Improved Method for the Identification and ... - ACS Publications

Aug 21, 2017 - the training and test sets, respectively, bolstering the good predictive ability of this model. Furthermore, we ... detection software ...
1 downloads 24 Views 3MB Size
Article pubs.acs.org/jcim

Improved Method for the Identification and Validation of Allosteric Sites Kun Song,† Xinyi Liu,† Wenkang Huang,† Shaoyong Lu,† Qiancheng Shen, Lu Zhang, and Jian Zhang* Department of Pathophysiology, Key Laboratory of Cell Differentiation and Apoptosis of National Ministry of Education, Shanghai Jiao Tong University, School of Medicine, Shanghai, 200025, China S Supporting Information *

ABSTRACT: Allosteric regulation induced by modulators binding to different, often distant, allosteric sites allows for exquisite control of protein functional activity. The structural diversity of allosteric sites endows allosteric modulators with high selectivity and low toxicity. Targeting allosteric sites, a novel tactic in drug discovery, has garnered much attention in the scientific community, and the identification of allosteric sites has become an important component of the development of allosteric drugs. Here we present AllositePro, a method which predicts allosteric sites in proteins by combining pocket features with perturbation analysis. The performance of AllositePro is superior to that of the other currently available methods. Using AllositePro, we predicted a novel allosteric site in cyclin-dependent kinase 2 (CDK2) and validated it by site-directed mutagenesis assay. Thus, the AllositePro method provides an effective way to identify allosteric sites and could be a useful strategy for allosteric drug discovery.



allosteric sites in proteins have currently been developed.6 For instance, Qi et al. built a coarse-grained two-state Go̅ model to identify allosteric sites.7 Founded on normal-mode analysis (NMA) to account for protein flexibility, Panjkovich et al. developed a web server called PARS to predict the location of allosteric sites.8 Recently, some elegant computational methods have been committed to probe allosteric dynamics and allosteric pocket detection. Berezovsky introduced a metric, called “binding leverage”, to measure the ability of a binding cavity to connect to the intrinsic movement of a system.9 Using a dynamics method, Mitternacht and Berezovsky developed a method named SPACER to locate biologically relevant allosteric sites.10 The same author developed a novel statistical mechanical model to quantify allosteric free energy to evaluate the configurational effect as the result of binding allosteric ligands to the protein.11 Then, they introduced some state-ofthe-art tools to predict allosteric sites and quest allostery, meanwhile, pointed out that from a functional site of interest, regulatory sites can be located in terms of the communication between them.12 The investigation of conformational changes has also been applied to probe allosteric binding. Taguchi et al.

INTRODUCTION Allostery, a major regulatory mechanism, is fundamentally important to many biological processes. Topographically, an allosteric site is an area of a protein distinct from an orthosteric site that can be targeted by allosteric modulators. Signal propagation from allosteric sites to different, often distant, orthosteric sites allows for fine-tuned control of protein functional activity.1 Due to the higher structural diversity conferred by allosteric sites compared to orthosteric sites or protein−protein interaction surfaces,2 allosteric modulators can exploit unique features of an allosteric site that distinguish the target protein from other homologous proteins. As a result, allosteric modulators have the potential to achieve higher selectivity, fewer side effects, and lower toxicity compared to orthosteric ligands.3 Despite extensive interest in targeting allosteric sites in drug development, experimentally identifying biologically relevant allosteric sites is challenging. In fact, allosteric sites were discovered adventitiously using small-molecules. Among the alternative approaches is the development of computational approaches to detect allosteric sites. Some great works on investigating allosteric binding4 and regulation5 promote the need of designing in silico predictive method to detect allosteric pockets. A minority of computational platforms to identify © 2017 American Chemical Society

Received: January 8, 2017 Published: August 21, 2017 2358

DOI: 10.1021/acs.jcim.7b00014 J. Chem. Inf. Model. 2017, 57, 2358−2363

Article

Journal of Chemical Information and Modeling

Figure 1. Distribution of three calculated site descriptors for allosteric sites and nonallosteric sites shown as boxplots. (left to right) Normalized maximum distance between two alpha sphere, normalized hydrophobicity density, flexibility score.

