Transplant-insert-constrain-relax-assemble (TICRA): protein-ligand complex structure modeling and application to kinases.

We introduce TICRA (transplant-insert-constrain-relax-assemble), a method for modeling the structure of unknown protein-ligand complexes using the X-ray crystal structures of homologous proteins and ligands with known activity. We present results from modeling the structures of protein kinase-inhibitor complexes using p38 and Lck as examples. These examples show that the TICRA method may be used prospectively to create and refine models for protein kinase-inhibitor complexes with an overall backbone rmsd of less than 0.75 Å for the kinase domain, when compared to published X-ray crystal structures. Further refinement of the models of the kinase domains of p38 and Lck in complex with their cognate ligands from the published crystal structures was able to improve the rmsd's of the model complexes to below 0.5 Å. Our results show that TICRA is a useful approach to the problem of structure-based drug design in cases where little structural information is available for the target proteins and the binding mode of active compounds is unknown.

The rational design of a novel potent analogue of the 5'-AMP-activated protein kinase inhibitor compound C with improved selectivity and cellular activity.

We have designed and synthesized analogues of compound C, a non-specific inhibitor of 5'-AMP-activated protein kinase (AMPK), using a computational fragment-based drug design (FBDD) approach. Synthesizing only twenty-seven analogues yielded a compound that was equipotent to compound C in the inhibition of the human AMPK (hAMPK) α2 subunit in the heterotrimeric complex in vitro, exhibited significantly improved selectivity against a subset of relevant kinases, and demonstrated enhanced cellular inhibition of AMPK.

Molecular dynamics simulations on discoidal HDL particles suggest a mechanism for rotation in the apo A-I belt model.

Apolipoprotein A-I (apo A-I) is the major protein component of high-density lipoprotein (HDL) particles. Elevated levels of HDL in the bloodstream have been shown to correlate strongly with a reduced risk factor for atherosclerosis. Molecular dynamics simulations have been carried out on three separate model discoidal high-density lipoprotein particles (HDL) containing two monomers of apo A-I and 160 molecules of palmitoyloleoylphosphatidylcholine (POPC), to a time-scale of 1ns. The starting structures were on the basis of previously published molecular belt models of HDL consisting of the lipid-binding C-terminal domain (residues 44-243) wrapped around the circumference of a discoidal HDL particle. Subtle changes between two of the starting structures resulted in significantly different behavior during the course of the simulation. The results provide support for the hypothesis of Segrest et al. that helical registration in the molecular belt model of apo A-I is modulated by intermolecular salt bridges. In addition, we propose an explanation for the presence of proline punctuation in the molecular belt model, and for the presence of two 11-mer helical repeats interrupting the otherwise regular pattern of 22-mer helical repeats in the lipid-binding domain of apo A-I.

Double belt structure of discoidal high density lipoproteins: molecular basis for size heterogeneity.

We recently proposed an all-atom model for apolipoprotein (apo) A-I in discoidal high-density lipoprotein in which two monomers form stacked antiparallel helical rings rotationally aligned by interhelical salt-bridges. The model can be derived a priori from the geometry of a planar bilayer disc that constrains the hydrophobic face of a continuous amphipathic alpha helix in lipid-associated apoA-I to a plane inside of an alpha-helical torus. This constrains each apoA-I monomer to a novel conformation, that of a slightly unwound, curved, planar amphipathic alpha 11/3 helix (three turns per 11 residues). Using non-denaturing gradient gel electrophoresis, we show that dimyristoylphosphocholine discs containing two apoA-I form five distinct particles with maximal Stokes diameters of 98 A (R2-1), 106 A (R2-2), 110 A (R2-3), 114 A (R2-4) and 120 A (R2-5). Further, we show that the Stokes diameters of R2-1 and R2-2 are independent of the N-terminal 43 residues (the flexible domain) of apoA-I, while the flexible domain is necessary and sufficient for the formation of the three larger complexes. On the basis of these results, the conformation of apoA-I on the R2-2 disc can be modeled accurately as an amphipathic helical double belt extending the full length of the lipid-associating domain with N and C-terminal ends in direct contact. The smallest of the discs, R2-1, models as the R2-2 conformation with an antiparallel 15-18 residue pairwise segment of helixes hinged off the disc edge. The conformations of full-length apoA-I on the flexible domain-dependent discs (R2-3, R2-4 and R2-5) model as the R2-2 conformation extended on the disc edge by one, two or three of the 11-residue tandem amphipathic helical repeats (termed G1, G2 and G3), respectively, contained within the flexible domain. Although we consider these results to favor the double belt model, the topographically very similar hairpin-belt model cannot be ruled out entirely.

