Description
OmegaFold is a cutting-edge AI model designed for high-resolution de novo protein structure prediction, directly from primary amino acid sequences. This release code provides access to the model described in the paper "High-resolution de novo structure prediction from primary sequence." The project aims for continuous optimization, focusing on reducing GRAM requirements for inferencing long proteins and releasing more powerful models.
Read More
Key advancements in OmegaFold include significant GRAM reduction, enabling the inference of protein sequences up to 4096 residues on an NVIDIA A100 GPU with 80 GB of memory when using specific `--subbatch_size` settings. The model's sensitivity to `--subbatch_size` is highlighted, offering a trade-off between computation time and memory usage. Users can adjust this parameter to manage GPU memory limitations, with a suggestion to halve the value if memory issues arise.
For macOS users, OmegaFold supports MPS (Apple Silicon) acceleration, provided the latest nightly version of PyTorch is installed. The current code requires macOS users to clone the repository and run the model using `python main.py`. The setup process involves installing the tool via pip or cloning the repository and running `python setup.py install`. The input format is a standard FASTA file, and the output includes one PDB file per sequence, with confidence values embedded in the b-factor column of the PDB files.
OmegaFold offers flexibility in balancing prediction quality and resource usage. Users can modify `--subbatch_size` to trade computation time for GRAM, or adjust `--num_cycle` to trade computation time for average prediction quality. Further details and options can be accessed via the `omegifold --help` command. The project encourages citation of the associated research paper for its utility.
The developers are committed to improving ease of use and expanding the model's capabilities. Future updates are expected to further enhance performance and accessibility for a wider range of users and computational environments. This open-source release facilitates research and development in structural biology and drug discovery.
- CategoryScientific Discovery & Lab Tools
- Pricing ModelOpen source
- DeploymentCloud, Self-hosted
- PlatformsmacOS
OmegaFold's Core Features
High-resolution de novo protein structure prediction from primary sequence
Significant GRAM reduction for inferencing long proteins
Supports protein sequences up to 4096 residues on NVIDIA A100 (80GB)
Adjustable `--subbatch_size` for memory-time trade-off
MPS (Apple Silicon) acceleration support for macOS users
Confidence values provided in PDB b-factor column
Option to trade computation time for prediction quality via `--num_cycle`
Command-line interface for running predictions
Open-source release code available on GitHub
Model 2 available via `--model 2` command-line argument
Requires PyTorch and biopython for basic functionality
Getting Started with OmegaFold
Set up environment: Install OmegaFold via pip or clone repository and run setup.py
Prepare input: Create a FASTA file with protein sequences
Run prediction: Execute `omegifold INPUT_FILE.fasta OUTPUT_DIRECTORY` or `python main.py INPUT_FILE.fasta OUTPUT_DIRECTORY`
Configure parameters: Adjust `--subbatch_size` for memory optimization or `--num_cycle` for quality trade-off
Access results: Retrieve generated PDB files from the output directory
Utilize macOS acceleration: Install latest PyTorch nightly and use `python main.py`
Explore options: Run `omegifold --help` for detailed command-line arguments
OmegaFold's Use Cases
- Protein Structure Prediction
- Structural Biology Research
- Drug Discovery
- Bioinformatics Analysis
- Computational Chemistry
- Genomic Research
