Prediction of Toxicity using the Tox21 Dataset

This repository represents a reproducible machine learning pipeline that predicts molecular toxicity across 12 different biological targets using the Tox21 dataset. Each target is independently modelled as a binary classification task, with models trained on engineered descriptors, and interpreted using SHAP values for mechanistic insight.

Description of the Dataset.

The Tox21 dataset is a benchmark toxicology dataset jointly released by the NIH, EPA and FDA. It contains:

• SMILES representations of ~8000 compounds.
• Binary activity labels, for 12 targets:
  • Nuclear Receptor Pathways: NR-AR, NR-AR-LBD, NR-AhR, NR-ER, NR-ER-LBD, NR-PPAR-γ
  • Stress Response Pathways: SR-ARE, SR-ATAD5, SR-HSE, SR-MMP, SR-p53

Target	Full Name	Pathway Type	Biological Relevance
NR-AR	Androgen Receptor	Nuclear Receptor	Regulates development of male characteristics; key in prostate cancer and endocrine disruption.
NR-AR-LBD	Androgen Receptor Ligand Binding Domain	Nuclear Receptor	Focuses on ligand binding effects; disruption can indicate endocrine interference.
NR-AhR	Aryl Hydrocarbon Receptor	Nuclear Receptor	Mediates xenobiotic metabolism; activated by dioxins and environmental toxins.
NR-ER	Estrogen Receptor	Nuclear Receptor	Regulates female reproductive system; critical in breast cancer and hormonal balance.
NR-ER-LBD	Estrogen Receptor Ligand Binding Domain	Nuclear Receptor	Tests direct binding to ER; useful in detecting estrogenic or anti-estrogenic effects.
NR-PPAR-γ	Peroxisome Proliferator-Activated Receptor Gamma	Nuclear Receptor	Involved in lipid metabolism, insulin sensitivity, and adipocyte differentiation.
SR-ARE	Antioxidant Response Element	Stress Response	Measures activation of Nrf2 pathway; key in cellular defense against oxidative stress.
SR-ATAD5	ATPase Family AAA Domain-Containing Protein 5	Stress Response	Involved in DNA replication stress response; linked to genomic stability and repair.
SR-HSE	Heat Shock Element	Stress Response	Monitors heat shock response; triggers expression of chaperone proteins under stress.
SR-MMP	Mitochondrial Membrane Potential	Stress Response	Indicates mitochondrial integrity; disruption implies apoptosis or metabolic damage.
SR-p53	Tumor Protein p53	Stress Response	Guardian of the genome; regulates DNA damage response, cell cycle, and apoptosis.

Objectives of the Project

• Performing robust exploratory and statistical analysis of toxicity labels and molecular descriptors

• Handling missing data using appropriate imputation strategies (KNN and MICE)

• Constructing informative molecular descriptors using cheminformatics tools (e.g., RDKit)

• Training individual binary classification models for each toxicity endpoint

• Generating model explanations using SHAP to identify key drivers of toxicity

• Exporting reproducible models for deployment and future predictions

Overview of the Pipeline

Exploratory Data Analysis (EDA)

• Visualized distribution of different molecular features and class imbalances.

• Utilized histograms, boxplots, scatterplots and correlation matrices to analyze inter-target label relationships.

Data Cleaning and Imputation

• Applied K-Nearest Neighbors (KNN) imputation and Multivariate Chain Equations (MICE) to address missing molecular descriptors and labels

• KNN imputation leverages the principle that structurally or physicochemically similar compounds tend to exhibit comparable bioactivity and toxicity profiles.

• MICE captures the inherent correlations among molecular properties (e.g., LogP, TPSA, H-bonding), aligning with how biological activity arises from the combined influence of multiple features.

• By conditioning each imputed value on all others, MICE maintains biologically plausible descriptor combinations that reflect realistic ADME behavior.

• Such entries often represent chemically atypical or non-druglike molecules and may introduce noise, undermining both model fidelity and biological insight.

Feature Engineering

• Derived descriptors from SMILES using RDKit: Molecular weight, TPSA, LogP, H-bond donors/acceptors, rotatable bonds, etc.

• Normalized and transformed of continuous features

Model Development

• Modeled each of the 12 toxicity tasks as an independent classification problem

• Used ensemble-based classifiers (Random Forest, XGBoost) with stratified sampling

• Conducted hyperparameter tuning via GridSearchCV

• Models trained using balanced class weights or SMOTE

Model Evaluation

• Metrics: ROC-AUC, F1-score, Precision, Recall, Accuracy

• Outputs saved as classification reports and confusion matrices.

Model Explainability

Applied SHAP (SHapley Additive exPlanations) to all models

Dependencies:

pip install -r requirements.txt

Future Work

• Extend to multi-label architectures

• Integrate graph-based molecular encodings (e.g., GNNs)

• Compare performance against deep learning baselines

• Publish models via web API or Streamlit