MolScreen: GNN for Molecular Toxicity Screening

MolScreen is a Graph Isomorphism Network with Edge features (GINEConv) trained on the Tox21 dataset for multi-task molecular toxicity prediction. It combines GNN-based predictions with gradient-based atom attribution and LLM-generated triage notes to make toxicity screening interpretable and actionable.

Developed by Veer Goradia.

Model Description

Drug toxicity is one of the leading causes of late-stage clinical trial failure, costing billions in wasted development. MolScreen addresses two key gaps in existing approaches: most models give a probability but no explanation, and none integrate a human-readable summary for non-expert decision makers.

Architecture:

• GINEConv backbone — uses edge features (bond type, aromaticity, conjugation) that plain GCN/GIN discard
• Multi-task classification across 12 Tox21 endpoints simultaneously
• Masked BCE loss for Tox21's missing labels (5,800–7,200 valid labels per task)
• Gradient-based atom attribution: highlights which molecular substructures drove the prediction
• LLM triage layer: generates a 3-sentence actionable summary (SMILES + flagged probabilities + top-3 attributed atoms per task)

Total parameters: ~374,000
Inference time: ~0.7 seconds end-to-end (excluding LLM API call)
Training time: ~6 minutes on laptop CPU (60 epochs)

Performance

Random Split (10-seed validation)

Metric	Value
Mean ROC-AUC	0.8442 ± 0.0073

Scaffold Split (10-seed validation)

Metric	Value
Mean ROC-AUC	0.7716 ± 0.0036

Baseline Comparison (Scaffold Split)

Model	ROC-AUC
MolScreen (GINEConv)	0.7716
Plain GCN	0.7481
Random Forest (Morgan FP)	0.7465
XGBoost (Morgan FP)	0.7378
Plain GIN	0.7249
Chen et al. SSL-GCN (2021)	0.7570

MolScreen beats the best non-pretrained baseline (GCN) by 2.4 points. The GINEConv edge-awareness advantage over plain GIN (+4.7 points) confirms that bond-level features matter for generalization.

Per-Task ROC-AUC (Random Split, representative run)

Endpoint	AUC
NR-AR	0.8518
NR-AR-LBD	0.9119
NR-AhR	0.8938
NR-Aromatase	0.8788
NR-ER	0.7486
NR-ER-LBD	0.8400
NR-PPAR-gamma	0.8918
SR-ARE	0.8125
SR-ATAD5	0.8762
SR-HSE	0.7374
SR-MMP	0.9156
SR-p53	0.8713

Training Data

• Dataset: Tox21 (7,823 valid compounds, 12 toxicity endpoints)
• Assay categories: Nuclear receptor panel (NR-) and stress response panel (SR-)
• Class imbalance: 2.6%–16% positive rate per task
• Splits: Random split (80/10/10) and scaffold split (chemotype-based, harder generalization test)

Intended Use

• Early-stage drug candidate toxicity screening
• Prioritizing which compounds advance to lab testing
• Interpretability analysis via atom-level attribution maps
• Research on GNN-based molecular property prediction

How to Use

import torch
from huggingface_hub import hf_hub_download

# Load model from HuggingFace
model_path = hf_hub_download(repo_id="vgoradia/MolScreen", filename="molscreen_best.pt")
model = torch.load(model_path, map_location='cpu')
model.eval()

# Or try the live Streamlit app — no code needed

Live Demo

Try the interactive app (draw a molecule, get toxicity predictions + atom attribution):
MolScreen Streamlit App

Publications & Acceptances

• NeurIPS 2026 AIDaR Workshop — ACCEPTED (poster, Paris, December 12 2026)
• IEEE BIBM 2026 AIPBDA Workshop — SUBMITTED (Paper ID S49203)
• Regeneron STS 2027 — In Progress (November 2026)
• AAAI-27 AISI Track — Submitted (Submission #598)

GitHub

github.com/vgoradia/mol-screen

Citation

@misc{goradia2026molscreen, title={MolScreen: Interpretable Multi-Task Molecular Toxicity Screening with Graph Neural Networks and LLM Triage}, author={Goradia, Veer}, year={2026}, note={NeurIPS 2026 AIDaR Workshop} }

License

Apache 2.0

Contact

Veer Goradia · vgoradia07@gmail.com