found that the change of fluctuation around the ligand can be a significant identifier to discriminate dynamics-driven allosteric sites from other sites.13 However, the accuracy of these in silico approaches in the identification of allosteric sites in general proteins still leaves much to be improved. More importantly, none of these methods have been sculpted to determine as-yet undiscovered allosteric sites in practice. In the study, using a combination of structure features and NMA-based approaches, we developed AllositePro to predict the location of allosteric sites in proteins. In benchmark studies, AllositePro can identify 52% and 63% known allosteric sites in the training and test sets, respectively, bolstering the good predictive ability of this model. Furthermore, we have verified the predictions of this approach experimentally, through the discovery of unknown allosteric sites in cyclin dependent kinase 2 (CDK2). Overall, we believe that AllositePro is effective and useful for the allosteric community to predict allosteric sites in proteins, thereby contributing to allosteric drug discovery.



iterations to obtain a more accurate pocket volume result, and other parameters were set by default. Redundant pockets were removed. The pockets with the largest overlap with allosteric modulators were labeled as the true allosteric sites. In total, 1510 sites (147 known allosteric sites and 1363 other sites) were predicted. Twenty-five site descriptors (Table S2) were calculated for each site using Fpocket.17 Details about the pocket descriptors can be found in the manual of Fpocket. Feature-Based Method. Here, we built the pocket feature model using a logistic regression method. The R package was used for feature selection, regression analysis and statistical tests. In order to select the suitable model of descriptors, all the possible combinations were generated between the variables. To overcome the class-imbalance problem in training the machine learning model, nonallosteric negative pockets were sampled randomly by the ratio of 2:1 to the allosteric pockets. Logistic regression was used to study the association between each feature and the outcome variables of known allosteric sites (coded 1) and nonallosteric sites (coded 0). The predicted response (probability of an outcome being allosteric) using pocket features, called Pfeature, is calculated as

MATERIALS AND METHODS

Data Collection. The allosteric sites used to construct AllositePro were collected from the ASBench data set.14 ASBench is a high-quality benchmarking data set of known allosteric sites determined by crystal structures, including a “Core set” composed of 235 representative crystal complexes for allosteric sites and a “Core-Diversity” set with 147 structurally diverse allosteric sites sharing a pairwise Pocket Similarity score (PS-score)15 of less than 0.5. To avoid redundancy, 147 allosteric sites from 127 allosteric proteins in the Core-Diversity set were used in this study. The protein set contains a large diversity of proteins, including transferases, hydrolases, oxidoreductases, and transcription factors. Allosteric sites were defined as the binding sites for allosteric modulators. An external test set was collected from the Protein Data Bank16 using the protocol which is the same as the data processing workflow in ASBench. Finally, 24 novel allosteric sites were carefully established from PDB as the external data set (Table S1). Pocket Detection. The original structures of 127 allosteric proteins were retrieved from PDB. The functional chains of proteins with allosteric sites were saved as a PDB format after removing water molecules, ligands, and redundant chains. Fpocket17 was used to detect binding pockets on the protein surface. Fpocket is a very fast open source protein pocket detection software package based on Voronoi tessellation. The parameter −v in the Fpocket calculation was set to 10 000

Pfeature =

1 1+e

−(β0 + β1X1+ ... +βnX n)

(1)

where β is the intercept and coefficients of the pocket descriptors (X1, ... Xn). All covariates should be statistically significant with an alpha less than or equal to 0.05. The goodness of fit of the full model should also be significant (i.e., p ≤ 0.05). Considering the imbalanced data set of positives and negatives, we chose the MCC (Matthews correlation coefficient) as the evaluation method in the variable selection procedure. Finally, the three descriptors (Normalized maximum distance between two alpha spheres, Normalized hydrophobicity Density, Flexibility score) with the highest MCC score were selected to build the logistic regression model. These descriptors are consistent with properties already described as important in differentiating allosteric pockets and nonallosteric pockets.18−20 Allosteric pockets tend to be larger, more hydrophobic, and more stable (Figure 1). The resulting logistic regression function is represented by eq 1 and the parameters in Table 1. Perturbation-Based Method. In this study, normal-mode analysis (NMA), an efficient way to study probable cooperative motions of biomolecules,21 was used to evaluate protein dynamic changes triggered by allosteric ligands. We used the 2359

