The construction of new lead compounds for a specific target is perceived as the bottle neck for drug design. To address this challenge, AI approaches using 3D information for molecular representations recently see a trend of rapid growth. However, the power of such approach cannot be fully released until an effective way of presenting molecule 3D structure to convolutional neural networks is identified other than the direct but information-sparse way of locating an atom simply at one voxel. Although several attempts of extending an atom to fill neighbor voxels have been discussed, such as gaussian blur and wave transform, molecule generation engines are still hunger for better fuels which can represent the nature of the atoms more closely in terms of physical and chemical properties. We for the first time report an experimental electron density-based molecule 3D structure representation method to tackle this challenge and thereby benefit molecule generation. In our work, we build two models: the first one, developed by using 3DCNN and GAN, can generate the electron density (ED) of the content on top of the ED of a given pocket by employing ED topology features which can reflect crucial physical and chemical properties of the pocket, such as shape and electrostatics; the second model is created to build a variety of molecules based on the content ED generated by the first model. Computed electron density maps for up to 100,000,000 3D conformations of ligands from Enamine are used for the pretrain of our models and over 120,000 experimental electron density maps from PDB are used for the fine-tuning. The enrichment of active compounds with favorable binding conformation is demonstrated by applying our models in several drug targets.

Pocket-based 3D molecule generation model powered by deep learning and molecular representations derived from experimental electron density

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.