Virtual Screening of Abl Inhibitors from Large Compound Libraries by


Virtual Screening of Abl Inhibitors from Large Compound Libraries by...

1 downloads 132 Views 1MB Size

J. Chem. Inf. Model. 2009, 49, 2101–2110

2101

Virtual Screening of Abl Inhibitors from Large Compound Libraries by Support Vector Machines X. H. Liu,† X. H. Ma,†,⊥ C. Y. Tan,‡ Y. Y. Jiang,‡ M. L. Go,§ B. C. Low,⊥ and Y. Z. Chen*,† Bioinformatics and Drug Design Group, Department of Pharmacy, Centre for Computational Science and Engineering, National University of Singapore, Blk S16, Level 8, 3 Science Drive 2, Singapore 117543, The Key Laboratory of Chemical Biology, Guangdong Province, The Graduate School at Shenzhen, Tsinghua University, Shenzhen, Guangdong, P. R. China, Department of Pharmacy, National University of Singapore, BLK S4, 18 Science Drive 4, Singapore 117543, and Department of Biological Science, National University of Singapore, Blk S2, Level 5, Science Drive 4, Singapore 117543 Received April 14, 2009

Abl promotes cancers by regulating cell morphogenesis, motility, growth, and survival. Successes of several marketed and clinical trial Abl inhibitors against leukemia and other cancers and appearances of reduced efficacies and drug resistances have led to significant interest in and efforts for developing new Abl inhibitors. In silico methods of pharmacophore, fragment, and molecular docking have been used in some of these efforts. It is desirable to explore other in silico methods capable of searching large compound libraries at high yields and reduced false-hit rates. We evaluated support vector machines (SVM) as a virtual screening tool for searching Abl inhibitors from large compound libraries. SVM trained and tested by 708 inhibitors and 65 494 putative noninhibitors correctly identified 84.4 to 92.3% inhibitors and 99.96 to 99.99% noninhibitors in 5-fold cross validation studies. SVM trained by 708 pre-2008 inhibitors and 65 494 putative noninhibitors correctly identified 50.5% of the 91 inhibitors reported since 2008 and predicted as inhibitors 29 072 (0.21%) of 13.56M PubChem, 659 (0.39%) of 168K MDDR, and 330 (5.0%) of 6 638 MDDR compounds similar to the known inhibitors. SVM showed comparable yields and substantially reduced falsehit rates against two similarity based and another machine learning VS methods based on the same training and testing data sets and molecular descriptors. These suggest that SVM is capable of searching Abl inhibitors from large compound libraries at low false-hit rates. INTRODUCTION

Abl plays key roles in cancers by regulating morphogenesis and motility and by promoting cell growth and survival via Bcr-Abl mediated activation of Src-family kinases and PI3K, Ras, Myc, c-jun, and STAT pathways.1 Abl inhibitors are effective in the treatment of leukemia and in clinical trials of other cancers.2-4 In some cases, these inhibitors show negligible activity against common mutations and modest effects in advanced cancer phases, and some patients develop resistance associated with Abl kinase domain mutations.4 The successes and problems of these inhibitors have raised significant interest in and has led to intensifying efforts for discovering new Abl inhibitors.4,5 Several in silico methods have been used for facilitating the search and design of Abl inhibitors, which include pharmacophore,6 QSAR,7 scaffold assembly,8 molecular docking,9,10 and their combinations.11,12 These in silico methods have shown impressive capability in the identification of potential Abl inhibitors, but their applications may be affected by such problems as the vastness and sparse nature of chemical space that needs to * Corresponding author: Telephone: 65-6874-6877. Fax: 65-6774-6756. E-mail: [email protected]. † Bioinformatics and Drug Design Group, Department of Pharmacy, National University of Singapore. ⊥ Department of Biological Science, National University of Singapore. ‡ Tsinghua University. § Department of Pharmacy, National University of Singapore.

