HyFactor

HyFactor generates molecular structures using a graph-based architecture that encodes the number of hydrogens on heavy atoms to enable molecular graph generation and reconstruction in cheminformatics.

Key Features:

• Hydrogen-labeled graph representation: Encodes the number of hydrogens attached to heavy atoms instead of explicit bond types, analogous to the InChI linear notation.
• DEFactor-derived architecture: Architecture inspired by the DEFactor design and explicitly compared to ReFactor (an implementation of DEFactor).
• Parameter efficiency: Uses on average approximately 20% fewer fitting parameters than ReFactor while maintaining comparable validity, uniqueness, and reconstruction rates.
• Benchmarking datasets: Developed and benchmarked using ZINC 250K, MOSES, and ChEMBL datasets.
• Training-set similarity: Produces structures more similar to training-set molecules than ReFactor, avoiding ReFactor's tendency to generate open-chain analogues of cyclic structures.
• Data augmentation impact: Application of data augmentation reduces reconstruction error for heavy molecules.

Scientific Applications:

• Molecular graph generation: Generation of molecular graphs for generative chemistry applications with assessment of validity, uniqueness, and reconstruction rates.
• Model benchmarking and comparison: Comparative evaluation of graph-based generative architectures using ZINC 250K, MOSES, and ChEMBL.
• Representation studies: Investigation of the effects of hydrogen-count-based encoding (InChI-like) on structural fidelity and reconstruction performance.

Methodology:

Implements a graph-based architecture inspired by DEFactor that represents molecules as hydrogen-labeled graphs (hydrogen counts on heavy atoms); development and benchmarking were performed on ZINC 250K, MOSES, and ChEMBL; compared performance against ReFactor; applied data augmentation to reduce reconstruction error; parameter counts are ~20% lower than ReFactor.