Atomic structures of human dihydrofolate reductase complexed with NADPH and two lipophilic antifolates at 1.09 a and 1.05 a resolution.

The crystal structures of two human dihydrofolate reductase (hDHFR) ternary complexes, each with bound NADPH cofactor and a lipophilic antifolate inhibitor, have been determined at atomic resolution. The potent inhibitors 6-([5-quinolylamino]methyl)-2,4-diamino-5-methylpyrido[2,3-d]pyrimidine (SRI-9439) and (Z)-6-(2-[2,5-dimethoxyphenyl]ethen-1-yl)-2,4-diamino-5-methylpyrido[2,3-d]pyrimidine (SRI-9662) were developed at Southern Research Institute against Toxoplasma gondii DHFR-thymidylate synthase. The 5-deazapteridine ring of each inhibitor adopts an unusual puckered conformation that enables the formation of identical contacts in the active site. Conversely, the quinoline and dimethoxybenzene moieties exhibit distinct binding characteristics that account for the differences in inhibitory activity. In both structures, a salt-bridge is formed between Arg70 in the active site and Glu44 from a symmetry-related molecule in the crystal lattice that mimics the binding of methotrexate to DHFR.

Combination of a naive Bayes classifier with consensus scoring improves enrichment of high-throughput docking results.

We have previously shown that a machine learning technique can improve the enrichment of high-throughput docking (HTD) results. In the previous cases studied, however, the application of a naive Bayes classifier failed to improve enrichment for instances where HTD alone was unable to generate an acceptable enrichment. We present here a protocol to rescue poor docking results a priori using a combination of rank-by-median consensus scoring and naive Bayesian categorization.

Finding more needles in the haystack: A simple and efficient method for improving high-throughput docking results.

The technology underpinning high-throughput docking (HTD) has developed over the past few years to where it has become a vital tool in modern drug discovery. Although the performance of various docking algorithms is adequate, the ability to accurately and consistently rank compounds using a scoring function remains problematic. We show that by employing a simple machine learning method (naïve Bayes) it is possible to significantly overcome this deficiency. Compounds from the Available Chemical Directory (ACD), along with known active compounds, were docked into two protein targets using three software packages. In cases where HTD alone was able to show some enrichment, the application of naïve Bayes was able to improve upon the enrichment. The application of this methodology to enrich HTD results can be carried out without a priori knowledge of the activity of compounds and results in superior enrichment of known actives compared to the use of scoring methods alone.

Enrichment of extremely noisy high-throughput screening data using a naïve Bayes classifier.

The noise level of a high-throughput screening (HTS) experiment depends on various factors such as the quality and robustness of the assay itself and the quality of the robotic platform. Screening of compound mixtures is noisier than screening single compounds per well. A classification model based on naïve Bayes (NB) may be used to enrich such data. The authors studied the ability of the NB classifier to prioritize noisy primary HTS data of compound mixtures (5 compounds/well) in 4 campaigns in which the percentage of noise presumed to be inactive compounds ranged between 81% and 91%. The top 10% of the compounds suggested by the classifier captured between 26% and 45% of the active compounds. These results are reasonable and useful, considering the poor quality of the training set and the short computing time that is needed to build and deploy the classifier.

Small molecule chemokine mimetics suggest a molecular basis for the observation that CXCL10 and CXCL11 are allosteric ligands of CXCR3.

BACKGROUND AND PURPOSE: The chemokine receptor CXCR3 directs migration of T-cells in response to the ligands CXCL9/Mig, CXCL10/IP-10 and CXCL11/I-TAC. Both ligands and receptors are implicated in the pathogenesis of inflammatory disorders, including atherosclerosis and rheumatoid arthritis. Here, we describe the molecular mechanism by which two synthetic small molecule agonists activate CXCR3. EXPERIMENTAL APPROACH: As both small molecules are basic, we hypothesized that they formed electrostatic interactions with acidic residues within CXCR3. Nine point mutants of CXCR3 were generated in which an acidic residue was mutated to its amide counterpart. Following transient expression, the ability of the constructs to bind and signal in response to natural and synthetic ligands was examined. KEY RESULTS: The CXCR3 mutants D112N, D195N and E196Q were efficiently expressed and responsive in chemotaxis assays to CXCL11 but not to CXCL10 or to either of the synthetic agonists, confirmed with radioligand binding assays. Molecular modelling of both CXCL10 and CXCR3 suggests that the small molecule agonists mimic a region of the '30s loop' (residues 30-40 of CXCL10) which interacts with the intrahelical CXCR3 residue D112, leading to receptor activation. D195 and E196 are located in the second extracellular loop and form putative intramolecular salt bridges required for a CXCR3 conformation that recognizes CXCL10. In contrast, CXCL11 recognition by CXCR3 is largely independent of these residues. CONCLUSION AND IMPLICATIONS: We provide here a molecular basis for the observation that CXCL10 and CXCL11 are allosteric ligands of CXCR3. Such findings may have implications for the design of CXCR3 antagonists.

