AI Protocol for Retrieving Protein Dynamic Structures from Two-Dimensional Infrared Spectra

🔔 News

Our latest dataset now contains 49,547 different proteins, all sourced from RCSB and SWISSPROT (AFDB-SWISSPROT). You can find them in the Quick Start section below, it includes the Two-Dimensional Infrared Spectroscopy (2DIR) data and PDB data for all proteins. You will need to randomly split the dataset into training and test sets yourself! Training the model from scratch takes approximately five minutes!

We tested the model's ability to transfer to proteins with 100 to 150 residues, and the complete dataset link is: this link.

2DIR：Predicting protein dynamic structures using two-dimensional infrared spectroscopy, with unknown structures.

Requirements

Operating System: Linux (Recommended)
No non-standard hardware is required.

Getting started

To get started using 2DIR, clone the repo:

git clone https://github.com/ZhuLvs/2DIR.git
cd 2DIR

Install Dependencies

conda create -n 2DIR python=3.8
conda activate 2DIR
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

pip install -r ./install/requirements.txt

Quick Start

🚀 First, you need to download the training dataset for 2DIR. this link and save it under the data directory. Training dataset： data/2DIR data/contact (distance map), Validation dataset： data/valA data/valB (distance map). Additional test datasets can be added by the user!

The PDBFliess dataset link is this link.

Then, you calculate the distance matrix (distance map) between the CA atoms of each residue in the protein PDB structure file. The calculation script can be found in the helper_scripts directory.

You may manually modify the parameters in model/main.py.

Preparation before starting

You need to download the 2DIR and PDB data using the link above, then unzip and place them in the data folder (data/2DIR/xxx.png, data/PDB/xxx.pdb). After that, run the script below, and it will automatically complete the data preprocessing.

bash Preprocessing.sh

Training

Before training begins, the protein residue distance matrix needs to be padded to ensure uniform size, which facilitates model processing, accelerates training, and so on. The padding code can be found in the helper_scripts directory and can be modified as needed（Running Preprocessing.sh will handle this process）.

Known Length Protein

bash train.sh

Unknown Length Protein

bash train_unknown_protein.sh

Inference

After inference is complete, the predicted results need to be trimmed based on the protein length, following the format provided in data/output.txt.

Known Length Protein

bash test.sh

Unknown Length Protein

For proteins with unknown lengths, you need to run model/pre_length.py to predict the protein length, and then refer to the scripts in the helper_scripts directory for trimming and processing.

bash test_len.sh

To generate the protein backbone structure from the protein residue distance matrix, please use the gradient descent algorithm available in the PyRosetta protocols, providing the predicted residue distances from the model and constraints from residue_constants.py.

For obtaining the amino acid sequence of unknown proteins, it is recommended to use the backbone structure as input for ProteinMPNN. The model typically converges after around 300 epochs.