DOI: 10.1021/acs.jcim.7b00014 J. Chem. Inf. Model. 2017, 57, 2358−2363

Article

Journal of Chemical Information and Modeling PAllosite = 0.8Pfeature + 0.2PNMA

Table 1. Constants of Logistic Regression Model feature intercept normalized maximum distance between two alpha sphere normalized hydrophobicity density flexibility

coefficient

std error

−1.8001

0.3595

1.7490

0.4040

2.2458

0.4351

−3.8057

0.9871

Quality Criteria. The model quality was estimated using eqs 7−10. Sensitivity assesses the ability to predict allosteric sites, while specificity assesses the ability to predict nonallosteric sites. In this work, another important criterion, Matthew’s correlation coefficient (MCC), was used to estimate model quality, correcting for the imbalance between allosteric sites and other sites in the data set.

p-value 5.52 × 10−7 1.50 × 10−5 2.45 × 10−7 1.16 × 10−4

elastic network model in our NMA calculations. In this model, protein structure is simplified to a network of Cα atoms, and the lowest-energy conformation is given by the initial crystal structure. The interactions between pairwise particles within a cutoff distance (Rc = 8 Å) are modeled as Hookean springs based on a harmonic potential.22 Potential energy V is expressed as follows: 1 V= ∑ k 0(dij − dij0)2 2 d 0< R ij

8π 2 (ΔR i)2 3

(3)

M

∑ j=1

aij 2 ωj 2

specificity =

TN TN + FP

(8)

precision =

TN TN + FN

(9)

where TP are true positives, FP, false positives, TN, true negatives and FN, false negatives. Enzyme Expression. Wild type pUHD-CDK2 was purchased from Addgene. The pGEX-2T-CAK1 and the pGEX-KG -cyclin A2 (comprising residues 173−432) were presented by Dr. Philipp Kaldis (National Cancer Institute). The DNA sequence of the regulatory domain of cyclin A (a.a173−a.a432, cyclin A2) and CDK2 were amplified by PCR. Both of them were cloned into PET28a (+) vector with an Nterminal His-tag. Mutant CDK2 plasmid including Arg150, Arg169, and Tyr180 was designed by the method of sitedirected mutagenesis; all of these amino acids were mutated into Ala. Besides all, combined mutation in the position of Arg169 and Tyr180 were also changed into Ala. All the recombinant plasmids were transfected into E. coli BL21. For CDK2 and cyclinA2, bacteria were incubated in 2XYT medium at 37 °C, and CDK2 and cyclinA2 were induced with 0.1 mM isopropyl-D-thiogalactoside (IPTG) when the optical density was between 0.6 and 0.8. Both of them were incubated at 16 °C for 18 h. For CAK1, when the optical density was 0.9, and then the bacterium were incubated at 16 °C for 1.5 h; then, CAK1P were induced with 0.4 mM IPTG and further incubated for 29.5 h at 16 °C. Precipitates were gathered after centrifugation at 4500 rpm for 20 min. Enzyme Purification. Resuspend and ultrasonic cracked bacterial cell pellet in lysis buffer (300 mM NaCl, 20 mM Tris/ HCl pH 8.0, 10 mM imidazole, 1 mM PMSF, and 5% NP-40). The supernatant of CDK2 and cyclin A2 after centrifugation (1 h at 20 000g) were purified by Ni-NTA agarose (GE LifeSciences), proteins were eluted with imidazole of various consistencies (20−400 mM) in 100 mM NaCl, 20 mM Tris/ HCl pH 7.5. GST-tagged CAK1 was purified by GST-affinity column chromatography (GE LifeSciences). To activate the CDK2-cyclin A2 complex, first of all, 2 mg of CDK2, 1 mg GST-CAK1, 5 mM ATP was mixed in 5 mL solution of 50 mM Tris/HCl pH 7.5 and 10 mM MgCl2. The mixture was incubated for 1.5 h at 25 °C, then used gel chromatography (Superdex 200 (10/300 GL) column) to remove the GSTCAK 1p, then added 2 mg of cyclin A2 to the mixture overnight at 4 °C. Purification of the activated phosphorylated CDK2cyclin A2 complex by gel chromatography (Superdex 200 (10/

where (ΔRi) represents the mean-square fluctuation of the atomic coordinates at position i: KBT mi

(7)

TP × TN − FP × FN

2

(ΔR i)2 =

TP TP + FN

(TP + FP) × (TP + FN) × (TN + FP) × (TN + FN) (10)

where k0 is the force constant, dij is the Euclidean distance between Cα atoms i and j and d0ij is the equilibrium distance between Cα atoms i and j. We computed the theoretical B-factor for estimating protein flexibility, and it is defined as Bi =

sensitivity =

MCC =

(2)

c

(6)

