Dedenser: A Python Package for Clustering and Downsampling Chemical Libraries

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Chemical library screening is crucial in drug discovery.
• Diverse screening sets are common but can be costly and inefficient due to overrepresentation.
• Efficient sampling of chemical space is needed to reduce costs and improve discovery.

Purpose of the Study:

• To develop a computational tool, Dedenser, for downsampling chemical clusters.
• To enable efficient and representative sampling of large chemical libraries.
• To reduce the cost burden associated with early-stage drug discovery screening.

Main Methods:

• Dedenser utilizes Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) to identify clusters in 3D chemical point clouds.
• Poisson disk sampling is applied to downsample clusters based on volume or density.
• The package includes tools for generating chemical point clouds, QSAR descriptor calculations (Mordred), and 3D embedding/visualization (UMAP).

Main Results:

• Dedenser effectively reduces the size of chemical clusters while preserving the overall topology and distribution of chemical space.
• The tool provides both command-line and graphical user interfaces for ease of use.
• It enables the generation of representative, evenly distributed subsets of molecules from larger collections.

Conclusions:

• Dedenser offers a valuable solution for triaged sampling of chemical libraries, facilitating more cost-effective and efficient drug discovery.
• The open-source nature of Dedenser promotes community adoption and further development.
• This tool supports selecting representative molecules and ensuring even distribution within chemical clusters, moving beyond single-molecule representation.