HyFactor: A Novel Open-Source, Graph-Based Architecture for Chemical Structure Generation

Abstract

Graph-based architectures are becoming increasingly popular as a tool for structure generation. Here, we introduce novel open-source architecture HyFactor in which, similar to the InChI linear notation, the number of hydrogens attached to the heavy atoms was considered instead of the bond types. HyFactor was benchmarked on the ZINC 250K, MOSES, and ChEMBL data sets against conventional graph-based architecture ReFactor, representing our implementation of the reported DEFactor architecture in the literature. On average, HyFactor models contain some 20% less fitting parameters than those of ReFactor. The two architectures display similar validity, uniqueness, and reconstruction rates. Compared to the training set compounds, HyFactor generates more similar structures than ReFactor. This could be explained by the fact that the latter generates many open-chain analogues of cyclic structures in the training set. It has been demonstrated that the reconstruction error of heavy molecules can be significantly reduced using the data augmentation technique.

Construction of order-independent molecular fragments space with vector quantised graph autoencoder

“…The number of heavy atoms in molecules was no more than 50. The canonical order of atoms in molecules was obtained from a breadth-first search (BFS) algorithm, similar as it was done by Mercado et al 14 The curated ChEMBL data set was then split into a training set (80% of data or 1.3M molecules) and a test set (20% of data or 327K molecules) as it was done for the training of HyFactor architecture 8 .…”

“…The input graph is a hydrogen-count labelled graph (HLG) 8 where each atom is encoded in a vector with the following properties: element number, period, group, number of electrons on the last subshell + atom's charge, number of last shells, the total number of hydrogens, whether or not the atom is in a ring, number of neighbours and counts of single, double, triple and aromatic bonds near current atom. Each property is normalised with a natural logarithm for numerical stability.…”