(4)

where aij is the j component of atom i in normal mode i, and ωj is the frequency of the mode j. KB is Boltzmann’s constant, T is the temperature, and mi represents the mass of atom i in a mass-weighted elastic network model. M is the number of models used in the calculation (set to 100 here). The normalmode analysis of all systems were performed using Prody package.23 In order to simulate the binding of an effector to an allosteric site, a dummy molecule with seven points was used to represent the ligand.24,25 Alpha spheres generated by FPocket in each pocket were used to represent the surface point of proteins. Alpha spheres were clustered into seven groups by Euclidean distances, and the representatives of dummy points were sampled from the center of each group. For each protein, B-factor was calculated in both apo structure and structures bound to a dummy ligand. To quantify the significance of conformational changes, a Wilcoxon− Mann−Whitney test was used to measure protein flexibility differences in the presence and absence of an allosteric ligand. A p-value < = 0.05 was considered significant. The score of perturbation method PNMA is defined by the p-value in the Wilcoxon−Mann−Whitney test: 0.05 PNMA = 0.05 + p (5) In the final model, an allosteric site is identified by combining the two methods: 2360

DOI: 10.1021/acs.jcim.7b00014 J. Chem. Inf. Model. 2017, 57, 2358−2363

Article

Journal of Chemical Information and Modeling 300 GL) column) to remove the p-CDK2, cyclinA2. The protein was stored in solution containing 50 mM Tris/HCl (pH 7.5), 10 mM MgCl2, 2 mM DTT, 5% (v/v) DMSO. Enzyme Kinetics. Kinetics of enzyme-catalyzed reactions were tested by a commercially available luminescent-based method (ADP-Glo Max Assay, Promega).26 In this assay, detectable luminescence reflected the level of phosphorylate for CDK2/cyclin A2 complex to its substrate, here it was replaced by an synthetic peptide HHASPRK. To calculate Michaelis− Menten equation, standard curve should be made. Diluting the 10 mM ADP from ADP-Glo Max Assay kit to 120.00, 24.00, 4.80, and 0.96 μM of final concentration. Using a Synergy H4 Hybrid Microplate Reader (BioTek) to capture luminescence, carrying out linear regration and the equation is Y = 97.4x − 11.4, R2 = 1. Then diluting ATP to 600, 300, 150, 75, 37.5, 18.75, 9.375, 3.75, 1.875, 0.75, 0.375, 0.15, and 0.075 μM. The reaction was started in 384 plates with 0.15625 μmol activated CDK-cyclin A2 complex and 250 μM peptide substrate and different concentrations of the ATP mixture; reaction time is 10 min. According to the value of luminescence and the standard curve, velocity can be calculated as the Y axis and solution concentration of ATP was posted as the X axis. The Michaelis− Menten curve can fit the data with the use of GraphPad Prism 6.

Table 3. Prediction of Allosteric Sites by Various Methods on Recently Released Allosteric Sites (External Test Set)a

SE

SP

PRE

MCC

PARS Allosite AllositePro (5-CV)

147 147 147

912 1363 1363

0.286 0.395 0.517

0.739 0.882 0.865

0.150 0.266 0.292

0.019 0.234 0.300

SE

SP

PRE

MCC

168 261 261

0.375 0.500 0.625

0.756 0.858 0.885

0.180 0.245 0.333

0.099 0.264 0.388

the tool. For those proteins with multiple structures, it is recommended to select one with common conformation since different results would be predicted due to large conformational changes of the same protein. In AllositePro, the crystal structure of CDK2 (PDB code: 2C6K) was uploaded as the input structure. As a result, one pocket with a score of 0.648 was predicted to be a novel allosteric site. From the complex structure of CDK2−peptide−ATP,27 two residues at the predicted pocket, R150 and Y180, are at least 5 Å away from either ATP-binding site or substrate site. In addition, we discovered two novel allosteric inhibitors based on the pocket around the two residues and the inhibitors did not competitively affect substrate and ATP binding.28 To confirm the allosteric regulation of the pocket, further mutagenesis studies demonstrated that two mutations R150A and Y180A in the predicted allosteric site obviously attenuated the catalytic activity of CDK2, suggesting the potential use of the predicted allosteric pocket in the design of allosteric inhibitors of CDK2 (Figure 2).



