Generating 3D compound structure using machine learning: testing Auto3D
Generating 3D structure of the compounds from prior knowledge is a challenge. Machine learning models can be used to accurately predict 3D conformers of small molecules and larger compounds (e.g. AlphaFold). GeoMol , TorsionNet and alike methods are basically taking well-known knowledge-based methods for 3D generation on the next level. But none of the published models can be reproduced without any pretraining and/or easy to install, except for Auto3D . Auto3D was developed in the group of Professor Olexandr Isayev and published in 2022 including code . Auto3D automatically generates conformational space further optimized by atomistic neural network potentials (NNPs) such as ANI or AIMNet (also by the Isayev Lab). Tautomers and isomer enumeration are in my opinion is not necessary as part of the package, but they are nice addition. I tested for 10 random drug-like compounds from ChEMBL. For whatever reason running algorithm on CPU gave me abysmal running time of 450 secon