🔮 Drug Side Effect Predictor

This project aims to predict the potential side effects of a drug molecule based on its name (encoded chemical structure) and its biological targets. It utilizes Natural Language Processing (NLP) techniques and multi-label supervised learning.

📋 Table of Contents

1. Project Overview
2. Repository Structure
3. Installation
4. Data Pipeline
5. Modeling and Performance
6. Dashboard Usage
7. Contribution of each member

🚀 Project Overview

The project merges data from STITCH, SIDER (MedDRA), and DrugBank to create a comprehensive dataset. This dataset associates:

• Drug names.
• Anatomical Therapeutic Chemical (ATC) codes.
• Biological targets.
• Indications and side effects.

An optimized Logistic Regression (OneVsRest) model is trained to predict the presence or absence of the 300 most frequent side effects.

📂 Repository Structure

• Fusionetconcat.py: Script for initial merging of STITCH and DrugBank databases, featuring XML parsing and Fuzzy Matching.
• Firstcleaning.py: Script for cleaning and formatting ATC codes.
• data_cleaning.ipynb: Exploratory data analysis and final dataset cleaning.
• word_embedding.ipynb: Data encoding (TF-IDF on character n-grams) and model training/comparison.
• dashboard.py: Interactive Dash application to test the model in real-time.
• models_export/: Directory containing trained models and vectorizers in .pkl format.

🛠 Installation

1. Clone the repository.
2. Install the necessary dependencies:

pip install pandas numpy scikit-learn dash joblib matplotlib plotly rapidfuzz

3. Ensure you have the source data files or the generated cleaned_drug_side_effects.csv.

📊 Data Pipeline

1. Preprocessing

• XML Parsing: Extraction of targets and descriptions from the DrugBank database.
• Fuzzy Matching: Reconciling drug names between STITCH and DrugBank using the token_sort_ratio algorithm.
• Cleaning: Removal of missing values and filtering of non-exploitable columns such as IDs and long descriptions.

2. Encoding (Embedding)

• Biological Targets: Binary encoding via MultiLabelBinarizer.
• Molecule Names: Use of TF-IDF on characters (n-grams of 3 to 5 letters) to capture chemical roots like -zole or -meth.

🤖 Modeling and Performance

The project compared several approaches:

• Baseline (Logistic Regression): Optimized via GridSearchCV.
• Random Forest.
• Voting Classifier.

Result: The Logistic Regression model achieved the best performance with a Micro F1-Score of approximately 0.49 on the top 300 side effects. This is highly significant for a multi-label problem of this complexity.

🖥 Dashboard Usage

To launch the prediction interface:

python dashboard.py

Once launched, access http://127.0.0.1:8050/ in your web browser. You can:

1. Enter a molecule name (e.g., Aspirin).
2. Select one or more biological targets from the dropdown menu.
3. Instantly receive a list of potential side effects detected by the model.

---

👥 Contibutions of each member

Ghita Fadili : Dataset selection and merging, Scientific Papers & References
Thomas Jego : Model training, Ensemble models, Voting, Model Comparison
Victoire Dezombre : Data Exploration & Data Cleaning
Léandre durand-Terasson : Word Embedding (TF-IDF, syntactic similarity, MultiLabelBinarizer)