DeepFusion-DDG: Protein Stability Prediction

Deep learning models for predicting protein thermodynamic stability changes (ΔΔG) from single-point mutations using sequence and structure information.

Project Overview

This project develops fusion models combining:

• Sequence-based models: ESM2 (Evolutionary Scale Modeling)
• Structure-based models: Graph Neural Networks (GNNs)
• Baseline models: Traditional ML approaches

Dataset: FireProtDB filtered dataset with 141 proteins, 4,291 single-point mutations

Environment Setup

This project uses uv for fast, reproducible dependency management.

Prerequisites

• Python 3.10+
• uv package manager

Installation

# Create and activate virtual environment
uv venv DeepFusion-DDG --python 3.11
source DeepFusion-DDG/bin/activate  # On macOS/Linux
# DeepFusion-DDG\Scripts\activate   # On Windows

# Install dependencies
uv pip install -e .

Development Installation

uv pip install -e ".[dev]"

Project Structure

BMI707FinalProject/
├── 707_Files_for_Colab/
│   ├── Data_Frames/           # Filtered FireProtDB datasets
│   ├── WT_PDBs/               # Wild-type protein structures
│   └── Data_To_Aggregate/     # Graph dataset PKL files
├── Official Model Notebooks/   # Trained models and notebooks
│   ├── BMI_707_Official_Aggregate_Data_05_08_2025.ipynb
│   ├── BMI_707_Official_ESMB_Model_From_Aggregated_05_08_2025.ipynb
│   ├── BMI_707_Official_GNN_Model_From_Aggregated_05_08_2025.ipynb
│   └── BMI_707_Official_Fusion_Model_From_Aggregated_05_08_2025.ipynb
├── Official Model Performance and Saved Weights/
│   ├── Baseline Models/
│   ├── Sequence Only Model/
│   ├── Structure Only Model/
│   └── Fusion Model/
└── Updates_for_MIT_Media_Lab_Collaboration/
    └── explore_filtered_fireprot_db.ipynb  # Dataset exploration

Key Components

Datasets

• FireProtDB: Curated protein stability database
• ΔΔG values: Ground truth thermodynamic stability changes
• Protein structures: PDB files for wild-type proteins

Models

1. Baseline Models: Ridge, Random Forest, Gradient Boosting
2. ESM2 Model: Sequence-only transformer model
3. GNN Model: Structure-only graph neural network
4. Fusion Model: Combined sequence + structure approach

Usage

# Load the filtered dataset
import pandas as pd
df = pd.read_csv('707_Files_for_Colab/Data_Frames/df_fireprot_median_ddG_old.csv')

# Load trained models
import torch
model = torch.load('Official Model Performance and Saved Weights/Fusion Model/fusion_best_model.pt')

Dependencies

Core dependencies:

• PyTorch 2.0+
• PyTorch Geometric 2.3+
• BioPython 1.81+
• pandas, numpy, scikit-learn
• matplotlib, seaborn

See pyproject.toml for complete dependency list.

Collaboration

For MIT Media Lab collaboration, see Updates_for_MIT_Media_Lab_Collaboration/ directory.

Authors

BMI 707 Final Project Team

Date

October 2025