Design of a Compound Screening Collection Gavin Harper Cheminformatics, Stevenage

In the Past...  Scientists chose what molecules to make  They tested the molecules for relevant activity

Now...  We often screen a whole corporate collection – compounds  But we choose what’s in the collection  If the collection doesn’t have the right molecules in it –we fail

“Screen MORE”  Everything’ll be fine  We’ll find lots of hits  Not borne out by our experience

How do I design a collection? - 1  Pick the right kind of molecules –hits similar biological targets –computational (in-silico) model predicts activity at right kind of target for given class of molecules –exclude molecules that fail simple chemical or property filters known to be important for “drugs”  FOCUS!

How do I design a collection? - 2  Cover all the options  Pick as “diverse” a set of molecules as possible  If there’s an active region of chemical space, we should have it covered  DIVERSE SELECTION –opposite extreme to focused selection

Basic Idea of Our Model  Relate biological similarity to chemical similarity  Use a realistic objective –maximize number of lead series found in HTS  Build a mathematical model on minimal assumptions  How does our collection perform now in HTS? –relate this to our model  Learn what we need to make/purchase for HTS to find more leads

A “simple” model  Chemical space is clustered (partitioned) –there are various possible ways to do this  For a given screen, each cluster i has –a probability  i that it contains a lead  If we sample a random compound from a cluster containing a lead, the compound has –a probability  i that it shows up as a hit in the screen  If we find a hit in the cluster, that’s enough to get us to the lead

And in pictures... clusters containing leads  i = Pr(box i is orange)

 i = Pr(dot is green) Hit Non-Hit Lead

Constrained Optimization Problem  Suppose that we want to construct a screening collection of fixed size M  To maximize expected number of lead series found we have to

Solution  If we know very little (  i,  i equal for all i) –select the same number from each cluster - diversity solution  If e.g. we know some clusters are far more likely than others to contain leads for a target –select compounds only from these clusters - focused solution (filters)  But we also have a solution for all the situations in between, where there is a balance between diversity and focus

Immediate Impact  Improved “diversity” score  Use in assessing collections for acquisition  We have integrated this score into our Multi-Objective Library Design Package * Gillett et al., J. Chem. Inf. Comp. Sci. 2002, 42,

The Waldman Criteria (Sheffield, 1998) 1.Adding redundant molecules to a system does not change its diversity. 2.Adding non-redundant molecules to a system always increases its diversity. 3.Space-filling behaviour of diversity space should be preferred. 4. Perfect (i.e. infinite) filling of a finite descriptor space should result in a finite value for the diversity function. 5.If the dissimilarity or distance of one molecule to all others is increased, the diversity of the system should increase. However, as this distance increases to infinity, the diversity should asymptotically approach a constant value. * Waldman et al., J. Mol. Graphics Modell. 2000, 18,

What value should  take?  Determining a value of  is important. We can cluster molecules using a variety of methods.  Fortunately, there is a recent paper from Abbott which answers this question  In 115 HTS assays, with a TIGHT 2-D clustering –consistent: mostly varies between 0.2 and 0.4  This agrees well with our experience  In practice we will use this (Taylor-Butina) clustering with radius 0.85 and using Daylight fingerprints * Martin et al., J. Med. Chem. 2002, 45,

How Difficult Are Typical Screens? Probability of one-or-more leads Screen Difficulty

Improving compound acquisition  For a sample to be eligible to form part of the collection  It has to be of a minimum purity –determined by the QA project  It has to pass a set of agreed in silico filters –good starting points –developability  Multiple lead series per screen –Multiple chemotypes => 2D representation –Collection model provides rationale and design guidelines  Leads for all targets –3D Pharmacophore coverage

Probability of finding a lead for different acquisition strategies

Why don’t far more screens find leads as collections get bigger? performance with a collection of singletons

The model assists with a wide range of strategic questions  What contribution did compounds from different sources make last year?  How big should a focused set of compounds for a biological system be? –what should be in it?  How many compounds that fail the default filters should I include in the corporate collection?  A secure framework for investigating a wide range of strategic questions  Also easy to check sensitivity to changes in parameters

The model generates new challenges  There appear to be fundamental limits in the performance of screening  Can I find a better way of describing molecules? –can I increase  i but keep  i at a similar level?  Are there assumptions in the model that I could breach? –model assumes I can only find leads from “hits” in the same cluster –how good can I get at jumping BETWEEN clusters from original hit compounds?  Can I get better at identifying high  i clusters? –improve modelling and estimation of probability values

Acknowledgements  Stephen Pickett  Darren Green  Jameed Hussain  Andrew Leach  Andy Whittington * Harper et al., Combinatorial Chemistry and High Throughput Screening 2004, 7,