論文紹介 / GeoDiff: A Geometric Diffusion Model for Molecular Conformation Generation
第10回 全日本コンピュータビジョン勉強会 生成モデル縛り論文読み会 にて、
"GeoDiff: A Geometric Diffusion Model for Molecular Conformation Generation" [Xu et al., ICLR 2022]
のご紹介をさせていただきました。
novelty: first generative model for molecular conformation generation based on a diffusion framework • Very clearly written paper. • Impressive empirical results with state-of-the-art results on GEOM-Drugs and GEOM-QM9 datasets. • Weaknesses: • Most of the weaknesses reported by the reviewers seem to have been addressed in the rebuttal. • The idea of the work is highly novel. The authors propose the first generative model for molecular conformation generation based on a diffusion framework. This paper brings together recent ideas and methods (e.g. diffusion, SE(3) equivariance) to the established task of molecular conformation generation with impressive empirical results. All the reviewers agree on acceptance with high scores. 2022/05/15 13 (強調は発表者によるもの)
informative representation for molecules is the 3D geometry, also known as conformation, where atoms are represented as their Cartesian coordinates. 2022/05/15 16
group: 特殊ユークリッド群 • N-次元空間における剛体を動かす「剛体運動(rigid motions) 」 • 回転、平⾏移動が含まれる • In this work, we consider the SE(3) group, i.e., the group of rotation, translation in 3D space. • 論⽂中には、”roto-translational” という表現も頻出 • the roto-translational invariance, (e.g. atomic distances does not change when rotating the molecule) • (私のように)あまりその分野に明るくない読者にとっても分かりやすいよう、 ⽤語の説明が丁寧になされているなという印象 2022/05/15 20
2015; Ho et al., 2020) の進展に基づく • : Ground Truth conformations • : Sequence of latent variables • Diffusion probabilistic model (Sohl-Dickstein et al., 2015) は以下 の 2 つのプロセスで説明される • The forward diffusion process • The reverse generative process 2022/05/15 22
modified ELBO, named alignment and chain-rule approaches. We denote models learned by these two objectives as GEODIFF-A and GEODIFF-C respectively. 2022/05/15 36
molecule are not sufficient to characterize a conformation (bonds are rotatable), we extend the original molecular graph by adding auxiliary edges, i.e., virtual bonds, between atoms that are 2 or 3 hops away from each other, which is a widely-used technique in previous work (Simm & Hernandez-Lobato, 2020; Xu et al., 2021) for reducing the degrees of freedom in 3D coordinates. • Edge 情報だけでは 3D 化が難しいところ(2D に対する 3D の ⾃由度に起因する)については、補助的な仮想 edge を⾜すこ とで、形状を決められるようにしている。 2022/05/15 46