Datasetpublic

mushroom2368/opioid-receptor-ordinal

Opioid receptor ordinal prediction — handover package Predicting ligand activity at the four opioid receptors (mu, delta, kappa, NOP) by decomposing each endpoint into monotone ordinal thresholds rather than regressing a value. The encoding admits censored records, keeps incommensurable endpoints separate, and puts the decision boundary in the model output. What is here Path Contents HANDOVER.md Read this first — environment, structure, known pitfalls… See the full description on the dataset page: https://huggingface.co/datasets/mushroom2368/opioid-receptor-ordinal.

sourceHugging Facecc-by-4.0updated 19d agoView on Hugging Face

0likes53downloads

Dataset Card

Opioid receptor ordinal prediction — handover package

Predicting ligand activity at the four opioid receptors (mu, delta, kappa, NOP) by decomposing each endpoint into monotone ordinal thresholds rather than regressing a value. The encoding admits censored records, keeps incommensurable endpoints separate, and puts the decision boundary in the model output.

What is here

Path	Contents
`HANDOVER.md`	Read this first — environment, structure, known pitfalls
`scripts/`	85 numbered scripts, roughly in execution order
`DrugBAN/`	Model code, config, and the `ord_base` checkpoint
`data/processed/`	Curated training data (20,760 measurements, 9,838 compounds)
`data/reports/`	Manuscript, per-section notes, candidate tables
`figures/paper_v2/`	44 figures, PNG and PDF at 600 dpi

Reproducibility

bash

python3 scripts/81_reproduce_check.py    # six checks, all should pass

Verified: the check passes 6/6 inside this package, standalone.

Check	Value	Paper
Profile AUROC, strong vs weak kappa	0.9571	0.957
Cross-receptor Spearman (median)	0.5739	0.574
Shuffled-receptor control	0.2382	0.238
tau predicted vs measured	0.5136	0.514
Ladder monotonicity violations	0	0

Headline result and its limit

Given eight non-opioid GPCRs the model never saw, within-target Spearman between predicted and measured pKi is 0.574, against 0.238 when the same ligands are paired with the wrong receptor — so the model reads the sequence, and 0.310 of the correlation is attributable to the protein rather than to the ligand alone.

The number that governs use is the other one: AUROC is 0.957 in distribution and 0.642 outside it. Ranking a 141,217-compound library lifts the hit rate from 13.9% to 25% in the top 100. That is worth an experiment; it is not a discovery.

Environment

Do not upgrade torch or dgl.

python 3.10.14 · torch 2.2.1 · dgl 2.1.0+cu121 · rdkit 2022.09.5

Data provenance

ChEMBL (activity), COCONUT (natural products), RCSB PDB (crystal contacts). All values in data/reports/ are model predictions unless explicitly marked as measured. No candidate reported here has been tested experimentally.