Molecular Sets (MOSES): A Benchmarking Platform for Molecular Generation Models.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Generative models are increasingly used for molecular exploration and de novo drug design.
• A lack of standardized benchmarks hinders the comparison and ranking of different molecular generative models.
• Generated molecules have potential applications in virtual screening and training predictive models.

Purpose of the Study:

• To introduce a standardized benchmarking platform for molecular generative models, named Molecular Sets (MOSES).
• To provide a consistent framework for training and evaluating generative models in chemistry.
• To facilitate objective comparison and ranking of molecular generation techniques.

Main Methods:

• Development of the MOSES benchmarking platform, including curated training and testing datasets.
• Implementation of a suite of metrics to assess the quality, diversity, and validity of generated molecular structures.
• Comparative analysis of several existing molecular generation models using the MOSES framework.

Main Results:

• MOSES offers standardized datasets and evaluation metrics for molecular generative models.
• The platform enables reproducible benchmarking and comparison of model performance.
• Comparative results are provided as reference points for future research in generative chemistry.

Conclusions:

• The MOSES platform addresses the need for standardized evaluation in molecular generative modeling.
• It facilitates objective assessment and comparison of different generative approaches.
• MOSES aims to accelerate advancements in the field of generative chemistry and AI-driven drug discovery.