Introduction

Whether you work in the early phases/discovery, or are in the preclinical phase, preparing to take a compound through clinical trials, the problem of failure of candidates due to poor absorption, distribution, metabolism, elimination (ADME) or toxicity properties is a costly one. Some studies attribute over 60% of failures of drug candidates in development to problems with ADME/Tox. No method can predict the ADME/Tox properties of a compound with complete certainty. Nevertheless, to the extent we can predict these properties ahead of time, we can prioritize our discovery and development efforts to the compounds with greatest chance of success.

Problem:

Compounds from discovery screening are often poor candidates for lead optimization or preclinical testing because screening efforts focus on target affinity, while paying limited attention to ADME/Tox properties.

Solution:

Using the ChemTree sequential screening process, push ADME/Tox optimization earlier into the drug discovery pipeline so compounds are discovered that have positive target affinity, reasonable drug-like properties and are more likely to have acceptable ADME/Tox properties.

Traditionally, ADME/Tox optimization is left to medicinal chemists in the lead optimization phase of drug discovery. However, many times the optimization of ADME is in conflict with maintaining potency. For instance, absorption can usually be improved by reducing molecular weight, yet potency can often be improved by increasing molecular weight. From where does this conflict stem? The problem is: screening efforts have found a local maximum in terms of potency, which may in turn be far from any sort of optimum, in terms of ADME. The resolution of this conflict is found by simultaneously optimizing for potency, ADME, and non-toxicity.

What are the barriers to implementation of this solution, and how do we overcome them?

Obstacle 1: While HTS is cheap and efficient for getting hits against a target, ADME/Tox (ADMET) screens have historically been expensive and low throughput. While in vivo screening will always be low throughput, many in vitro screens are available for ADMET. Even if these screens are modest proxies for ADMET, we have a better chance of at least getting our compounds in the ballpark of good ADMET, allowing medicinal chemists to work their magic and perfect the compounds. Also, it is not necessary to do high throughput AMDET screening -- one can take the top few hundred or thousand compounds from HTS, and then do low throughput ADMET screening. The biggest challenge may very well be overcoming the silo mind set -- but this seems to be rapidly changing at a number of pharmaceutical companies, where cross-disciplinary workgroups from all pipeline stages collaborate and are responsible for bringing new drugs to market.

Obstacle 2: High-quality in silico predictive methods are needed to optimize compounds for multiple attributes simultaneously. Golden Helix solves this problem with ChemTree multivariate recursive partitioning. This technology enables you to build predictive QSAR models that describe structure activity relationships in terms of multiple screening targets. Hence, it is possible to optimize for target binding affinity, specifically, ADME and toxicity at once. Obstacle 3: In many cases, screening logistics that allow the "cherry-picking" of specific molecules for testing, based on in silico predictions, are not present. In some cases, the freezing of compounds in plates, versus tubes, is the problem. To select a particular molecule you must thaw out 96 or 384. An interim solution is to screen the whole plate, whereas a longer term solution is to acquire integrated tube store refrigeration and screening robotic equipment that allows quick random access to any compound of interest for screening. Example:

Above, we see a screen shot of ChemTree multivariate recursive partitioning analysis of 21 different assays of the same 869 compounds. In the first split, four compounds are found that are four standard deviations higher in potency than the overall average potency for a particular endpoint of interest. In the second split, we found 30 compounds that have a different activity profile, and a number of endpoints are up to one standard deviation higher or lower than the overall distribution.

Below, we have built a multiple tree model that enables us to understand structure activity relationships for a host of targets simultaneously.

The last screen shot shows how we can "cherry pick" compounds that have optimal properties based on our predictive model that was built from about two dozen screens.