Chai-1 is a multi-modal foundation model for molecular structure prediction that performs at the state-of-the-art across a variety of benchmarks. Chai-1 enables unified prediction of proteins, small molecules, DNA, RNA, glycosylations, and more.
For more information on the model's performance and capabilities, see our technical report.
# version on pypi:
pip install chai_lab==0.4.2
# newest available version (updates daily to test features that weren't released yet):
pip install git+https://github.com/chaidiscovery/chai-lab.gitThis Python package requires Linux, and a GPU with CUDA and bfloat16 support. We recommend using an A100 80GB or H100 80GB chip, but A10s and A30s should work for smaller complexes. Users have also reported success with consumer-grade RTX 4090.
You can fold a FASTA file containing all the sequences (including modified residues, nucleotides, and ligands as SMILES strings) in a complex of interest by calling:
chai fold input.fasta output_folderBy default, the model generates five sample predictions, and uses embeddings without MSAs or templates. For additional information about how to supply MSAs and restraints to the model, see the documentation below, or run chai fold --help.
The main entrypoint into the Chai-1 folding code is through the chai_lab.chai1.run_inference function. The following script demonstrates how to programmatically provide inputs to the model, and obtain a list of PDB files for downstream analysis:
python examples/predict_structure.pyTo get the best performance, we recommend running the model with MSAs. The following script demonstrates how to provide MSAs to the model by calling out to an MSA server:
python examples/msas/predict_with_msas.pyFor further instructions, see "How can MSAs be provided to Chai-1?" below.
Where are downloaded weights stored?
By default, weights are automatically downloaded and stored in /downloads (usually that's within site-packages). In cases where you want to control the download location (e.g. on a mounted drive in Docker), you can use the CHAI_DOWNLOADS_DIR envvar to control the download location. For example:
CHAI_DOWNLOADS_DIR=/tmp/downloads python ./examples/predict_structure.py How can MSAs be provided to Chai-1?
Chai-1 supports MSAs provided as an aligned.pqt file. This file format is similar to an a3m file, but has additional columns that provide metadata like the source database and sequence pairing keys. We provide code to convert a3m files to aligned.pqt files. For more information on how to provide MSAs to Chai-1, see this documentation.
For user convenience, we also support automatic MSA generation via the ColabFold MMseqs2 server via the --msa-server flag. As detailed in the ColabFold repository, please keep in mind that this is a shared resource. Note that the results reported in our preprint and the webserver use a different MSA search strategy than MMseqs2, though we expect results to be broadly similar.
How can I customize the inputs to the model further?
For more advanced use cases, we also expose the chai_lab.chai1.run_folding_on_context, which allows users to construct an AllAtomFeatureContext manually. This allows users to specify their own templates, MSAs, embeddings, and constraints. We currently provide an example of how to construct an embeddings context as well as an MSA context, and will be releasing helper methods to build template contexts soon.
We provide a web server so you can test the Chai-1 model right from your browser, without any setup.
Chai-1 uniquely offers the ability to fold complexes with user-specified "restraints" as inputs. These restraints specify inter-chain contacts at various resolutions that are used to guide Chai-1 in folding the complex. See restraints documentation for details.
Found a 🐞? Please report it in GitHub issues.
We welcome community testing and feedback. To share observations about the model's performance, please reach via GitHub discussions, or via email.
We use devcontainers in development, which helps us ensure we work in identical environments. We recommend working inside a devcontainer if you want to make a contribution to this repository.
Devcontainers work on local Linux setup, and on remote machines over an SSH connection.
Since this is an initial release, we expect to make some breaking changes to the API and are not guaranteeing backwards compatibility. We recommend pinning the current version in your requirements, i.e.:
chai_lab==0.4.2
If you find Chai-1 useful in your research or use any structures produced by the model, we ask that you cite our technical report:
@article{Chai-1-Technical-Report,
title = {Chai-1: Decoding the molecular interactions of life},
author = {{Chai Discovery}},
year = 2024,
journal = {bioRxiv},
publisher = {Cold Spring Harbor Laboratory},
doi = {10.1101/2024.10.10.615955},
url = {https://www.biorxiv.org/content/early/2024/10/11/2024.10.10.615955},
elocation-id = {2024.10.10.615955},
eprint = {https://www.biorxiv.org/content/early/2024/10/11/2024.10.10.615955.full.pdf}
}
You can also access this information by running chai citation.
Additionally, if you use the automatic MMseqs2 MSA generation described above, please also cite:
@article{mirdita2022colabfold,
title={ColabFold: making protein folding accessible to all},
author={Mirdita, Milot and Sch{\"u}tze, Konstantin and Moriwaki, Yoshitaka and Heo, Lim and Ovchinnikov, Sergey and Steinegger, Martin},
journal={Nature methods},
year={2022},
}
Chai-1 is released under an Apache 2.0 License (both code and model weights), which means it can be used for both academic and commerical purposes, including for drug discovery.
See LICENSE.
To discuss partnership and access to new internal capabilities, reach us via email.