ToxPredict

Model Performance Comparision
Confusion Matrix
Final Model performance metrics
ROC CURVE(XGBOOST)

Technologies Used

Programming Languages

Python — core language for data processing, feature engineering, and model building

Libraries and Frameworks

Cheminformatics

RDKit — used to:
- Convert SMILES strings into molecular structures
- Generate Morgan fingerprints (2048-bit vectors)
- Compute molecular descriptors (e.g., molecular weight, logP)

Machine Learning

XGBoost — gradient boosting model used for classification
- Handles small datasets effectively
- Robust to class imbalance
- Provides feature importance

Data Processing

Pandas — dataset creation, cleaning, and manipulation
NumPy — numerical operations and vector handling

Explainability

SHAP (SHapley Additive exPlanations) — used to interpret model predictions
- Identifies which molecular features contribute to toxicity

Data Sources / APIs

FDA Adverse Event Reporting System (FAERS)
- Used as inspiration to identify pulmonary toxicity signals
PubChem
- Source for canonical SMILES and molecular structures
ChEMBL
- Used to expand dataset with real drug molecules programmatically

Model Pipeline

SMILES → RDKit → Morgan Fingerprints → XGBoost → Risk Score → SHAP Explanation

Frontend / Interface

Streamlit — lightweight web interface for:
- Inputting SMILES strings
- Displaying risk predictions
- Showing explanation outputs

Development Environment

Jupyter Notebook / Google Colab — experimentation and model training
Git & GitHub — version control and project management

Key Techniques Used

Molecular fingerprinting
Supervised classification
Class imbalance handling
Model explainability
Data curation and validation

Summary

The system integrates cheminformatics, machine learning, and explainability into a single pipeline that transforms raw chemical structures into interpretable lung toxicity predictions.