OrcaFold

A state-of-the-art protein structure prediction system implementing advanced deep learning architectures inspired by AlphaFold.

Core Components

Evoformer Module

• Advanced MSA (Multiple Sequence Alignment) processing
• Pair representation handling
• Multi-head attention mechanisms
• Iterative refinement pipeline

Structure Module

• 3D coordinate prediction
• Backbone and side chain refinement
• Atomic structure assembly
• Iterative structure optimization

Supporting Systems

• High-performance data pipeline
• Configuration management
• Confidence scoring (pLDDT & TM-score)

Project Structure

OrcaFold/
├── orcafold/
│   ├── __init__.py
│   ├── config/
│   │   ├── __init__.py
│   │   ├── model_config.py
│   │   └── train_config.py
│   ├── modules/
│   │   ├── __init__.py
│   │   ├── evoformer/
│   │   │   ├── __init__.py
│   │   │   ├── msa_processor.py
│   │   │   ├── pair_processor.py
│   │   │   └── attention.py
│   │   ├── structure/
│   │   │   ├── __init__.py
│   │   │   ├── backbone.py
│   │   │   ├── sidechain.py
│   │   │   └── recycling.py
│   │   └── confidence/
│   │       ├── __init__.py
│   │       ├── plddt.py
│   │       └── tm_score.py
│   ├── data/
│   │   ├── __init__.py
│   │   ├── pipeline.py
│   │   ├── msa_tools.py
│   │   └── templates.py
│   └── utils/
│       ├── __init__.py
│       ├── geometry.py
│       ├── cuda_utils.py
│       └── visualization.py
├── scripts/
│   ├── train.py
│   ├── predict.py
│   └── evaluate.py
├── tests/
│   ├── __init__.py
│   ├── test_evoformer.py
│   ├── test_structure.py
│   └── test_pipeline.py
├── examples/
│   ├── single_prediction.py
│   ├── batch_processing.py
│   └── visualization.py
├── docs/
│   ├── architecture.md
│   ├── installation.md
│   ├── pipeline.md
│   └── api.md
├── requirements/
│   ├── base.txt
│   ├── dev.txt
│   └── gpu.txt
├── setup.py
├── README.md
└── LICENSE

Module Details

1. Evoformer (orcafold/modules/evoformer/)

• MSA Processor: Handles multiple sequence alignment processing
• Pair Processor: Manages residue pair representations
• Attention Mechanisms: Implements various attention patterns
  • Row-wise attention
  • Column-wise attention
  • Triangle multiplication updates

2. Structure Module (orcafold/modules/structure/)

• Backbone Generator: Creates initial backbone trace
• Side Chain Placement: Predicts side chain conformations
• Structure Refinement: Iteratively improves predictions
• Recycling Handler: Manages prediction recycling

3. Data Pipeline (orcafold/data/)

• MSA Generation: Interfaces with JackHMMER/HHblits
• Template Search: Finds and processes template structures
• Feature Processing: Prepares features for model input

4. Configuration (orcafold/config/)

• Model architecture settings
• Training parameters
• Runtime configurations
• Prediction modes (monomer/multimer)

5. Confidence Metrics (orcafold/modules/confidence/)

• pLDDT score calculation
• TM-score estimation
• Per-residue confidence metrics

Installation

# Clone the repository
git clone https://github.com/yourusername/OrcaFold.git
cd OrcaFold

# Create conda environment
conda create -n orcafold python=3.8
conda activate orcafold

# Install dependencies
pip install -r requirements/base.txt
pip install -r requirements/gpu.txt  # for CUDA support

# Install in development mode
pip install -e .

Quick Start

from orcafold import OrcaFold
from orcafold.data import Pipeline

# Initialize pipeline and model
pipeline = Pipeline()
model = OrcaFold(device='cuda')

# Prepare input data
features = pipeline.process_sequence('SEQUENCE.fasta')

# Generate prediction
structure, confidence = model.predict(features)

# Save results
structure.save('predicted_structure.pdb')

Performance Optimization

• CUDA-accelerated computations
• Mixed precision training
• Efficient MSA processing
• Optimized attention mechanisms

Contributing

Please read CONTRIBUTING.md for details on our code of conduct and the process for submitting pull requests. Bugs: visualizations

License

This project is licensed under the MIT License - see LICENSE for details.