Inspired by ColabFold by Sergey O.
Visualization inspired by pymol-color-alphafold.
Thanks to ESMFold by Meta and the API.
Fast access to AlphaMissense predicted Human proteins provided by hegelab.

This project enables structure and domain prediction directly within the PyMOL visualization software.

SO, download and install PyMOL from the official website.

You can directly run the following command in >PyMOL

run https://raw.githubusercontent.com/JinyuanSun/PymolFold/refs/heads/main/run_plugin.py

If you see the following, installation was successful:

PymolFold utilizes APIs from ESM3 and NVIDIA Boltz2.
You need to obtain API tokens from both:

• ESM3 API
• NVIDIA Boltz2

After obtaining your tokens, set them as environment variables:

• ESM_API_TOKEN
• NVCF_API_KEY
  Both must be uppercase and contain underscores.

How to set environment variables: Again, in >PyMOL, you can execute following command step by step

import os
print(os.environ["ESM_API_TOKEN"])
print(os.environ["NVCF_API_KEY"]) # if error occurs, it means you haven't set the key right

set_api_key ESM_API_TOKEN [, your_esm_api_key]
set_api_key NVCF_API_KEY [, your_nvcf_api_key]

print(os.environ["ESM_API_TOKEN"]) # it should print out what you just set
print(os.environ["NVCF_API_KEY"])

PymolFold provides several features: esm3, bfold, boltz2, color_plddt and pxmeter_align.

Use the convenient esm3 command:

esm3 sequence [, name]
# Example:
esm3 MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG

Use the bfold command for monomer folding with MSA support:

bfold sequence [, name]
# Example:
bfold MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG

This command:

• Uses Boltz2 for structure prediction
• Automatically fetches MSA from ColabFold via NVIDIA API
• Runs directly without launching the Streamlit UI
• Outputs results to your working directory (set with set_workdir)

Note: This is specifically for monomer proteins. For complexes, DNA, RNA, or ligands, use the boltz2 command below.

For more complex predictions, use the boltz2 command.
Due to the number of required inputs, a web interface is provided (inspired by NVIDIA Boltz2).
Currently, conditional prediction is not supported, but may be added in the future.

To launch the web interface, enter the following in the PyMOL command line:

boltz2

You can run the provided example:

When using CCD code, you can check all the existed CCDs under pymolfold/gui/ccd_keys.json.

If you want to predict a head-tail amide bonded cyclic peptide, you may set cyclic on.

If you want to use MSA as an auxiliary information, you may set Add MSA on. It will automatically query colab design msa search via NVIDIA API.

After clicking Run on the web page, wait about 6 seconds (depending on protein size), and the structure will appear in PyMOL!

After prediction, enter the following to view pLDDT scores for the predicted structure:

color_plddt

We utilized PXMeter to evaluate the differences between predicted structures and reference structures. PXMeter(0.1.4) now only supports PPI analysis, and more details can be seen in their repo. But unfortunately, we copied the repo and refine it since the python version may conflict with the one of PyMOL.

But how to use in PyMolFold?

pxmeter_align obj_real_structure_name, obj_pred_structure_name

It takes around 20s to loading when you first time using this method.

After running the script above, you will get the metrics in csv and png format under the folder you setted (if not set, it will generate in the root path). You can use the exmaple files under pymolfold/example/, and the results should be exactly the same as pymolfold/example/metrics.

Version If you are interesting in the source code, you can install pymolfold directly by pip install pymolfold.

Info
The PymolFold service is running on a A5000 instance (cost $100 a week), and the sequence length is limited to 1000aa.

Issues and Errors
If you encounter any errors or issues while using this project, please don't hesitate to open an issue here on GitHub. Your feedback helps us improve the project and make it more user-friendly for everyone.

PymolFold Server: A Shared Resource
Please note that the PymolFold server is a shared resource, and I request you to use it responsibly. Do not abuse the server, as it can affect the availability and performance of the service for other users.

21Sept2025: Refactor PyMOLFold, deleting unrelated module, adding boltz2 using NVIDIA API
17Jan2025: Add `esm3` to use ESM-3 for folding.
21Aug2023: As the ESMFold API is not stable, the job will be sent to PymolFold server if the job failed.
11Apr2023: `pf_plugin.py` is the PyMOL plugin and the `pf_pkg.py` is a pymol-free python package.