DISCUSSION As shown in Figure 1, the properties of allosteric sites range widely. The feature-based method can only cover a part of allosteric sites. We used perturbation-based method to improve model accuracy. In the ASBench test, we have found six cases that can be identified by ASBench mainly contributed by perturbation method, which have a score lower than 0.5 in feature-based model (Figure 3). For example, Ribonucleotide reductase (RNR) is an enzyme catalyzing the biosynthesis of deoxyribonucleotides from the corresponding ribonucleotides, which play crucial roles in DNA synthesis.29 This protein is regulated allosterically by the binding of a ATP to control general activity, as shown in Figure 3, the red pocket matches the binding site of ATP. In our analysis, due to small pocket size (normalized maximum distance between two alpha spheres: 0.419) and high flexibility (0.489), the feature-based method returns a probability of 0.446, but in the normal-mode analysis mode RNR significantly affects overall protein flexibility if occupied by a ligand and the perturbation score is 0.893. After combining the logistic regression score and NMA score, the allosteric site in RNR can be identified in AllositePro. Likewise, another five allosteric sites in ASBench have been found by combination method in Caspase-7, Ribonucleoside-diphosphate reductase, UDP-glucose 6-dehydrogenase, Pyruvate carboxylase, and POL polyprotein (Figure 3). The novel allosteric pocket was identified on a representative conformation of CDK2 (PDB ID 2C6K) using our method. Meanwhile, we used the method on other different conformations of CDK2 (e.g., 3PXZ and 5ANE) but failed to find potential allosteric sites. Nussinov et al. showed that the proportion of conformations in allosteric state is actually quite

Table 2. Prediction of Allosteric Sites by Various Methods on Allosteric Benchmarking Set Core-Diversity of ASBencha neg

neg

24 24 24

Abbreviations: SE sensitivity; SP specificity; PRE precision; MCC Matthew’s correlation coefficient. Detailed information is given in the Quality Criteria section.

RESULTS Crystal structures of 147 nonredundant allosteric sites from the Core-Diversity set of ASBench were used to train the model. Meanwhile, 24 novel allosteric sites after ASBench release were carefully selected from Protein Data Bank (PDB) as test set. Using feature-based regression combined with perturbation, a model for allosteric site identification was trained and tested. The performance of AllositePro was evaluated and compared to the other widely used methods for identifying allosteric sites, including Allosite and PARS. Using the Core-Diversity set of ASBench, the accuracy of the 5-fold cross-validation by the Matthews Correlation Coefficient (MCC) in AllositePro was 0.300 (76/147 allosteric sites successfully predicted; Table 2),

pos

pos

a



methods

methods PARS Allosite AllositePro

a

Abbreviations: SE sensitivity; SP specificity; PRE precision; MCC Matthew’s correlation coefficient; 5-CV 5-fold cross-validation. Detailed information is given in the Quality Criteria section.

which was 28.2% higher than Allosite and PARS, showing a remarkable improvement in prediction. Meanwhile, the sensitivity and precision of AllositePro increased >30% and >14% compared to the other methods, respectively, indicating a good ability to identify true allosteric sites. To further validate the reliability of AllositePro, 24 novel allosteric sites recently released from PDB were tested using each method. AllositePro found 15 sites, a similar success rate to its performance in the cross-validation (Table 3) and a superior success rate compared to the other two methods (12 sites successfully predicted by Allosite and 9 by PARS). To investigate AllositePro in practice, CDK2 was selected as a case to verify if a novel allosteric site could be discovered by 2361

DOI: 10.1021/acs.jcim.7b00014 J. Chem. Inf. Model. 2017, 57, 2358−2363

Article

Journal of Chemical Information and Modeling

Figure 2. Prediction and validation of allosteric site in CDK2. (A) The workflow of AllositePro. (B) View of the novel allosteric site of CDK2 predicted by AllositePro. The predicted allosteric site is represented by mesh. The green pocket represents the substrate site (orthosteric site), and the ATP is shown in stick mode. Two key residues in the predicted allosteric site Arg150 and Tyr180 were colored pink. (C) Reaction rate versus ATP concentration plot of wild-type, R150A, and Y180A CDK2 kinase activity.

