High-resolution de novo protein structure prediction from amino acid sequences using deep learning.
OmegaFold is a deep learning-based tool for predicting high-resolution 3D protein structures directly from their amino acid sequences. It solves the problem of structure prediction for novel proteins where traditional homology modeling fails, enabling researchers to understand protein function and interactions without experimental data. The model operates de novo, requiring only the primary sequence as input.
Computational biologists, bioinformaticians, and structural biology researchers who need to predict protein structures for novel sequences, orphan proteins, or engineered proteins without relying on evolutionary information.
OmegaFold provides accurate de novo structure prediction without requiring multiple sequence alignments, making it uniquely suited for novel proteins. Its optimized memory usage allows handling long sequences, and it offers configurable trade-offs between resource consumption and prediction quality.
OmegaFold Release Code
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Predicts 3D structures directly from amino acid sequences without relying on evolutionary data, enabling accurate modeling for novel proteins where homology-based methods fail.
Optimized to handle sequences up to 4096 residues on high-memory GPUs like NVIDIA A100s, as noted in the README's update on GRAM reduction.
Allows adjustment of subbatch size and cycle count to balance memory usage, speed, and prediction quality, providing flexibility for different hardware setups.
Supports Linux with CUDA and macOS with Apple Silicon MPS acceleration via PyTorch nightly, broadening accessibility across common research environments.
Outputs per-residue confidence values stored as b-factors in PDB files, helping users assess prediction reliability without additional tools.
Requires installing the latest nightly version of PyTorch and cloning the repository, rather than a simple pip install, adding steps and potential instability.
The README admits there is no rule of thumb for setting subbatch_size, making optimization trial-and-error and inefficient for users.
Even with GRAM reductions, long sequences necessitate GPUs with substantial memory, limiting accessibility for labs with constrained resources.
Notes in the README warn that some comments might be out-of-date, which can lead to confusion and errors during setup or usage.