MedSafeAI

An AI-powered medication safety platform that predicts drug-drug interactions and adverse drug reactions using graph neural networks, molecular fingerprints, and large language models.

Live Demo: medsafe-ai-alpha.vercel.app

---

Overview

MedSafeAI is a full-stack clinical decision support tool built to help patients and caregivers make safer medication decisions. It offers two core capabilities:

• Polypharmacy Interaction Checker — Evaluates the safety of adding a new drug to a patient's existing medication list, returning a risk score, tier, and AI-generated explanation.
• Adverse Drug Reaction (ADR) Predictor — Predicts side effects for a single drug using a knowledge-graph-based model, grouped by body system category with confidence scores.

This project was developed as part of the S6 Mini Project — B.Tech Computer Science & Engineering.

---

Features

Feature	Description
⚠️ Polypharmacy Interaction Check	Detects dangerous drug combinations from a patient's medication list
🧬 ADR Side Effect Prediction	Predicts adverse reactions for a single drug by body system
📊 Safety Scoring	Returns a 0–100 safety score with risk tiers: ✅ SAFE / ⚠️ CAUTION / 🚫 AVOID
🤖 AI Explanations	Llama 3.3 70B (via Groq) generates plain-language clinical explanations
🔍 Drug Search	Fuzzy search endpoints for both polypharmacy and ADR drug databases
🖥️ Modern UI	React + Vite frontend with Tailwind CSS and Framer Motion animations

---

Tech Stack

Frontend (client/)

Technology	Version	Purpose
React	19	UI framework
TypeScript	5.9	Type safety
Vite	8	Build tool & dev server
Tailwind CSS	4	Styling
Framer Motion	12	Animations
Axios	1.13	API calls
React Router	7	Client-side routing

Backend (server/)

Technology	Version	Purpose
FastAPI	0.135	REST API framework
Uvicorn	0.41	ASGI server
PyTorch	2.10	Deep learning inference
PyTorch Geometric	2.7	Graph neural networks
RDKit	2025.9	SMILES parsing & molecular fingerprints
Transformers (HuggingFace)	latest	BioBERT embeddings
Groq SDK	0.28	LLM API (Llama 3.3 70B)
Scikit-learn	1.8	ML utilities
SHAP	0.51	Model explainability
Pydantic	2.12	Request/response validation

Datasets

Dataset	Usage
TWOSIDES	42M FDA adverse event reports for polypharmacy training
DrugBank	Drug metadata, protein targets, known interactions
SIDER	Drug side effect database for ADR model training
Reactome	Biological pathway data
PubChem	SMILES molecular structures

---

Project Structure

S6-MiniProject/
├── client/                        # React + Vite frontend (TypeScript)
│   ├── public/
│   ├── src/
│   │   ├── api/                   # Axios API call functions
│   │   ├── components/            # Reusable UI components
│   │   ├── pages/
│   │   │   ├── Landing.tsx        # Home / landing page
│   │   │   ├── PolypharmacyPredictor.tsx   # Drug interaction checker UI
│   │   │   └── ADRPredictor.tsx   # Side effect predictor UI
│   │   ├── types.ts               # TypeScript type definitions
│   │   ├── App.tsx
│   │   └── main.tsx
│   ├── package.json
│   └── vite.config.ts
│
└── server/                        # FastAPI backend (Python)
    ├── api.py                     # Main FastAPI app & route definitions
    ├── schemas.py                 # Pydantic request/response models
    ├── requirements.txt
    ├── models/
    │   ├── poly_model.py          # MLP model for polypharmacy (TWOSIDES)
    │   └── adr_model.py          # BioBERT + HeteroGCN for ADR prediction
    ├── services/
    │   ├── poly_service.py        # Fingerprint generation, scoring, Groq AI
    │   └── adr_service.py        # ADR inference & Groq AI explanation
    ├── training_scripts/          # Offline model training code
    ├── best_model.pt             # Trained polypharmacy MLP weights
    ├── adr_model.pt              # Trained ADR HeteroGCN weights
    ├── fingerprint_dict.pkl      # Precomputed Morgan fingerprints (TWOSIDES drugs)
    ├── top_effects.pkl           # Top adverse effects labels
    └── adr_mappings.pkl         # Drug/SE/target/pathway index mappings

---

Getting Started

Prerequisites

• Node.js v18+
• Python 3.10+
• pip
• A Groq API key (free tier available)

---

1. Clone the Repository

git clone https://github.com/niranjanasiju/S6-MiniProject.git
cd S6-MiniProject

---

2. Backend Setup

cd server

# Create and activate a virtual environment
python -m venv venv
source venv/bin/activate        # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Set up environment variables
# Create a .env file and add your Groq API key:
# GROQ_API_KEY=your_key_here

# Start the FastAPI server
uvicorn api:app --reload --host 0.0.0.0 --port 8080

The API will be available at: http://localhost:8080
Interactive API docs (Swagger UI): http://localhost:8080/docs

---

3. Frontend Setup

cd client

# Install dependencies
npm install

# Start the development server
npm run dev

The app will be available at: http://localhost:5173

---

API Reference

General

Method	Endpoint	Description
GET	/	Health check

Polypharmacy Interaction Checker

Method	Endpoint	Description
GET	/drugs/search?q={query}	Search polypharmacy drug database
POST	/predict	Check interaction safety for a drug combination

POST /predict — Example Request:

{
  "current_drugs": ["warfarin", "metformin"],
  "new_drug": "aspirin",
  "age": 65,
  "gender": "male"
}

Response includes: safety score (0–100), risk tier, predicted side effects per drug pair, and an AI-generated clinical explanation.

---

ADR Predictor

Method	Endpoint	Description
GET	/adr/drugs/search?q={query}	Search ADR drug database
POST	/predict-adr	Predict side effects for a single drug

POST /predict-adr — Example Request:

{
  "drug_name": "aspirin",
  "threshold": 0.5
}

Response includes: side effects by body system with confidence scores, biological protein targets, affected pathways, and an AI explanation.

---

Risk Tiers

Score	Tier	Meaning
≥ 70	SAFE	Low interaction risk
45 – 69	CAUTION	Moderate risk — consult a doctor
< 45	AVOID	High risk — potentially dangerous combination

---

🔬 How It Works

Polypharmacy Interaction Model

Drug A + Drug B (names)
        │
        ▼
Morgan Fingerprints (2048-dim, from fingerprint_dict.pkl)
        │
        ▼
Concatenated 4096-dim feature vector
        │
        ▼
MLP Classifier  ←  Trained on TWOSIDES (42M FDA adverse event reports)
        │
        ▼
Effect probabilities → Safety Score (0–100) + Risk Tier
        │
        ▼
Llama 3.3 70B via Groq → Plain-language clinical explanation

ADR Prediction Model

Drug name → SMILES (PubChem) → BioBERT embeddings
        │
        ▼
Heterogeneous Knowledge Graph
  Nodes : Drug · Side Effect · Protein Target · Biological Pathway
  Edges : treats · targets · in_pathway · causes (+ reverses)
        │
        ▼
HeteroGCN  ←  Trained on SIDER + DrugBank + Reactome
        │
        ▼
Side effect predictions by body system + confidence scores
        │
        ▼
Llama 3.3 70B via Groq → Clinical AI explanation

---

Disclaimer

MedSafeAI is an academic research project and is not intended for real clinical use. It should not be used as a substitute for professional medical advice, diagnosis, or treatment. Always consult a qualified healthcare provider before making any medication decisions.

---

License

This project is for academic purposes only. All rights reserved by the respective authors.