Predict toxicity. Save lives. Faster.

CodeCure Biohackathon · IIT BHU · Track A

Predict Drug Toxicity With AI Precision

Screen any molecule across 12 biological targets in milliseconds. Built on Tox21 — the gold standard toxicity dataset.

0Molecules Trained
0Assay Targets
0Best AUC

$2.6 Billion

Average cost of a failed drug due to unexpected toxicity

Years of Testing

Traditional toxicity testing takes 3-5 years per compound through animal studies and lab work

High Failure Rate

90% of drug candidates fail — toxicity is the #1 reason for late-stage failures

Massive Costs

Each failed compound wastes millions before toxicity is discovered

ToxPredict changes this.

Try It Live

Type any drug name and get instant toxicity predictions

vs
Quick test:

See It In Action

toxpredict.streamlit.app
Nitro group detected
Aromatic amine detected
Aspirin
Estradiol
Caffeine
Launch the full app to test any molecule

State-of-the-Art Performance

9 of 12 toxicity targets exceed the 0.75 industry benchmark threshold

0.75 Benchmark
0Best AUC Score
0Features Used
0Molecules Trained
0Models in Ensemble

Built Like a Real Pharmaceutical AI

We extract three complementary types of molecular features:

  • Morgan Fingerprints (2048-bit): Encode the circular atomic neighborhood structure — the same approach used in production pharmaceutical AI
  • MACCS Keys (166 features): 166 chemist-designed structural keys encoding specific molecular patterns
  • Toxicophore Flags (10 features): Binary detection of 10 known toxic chemical substructures

Total: 2,224 features per molecule

XGBoost
Random Forest
Logistic Regression
Soft Voting
Toxicity Probability

Soft voting combines probability outputs from all three models — this approach almost always outperforms any single model and is more robust to individual model weaknesses.

We use 5-fold Stratified Cross Validation to ensure our metrics are reliable and not overfitted to a single train/test split. This is the standard validation approach in published pharmaceutical ML research.

SMILES
Features
Models
Prediction

What Our Model Measures — And What It Doesn't

Scientific honesty is part of good science

What We Predict

  • Nuclear receptor disruption (NR-AR, NR-ER, etc.)
  • Cellular stress responses (SR-ARE, SR-HSE)
  • DNA damage signaling (SR-p53, SR-ATAD5)
  • Mitochondrial toxicity (SR-MMP)
  • Aryl hydrocarbon receptor activation (NR-AhR)

Outside Our Scope

  • Addiction and opioid receptor binding
  • Chronic low-dose toxicity (e.g. benzene → leukemia)
  • Teratogenicity (e.g. thalidomide birth defects)
  • Neurotoxicity and CNS effects
  • Organ-specific long-term toxicity
"Heroin scores low on our model — not because it's safe, but because its danger comes from opioid receptor binding, which Tox21 doesn't measure. A complete drug safety system requires multiple complementary assay panels. This is a known limitation of all Tox21-based models."— ToxPredict Team

Why This Matters — Real Drug Safety Failures

1957

Thalidomide

Prescribed to pregnant women for morning sickness — caused 10,000+ birth defects worldwide. Teratogenicity was not tested.

1997

Fen-Phen

Weight loss drug withdrawn after causing fatal heart valve damage in thousands of patients.

2004

Vioxx

Painkiller withdrawn — linked to 60,000+ cardiac deaths. Cardiovascular risk missed in trials.

Future

AI Prevention

AI toxicity screening like ToxPredict catches cellular danger signals before compounds reach human trials.

Built at CodeCure · IIT BHU Biohackathon

Track A — Drug Toxicity PredictionOrganized by IIT BHU
DB

Dhruv Bajpai

AI/ML Developer · CodeCure Biohackathon 2025

bajpaidhruv2018
PythonXGBoostRDKitSHAPStreamlitScikit-learnPlotly3Dmol.js