Hello fellow computational chemists!

I have encountered a challenge that I'd like to discuss and seek advice on. Suppose I have a substantial dataset in SDF format containing thousands of molecules. My goal is to identify the major chemical series, pharmacophores, or scaffolds within this dataset. The primary motivation behind this task is to reduce a large dataset into a more manageable one that retains representative structures. This curated dataset will then serve as a reference for subsequent hit selection processes in my research.

The overarching objective is to ensure that my hit selection process captures sufficient chemical diversity from the larger dataset for subsequent in vitro testing. This is crucial for the success of my research, as it helps guarantee a broad exploration of chemical space and increases the likelihood of discovering novel compounds with desirable biological activity.

I would greatly appreciate any insights, suggestions, or experiences you might have to share regarding this process. Have you encountered similar challenges in your computational chemistry work? Are there specific tools or techniques you've found particularly useful in identifying key chemical features in large datasets? Your input will be invaluable in helping me streamline my research and improve the diversity and representativeness of my hit selection process.

Thank you in advance for your contributions to this discussion!

  • 1
    $\begingroup$ I recommend starting filtering from ADME properties first, using a set (3-5) of properties to filter (not only the Lipinski's rule, for example). $\endgroup$
    – Camps
    Sep 19, 2023 at 16:17

1 Answer 1


To identify chemical series effectively, consider utilizing clustering techniques based on the available information at your disposal. Depending on your dataset, incorporating bioactivity or toxicity data into the clustering process can add valuable insights. Alternatively, when such data is limited, structural methods, such as molecular descriptors, molecular fingerprints or substructure matching, can be employed.

You may read a bit about chemical space measures. The Coverage Score (10/grnnrm) leverages molecular fingerprints to generate a subset of molecules from a broader collection. This subset attempts to efficiently sample the available chemical space. Additionally, you might check the #Circles metric (10.48550/arXiv.2112.12542), which offers an approach to comparing two sets of molecules. It's versatile and can be applied to any generic set of properties. Beyond evaluating which set better samples the chemical space, it quantifies the impact of adding a single molecule to the overall sampling.

  • $\begingroup$ Thank you, that was a pretty good answer! Would you recommend any article to know more about differences in fingerprint and 2D structure clustering methods? $\endgroup$
    – Poccia
    Sep 26, 2023 at 12:56

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .