Problem

Many screens get few or no hits, and it is difficult to justify screening many more compounds.

Solution

If the assay is valid, and we have a distribution of activities among the inactives, then we can build a ChemTree predictive model to select a new round of compounds to screen with the high potency tails of the distribution greatly enhanced. These compounds can either be selected from in-house collections and third party vendors, or built through selective combinatorial synthesis.

Example

We took a set of 31,693 compounds screened for anti-HIV activity, and selected three concentration thresholds for activity: <1e-6M, <1e-7M, and 1e-8M. We randomly choose 1000 compounds and their activity is shown below in the first row. Note there are no compounds with activity >=1e-8. However, when we built a ChemTree model, and "cherry-picked" another 1000 compounds in focused screen 1, we obtained 15 hits. Several additional iterations found more hits until we had found 76% of the most active compounds screening only 13% of the whole library. In practice, the search space of 31,693 compounds is quite small to be doing so many rounds of sequential screening. By using combinatorial libraries, we can obtain improved hit rates with each additional round of sequential screening.

Sequential screening process