I have an interest to predictive model build with 3D compound information. Pytorch3d and open3d seems attractive package for me. However, to use the package, I need to convert molecular information to 3D data such as pointcloud etc.
At first I tried it to use openbabel because recent version of openbabel can convert molecule from MDL molformat to VDW pointcloud xyz format. It worked well but it was difficult to add color information such as atom kind.
So I gave up to use openbabel and shifted pymol.
I installed open source pymol from source code to my python3.7 env. And tried it. Fortunately pymol can save surface 3D data as wrl format. OK sounds good ….. Wait! Unfortunately open3d can’t read wrl format file so I need to convert wrl file to pyl (point cloud format) format.
Finally I found the solution. Use meshlab! It is easy to install meshlab to linux. Just use apt. ;)
$sudo apt -yV install meshlab
Type meshlab in the terminal meshlab application will launch. And the app can convert file format. But I would like to it from python. So I used meshlabserver from subprocess.
meshlabserver is useful for batch process. Following code is an example for reading 3d mesh data generated from rdkit.
At first, load molecule from SMILES and generate 3D conformer. The save it as MDL mol format. Then load file from pymol from library mode.
Then call ‘show surface’ and change quality of surface.
Finally save surface as wrl format. It is easy. Just call ‘save hogehoge.wrl’ in pymol. ;)
import subprocess import open3d as o3d from rdkit import Chem from rdkit.Chem import AllChem import pymol mol = Chem.MolFromSmiles('c1ccncc1C') mol = Chem.AddHs(mol) AllChem.EmbedMolecule(mol, randomSeed=794) # to generate same coordination of molecule Chem.MolToMolFile(mol, 'mol.mol') pymol.cmd.load('mol.mol') pymol.cmd.show('surface') pymol.cmd.set('surface_quality', '3') pymol.cmd.save('surface.wrl')
Now I could get molecular surface data as wrl. Then convert the file format. Call meshserver command with -i inputfile -o outputfile and -om vc option.
-om vc means add vertex color information to the outputfile. Following code I would like to get ply format. So -o is surface.ply.
subprocess.call('meshlabserver -i surface.wrl -o surface.ply -om vc'.split())
Readly. Open3d can read ply format. (pytorch3d can read the file format too!)
giom=o3d.io.read_point_cloud('surface.ply') giom.has_colors() > True
Open3d has visualization method.
After calling the command above, another window will open and I can see interactive 3d pointcloud image of query molecule.
I tried several condition of surface quality settings. By using default pymol setting, very sparse data is generated.
Now ready to dive into 3D information based deep learning. ;)