be searched, the complexity and flexibility of target structures, the difficulties in accurately estimating binding affinity and solvation effects, and the limited diversity of training active compounds.13-15 Therefore, it is desirable to explore other in silico methods that complement these methods with expanded coverage of chemical space, increased screening speed, and reduced false-hit rates without necessarily relying on the modeling of target structural flexibility, binding affinity, and salvation effects. A ligand-based virtual screening (VS) method, support vector machines (SVM), has been explored as such a method that produces high yields and low false-hit rates in searching active agents of single and multiple mechanisms from large compound libraries (i.e., with an expanded applicability domain)16 and in identifying active agents of diverse structures.16-20 Good VS performance can also be achieved by SVM trained from sparsely distributed active compounds.21 SVM classifies active compounds based on differentiating physicochemical profiles between active and inactive compounds rather than on structural similarity to active compounds per se, which has the advantage of not relying on the accurate computation of structural flexibility, activity-related features, binding affinity, and solvation effects. Moreover, the fast speed and expanded applicability domain of SVM enables an efficient search of a vast chemical space. Therefore, SVM may be a potentially useful VS tool

10.1021/ci900135u CCC: $40.75  2009 American Chemical Society Published on Web 08/18/2009

2102

J. Chem. Inf. Model., Vol. 49, No. 9, 2009

to complement other in silico methods for searching ABL inhibitors from large libraries. In this work, we developed a SVM VS model for identifying Abl inhibitors and evaluated its performance by both a 5-fold cross validation test and a large compound database screening test. In the 5-fold cross validation test, a data set of Abl inhibitors and noninhibitors was randomly divided into five groups of approximately equal size, with four groups used for training a SVM VS tool and one group used for testing it; the test process is repeated for all five possible compositions to derive an average VS performance. In the large database screening test, a SVM VS tool was developed by using Abl inhibitors published before 2008, its yield (percent of known inhibitors identified as virtualhits) was estimated by using Abl inhibitors reported since 2008 and not included in the training data sets. Virtual hits and false-hit rate in searching large libraries were evaluated by using 13.56 M PubChem, 168K MDDR, and 6 638 MDDR compounds similar in structural and physicochemical properties to the known Abl inhibitors. PubChem and MDDR contain high percentages of inactive compounds significantly different from the Abl inhibitors, and the easily distinguishable features may make VS enrichments artificially good.22 Nonetheless, certain percentages of PubChem and MDDR compounds are kinase inhibitors or are similar to known Abl inhibitors. For instance, about 1 500 MDDR and 10 000 PubChem compounds are kinase inhibitors, and 6 638 MDDR compounds are similar to at least one known Abl inhibitor. Therefore, VS performance may be more strictly tested by using these and other compounds that resemble the physicochemical properties of the known Abl inhibitors, so that enrichment is not simply a separation of trivial physicochemical features.23 To further evaluate whether our SVM VS tool predicts Abl inhibitors and noninhibitors rather than membership of certain compound families, distribution of the predicted active and inactive compounds in the compound families was analyzed. Moreover, VS performance of SVM was compared to those of two similar VS methods, Tanimoto similarity searching and k nearest-neighbor (kNN), and to an alternative but equally popular machine learning method, probabilistic neural network (PNN), which is based on the same training and testing data sets (same sets of PubChem and MDDR compounds) and molecular descriptors. In a study that compares the performance of SVM to 16 classification and nine regression methods, it has been reported that SVMs show mostly good performances both on classification and regression tasks, but other methods proved to be very competitive.24 Therefore, it is useful to evaluate the VS performance of SVM in searching large compound libraries by comparison with those of similarity based approaches and other typical machine learning methods. METHODS

Compound Collections and Construction of Training and Testing Data Sets. A total of 708 Abl inhibitors, with IC50 < 50 µM, were collected from the literatures11,25-27 and the BindingDB database.28 The inhibitor selection criterion of IC50 < 50 µM was used because it covers most of the reported HTS and VS hits.29,30 The structures of

LIU

ET AL.

