|
|
|
|
|
|
|
|
Dr. Robert Clark is currently Sr. Director of Research at Tripos, Inc., in St. Louis MO. Tripos produces computational chemistry and cheminformatics software, including Sybyl, Unity and SARNavigator, among others . The company also undertakes contract and collaborative research projects entailing development of new data analysis methods for drug discovery, customized data-sharing systems and library design, and carries out synthesis of generalized discovery and customized lead follow-up libraries.
Dr. Clark received his bachelor's and master's degrees from Ohio University in chemistry, and went on to earn his PhD in biochemistry from Cornell University, minoring in statistics and biometry. He then spent two years doing post-doctoral work in plant bioenergetics at Brookhaven National Laboratories. Dr. Clark came to St. Louis in 1984, where he spent the next ten years with Monsanto. His initial responsibilities involved research on plant physiology and herbicide modes of action, but he spent most of the last six of those years carrying out SAR-guided herbicide and fungicide synthesis.
Since joining Tripos in 1994, he has drawn on this broad range of experience in discovering and developing biologically active small molecules to expand the spectrum of software tools and analysis methodology available to others in the field. He has played a key role in the expansion of Tripos' world-renowned products for elucidating quantitative structure-activity relationships (QSARs) and molecular diversity, and has published extensively in these and related areas. Dr. Clark now conducts and directs research into new techniques for advanced QSAR, with a particular focus on those used in dealing with data from high-throughput screening (HTS) programs.
|
|
The "Structures" in Structure-Activity Relationships
Bob Clark, Tripos, Inc.
Most in silico ADME/Tox work is based, in one way or another, on analyses of structure-activity relationships (SARs). People carrying out such work are often very sensitive to the limitations of their assay data, as they should be. They rarely consider uncertainties involving molecular structure, however. Such studies generally presume that a single structure adequately reflects each particular compound under consideration, but this assumption is dangerous when there is potential for protonation, deprotonation, tautomerization or stereochemical ambiguity. Even bond isomerization can be problematic in some circumstances, particularly where molecular similarity or fragment analysis is involved. Moreover, in many cases where a single structure can adequately represent a compound, the appropriate structure to consider is different in different contexts. Hence there is a real need to consider structure as a very high-dimensional problem that extends well beyond simple molecular connectivity.
|
The Challenges of ADME/Tox Prediction
Bob Clark, Tripos, Inc.
Absorption, distribution and excretion problems with drug candidates have been greatly reduced by the wide-spread adoption of Lipinski’s “Rule of 5” and similar rules of thumb for blood-brain barrier penetration. Problems not identified by such relatively simple filters continue to contribute to disappointing clinical trials for drug candidates from smaller companies, but their importance to large pharmaceutical companies has gone down as relevant in vitro and ex vivo screens have become more widely available and more reliable. As a result, issues involving metabolism and toxicology – with toxicity often being associated with one or more metabolites – have become all the more important. Such concerns are apt to be exacerbated by two trends in drug development: the tendency to look for blockbuster drugs that ameliorate but do not cure chronic disorders; and the recognition that many drugs – particularly inhibitors of regulatory kinases and those targeting mental disorders – must be somewhat promiscuous to be effective.
Broadly speaking, predictive metabotox approaches fall into two classes: those based on molecular similarity and those based on mechanism. In a sense, similarity based methods are non-parametric, in that they make minimal assumptions about how and where problems might occur. Mechanistic models, in contrast, are parametric, in that conformity to the model assumptions is critical to predictive performance and reliability. Each class has generic strengths and weaknesses, particularly as regards scope of applicability and the kind of errors that are likely to dominate performance. These will be discussed in some detail, as will limitations shared by both.
|
|
|
|
|
|
|
|
|
|
|