Computational approach to de novo discovery of fragment binding for novel protein states.

In silico fragment-based drug discovery has become an integral component of the new fragment-based approach that has evolved over the past decade. Protein structure of high quality is essential in carrying out computational designs, and protein flexibility has been shown to impact prospective designs or docking experiments. Here we introduce methodology to calculate protein normal modes and protein molecular dynamics in torsion space which enable the development of multiple protein states to address the natural flexibility of proteins. We also present two fragment-based sampling methods, grand canonical Monte Carlo and systematic sampling, which are used to study protein-fragment interactions by generating fragment ensembles and we discuss the process by which these ensembles are linked to design ligands.

Machine learning algorithms for the prediction of hERG and CYP450 binding in drug development.

IMPORTANCE OF THE FIELD: The cost of developing new drugs is estimated at approximately $1 billion; the withdrawal of a marketed compound due to toxicity can result in serious financial loss for a pharmaceutical company. There has been a greater interest in the development of in silico tools that can identify compounds with metabolic liabilities before they are brought to market. AREAS COVERED IN THIS REVIEW: The two largest classes of machine learning (ML) models, which will be discussed in this review, have been developed to predict binding to the human ether-a-go-go related gene (hERG) ion channel protein and the various CYP isoforms. Being able to identify potentially toxic compounds before they are made would greatly reduce the number of compound failures and the costs associated with drug development. WHAT THE READER WILL GAIN: This review summarizes the state of modeling hERG and CYP binding towards this goal since 2003 using ML algorithms. TAKE HOME MESSAGE: A wide variety of ML algorithms that are comparable in their overall performance are available. These ML methods may be applied regularly in discovery projects to flag compounds with potential metabolic liabilities.

SU11248 (sunitinib) directly inhibits the activity of mammalian 5'-AMP-activated protein kinase (AMPK).

AMPK has been termed the fuel sensor of mammalian cells because it directly responds to the depletion of the fuel molecule ATP. In previous work, we found that AMPK is strongly activated by tumor-like hypoxia and glucose deprivation, independently of the oxygen response system associated with HIF-1. We also observed high levels of AMPK activity in tumor cells in vivo, using different model tumors. These findings suggested the hypothesis that modulation of AMPK activity could have therapeutic value for the treatment of solid tumors. To investigate this hypothesis, we have been conducting a SAR study of potential small-molecule modulators of AMPK activity. Here we report that the chemotherapeutic drug SU11248 (sunitinib) is at least as potent an inhibitor of AMPK as compound C, which is a commonly used experimental direct inhibitor of the enzyme. We also provide a computational model of the binding pose of SU11248 to an AMPKα subunit, which suggests a structural basis for the affinity of the drug for the ATP site of the catalytic domain. The ability of SU11248 to inhibit AMPK has potential clinical significance--there may be populations of SU11248-treated patients in which AMPK activity is inhibited in normal as well as in tumor tissue.

Bayesian modeling in virtual high throughput screening.

Naïve Bayesian classifiers are a relatively recent addition to the arsenal of tools available to computational chemists. These classifiers fall into a class of algorithms referred to broadly as machine learning algorithms. Bayesian classifiers may be used in conjunction with classical modeling techniques to assist in the rapid virtual screening of large compound libraries in a systematic manner with a minimum of human intervention. This approach allows computational scientists to concentrate their efforts on their core strengths of model building. Bayesian classifiers have an added advantage of being able to handle a variety of numerical or binary data such as physicochemical properties or molecular fingerprints, making the addition of new parameters to existing models a relatively straightforward process. As a result, during a drug discovery project these classifiers can better evolve with the needs of the projects from general models in the lead finding stages to increasingly precise models in the lead optimization stages that are of particular interest to a specific medicinal chemistry team. Although other machine learning algorithms abound, Bayesian classifiers have been shown to compare favorably under most working conditions and have been shown to be tolerant of noisy experimental data.

Technique for generating three-dimensional alignments of multiple ligands from one-dimensional alignments.

