gokulalgates/nrtoxpred-models

NR-ToxPred Models

Pre-trained machine learning models for predicting the binding activity of small molecules against nine human nuclear receptors (NRs).

These models are used by the NR-ToxPred GUI application — a desktop app that requires no coding experience.

---

What this repository contains

Folder	Contents
MODELS/morgan/	SVM classifiers trained on Morgan (ECFP6) fingerprints — one per receptor
MODELS/MACCS/	SVM classifiers trained on MACCS Keys — one per receptor
MODELS/ARclasses.npy	Label encoder (Active / Inactive)
X_train/	Training set SMILES used for Applicability Domain assessment

SuperLearner ensemble models are not included here due to their size (1–1.5 GB each).

---

Receptors covered

Receptor	Full Name
AR	Androgen Receptor
ERA	Estrogen Receptor Alpha
ERB	Estrogen Receptor Beta
FXR	Farnesoid X Receptor
GR	Glucocorticoid Receptor
PPARD	Peroxisome Proliferator-Activated Receptor Delta
PPARG	Peroxisome Proliferator-Activated Receptor Gamma
PR	Progesterone Receptor
RXR	Retinoid X Receptor

---

How to use

Option A — Desktop GUI (recommended, no coding needed)

Download the NR-ToxPred GUI from GitHub and run the installer. The app will download these models automatically on first launch.

👉 NR-ToxPred GUI on GitHub

Option B — Python (programmatic use)

from huggingface_hub import hf_hub_download
import pickle, numpy as np
from rdkit import Chem
from rdkit.Chem import AllChem

# Download a model
model_path = hf_hub_download(
    repo_id="gokulalgates/nrtoxpred-models",
    filename="MODELS/morgan/ARsvm_best.model",
    repo_type="model",
)

# Load model
model = pickle.load(open(model_path, "rb"))

# Generate Morgan fingerprint (ECFP6, 1024 bits)
mol = Chem.MolFromSmiles("CC(C)(c1ccc(O)cc1)c1ccc(O)cc1")  # bisphenol A
fp = AllChem.GetMorganFingerprintAsBitVect(mol, radius=3, nBits=1024)
X = np.array(fp).reshape(1, -1)

# Predict
label_enc = {0: "Inactive", 1: "Active"}
pred = model.predict(X)[0]
print(f"AR prediction: {pred}")

---

Model details

Property	Value
Algorithm	Support Vector Machine (SVM)
Fingerprints	Morgan ECFP6 (radius=3, 1024 bits) and MACCS Keys (167 bits)
Framework	scikit-learn 0.23.2
Task	Binary classification (Active / Inactive)
Applicability Domain	Tanimoto fingerprint similarity to training set

---

Applicability Domain

Each prediction comes with a reliability label:

• Reliable — the compound is similar (Tanimoto ≥ 0.25) to at least one training set compound
• Unreliable — the compound lies outside the training chemical space; interpret with caution

The X_train/ folder contains the training set SMILES used to compute these assessments.

---

Citation

If you use these models in your research, please cite:

Predicting the binding of small molecules to nuclear receptors using machine learning. Brief Bioinform. 2022 May 13;23(3):bbac114. doi: 10.1093/bib/bbac114

---

License

MIT License