AV_Algos

Drug Repurposing for Covid 19

Comment

🧪 AI-Driven Drug Repurposing & Molecular Generation

🌟 Inspiration

The COVID-19 pandemic exposed the urgent need for faster drug discovery methods. With no specific cure and new variants emerging rapidly, traditional drug development—which takes 5-10 years and millions of dollars—was not feasible.

We were inspired by:

  • The success of drug repurposing in past pandemics.
  • The advancements in AI & Machine Learning (ML) for computational drug discovery.
  • The potential of Generative Adversarial Networks (GANs) to design new drugs with higher efficacy.

This project was born from a desire to accelerate drug discovery using AI while keeping costs low and making treatments more accessible.

🔬 What It Does

Our solution integrates machine learning and AI-driven molecular generation to:

  1. Identify repurposed drugs that can inhibit the SARS-CoV-2 main protease.
  2. Predict the binding affinity of any given drug molecule using a Random Forest Regressor.
  3. Generate new molecular compounds with high binding affinity using a Generative Adversarial Network (GAN).
  4. Provide a user-friendly web application where researchers can:
    • Input a drug name or chemical formula
    • Get a predicted binding affinity score
    • Discover alternative drug candidates

🏗 How We Built It

1️⃣ Data Collection & Processing

  • Extracted data from ChEMBL and docking simulations.
  • Filtered and preprocessed 9001 molecules with valid SMILES representations.

2️⃣ Binding Affinity Predictor (Discriminator Model)

  • Used mol2vec embeddings to convert molecules into numerical fingerprints.
  • Developed a Random Forest Regressor trained on AutoDock Vina scores.
  • Evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and R² score.

3️⃣ Drug Design Using Generative Adversarial Network (GAN)

  • Designed a Stack-Augmented Recurrent Neural Network (RNN) to generate molecular structures.
  • Used GRU (Gated Recurrent Units) to handle sequential SMILES data.
  • Set hidden layers = 1500, depth = 200 layers, and learning rate = 0.01.
  • The discriminator (Random Forest Regressor) assessed their binding affinity.
  • Generated 10,000 molecules, out of which 6,321 were valid.

4️⃣ Web Application

  • Deployed a Flask/Django backend with a React frontend.
  • Allows users to search for a drug by name or chemical formula and retrieve predictions.

🚧 Challenges We Ran Into

  1. Handling Large Molecular Data

    • Processing thousands of molecules and converting them into meaningful vectors was computationally intensive.
    • Solution: Used high-performance computing (HPC) and optimized memory management.

  2. Optimizing the GAN Model

    • Initially, the model generated invalid molecular structures.
    • Solution: Fine-tuned the RNN layers, stack depth, and learning rate to improve molecular validity.

  3. Ensuring Model Generalization

    • The Random Forest Regressor needed to generalize well on unseen molecules.
    • Solution: Used cross-validation, hyperparameter tuning, and feature selection.

  4. Deploying the Web Application

    • Making the model accessible via a simple web interface required integrating ML predictions into a user-friendly UI.
    • Solution: Used Flask/Django for API and React for frontend.

🏆 Accomplishments That We're Proud Of

✅ Successfully trained a Random Forest Regressor to predict drug binding affinity.
✅ Developed a GAN model that can generate new drug-like molecules.
✅ Built a fully functional web application to make AI-driven drug discovery accessible.
✅ Generated 6,321 valid molecular compounds, some with higher predicted binding affinity than existing drugs.
✅ Optimized AI algorithms to work efficiently on large molecular datasets.

📚 What We Learned

🔬 Drug Discovery: How AI can accelerate molecular screening & drug repurposing.
💡 Molecular Fingerprinting: Using mol2vec embeddings for molecular structure representation.
🧠 Deep Learning: Training a GAN model for molecular generation.
⚡ High-Performance Computing (HPC): Handling large-scale molecular datasets.
🌐 Web Deployment: Integrating ML models into a real-world application.

🚀 What's Next for AV_Algos

🔹 Laboratory Testing: Synthesizing and testing AI-generated molecules for real-world efficacy.
🔹 Integration with High-Throughput Screening: Combining AI models with wet lab experiments to accelerate drug discovery.
🔹 Improved Molecular Generation: Exploring transformers instead of RNNs for better SMILES sequence generation.
🔹 Expansion to Other Diseases: Applying the same methodology to cancer, HIV, and other viral infections.
🔹 Enhancing the Web Application: Adding features like drug similarity searches and real-time docking simulations.

📌 Key Takeaway:
💡 AI-driven drug repurposing and molecular generation can significantly accelerate the discovery of new treatments for emerging diseases like COVID-19.
🌍 Our web app makes this technology accessible to researchers worldwide.

Built With

Share this project:

Updates