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

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning for drug discovery

Background:

• Graph-based architectures are gaining traction for molecular structure generation.
• Existing methods often rely on bond types, which can limit representation.
• There is a need for efficient and accurate molecular generation tools.

Purpose of the Study:

• Introduce HyFactor, a novel open-source graph-based architecture for molecular structure generation.
• Benchmark HyFactor against the ReFactor (DEFactor) architecture.
• Evaluate the performance of HyFactor in terms of parameter efficiency and generation quality.

Main Methods:

• Developed HyFactor, a graph-based architecture considering hydrogen atom counts instead of bond types.
• Benchmarked HyFactor and ReFactor on ZINC 250K, MOSES, and ChEMBL datasets.
• Analyzed model fitting parameters, validity, uniqueness, reconstruction rates, and structural similarity.

Main Results:

• HyFactor models have approximately 20% fewer fitting parameters than ReFactor models.
• Both architectures demonstrated similar validity, uniqueness, and reconstruction rates.
• HyFactor generated structures more similar to the training set compounds compared to ReFactor.

Conclusions:

• HyFactor presents a more parameter-efficient alternative for molecular structure generation.
• The novel approach of considering hydrogen atoms offers comparable performance to existing methods.
• HyFactor shows potential for generating molecules with higher structural fidelity to training data.