This repository contains the author's implementation for the paper:
PointSite: a point cloud segmentation tool for identification of protein ligand binding atoms [bioRxiv]
Created by Zhen Li, Xu Yan and Sheng Wang

Tested with CUDA 10.0, Ubuntu 18.04, Python 3.6 with Conda and PyTorch 1.3.

git clone --recursive https://github.com/PointSite/PointSite_Inference.git
cd PointSite_Inference/
./install.sh

WARNING: To install the package successfully, users shall use the latest version Anaconda, such as Anaconda2-2019.10.

python inference.py 
--gpu: GPU index, if you have not GPU, just ignore it
--output: output root (required)
--data: data root, only support .xyz file (required)
--select_list: TXT file for selected protein name, default None
--num_vote: voting number in inference (default 25, larger number can archieve more stable and high performance)

conda activate pointsite_inference
chmod +x ./util/PDB_Tool
chmod +x ./util/PDB_To_XYZ
python inference.py --output blind_out --data example/blind --select_list example/blind_list
conda deactivate

Note that the above input data (in '.XYZ' format) contain the ground-truth label of binding atoms. Run below script for identifying binding atoms on unlabeled data in '.PDB' files.

chmod +x ./pointsite_run.sh
./pointsite_run.sh example/blind_list example/blind blind_out `pwd`

You will get .obj file in output folder, please use MeshLab to visualize.

Users may find the training data here;
Users may find the test data here;
Users may find the evaluation results here.

3D Semantic Segmentation with Submanifold Sparse Convolutional Networks, CVPR 2018 facebookresearch/SparseConvNet