DynamicBindHPC is a fork of DynamicBind, which adds enables you to run DynamicBind on HPC systems using Singularity and Slurm.
For more details about DynamicBind itself, we refer to the DynamicBind Github and the original publication.

• Singularity
• Slurm (There is a --no_slurm mode, but using Slurm is highly recommended)

1. Clone the repository and navigate to it

   git clone https://github.com/Jnelen/DynamicBindHPC
   

   cd DynamicBindHPC
   

2. Run a test example to be prompted to automatically download the Singularity image (~5 GB) and again to download and install the model weights (~450 MB). Additionally, required cache look-up tables for SO(2) and SO(3) distributions will have to be generated. (This only needs to happen once and usually takes around 15 minutes).
   The --no_slurm flag is optional here, but makes it easier to track the progress.

   python inferenceVS.py -p data/origin-1qg8.pdb -l data/ -out TEST -j 1 --no_slurm
   

   Or if you have access to a GPU, you can also add the -gpu tag like this:

   python inferenceVS.py -p data/origin-1qg8.pdb -l data/ -out TEST -j 1 -gpu --no_slurm
   

You can also download the Singularity image manually:

wget --no-check-certificate -r "https://drive.usercontent.google.com/download?id=1QFfsqvEdDJVLUvh0kVhksG8sIsrGenWO&confirm=t" -O singularity/DynamicBindHPC.sif

alternatively, you can build the singularity image yourself using:

singularity build singularity/DynamicBindHPC.sif singularity/DynamicBindHPC.def

The main file to use is inferenceVS.py. It has the following options/flags:

• -p, -r, --protein_path: Path to the protein/receptor .pdb file.

• -l, --ligand: The path to the directory of (separate) mol2/sdf ligand files.

• -o, --out, --out_dir: Directory where the output structures will be saved to.

• -j, --jobs: Number of jobs to use.

• -qu, --queue: On which node to launch the slurm jobs. The default value is the default queue for the user. Might need to be specified if there is no default queue configured.

• -m, --mem: How much memory to use for each job. The default value is 4GB.

• -gpu, --gpu: Use GPU resources. This will accelerate docking calculations if a compatible GPU is available.

• -c, --cores: How many cores to use for each job. The default value is 1 when used with the GPU option enabled, otherwise it defaults to 4 cores.

• -n, --num_outputs, --samples_per_complex: How many structures to output per compound. The default value is 1.

• --model: DynamicBind supports two models: ema_inference_epoch314_model.pt and pro_ema_inference_epoch138_model.pt. The default model is the same as the one used in the paper (ema_inference_epoch314_model.pt)

• --remove_hs: Remove the hydrogens in the final output structures.

• --no_slurm: Don't use slurm to handle the resources. This will run all samples in interactive mode. The --gpu and -c options will still work to use a gpu and set the number of CPU cores. However, other Slurm arguments such as the amount memory, time limit, ... will be ignored.

• --no_clean: Don't clean the input protein structure. Not recommended unless you properly prepared the input protein structure (removed ligands, waters, ...)

• --save_visualisation: Save the output ligand and protein files. These files can be used to generate an animation with the movie_generation.py script.

• -h, --help: Show the help message and exit.

To run a DynamicBindHPC calculation

python inferenceVS.py -p data/origin-1qg8.pdb -l data/ -out TEST -j 1 -gpu --no_slurm --save_visualisation

And to generate the different steps from the diffusion process you can do:

python movie_generation.py VS_DB_TEST_.../complexes/1opj_STI_A/

It is also possible to run this with a gpu, which speeds up the calculations significantly. You can specify a device as well (which sets the CUDA_VISIBLE_DEVICES to your input):

python movie_generation.py outputDir/complexes/complex_of_interest/ --device 1 --gpu  

A rank1_receptor_reverseprocess_relaxed.pdb and rank1_ligand_reverseprocess_relaxed.sdf will be output in the same directory as the original input. These contain the states of the different diffusion steps.

Note: I found that often in the first steps of the animation, the poses can clash with the protein backbone. However, I run into the exact same problems when running native DynamicBind with the same inputs. I raised an issue about this on their GitHub, and if a fix is made, I will also try to update DynamicBindHPC accordingly.

MIT