PS4 is the largest open-source dataset for Protein Single Sequence Secondary Structure prediction.
This repository contains the data itself, methods to validate and add new samples to the dataset, and a toolkit for developing and evaluating state-of-the-art secondary structure prediction models.
For more information, see the official paper.
If you're primarily interested in using the pretrained models to predict secondary structure on your own sequences, you can use this official Hugging Face Space to do so with no setup.
The core dataset is contained in ps4_data/data/data.csv. This contains 18,731 proteins with their PDB code, index of the first residue in their respective DSSP file, their residue sequence and 9-category secondary structure sequence (including polyproline helices).
The train/test split can be found in ps4_data/data/chain_ids.npz.
After cloning the repository, cd into the root directory and run:
chmod a+rx install.sh
./install.sh
This will make the install script executable and then proceed to run it, which will:
- Download all required git submodules
- Install all PyPI module dependencies
- Build and install Mega
- Build and install the
ps4-rsRust package, for preprocessing samples when making new additions to PS4, using maturin.
Building ps4-rs requires Rust to be installed. Please see the official Rust language documentation for installation instructions.
Pretrained weights are made available via git LFS. To evaluate with trained weights or tain a new model, first generate the pytorch-ready dataset:
python main.py --gen_dataset
Generating the pytorch-ready dataset involves extracting embeddings from a large pretrained protein language model and may take a few hours. It is highly recommended to run this process on GPU.
Once --gen_dataset has been run, you can train a new model from scratch. To train PS4-Mega:
python main.py --train --mega
To train PS4-Conv:
python main.py --train --conv
To train a custom model, add your model code and make the relevant adjustments to ps4_train/train.py.
Once you've run --gen_dataset, you can easily evaluate pretained models on the PS4 test set or the CB513. For example, to evaluate PS4-Mega on CB513:
python main.py --eval --cb513 --mega
Or to evaluate PS4-Conv on the PS4 test set:
python main.py --eval --ps4 --conv
You can optionally specify a path to your own model weights as a final argument to python main.py --eval to evaluate a custom model.
You can use our pretrained models to predict secondary structure on new sequences listed in a FASTA file:
python main.py --sample <fasta_path>
Or to specify to use PS4-Conv instead of PS4-Mega:
python main.py --sample <fasta_path> --conv
Sampling also involves extracting embeddings from a large pretrained protein language model. It is recommended to run this process on a machine with at least 16GB of RAM.
With the help of the community, we can continue to grow PS4 and ensure a high-quality, non-redundant dataset. To preprocess new samples to add them to the PS4, run the following to ensure non-redundancy against the rest of the PS4 dataset and the CB513 using the ps4-rs package:
python extend_ps4.py <in_path> <out_path>
where <in_path> is replaced by the path to a folder containing DSSP files of proteins you'd like to add to the dataset, and <out_path> is the file you'd like to save the new non-redundant samples to, and must end in .csv.
To share new samples with the community, open a pull request!