Awesome

SCUBA-D

SCUBA-D: De novo protein design with a denoising diffusion network independent of pre-trained structure prediction models

SCUBA-D is a method based on deep learning for generating protein structure backbone. Here we published the source code and the demos for SCUBA-D(version 1.0).

To run SCUBA-D, clone this GitHub repository and install Python.

Requirements

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

Install Dependencies

conda create -n SCUBA-D python=3.8
conda activate SCUBA-D

pip install -r ./install/requirements.txt
bash ./install/postInstall.sh

Optimal, the SCUBA(SCUBA-sketch) software can be downloaded and installed from zenodo: https://doi.org/10.5281/zenodo.10947360 or github: https://github.com/USTCwangsheng/pySCUBA if you wish to utilize it for generating an initial structure(sketch) as input for SCUBA-D.

Quick start

First, you need to check whether the weights file checkpoint_clean.pt is in the /savedir/priorDDPM_full_ESM_GAN/checkpoint/ path, if not, you need to download the weights file from https://biocomp.ustc.edu.cn/servers/downloads/checkpoint_clean.pt and save it under the /savedir/priorDDPM_full_ESM_GAN/checkpoint/ path.

Input for run.sh is a text file that contains controls for each design. The following options are suggested to be manually changed according to your needs:

• CUDA_VISIBLE_DEVICES - visible GPU idx
• test_list - list of designs
• batch_size - number of designs for each batch

Control Templates

All controls of the design are set in a JSON file. For convenience, we have provided some templates for different usages as references:

• demo_json/gen_from_noise/gen_from_all_noise.json - unconditional generation from noise
• demo_json/gen_from_noise/gen_from_all_sstype.json - generation with provided secondary structure as prior
• demo_json/gen_from_noise/gen_from_noise_partial_fixed.json - unconditional generation but fix some coordinates of motifs
• demo_json/refine_prior/structure_refine.json - refine the provided structure. We used the SCUBA suite to generate input files. SCUBA can be installed from https://github.com/USTCwangsheng/pySCUBA or https://doi.org/10.5281/zenodo.10947360.

Users can also create their own templates by changing the parser in protdiff/dataset/refine_dataset_par.py.

To provide more specific details on how to use the provided templates, see demo_json/README.md for more details.

Inference

Sample protein with

bash run.sh

Output example in target dir results:

gen_from_all_noise.json
gen_from_all_noise_batch_0.pdb

It takes approximately 30 seconds for SCUBA-D to generate a protein backbone of 100 amino acids from noise on a computer equipped with an RTX 3090 graphics card. It takes about 5 minutes to run the entire demo.