Skip to content

Latest commit

 

History

History
104 lines (74 loc) · 4.8 KB

README.md

File metadata and controls

104 lines (74 loc) · 4.8 KB

Mini AlphaFold 🧬

This repository provides a simplified, educational implementation of AlphaFold, the groundbreaking protein structure prediction model. It's perfect for learning core concepts of protein folding and exploring associated machine learning techniques. It includes a toy protein dataset and the building blocks for a simplified model.

Getting Started 🚀

1. Setup 🛠️

  • Virtual Environment: Create and activate a virtual environment (recommended):

    python3 -m venv venv
source venv/bin/activate

  • Install Dependencies: Install required packages (NumPy, SciPy, Matplotlib, tqdm, and PyTorch):

    pip install -r requirements.txt

2. Generate the Toy Dataset 🧪

python minifold_dataset.py

This script generates a toy protein dataset with a reduced amino acid alphabet and a simplified energy function. The data is split into train, validation, and test sets and saved in the protein_dataset directory. A usage_example.py file is also generated within this directory.

3. Preprocess the Data ⚙️

python preprocess_dataset.py

This script preprocesses the generated dataset, including padding, relative position encoding, normalization, and caching. The preprocessed data is saved in the preprocessed_protein_dataset directory (.pt files). Dataset statistics are also saved.

4. Loading the Preprocessed Data (PyTorch) ⬇️

import torch
from torch.utils.data import DataLoader
from preprocessed_protein_dataset import PreprocessedProteinDataset

# Load datasets
train_dataset = PreprocessedProteinDataset.load('preprocessed_protein_dataset/train')
val_dataset = PreprocessedProteinDataset.load('preprocessed_protein_dataset/val')
test_dataset = PreprocessedProteinDataset.load('preprocessed_protein_dataset/test')

# Create DataLoaders
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=32, shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=32, shuffle=False)

# Example: Iterate through the train loader ♻️
for batch in train_loader:
    sequence = batch['sequence']          # Padded sequence indices
    coords = batch['coordinates']         # 2D coordinates
    distance_matrix = batch['distance_matrix'] # Padded distance matrix
    position_encoding = batch['position_encoding'] # Padded position encoding
    energy = batch['energy']            # Protein energies
    length = batch['length']            # Original sequence lengths
    properties = batch['properties']      # Hydrophobic and charged ratios

    # Process the batch (e.g., feed to your model)
    print(f"Sequence shape: {sequence.shape}")
    print(f"Coordinates shape: {coords.shape}")
    # ... process other features ...

Model Components 📦

  • mini_alphafold.py: The main model architecture (SimplifiedAlphaFold). Integrates the Pairformer and Diffusion modules to predict protein structures. Accepts sequences, distance matrices, positional encodings, and other properties as input, outputting 3D coordinates. Includes a training_step function.

  • pairformer.py: Implements a simplified Pairformer block using Triangle Attention. Focuses on pairwise amino acid interactions, using distances between residues for more informed attention calculations.

  • simplediffusion.py: A simplified diffusion model. Generates protein structures by denoising random Gaussian noise conditioned on sequence embeddings and timesteps.

Project Roadmap 🗺️

This project is under active development! Dataset generation and preprocessing are complete. Model training and inference are next.

  • train.py: (Planned) Training script for mini_alphafold.py. Handles training loops, loss calculation, optimization, and checkpointing.
  • inference.py: (Planned) Inference script to predict 3D structures of new sequences.
  • Model Refinement: (Future) Advanced loss functions, attention mechanisms, and other improvements.
  • Dataset Expansion: (Future) Larger amino acid alphabet and/or more realistic energy functions.
  • Evaluation Metrics: (Future) Implement metrics to assess model performance.

Repository Contents 📂

  • minifold_dataset.py: Dataset generation
  • preprocess_dataset.py: Dataset preprocessing
  • mini_alphafold.py: Simplified AlphaFold model
  • train.py: (Planned) Training script
  • inference.py: (Planned) Inference script
  • protein_dataset/: Raw dataset
  • preprocessed_protein_dataset/: Preprocessed dataset
  • pairformer.py: Pairformer module
  • simplediffusion.py: Diffusion module
  • requirements.txt: Project dependencies

Contributing 🤝

Contributions and suggestions are welcome! This educational project is a simplified starting point for understanding protein folding.