LigPCDS: Labeled Dataset of X-ray Protein Ligand Images in 3D Point Cloud and Validated Deep Learning Models

The difference electron density from X-ray protein crystallography was used to create the first dataset of labeled ligand images in 3D point clouds, named LigPCDS. The dataset contain 244,226 entries of free organic ligands containing 3D representations labeled with two major labeling approaches: SP-based and AtomSymbol-based. The data from free organic molecules (non-covalent ligands) was retrieved from the Research Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB PDB) in december 2019 with resolutions ranging from 1.5 to 2.2 Å. The ligand images (blobs) were interpolated from their calculated difference electron density map in a 3D grid-like bounding box, around their atomic positions, and stored in point clouds. These ligand grid representations were further processed to retrive the final ligands representation in 3D point clouds using a mask of the shape of the ligand. A grid spacing of 0.5 Å gave the best results. The density value of the grid points was used as feature. The labeling approach used the structure of the ligands to propose vocabularies of chemical classes based on the chemical atoms themselves and their cyclic substructures. These structure annotations were applied pointwise to the ligand 3D representations using an atomic sphere model. Four proposed vocabularies were validated by successfully training good performance deep learning models for the semantic segmentation of a stratified dataset from LigPCDS, using 78902 entries. The four validated deep learning models are: (i) the LigandRegion, composed by generic atoms of any type; (ii) the AtomCycle, composed by generic atoms outside cycles and generic cycles; (iii) the AtomC347CA56, composed by generic atoms outside cycles, not aromatic cycles of size 3 to 7 and aromatic cycles of size 5 and 6; and (iv) the AtomSymbolGroups, composed by the atoms symbols with groupings. The mean accuracy of these models in their cross-validation was between 49.7% [-19.4,20.2] and 77.4% [-11.7,12.1] in terms of Intersection over Union (mIoU) metric and between 62.4% [-18.8,19.7] and 87.0% [-8.4,8.8] in F1-score (mF1), confidence interval between squared brackets. The models i, ii and iii and the used labeled representations in 3D point cloud are contained in the SP-based record; and model iv and its used labeled representations are contained in the AtomSymbol-based record. The dataset and validated models may be used to tackle problems regarding known and unknown ligand building to drug discovery and fragment screening pipelines. The code used to create and validated the LigPCDS is available at the following repository: https://github.com/danielatrivella/np3_ligand This repository also contains the NP³ Blob Label application for ligand building using the validated deep learning models from LigPCDS.