rasayan-labs/rasayan-tox21-snn

Demo

Open Interactive Demo →

Rasayan Tox21 SNN Ensemble

A Self-Normalizing Neural Network (SNN) ensemble for predicting molecular toxicity across 12 Tox21 Challenge endpoints. Trained on the NIH Tox21 dataset with extensive feature engineering and rigorous cross-validation.

Model Description

This model predicts the probability of a molecule being active (toxic) against 12 biological targets from the Tox21 Challenge:

Nuclear Receptor Panel

Endpoint	Target	Description
NR-AR	Androgen Receptor	Male reproductive toxicity
NR-AR-LBD	AR Ligand Binding Domain	Direct AR modulation
NR-AhR	Aryl Hydrocarbon Receptor	Dioxin-like toxicity
NR-Aromatase	CYP19A1	Estrogen synthesis disruption
NR-ER	Estrogen Receptor	Endocrine disruption
NR-ER-LBD	ER Ligand Binding Domain	Direct ER modulation
NR-PPAR-gamma	PPARγ	Metabolic disruption

Stress Response Panel

Endpoint	Target	Description
SR-ARE	Antioxidant Response Element	Oxidative stress
SR-ATAD5	ATAD5	DNA damage response
SR-HSE	Heat Shock Element	Protein folding stress
SR-MMP	Mitochondrial Membrane Potential	Mitochondrial toxicity
SR-p53	Tumor Protein p53	Genotoxicity

Architecture

Component	Specification
Type	Self-Normalizing Neural Network
Ensemble	10 models (top from 40-fold CV)
Hidden Layers	8 layers × 768 units
Activation	SELU
Regularization	AlphaDropout (0.1)
Output	Sigmoid (12 endpoints)
Parameters	~19M total

Molecular Features (11,377 dimensions)

Feature Type	Dimensions	Description
ECFP6	8,192	Extended-connectivity fingerprints (radius 3)
MACCS Keys	167	Structural keys for substructure screening
RDKit Descriptors	208	Physicochemical properties
Toxicophores	1,868	SMARTS-based toxicity alerts
Structural Filters	815	PAINS, BRENK, NIH, ZINC alerts
Target Similarity	127	Tanimoto similarity to known ligands

Training

Parameter	Value
Dataset	Tox21 Challenge (7,831 compounds)
Validation	40-fold Stratified CV
Epochs	300
Batch Size	256
Optimizer	AdamW (lr=1e-4, weight_decay=0.01)
Loss	Focal Loss (γ=2.5, α=0.25)
Regularization	Label Smoothing (0.1), Mixup (α=0.2)
CV AUC	0.882 ± 0.021

Usage

With the Inference API

import requests

response = requests.post(
    "https://rasayan-labs-rasayan-tox21.hf.space/predict",
    json={"smiles": ["CCO", "c1ccccc1"]}
)

predictions = response.json()["predictions"]
for smiles, scores in predictions.items():
    print(f"{smiles}:")
    for target, prob in sorted(scores.items(), key=lambda x: -x[1])[:3]:
        print(f"  {target}: {prob:.1%}")

Direct Model Loading

import torch
import json

checkpoint = torch.load("ensemble.pt", map_location="cpu")

scalers = checkpoint["scalers"]
feature_indices = checkpoint["feature_indices"]
models = checkpoint["models"]

print(f"Loaded {len(models)} ensemble members")

Files

File	Description
ensemble.pt	PyTorch checkpoint with 10 models + scalers
config.json	Model configuration
toxicophores_validated.json	1,868 toxicophore SMARTS patterns
target_ligands_validated.json	Reference ligands for similarity

Intended Use

This model is intended for:

• Early-stage drug discovery toxicity screening
• Prioritization of compounds for experimental testing
• Educational purposes in computational toxicology

Limitations

• Trained on Tox21 assay data which may not capture all toxicity mechanisms
• Performance may vary for chemical spaces outside the training domain
• Should not replace experimental validation

Citation

@misc{rasayan-tox21-2026,
  author = {Rasayan Labs},
  title = {Rasayan Tox21 SNN Ensemble},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/rasayan-labs/rasayan-tox21-snn}
}

License

Apache 2.0

---

Built by Rasayan Labs