Figure 1. The structures of representative Abl inhibitors.

representative Abl inhibitors are shown in Figure 1. As few noninhibitors have been reported, putative noninhibitors were generated by using our method for generating putative inactive compounds.16,21 This method requires no knowledge of known inactive and active compounds of other target classes, which enables more expanded coverage of the “noninhibitor” chemical space. Although the yet to be discovered inhibitors are likely distributed in some of these noninhibitor families, a substantial percentage of these inhibitors are expected to be identified as inhibitors rather than noninhibitors, even though representatives of their families are putatively assigned as noninhibitors.16 The 13.56M PubChem and 168K MDDR compounds were grouped into 8 423 compound families by clustering them in the chemical space defined by their molecular descriptors.31,32 The number of generated families is consistent with the 12 800 compound-occupying neurons (regions of topologically close structures) for 26.4M compounds of up to 11 atoms33 and the 2,851 clusters for 171,045 natural products.34 The collected Abl inhibitors are distributed in 221 families. Because of the extensive efforts in searching kinase inhibitors from known compound libraries, the number of undiscovered Abl inhibitor families in PubChem and MDDR databases is expected to be relatively small, most likely no more than several hundred families. The ratio of the discovered and undiscovered inhibitor families (hundreds) and the families that contain no known inhibitor of each kinase (8 423 based on the current versions of PubChem and MDDR) is expected

VIRTUAL SCREENING

OF

ABL INHIBITORS

J. Chem. Inf. Model., Vol. 49, No. 9, 2009 2103

Table 1. Molecular Descriptors Used in This Work descriptor class simple molecular properties

78

number of descriptors in class

descriptors

18

number of C,N,O,P,S, number of total atoms, number of rings, number of bonds, number of non-H bonds, molecular weight, number of rotatable bonds, number of H-bond donors, number of H-bond acceptors, number of five-member aromatic rings, number of six-member aromatic rings, number of N heterocyclic rings, number of O heterocyclic rings, number of S heterocyclic rings. Sanderson electronegativity, molecular polarizability, aLogp Schultz molecular topological index, Gutman molecular topological index, Wiener index, Harary index, gravitational topological index, molecular path count of length 1-6, total path count, Balaban Index J, 0-2th valence connectivity index, 0-2th order delta chi index, Pogliani index, 0-2th solvation connectivity index, 1-3th order Kier shape index, 1-3th order kappa alpha shape index, Kier molecular flexibility index, topological radius, graph-theoretical shape coefficient, eccentricity, centralization, Logp from connectivity. sum of estate of atom type sCH3, dCH2, ssCH2, dsCH, aaCH, sssCH, dssC, aasC, aaaC, sssC, sNH3, sNH2, ssNH2, dNH, ssNH, aaNH, dsN, aaN, sssN, ddsN, aOH, sOH, ssO, sSH; sum of estate of all heavy atoms, all C atoms, all hetero atoms, sum of estate of H-bond acceptors, sum of H estate of atom type HsOH, HdNH, HsSH, HsNH2, HssNH, HaaNH, HtCH, HdCH2, HdsCH, HaaCH, HCsats, HCsatu, Havin, sum of H estate of H-bond donors

chemical properties79 molecular connectivity and shape78,80

3 35

electro-topological state78,81

42

to be 6%) of virtual hits antineoplastic signal transduction inhibitor tyrosine-specific protein kinase inhibitor antiarthritic antiangiogenic

Substantial percentages of the MDDR virtual-hits belong to the classes of antineoplastic, signal transduction inhibitors, tyrosine-specific protein kinase inhibitors, antiarthritic, and antiangiogenic (Table 4, details in Evaluation of SVM Identified MDDR Virtual-Hits section). As some of these virtual-hits may be true Abl inhibitors, the false-hit rate of our SVM is at most equal to and likely less than the virtualhit rate. Hence, the false-hit rate is