Direct Molecular Conformation Generation

Jinhua Zhu, Yingce Xia, Chang Liu, Lijun Wu, Shufang Xie, Wengang Zhou, Tao Qin, Houqiang Li, Tie-Yan Liu

Keywords: Molecular Conformation Generation

Abstract: Molecular conformation generation, which is to generate 3 dimensional coordinates of all the atoms in a molecule, is an important task for bioinformatics and pharmacology. Most existing machine learning based methods first predict interatomic distances and then generate conformations based on them. This two-stage approach has a potential limitation that the predicted distances may conflict with each other, e.g., violating the triangle inequality. In this work, we propose a method that directly outputs the coordinates of atoms, so that there is no violation of constraints. The conformation generator of our method stacks multiple blocks, and each block outputs a conformation which is then refined by the following block. We adopt the variational auto-encoder (VAE) framework and use a latent variable to generate diverse conformations. To handle the roto-translation equivariance, we adopt a loss that is invariant to rotation and translation of molecule coordinates, by computing the minimal achievable distance after any rotation and translation. Our method outperforms strong baselines on four public datasets, which shows the effectiveness of our method and the great potential of the direct approach. The code is released at \url{https://github.com/DirectMolecularConfGen/DMCG}.

One-sentence Summary: We propose a method for molecular conformation generation, that directly outputs the coordinates of atoms instead of generating the interatomic distances first.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2202.01356/code)