Predicting Molecular Interactions from Protein Binding Images

📌 Project Overview

This project presents a deep learning–based framework for predicting protein–ligand interactions using protein binding images derived from the Protein Data Bank (PDB).
The proposed approach employs a hybrid architecture combining 3D Convolutional Neural Networks (3D-CNN) and VGG16 to capture both spatial and structural features of protein–ligand complexes.

The model performs binary classification to determine whether a ligand binds or does not bind to a target protein.

---

🧬 Motivation

Protein–ligand interactions are fundamental to:

• Drug discovery
• Molecular medicine
• Protein engineering

Traditional methods such as X-ray crystallography, NMR spectroscopy, cryo-electron microscopy, and molecular docking are accurate but:

• Time-consuming
• Computationally expensive
• Difficult to scale

This project provides an automated, scalable, and AI-driven alternative using image-based deep learning techniques.

---

🏗️ Proposed Methodology

The system follows a dual-branch hybrid architecture:

🔹 3D-CNN Branch

• Input: 32 × 32 × 32 voxel grids
• Learns fine-grained volumetric and spatial features
• Captures atomic-level interactions in protein binding pockets

🔹 VGG16 Branch (Transfer Learning)

• Input: 224 × 224 protein binding images
• Pre-trained on PyMol
• Extracts high-level structural and visual features

🔹 Feature Fusion & Classification

• Features from both branches are concatenated
• Fully connected layers integrate information
• Output: Binding / Non-Binding

---

📂 Dataset

• Source: Protein Data Bank (PDB)
• Total samples: 120
  • Binding: 60
  • Non-Binding: 60
• Data representations:
  • 3D voxel grids
  • 2D protein binding images

⚠️ Due to size constraints, raw PDB files and processed voxel data are not included in this repository.

---

📊 Experimental Results

Model	Accuracy	Precision	Recall	F1-score
3D-CNN	66%	0.71	0.55	0.62
VGG16	50%	0.50	0.22	0.30
3D-CNN + VGG16 (Hybrid)	77.78%	0.85	0.78	0.77

✅ The hybrid model outperforms individual models across all evaluation metrics.

---

⚙️ Technologies Used

• Python
• TensorFlow / Keras
• NumPy
• OpenCV
• Biopython
• Scikit-learn
• Matplotlib
• PyMol
• Google Colab (GPU support)

---

👩‍💻 Team

• B Sai Swaroop
• Mythri J Reddy

📜 License

This project is licensed under the MIT License.

---

🚀 Future Scope

• Scaling to larger and diverse protein–ligand datasets
• Incorporating molecular dynamics and attention-based fusion
• Extending predictions to binding affinity estimation
• Improving generalization using advanced data augmentation