Comprehensive structural and functional characterization of the human kinome by protein structure modeling and ligand virtual screening

Abstract

The growing interest in the identification of kinase inhibitors, promising therapeutics in the treatment of many diseases, has created a demand for the structural characterization of the entire human kinome. At the outset of the drug development process, the lead-finding stage, approaches that enrich the screening library with bioactive compounds are needed. Here, protein structure based methods can play an important role, but despite structural genomics efforts, it is unlikely that the three-dimensional structures of the entire kinome will be available soon. Therefore, at the proteome level, structure-based approaches must rely on predicted models, with a key issue being their utility in virtual ligand screening. In this study, we employ the recently developed FINDSITE/Q-Dock ligand homology modeling approach, which is well-suited for proteome-scale applications using predicted structures, to provide extensive structural and functional characterization of the human kinome. Specifically, we construct structure models for the human kinome; these are subsequently subject to virtual screening against a library of more than 2 million compounds. To rank the compounds, we employ a hierarchical approach that combines ligand- and structure-based filters. Modeling accuracy is carefully validated using available experimental data with particularly encouraging results found for the ability to identify, without prior knowledge, specific kinase inhibitors. More generally, the modeling procedure results in a large number of predicted molecular interactions between kinases and small ligands that should be of practical use in the development of novel inhibitors. The data set is freely available to the academic community via a user-friendly Web interface at http://cssb.biology.gatech.edu/kinomelhm/ as well as at the ZINC Web site ( http://zinc.docking.org/applications/2010Apr/Brylinski-2010.tar.gz ).

Figure 1

Availability of the ligand-bound and ligand-free crystal structures for the human kinome. Inset: Histogram of the number of abstracts published since 1995 selected from the PubMed using following queries: (“inhibitor”[Text Word]) AND (“YEAR/01/01”[Publication Date]: “YEAR/12/31”[Publication Date]) and ((“inhibitor”[Text Word]) AND (“kinase”[Text Word])) AND (“YEAR/01/01”[Publication Date]: “YEAR/12/31”[Publication Date]).

Figure 2

Hierarchical approach to structural and functional characterization of proteins using homology modeling techniques.

Figure 3

Accuracy of kinase structure modeling using TASSER. Global Cα RMSD (A) and TM-score (B) are calculated versus ligand-bound (holo) and ligand-free (apo) structural forms of the target proteins. Local Ca and all-atom RMSD calculated over the binding residues are shown in C.

Figure 4

ATP-binding pocket detection by FINDSITE. The results are presented as the cumulative fraction of kinase targets with a distance between the center of mass of an inhibitor in the crystal complex and the center of the predicted binding sites, less than or equal to the distance displayed on the x axis. Open circles show the results for a non-redundant (nr) dataset with respect to the target proteins. Gray area corresponds to randomly selected patches on the protein surface. Inset: Matthew’s correlation coefficient calculated for the predicted binding residues.

Figure 5

Docking accuracy of the ligand homology modeling approach applied to the human kinome. Fraction of binding residues (A) and specific protein-ligand contacts (B) predicted by FINDSITE^LHM, Q-Dock^LHM and AMMOS is compared to the ligand poses directly transferred from the crystal structures as well as to ligands randomly placed into the binding pockets.

Figure 6

Low-resolution docking/refinement by ligand homology modeling using protein models as the target receptors. A – CDK2, 1oiq; B – PIM1, 1yxx; C – FGFR2, 1oec and D – CDK2, 2btr. Left, middle: Inhibitor binding poses predicted by FINDSITE^LHM and Q-Dock^LHM (solid sticks, colored by atom type) are compared to the crystal structures (transparent sticks). Protein models (binding residues colored in red) are superposed onto the crystal structures of the target kinases (binding residues colored in orange). Right: correlation of the Q-Dock energy score and RMSD from the crystal binding pose for the ligand conformations sampled using Replica Exchange Monte Carlo (REMC). The red line highlights low-energy conformations for the broad range of RMSD values.

Figure 7

Crystal structures of several protein kinases complexed with staurosporine (STU) and ATP. A – CDK2 (STU: 1aq1, ATP: 1b38), B – GSK3B (STU: 1q3d, ADP: 1j1c), C – LCK (STU: 1qpd, ANP: 1qpc), D – PIM1 (STU: 1yhs, AMP: 1yxu), E – PDK1 (STU; 1oky, ATP: 1h1w), F – MAPKAPK2 (STU: 1nxk, ADP: 1ny3). STU, the set ATP/ADP/AMP/ANP and selected binding residues are colored in green, red and blue, respectively.

Figure 8

Modeled structures of protein kinases bound to staurosporine (STU) and ATP. A – CDC2, B – Erk1, C – FGR, D – LYN, E – PKACa, F – PKCa, G – PKCg, H – PKG1, I – smMLCK. STU, ATP and selected binding residues are colored in green, red and blue, respectively. ATP and STU ranks and Z-scores from virtual screening using Q-Dock^LHM against modeled kinase structures are given.

Figure 9

Performance of virtual screening on the BindingDB dataset. Active compounds are sorted by increasing rank reported by FINDSITE fingerprints (ligand-based screening), Q-Dock^LHM (structure-based screening, low-resolution) and AMMOS (structure-based screening, high-resolution). Inset: ATP ranks for all protein kinases; for FINDSITE, the ranks in the KEGG compound library are used.

Figure 10

Virtual screening for protein kinase C inhibitors. The enrichment behavior for FINDSITE (molecular fingerprints), Q-Dock^LHM (total energy score and the pocket-specific component) and AMMOS (all-atom scoring) is compared to a random ligand selection for different isoenzymes of PKC.

Figure 11

Prediction of the PKC isoenzyme selectivity of known PKC inhibitors from MDDR. A – three-state binding assignment of good (IC₅₀ <100 nM), weak (100nM < IC₅₀ < 1 μM) and non-binders (IC₅₀ >1 μM) by machine learning. B – number of MDDR compounds predicted to inhibit different PKC isoforms with IC₅₀ <100 nM, C – number of hits returned by the Google search engine (http://www.googlefight.com/) using different PKC isoenzyme inhibitors as the query phrases.

Figure 12

Docking times for FINDSITE^LHM, Q-Dock^LHM and AMMOS. Boxes end at the quartiles Q₁ and Q₃; a horizontal line in a box is the median. “Whiskers” point at the farthest points that are within 3/2 times the interquartile range. Outliers and suspected outliers are presented as solid and blank circles, respectively.