Eric J. Martin

The discovery of tetrahydro-β-carbolines as inhibitors of the kinesin Eg5

A series of tetrahydro-beta-carbolines were identified by HTS as inhibitors of the kinesin Eg5. Molecular modeling and medicinal chemistry techniques were employed to explore the SAR for this series with a focus of removing potential metabolic liabilities and improving cellular potency.

Conformational Analysis of Macrocycles: Finding What Common Search Methods Miss

As computational drug design becomes increasingly reliant on virtual screening and on high-throughput 3D modeling, the need for fast, robust, and reliable methods for sampling molecular conformations has become greater than ever. Furthermore, chemical novelty is at a premium, forcing medicinal chemists to explore more complex structural motifs and unusual topologies. This necessitates the use of conformational sampling techniques that work well in all cases. Here, we compare the performance of several popular conformational search algorithms on three broad classes of macrocyclic molecules. These methods include Catalyst, CAESAR, MacroModel, MOE, Omega, Rubicon and two newer self-organizing algorithms known as stochastic proximity embedding (SPE) and self-organizing superimposition (SOS) that have been developed at Johnson & Johnson. Our results show a compelling advantage for the three distance geometry methods (SOS, SPE, and Rubicon) followed to a lesser extent by MacroModel. The remaining techniques, particularly those based on systematic search, often failed to identify any of the lowest energy conformations and are unsuitable for this class of structures. Taken together with our previous study on drug-like molecules (Agrafiotis, D. K.; Gibbs, A.; Zhu, F.; Izrailev, S.; Martin, E. Conformational Sampling of Bioactive Molecules: A Comparative Study. J. Chem. Inf. Model., 2007, 47, 1067-1086), these results suggest that SPE and SOS are two of the most robust and universally applicable conformational search methods, with the latter being preferred because of its superior speed.

Kinase-Kernel Models: Accurate In silico Screening of 4 Million Compounds Across the Entire Human Kinome

Reliable in silico prediction methods promise many advantages over experimental high-throughput screening (HTS): vastly lower time and cost, affinity magnitude estimates, no requirement for a physical sample, and a knowledge-driven exploration of chemical space. For the specific case of kinases, given several hundred experimental IC(50) training measurements, the empirically parametrized profile-quantitative structure-activity relationship (profile-QSAR) and surrogate AutoShim methods developed at Novartis can predict IC(50) with a reliability approaching experimental HTS. However, in the absence of training data, prediction is much harder. The most common a priori prediction method is docking, which suffers from many limitations: It requires a protein structure, is slow, and cannot predict affinity. (1) Highly accurate profile-QSAR (2) models have now been built for roughly 100 kinases covering most of the kinome. Analyzing correlations among neighboring kinases shows that near neighbors share a high degree of SAR similarity. The novel chemogenomic kinase-kernel method reported here predicts activity for new kinases as a weighted average of predicted activities from profile-QSAR models for nearby neighbor kinases. Three different factors for weighting the neighbors were evaluated: binding site sequence identity to the kinase neighbors, similarity of the training set for each neighbor model to the compound being predicted, and accuracy of each neighbor model. Binding site sequence identity was by far most important, followed by chemical similarity. Model quality had almost no relevance. The median R(2) = 0.55 for kinase-kernel interpolations on 25% of the data of each set held out from method optimization for 51 kinase assays, approached the accuracy of median R(2) = 0.61 for the trained profile-QSAR predictions on the same held out 25% data of each set, far faster and far more accurate than docking. Validation on the full data sets from 18 additional kinase assays not part of method optimization studies also showed strong performance with median R(2) = 0.48. Genetic algorithm optimization of the binding site residues used to compute binding site sequence identity identified 16 privileged residues from a larger set of 46. These 16 are consistent with the kinase selectivity literature and structural biology, further supporting the scientific validity of the approach. A priori kinase-kernel predictions for 4 million compounds were interpolated from 51 existing profile-QSAR models for the remaining >400 novel kinases, totaling 2 billion activity predictions covering the entire kinome. The method has been successfully applied in two therapeutic projects to generate predictions and select compounds for activity testing.

Toward the Validation of Maternal Embryonic Leucine Zipper Kinase: Discovery, Optimization of Highly Potent and Selective Inhibitors, and Preliminary Biology Insight.

MELK kinase has been implicated in playing an important role in tumorigenesis. Our previous studies suggested that MELK is involved in the regulation of cell cycle and its genetic depletion leads to growth inhibition in a subset of high MELK-expressing basal-like breast cancer cell lines. Herein we describe the discovery and optimization of novel MELK inhibitors 8a and 8b that recapitulate the cellular effects observed by short hairpin ribonucleic acid (shRNA)-mediated MELK knockdown in cellular models. We also discovered a novel fluorine-induced hydrophobic collapse that locked the ligand in its bioactive conformation and led to a 20-fold gain in potency. These novel pharmacological inhibitors achieved high exposure in vivo and were well tolerated, which may allow further in vivo evaluation.

FOCUS — Development of a Global Communication and Modeling Platform for Applied and Computational Medicinal Chemists

Communication of data and ideas within a medicinal chemistry project on a global as well as local level is a crucial aspect in the drug design cycle. Over a time frame of eight years, we built and optimized FOCUS, a platform to produce, visualize, and share information on various aspects of a drug discovery project such as cheminformatics, data analysis, structural information, and design. FOCUS is tightly integrated with internal services that involve-among others-data retrieval systems and in-silico models and provides easy access to automated modeling procedures such as pharmacophore searches, R-group analysis, and similarity searches. In addition, an interactive 3D editor was developed to assist users in the generation and docking of close analogues of a known lead. In this paper, we will specifically concentrate on issues we faced during development, deployment, and maintenance of the software and how we continually adapted the software in order to improve usability. We will provide usage examples to highlight the functionality as well as limitations of FOCUS at the various stages of the development process. We aim to make the discussion as independent of the software platform as possible, so that our experiences can be of more general value to the drug discovery community.