We describe and demonstrate a method for the simultaneous, fully flexible alignment of multiple molecules with a common biological activity. The key aspect of the algorithm is that the alignment problem is first solved in a lower dimensional space, in this case using the one-dimensional representations of the molecules. The three-dimensional alignment is then guided by constraints derived from the one-dimensional alignment. We demonstrate using 10 hERG channel blockers, with a total of 72 rotatable bonds, that the one-dimensional alignment is able to effectively isolate key conserved pharmacophoric features and that these conserved features can effectively guide the three-dimensional alignment. Further using 10 estrogen receptor agonists and 5 estrogen receptor antagonists with publicly available cocrystal structures we show that the method is able to produce superpositions comparable to those derived from crystal structures. Finally, we demonstrate, using examples from peptidic CXCR3 agonists, that the method is able to generate reasonable binding hypotheses.

Library fingerprints: a novel approach to the screening of virtual libraries.

We propose a novel method to prioritize libraries for combinatorial synthesis and high-throughput screening that assesses the viability of a particular library on the basis of the aggregate physical-chemical properties of the compounds using a naïve Bayesian classifier. This approach prioritizes collections of related compounds according to the aggregate values of their physical-chemical parameters in contrast to single-compound screening. The method is also shown to be useful in screening existing noncombinatorial libraries when the compounds in these libraries have been previously clustered according to their molecular graphs. We show that the method used here is comparable or superior to the single-compound virtual screening of combinatorial libraries and noncombinatorial libraries and is superior to the pairwise Tanimoto similarity searching of a collection of combinatorial libraries.

Improved naïve Bayesian modeling of numerical data for absorption, distribution, metabolism and excretion (ADME) property prediction.

We have implemented a naïve Bayesian classifier which models continuous numerical data using a Gaussian distribution. Several cases of interest in the area of absorption, distribution, metabolism, and excretion prediction are presented which demonstrate that this approach is superior to the implementation of naïve Bayesian classifiers in which continuous chemical descriptors are modeled as binary data. We demonstrate that this enhanced performance, upon comparison with other implementations, is independent of the descriptor sets chosen. We also compare the performance of three implementations of naïve Bayesian classifiers with other previously described models.

Application of machine learning to improve the results of high-throughput docking against the HIV-1 protease.

We have previously reported that the application of a Laplacian-modified naive Bayesian (NB) classifier may be used to improve the ranking of known inhibitors from a random database of compounds after High-Throughput Docking (HTD). The method relies upon the frequency of substructural features among the active and inactive compounds from 2D fingerprint information of the compounds. Here we present an investigation of the role of extended connectivity fingerprints in training the NB classifier against HTD studies on the HIV-1 protease using three docking programs: Glide, FlexX, and GOLD. The results show that the performance of the NB classifier is due to the presence of a large number of features common to the set of known active compounds rather than a single structural or substructural scaffold. We demonstrate that the Laplacian-modified naive Bayesian classifier trained with data from high-throughput docking is superior at identifying active compounds from a target database in comparison to conventional two-dimensional substructure search methods alone.

Comparative models for human apolipoprotein A-I bound to lipid in discoidal high-density lipoprotein particles.

We have constructed a series of models for apolipoprotein A-I (apo A-I) bound to discoidal high-density lipoprotein (HDL) particles, based upon the molecular belt model [Segrest, J. P., et al. (1999) J. Biol. Chem. 274, 31755-31758] and helical hairpin models [Rogers, D. P., et al. (1998) Biochemistry 37, 11714-11725], and compared these with picket fence models [Phillips, J. C., et al. (1997) Biophys. J. 73, 2337-2346]. Molecular belt models for discoidal HDL particles with differing diameters are presented, illustrating that the belt model can explain the discrete changes in HDL particle size observed experimentally. Hairpin models are discussed for the binding of apo A-I to discoidal HDL particles with diameters identical to those for the molecular belt model. Two models are presented for the binding of three monomers of apo A-I to a 150 A diameter discoidal HDL particle. In one model, two monomers of apo A-I bind to the exterior of the HDL particle in an antiparallel belt, with a third monomer of apo A-I bound to the disk in a hairpin conformation. In the second model, all three monomers of apo A-I are bound to the discoidal HDL particle in a hairpin conformation. Previously published experimental data for each model are reviewed, with FRET favoring either the belt or hairpin models over the picket fence models for HDL particles with diameters of 105 A. Naturally occurring mutations appear to favor the belt model for the 105 A particles, while the 150 A HDL particles favor the presence of at least one hairpin.