concentration range actually tested and are not measurements. The assay covers roughly pEC50 4.3 to 8.3; inactive fits cluster near 2.2, far weaker than the assay could observe. Treat them as "no activity detected up to the highest dose tested", not as potency estimates.

The pEC50_below_tested_range column flags every such row. It also catches 178 compounds marked active in ahr_drc.csv whose fitted EC50 still falls above the top dose. Nothing has been capped or removed -- decide the treatment that suits your use.

Quality filtering

Curves with no fitted pEC50 were dropped. CDD's QC_status / QC_flags fields are not exposed through Delta Sharing, and experiments.filters is empty for these runs; on the comparable PXR dataset a QC failure coincides exactly with a missing fit, so this reproduces that filter. activity_status is deliberately not used for filtering -- it reports activity, not data quality.

The qc_* columns reconstruct the CDD failure modes that can be recomputed, but their thresholds are not documented, so they are advisory only. In particular qc_ci_too_wide fires on most of the counter screen simply because inactive compounds have poorly constrained curves.

Replicate pooling

Where a compound has several curves, estimates are averaged and standard errors pooled in quadrature, with confidence intervals recomputed as estimate +/- 1.96 x pooled SE. Compounds whose replicate pEC50s disagreed by more than 1.5 log units were dropped. In this release every compound has a single curve (n_replicates == 1), so pooling changed nothing.

Structure flags

Structures are standardised and checked during curation — salts, mixtures, metal complexes, structures RDKit cannot process, and IDs mapping to more than one structure. 8 compounds are flagged, and all of them remain in the data: nothing was excluded on structural grounds.

The flag columns themselves are not exported, to keep the layout aligned with PXR, so the files carry no marker for these compounds. Re-run notebooks/ahr.ipynb section 8 to list them. Only the vendor-supplied SMILES is released; the standardised SMILES and InChIKey are used for the checks and then dropped.

Coverage

Only runs shared with OpenADMET are included. Several earlier AHR runs (assay AY-121, the Chembridge series, and the HTS onboarding pilots) exist in the experiment tracker but are not shared through Delta Sharing, so they are absent here.