Euclidean chemical spaces from molecular fingerprints: Hamming distance and Hempel’s ravens

Molecules are often characterized by sparse binary fingerprints, where 1s represent the presence of substructures and 0s represent their absence. Fingerprints are especially useful for similarity calculations, such as database searching or clustering, generally measuring similarity as the Tanimoto coefficient. In other cases, such as visualization, design of experiments, or latent variable regression, a low-dimensional Euclidian “chemical space” is more useful, where proximity between points reflects chemical similarity. A temptation is to apply principal components analysis (PCA) directly to these fingerprints to obtain a low dimensional continuous chemical space. However, Gower has shown that distances from PCA on bit vectors are proportional to the square root of Hamming distance. Unlike Tanimoto similarity, Hamming similarity (HS) gives equal weight to shared 0s as to shared 1s, that is, HS gives as much weight to substructures that neither molecule contains, as to substructures which both molecules contain. Illustrative examples show that proximity in the corresponding chemical space reflects mainly similar size and complexity rather than shared chemical substructures. These spaces are ill-suited for visualizing and optimizing coverage of chemical space, or as latent variables for regression. A more suitable alternative is shown to be Multi-dimensional scaling on the Tanimoto distance matrix, which produces a space where proximity does reflect structural similarity.

Effect of Chirality on Common in Vitro Experiments: An Enantiomeric Pair Analysis.

This analysis elucidates the impact of small molecule architecture on common in vitro ADME assays. In vitro parameters considered in this analysis included Caco-2 permeability/efflux, CYP3A4 inhibition, hERG inhibition, and rat microsomal extraction ratio (ER). The statistical significance and practical meaningfulness of chirality were determined by comparison of the distribution of enantiomers with the experimental variation distribution observed from duplicate measurements. Statistical tools were applied to characterize the role of molecular architecture on the outcome of a given in vitro assay. We found that CYP3A4 inhibition, hERG inhibition, Caco-2 permeability, and efflux are unlikely to be modulated by chirality. However, rat microsomal ER provides a statistically significant, and quantitatively meaningful, chance of being influenced by chirality.

Profile-QSAR and Surrogate AutoShim Protein-Family Modeling of Proteases

The 2D Profile-QSAR and 3D Surrogate AutoShim protein-family virtual screening methods were originally developed for kinases. They are the key components of an iterative medium-throughput screening alternative to expensive and time-consuming experimental high-throughput screening. Encouraged by the success with kinases, the S1-serine proteases were selected as a second protein family to tackle, based on the structural and SAR similarity among them, availability of structural and bioactivity data, and the current and future small-molecule drug discovery interest. A validation study on 24 S1-serine protease assay data sets from 16 unique proteases gave very promising results. Profile-QSAR gave a median R(ext)² = 0.60 for 24 assay data sets, and pairwise selectivity modeling on 60 protease pairs gave a median R(ext)² = 0.64, comparable to the performance for kinases. A 17-structure universal ensemble S1-serine protease surrogate receptor for Autoshim was developed from a collection of ~1500 X-ray structures. The predictive performance on 24 S1-serine protease assays was good, with a median R(ext)² = 0.41, but lower than had been obtained for kinases. Analysis suggested that the higher structural diversity of the protease structures, as well as smaller assay data sets and fewer potent compounds, both contributed to the decreased predictive power. In a prospective virtual screening application, 32 compounds were ordered from a 1.5 million archive and tested in a biochemical assay. Thirteen of the 32 compounds were active at IC₅₀ ≤ 10 μM, a 41% hit-rate. Three new scaffolds were identified which are being followed up with testing of additional analogues. A SAR similarity analysis for this target against 13 other proteases also indicated two potential protease targets which were positively and negatively correlated with the activity of the target protease.

Exploiting Structure–Activity Relationships in Docking

From the perspective of 2D chemical descriptors, error in docking activity predictions is separated into noise and systematic components. This error framework explains how fitting docking scores to a 2D-QSAR equation often improves accuracy as well as its logical limits. Intriguingly, in examined cases where multiple docking models (e.g., multiple crystal structures or multiple scoring functions) are available for an enzyme, the noise component of error dominates the difference between the more accurate and less accurate docking models. When this is true, the QSAR equation fit statistics can rank each docking-score sets accuracy in the absence of experimental activity data.

Gazing into the crystal ball; the future of computer-aided drug design

Twenty-five years is almost a full career for a scientist, but before looking to the future, we should ask what is really new in the last 25 years, i.e. since 1986? Surprisingly little! Here is a partial but still fairly good list of techniques routinely used by modellers: high throughput docking, high precision docking, free-energy calculations, quantum mechanics, molecular mechanics, distance geometry, molecular dynamics, statistical thermodynamics, conformational searching, scaffold morphing, solvation, QSPR, QSAR, bioavailability predictions, pharmacophores, protein modeling, de novo design, library design, chemical databases and searching, data analysis and visualization, virtual screening, chemometrics, interaction analysis using small molecule and protein x-rays, and FBDD. The majority of these techniques were introduced in the early to mid 1980s, and we think everything on the list except FBDD was introduced by the early 1990s (many techniques have been re-invented since; the collective memory of the literature seems to be under 10 years and falling). The biggest revolution in computational chemistry over the last 25 years was not a new computational technique, but rather the introduction of Beowulf clusters around 2000, which in just a few years increased processing power by about 1009 beyond Moore’s Law for many problems, i.e., it skipped at least a decade. This ‘‘suddenly’’ enabled application of a large number of the techniques from the 1980s to real systems.