Comparisons of Molecular Structure Generation Methods Based on Fragment Assemblies and Genetic Graphs

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Quantitative structure-property relationships (QSPRs) have long aided molecular property prediction.
• Automatic molecular structure generation is an emerging field with algorithm selection challenges.
• Factors influencing molecular generation include desired chemical diversity and construction level (atoms, fragments, patterns).

Purpose of the Study:

• To address the challenge of molecular structure generation by comparing fragment-based methods (FMs) and genetic-based methods (GMs).
• To introduce novel indices for evaluating molecular generation techniques.
• To assess generation methods without relying on a database of existing chemicals.

Main Methods:

• Revisiting and comparing fragment-based methods (FMs) and genetic-based methods (GMs) for molecular structure generation.
• Defining new indices to assess generation performance: coverage (explored data space), representativeness (data space exploration comparison), and generation specificity (ratio of molecules meeting requirements).
• Utilizing QSPR models to evaluate the properties of generated molecules.

Main Results:

• Developed and applied novel indices to quantitatively compare different molecular generation approaches.
• The new indices provide insights into the explored chemical space and the success rate of generating molecules with desired properties.
• Enabled comparison of FMs and GMs based on their ability to generate molecules satisfying specific property criteria.

Conclusions:

• The study provides a framework for evaluating and comparing molecular structure generation algorithms.
• Novel indices offer a robust method for assessing generation specificity and data space coverage.
• This work contributes to the advancement of automated molecular design by offering better tools for method selection and validation.