This is the idpGAN repository. IdpGAN is a machine-learning based conformational ensemble generator for coarse grained (CG) models of intrinsically disordered proteins (IDPs).

Details of idpGAN are described in https://doi.org/10.1038/s41467-023-36443-x.

To use idpGAN (implemented in PyTorch), you can run a Jupyter notebook illustrating idpGAN functionalities.

The notebook shows how to use the generator neural network of idpGAN to generate 3D structures of CG IDPs.

There are two ways in which you can run the notebook.

This is the easiest way. This option allows you to run the notebook remotely. Just reach the notebook at: idpGAN Colab notebook.

Running the initial cells will automatically install all the dependencies (please note that the installation may require a few minutes).

NOTE: make sure to use a GPU runtime to largely speed up the idpGAN conformation generation process. If you use the default runtime (running on CPU), the process could take several minutes. To use a GPU runtime:

• Use the Edit -> Notebook settings item in the main menu of the Colab page.
• Set the Hardware accelerator option to GPU.

You can also run the notebook on your machine. What you need to do is:

• Make sure to have NumPy, Matplotlib and PyTorch installed.
  • Optional: if you want to visualize 3D conformational ensembles in Jupyter, also install NGLview and MDTraj.
• Clone or download this repository on your system.
• Make sure to have the idpgan directory (the one found in the root of this repository) in your PYTHONPATH, so that you can import the idpgan library.
• Run the idpgan_experiments.ipynb notebook in the notebooks directory.
• In the notebook, there is the following line of code: data_dp = "data". It must point to the path on your system where the data directory of this repository is located (the directory contains data files and the generator weights). Make sure to edit the line so that it points to the right location.

We trained an idpGAN version on conformations extracted from MD simulations performed using a CG-based protein model developed in our research group. In the data directory of this repository, we have the following files with information on the training, validation and test sets of this version of idpGAN:

• idpgan_training_set.fasta: a FASTA file storing all the sequences used in the training set of idpGAN. All of them were obtained from DisProt.
• hbval_split_[01234].txt: files storing the names of the training set sequences used in the five validation partitions of the HB_val set in the idpGAN article.
• idptest.fasta: a FASTA file storing all the sequences of IDP_test, the test set of idpGAN.

We also have the following files, that allow you to run a demo of the generatative model on a small dataset:

• generator.pt: PyTorch weights for a pre-trained generator model. They are the same weights we used to generate conformations for the IDP_test proteins in the article.
• *.npy: NumPy array files storing the xyz coordinates for 5 x 1000 ns MD simulations for two IDP_test proteins and one poly-alanine molecule.

We also trained an idpGAN version on Cα traces extracted from all-atom simulations performed using the ABSINTH implicit solvent model, which was found to reproduce with good accuracy the experimental behavior of several IDPs. ABSINTH simulations can be performed using the CAMPARI software package. In the data directory of this repository, we have the following files for this version of idpGAN:

• idpgan_training_set.fasta: idpGAN training set sequences. For this version of the model, we used simulation data from all sequences with length <= 40 residues.
• abstest.fasta: this is the test set sequences that we used for evaluating this version of idpGAN. They are 15 sequences which were originally simulated with the ABSINTH potential in [Mao et al., 2010]. See the idpGAN article for more information.
• abs_generator.pt: PyTorch weights for a generator model pre-trained on ABSINTH simulation data.
• abs_selector.pt: PyTorch weights for a model for selecting the correct mirror images of xyz conformations generated by idpGAN.

You can run the idpGAN notebook remotely on Colab. You do not need to install any software on your system, you only need to first login on your Google account.

Otherwise you can run the idpGAN notebook (and library) locally on your machine. Any Python (version >= 3.8) with the above requirements should work well on all major operating systems (Windows, Mac, Linux). We developed and tested the code on Linux, Python (3.8.10) and PyTorch (1.7.1).

In all cases, if you plan to generate conformational ensembles with large number of snapshots (> 5000), we suggest to use PyTorch GPU support to largely speed up computational times.

• Janson G, Valdes-Garcia G, Heo L, Feig M. Direct Generation of Protein Conformational Ensembles via Machine Learning. BioRxiv (2022) doi: 10.1101/2022.06.18.496675

FeigLab, [email protected]

Michigan State University