Iclr2022 | spherenet and g-spherenet: autoregressive flow model for 3D molecular graph representation and molecular geometry generation

This article introduces Dezhou A & M University CSE system Shuiwang Ji professor （http://people.tamu.edu/~sji/） Team is ICLR2022 Two jobs received ：SphereNet And G-SphereNet.

1. Spherical Message Passing for 3D Molecular Graphs

Article address ：

https://openreview.net/forum?id=givsRXsOt9r

This work studies 3D graphs. stay 3DGNN And information transmission （message passing） in , How to express completely 3D Information , To get better data representation ？ Spherical information transmission （Spherical Message Passing-SMP) Method with distance , angle , Dihedral angle is used as input for information transmission .SMP Theoretically, it is nearly complete ,3D There is little loss of information .SphereNet hold SMP It is combined with the physical characteristics derived from Schrodinger equation , obtain SphereNet.SphereNet In three widely used datasets MD17,QM9,OC20 On the Internet SOTA Result , At the same time, it is also very efficient .

2. An Autoregressive Flow Model for 3D Molecular Geometry Generation from Scratch

Article address ：

https://openreview.net/forum?id=C03Ajc-NS5W

This paper presents a generation model that can generate three-dimensional molecular geometry from scratch . Put forward G-SphereNet Method , As shown in the figure below , This is the first known method to use autoregressive flow model to design and generate the three-dimensional geometric structure of molecules from scratch .G-SphereNet The method adopts the method of sequence generation , That is, only one atom is generated in each step .G-SphereNet By generating the distance , Angle and dihedral angle to indirectly determine the 3D Location . in addition , We use it SphereNet As a skeleton feature extraction network to effectively extract the three-dimensional information of molecules .