HyFactor: Hydrogen-count labelled graph-based defactorization Autoencoder

Graph-based architectures are becoming increasingly popular as a tool for structure generation. Here, we introduce a novel open-source architecture HyFactor which is inspired by previously reported DEFactor architecture and based on the hydrogen labeled graphs. Since the original DEFactor code was not available, its new implementation (ReFactor) was prepared in this work for the benchmarking purpose. HyFactor demonstrates its high performance on the ZINC 250K MOSES and ChEMBL data set and in molecular generation tasks, it is considerably more effective than ReFactor. The code of HyFactor and all models obtained in this study are publicly available from our GitHub repository: https://github.com/Laboratoire-de-Chemoinformatique/hyfactor

PDF

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Molecular Graph Generation MOSES HyFactor Validity 0.729 # 3
Internal Diversity 0.860 # 1
Scaffold similarity 0.862 # 1
Nearest neighbor similarity (SNN) 0.614 # 1
Frechet ChemNet Distance (FCD) 0.365 # 1

Methods


No methods listed for this paper. Add relevant methods here