This filter selects the most diverse compounds from large compound collections by eliminating of the most similar structures. The size of the input collection is decreased, while the maximum possible coverage of its represented chemical space is retained.

If you have limited experimental or computational resources, diversity selection is an unbiased way to limit the number of compounds to handle. Collecting compounds from different regions of the chemical space is an efficient strategy to maximize the diversity of the identified active scaffolds.

Using this filter you can either reduce the size of large (virtual) screening libraries, or select a diverse, representative set of your virtual hits.

It is recommended that you eliminate unwanted structures before a diversity selection, placing the filter after structural or phys-chem filters. This way you can avoid exotic structures or structures with exotic substituents remaining in the results, which would likely happen with this filter.

The following options can be used to control the diversity of the resulting collection.

• Similarity threshold: the maximum S similarity allowed in the diverse set. It is guaranteed that none of the resulting molecules are more similar than S.
• Number of most diverse molecules: the maximum N number of diverse molecules to be selected. The diversity selection algorithm will select the most diverse N molecules, unless the maximum allowed similarity is reached first.

If you don’t limit the selection, the full collection will be returned in a diversity order. This means that the top N molecules in the resulting collection will be the most diverse N molecules. The maximum similarity found at the Nth molecule refer to the diversity of the first N molecules: none of them are more similar than this value.

You can adjust the meaning of similarity and dissimilarity of molecules here, selecting the descriptor on which similarity scores are calculated. We use the Tanimoto coefficient (Jackard index)¹⁾ as the measure of similarity now, but you can chose between different chemical fingerprints as descriptors. We plan to introduce more descriptor types and more similarity measure types in the future.

• Molecular descriptor: the molecular descriptor used to represent chemical structures during the calculation

The default descriptor used is the linear fingerprint implemented in Open Babel ²⁾, which is similar to Daylight’s fingerprint and Chemaxon’s linear fingerprint, and the Tanimoto coefficient is calculated as the similarity of fingerprints.

If you have no suggestions to use another setup, you can rely on our choices. After implementation and evaluation of new fingerprints and metrics, the default setup can be changed. This can be tracked at the end of this document, in the Changelog section.

We use an optimized implementation of the stepwise elimination algorithm³⁾, which can be described as follows:

1. calculate the similarity matrix of the molecules in the input collection
2. process the matrix elements as follows:
   1. select the largest off-diagonal element in the similarity matrix
   2. eliminate one molecule of the most similar molecule pair randomly
   3. go to step I. if off-diagonal elements remained
3. sort the list of eliminated molecules by similarity values associated to the elimination steps in increasing order

During this process, the size of the collection is reduced and diversity increases. Each elimination step throws out a compound that has close analogues in the remaining set. In result, we get a single compound, and a list of compounds with decreasing similarity values, which can be interpreted as the increasing diversity of the remaining set.

After the algorithm finishes, structures are sorted by similarity values and are placed in the result collection. The first molecules in the resulted collection are the most dissimilar (most diverse) ones, and the length of the result list is determined using the diversity options: the maximum number of compounds to be selected and the maximum similarity values allowed between diverse compounds.

The diversity selection is freely accessible for every mcule user with a monthly limit of 10000 input compounds. The average run time for 10000 compouds is about 5 minutes. The usage of your diversity filter can be tracked on the user profile / limits. Our technologies allow effective processing of very large collections (~10M). If you want to exceed your limits, please contact us.

—