Representative Random Sampling of Chemical Space.

Area of Science:

• Computational chemistry and cheminformatics.
• Data science and machine learning applications in chemistry.

Background:

• The vastness of chemical space, the set of all possible molecules, prevents comprehensive exploration and leads to biased sampling in current databases.
• Existing methods for analyzing chemical space are limited by the inability to enumerate all molecules, hindering data-driven characterization.
• Substantial biases in current molecular databases limit their representativeness of the true chemical space.

Purpose of the Study:

• To develop a method for generating unbiased, representative random samples of chemical space without enumerating all molecules.
• To enable the estimation of the total number of molecules within any defined custom chemical space.
• To assess the representativeness of existing molecular databases and establish criteria for their size.

Main Methods:

• A novel approach for producing unbiased random samples of chemical space, applicable to molecules representable as graphs.
• Efficient algorithms that scale to molecules with up to 30 atoms.
• Methodology for estimating the size of custom chemical spaces.

Main Results:

• The developed method allows for unbiased sampling and size estimation of chemical space.
• The approach is computationally efficient, even for moderately sized molecules.
• Analysis of current databases reveals their representativeness and establishes a lower bound for database size to ensure representativeness.

Conclusions:

• This work provides a powerful tool for unbiased exploration and characterization of chemical space.
• The findings offer a criterion for building representative molecular databases crucial for data-driven discovery.
• The method facilitates a more accurate understanding of chemical diversity and database limitations.