Drug–Disease Association Prediction

Hybrid Machine Learning + Multi-Modal Deep Neural Network (MM-DNN)

Kaggle Notebook

Open in Kaggle

This project predicts whether a drug can treat a specific disease using chemical structure features, protein targets, and disease identity.
We evaluate traditional machine learning baselines and propose a multi-modal deep neural network (MM-DNN) that achieves the best performance.

---

Overview

Drug repositioning aims to find new therapeutic uses for existing drugs, offering a faster and more cost-effective alternative to traditional drug development.

This project includes:

• Dataset construction with negative sampling
• Feature engineering for drugs and diseases
• Baseline ML models (LR, RF, XGBoost)
• A Two-Tower MM-DNN for nonlinear drug–disease interaction learning

---

Dataset

Source: https://www.kaggle.com/datasets/ariasha/drug-repositioning

File	Description
drugsInfo.csv	Drug SMILES, targets, metadata
diseasesInfo.csv	Disease names and categories
mapping.csv	Known drug–disease associations

Dataset size:

• 1410 drugs
• 1573 diseases
• 42,200 positive associations

---

Problem Definition

Binary classification task:

• Input: Drug Features (Chemical Structure) + Disease Features (ID/Embedding)
• Output:
  • 1: Associated (Effective)
  • 0: Not Associated

---

Feature Engineering

Drug Features

• Morgan Fingerprints (1024 bits) from SMILES
• Protein Target Multi-Hot Encoding

Disease Features

• Machine Learning models: One-Hot
• MM-DNN: Embedding Layer (learns semantic disease vectors)

---

Negative Sampling Strategy

The dataset contains only positive associations.
To create a balanced dataset:

• For each drug, sample diseases not associated with it
• Maintain a 1:1 ratio of positive/negative samples
• Prevents trivial predictions (e.g., predicting all 1's)

---

Models Implemented

Baseline Machine Learning Models

• Logistic Regression
• Random Forest
• XGBoost

Multi-Modal DNN (MM-DNN)

Two-Tower neural architecture:

• Drug Tower: Dense layers on fingerprint + target features
• Disease Tower: Embedding layer
• Fusion Layer: Concatenate + Dense for final prediction

---

Results

ROC Curve

Confusion Matrix (MM-DNN)

PCA Visualization

XGBoost Feature Importance

---

Model Performance (AUC)

Model	AUC
Logistic Regression	0.8396
Random Forest	0.8273
XGBoost	0.7350 — Lower (struggles with sparse FP)
MM-DNN (Ours)	0.8508 — Best