Figure 1.

CFFN architecture. (A) Illustration of overall architecture of CFFN that integrates planar and stereo structural information for molecular property predictions. 2D and 3D information is respectively represented as common molecular graph (2D-G) and full connection graph (3D-G) including all atomic distance, angle and dihedral angle. The two types of molecular graphs are processed using PDN and MPNN networks, respectively, and fused together to form a final hybrid graph (updated 2D + 3D G), which makes final predictions toward certain molecular property. (B) Illustration of generating 3D-G using the MPNN network. Atomic information of a target atom is updated by aggregating information of all edges connecting to it, which is updated iteratively using atomic information. (C) Illustration of an interdigitated arrangement of PDN and MPNN that together forms a zipper-like structure. PDN, which processes 2D-G, and MPNN, which processes 3D-G, communicate via sharing atom information.