Biological Activity Prediction

An interactive Streamlit application and supporting machine learning assets for predicting the biological activity of compounds sourced from ChEMBL data. The project bundles cleaned datasets, pre-trained models (Gaussian Naive Bayes, XGBoost, and an Artificial Neural Network), and notebooks that document the full training and evaluation workflow.

Features

• Pre-trained models ready for immediate inference through the Streamlit UI.
• Batch predictions from user-supplied CSV files with automatic preprocessing and normalization.
• Visual insights including distribution plots and descriptive statistics for prediction confidences.
• Reproducible notebooks covering data preprocessing, model training, and comparative evaluation.
• Sample CSV files and a demo video to help you get started quickly.

Repository Layout

.
|- data/                  # Raw and split datasets derived from ChEMBL
|- models/                # Saved models, scaler, and development notebooks
|- test/                  # Example CSV inputs for quick experimentation
|- webapp/app.py          # Streamlit application entry point
`- DemoVideo.mp4          # Walkthrough of the deployed application

Requirements

• Python 3.9 or newer
• Recommended packages: streamlit, pandas, numpy, scikit-learn, joblib, tensorflow, matplotlib, seaborn, xgboost

Create and activate a virtual environment (optional but encouraged) and install dependencies:

python -m venv .venv
# Windows
. .venv\Scripts\activate
# macOS / Linux
source .venv/bin/activate

pip install streamlit pandas numpy scikit-learn joblib tensorflow matplotlib seaborn xgboost

If you maintain a requirements.txt, you can install everything at once with pip install -r requirements.txt.

Running the Streamlit App

1. Ensure the model files in models/ are accessible to the app. The current script loads them via absolute paths; update webapp/app.py so that MODEL_DIR points to your local models folder, for example:

   from pathlib import Path
   
   MODEL_DIR = Path(__file__).resolve().parents[1] / "models"
   scaler = joblib.load(MODEL_DIR / "scaler.pkl")
   gnb_model = joblib.load(MODEL_DIR / "gaussian_nb.pkl")
   xgb_model = joblib.load(MODEL_DIR / "xgboost.pkl")
   ann_model = load_model(MODEL_DIR / "ann.h5", compile=False)

2. Launch the app:

   streamlit run webapp/app.py

3. Open the URL shown in the terminal (defaults to http://localhost:8501).

Input Data Format

Your CSV file should at minimum include the columns shown below (see test/sample*.csv for ready-made examples):

Column	Description
Compounds	Numeric representation of compound-related measurements
Type	Categorical type (e.g., CELL-LINE, SMALL MOLECULE, etc.)
Tax ID	NCBI taxonomy identifier for the organism
Organism	Organism name matching the taxonomy identifier

Missing values are automatically filled with column means, categorical fields are label-encoded on the fly, and numeric columns are scaled using the stored scaler.pkl.

Notebooks

The models/ directory contains three Jupyter notebooks that document the end-to-end pipeline:

• data_preprocessing.ipynb - Cleaning, feature engineering, and train/test splitting.
• model_training.ipynb - Training GaussianNB, XGBoost, and ANN models.
• evaluate_models.ipynb - Benchmarking performance metrics and visual diagnostics.

Run them in order within a Jupyter environment if you need to retrain or extend the models.

Demo & Testing Assets

• DemoVideo.mp4 showcases the user experience and typical workflow.
• test/ includes several CSV files you can upload directly into the app to verify predictions.

Troubleshooting

• Model path errors: Ensure the absolute or relative paths in webapp/app.py match your local directory layout.
• Missing dependencies: Re-run the pip install command above and confirm your virtual environment is active.
• Streamlit port conflicts: Use streamlit run webapp/app.py --server.port 8502 to switch to a different port.

Next Steps

Possible extensions include integrating a REST API, refining feature engineering, or packaging the project with Docker for easier deployment.