Figure 3. View of the predicted allosteric sites of six proteins predicted by AllositePro whose perturbation method plays an important role in the prediction. Regions of allosteric sites are highlighted in red on the surface. Pseudoligands are shown as seven spheres.

small among conformation ensembles in many proteins.1,30 Here, allosteric sites in CDK2 may also emerge in the minority of conformations. To explore the impact of different conformations at the allosteric site in 2C6K, we compared the other conformations to 2C6K. The results showed that many conformations of CDK2 changed in various degrees, leading to disrupt the allosteric pocket. For example, some conformations displaced in backbone to vanish the allosteric site, such as 3PXZ; some conformations dramatically reoriented side chains of residues (180Y, 150R, 178K, 122R, 57E, or 51E) at the site, which induced the pocket into be shallow, such as 5ANE. AllositePro was trained by known allosteric sites with special space and physiochemical features, the large conforma-

tional changes different from allosteric state may perform less optimally on the prediction of allosteric sites. Taken together, the conformation of the protein plays a crucial role in the prediction of allosteric sites in our method, and it is better to find potential allosteric sites when using multiple diverse conformations of a protein by this method. In summary, we introduced AllositePro for the identification of allosteric sites by combining the feature-based regression and perturbation-based methods. Tests on the allosteric benchmarking data set as well as a successful application in CDK2 showed that AllositePro provides more accurate and more reliable predictions of protein allosteric sites. Therefore, AllositePro can help in allosteric drug design. 2362

DOI: 10.1021/acs.jcim.7b00014 J. Chem. Inf. Model. 2017, 57, 2358−2363

Article

Journal of Chemical Information and Modeling



(11) Guarnera, E.; Berezovsky, I. N. Structure-Based Statistical Mechanical Model Accounts for the Causality and Energetics of Allosteric Communication. PLoS Comput. Biol. 2016, 12, e1004678. (12) Guarnera, E.; Berezovsky, I. N. Allosteric sites: remote control in regulation of protein activity. Curr. Opin. Struct. Biol. 2016, 37, 1−8. (13) Taguchi, J.; Kitao, A. Dynamic profile analysis to characterize dynamics-driven allosteric sites in enzymes. Biophys Physicobiol. 2016, 13, 117−126. (14) Huang, W.; Wang, G.; Shen, Q.; Liu, X.; Lu, S.; Geng, L.; Huang, Z.; Zhang, J. ASBench: Benchmarking Sets for Allosteric Discovery. Bioinformatics 2015, 31, 2598−2600. (15) Gao, M.; Skolnick, J. APoc: Large-Scale Identification of Similar Protein Pockets. Bioinformatics 2013, 29, 597−604. (16) Berman, H. M. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235−242. (17) Le Guilloux, V.; Schmidtke, P.; Tuffery, P. Fpocket: An Open Source Platform for Ligand Pocket Detection. BMC Bioinf. 2009, 10, 168. (18) Li, X.; Chen, Y.; Lu, S.; Huang, Z.; Liu, X.; Wang, Q.; Shi, T.; Zhang, J. Toward an Understanding of the Sequence and Structural Basis of Allosteric Proteins. J. Mol. Graphics Modell. 2013, 40, 30−39. (19) Yang, J.-S.; Seo, S. W.; Jang, S.; Jung, G. Y.; Kim, S. Rational Engineering of Enzyme Allosteric Regulation through Sequence Evolution Analysis. PLoS Comput. Biol. 2012, 8, e1002612. (20) Lu, S.; Li, S.; Zhang, J. Harnessing Allostery: A Novel Approach to Drug Discovery. Med. Res. Rev. 2014, 34, 1242−1285. (21) Bahar, I.; Rader, A. J. Coarse-Grained Normal Mode Analysis in Structural Biology. Curr. Opin. Struct. Biol. 2005, 15, 586−592. (22) Tirion, M. M. Large Amplitude Elastic Motions in Proteins from a Single-Parameter, Atomic Analysis. Phys. Rev. Lett. 1996, 77, 1905− 1908. (23) Bakan, A.; Meireles, L. M.; Bahar, I. ProDy: Protein Dynamics Inferred from Theory and Experiments. Bioinformatics 2011, 27, 1575−1577. (24) Hert, J.; Keiser, M. J.; Irwin, J. J.; Oprea, T. I.; Shoichet, B. K. Quantifying the Relationships among Drug Classes. J. Chem. Inf. Model. 2008, 48, 755−765. (25) Ming, D.; Wall, M. E. Allostery in a Coarse-Grained Model of Protein Dynamics. Phys. Rev. Lett. 2005, 95, 198103. (26) Saalau-Bethell, S. M.; Berdini, V.; Cleasby, A.; Congreve, M.; Coyle, J. E.; Lock, V.; Murray, C. W.; O’Brien, M. A.; Rich, S. J.; Sambrook, T.; Vinkovic, M.; Yon, J. R.; Jhoti, H. Crystal Structure of Human Soluble Adenylate Cyclase Reveals a Distinct, Highly Flexible Allosteric Bicarbonate Binding Pocket. ChemMedChem 2014, 9, 823− 832. (27) Brown, N. R.; et al. The structural basis for specificity of substrate and recruitment peptides for cyclin-dependent kinases. Nat. Cell Biol. 1999, 1, 438−443. (28) Hu, Y.; Li, S.; Liu, F.; Geng, L.; Shu, X.; Zhang, J. Discovery of novel nonpeptide allosteric inhibitors interrupting the interaction of CDK2/cyclin A3 by virtual screening and bioassays. Bioorg. Med. Chem. Lett. 2015, 25, 4069−4073. (29) Eriksson, M.; Uhlin, U.; Ramaswamy, S.; Ekberg, M.; Regnström, K.; Sjöberg, B. M.; Eklund, H. Binding of Allosteric Effectors to Ribonucleotide Reductase Protein R1: Reduction of Active-Site Cysteines Promotes Substrate Binding. Structure 1997, 5, 1077−1092. (30) Nussinov, R.; Tsai, C.-J. Unraveling structural mechanisms of allosteric drug action. Trends Pharmacol. Sci. 2014, 35, 256−264.

ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.jcim.7b00014. Additional information about the external data set of AllositePro (PDF)



AUTHOR INFORMATION

Corresponding Author

*Phone: +86-21-63846590-776922. Fax: +86-21-64154900. Email: [email protected] (J.Z.). ORCID

Jian Zhang: 0000-0002-6558-791X Author Contributions †

The first four authors equally contribute to this work.

Notes

The authors declare no competing financial interest.



ACKNOWLEDGMENTS This work was funded in part by grants from the National Basic Research Program of China (973 Program) (2015CB910403), the National Natural Science Foundation of China (81630075, 81322046, 81473137, 81302698), the Shanghai Rising-Star Program (13QA1402300), the Shanghai Sailing Program (17YF1410600), the National Program for Support of Topnotch Young Professionals (2015), and the Program for New Century Excellent Talents in University.



REFERENCES

(1) Nussinov, R.; Tsai, C.-J. Allostery in Disease and in Drug Discovery. Cell 2013, 153, 293−305. (2) Jiang, H.; Deng, R.; Yang, X.; Shang, J.; Lu, S.; Zhao, Y.; Song, K.; Liu, X.; Zhang, Q.; Chen, Y.; Chinn, Y. E.; Wu, G.; Li, J.; Chen, G.; Yu, J.; Zhang, J. Peptidomimetic inhibitors of APC-Asef interaction block colorectal cancer migration. Nat. Chem. Biol. 2017, 13, 994−1001. (3) Shen, Q.; Wang, G.; Li, S.; Liu, X.; Lu, S.; Chen, Z.; Song, K.; Yan, J.; Geng, L.; Huang, Z.; Huang, W.; Chen, G.; Zhang, J. ASD v3.0: Unraveling Allosteric Regulation with Structural Mechanisms and Biological Networks. Nucleic Acids Res. 2016, 44, D527−D535. (4) Li, S.; Shen, Q.; Su, M.; Liu, X.; Lu, S.; Chen, Z.; Wang, R.; Zhang, J. Alloscore: a method for predicting allosteric ligand-protein interactions. Bioinformatics 2016, 32, 1574−1576. (5) Shen, Q.; Cheng, F.; Song, H.; Lu, W.; Zhao, J.; An, X.; Liu, M.; Chen, G.; Zhao, Z.; Zhang, J. Proteome-scale investigation of protein allosteric regulation perturbed by somatic mutations in 7,000 cancer genomes. Am. J. Hum. Genet. 2017, 100, 5−20. (6) Lu, S.; Huang, W.; Zhang, J. Recent Computational Advances in the Identification of Allosteric Sites in Proteins. Drug Discovery Today 2014, 19, 1595−1600. (7) Qi, Y.; Wang, Q.; Tang, B.; Lai, L. Identifying Allosteric Binding Sites in Proteins with a Two-State Go̅ Model for Novel Allosteric Effector Discovery. J. Chem. Theory Comput. 2012, 8, 2962−2971. (8) Panjkovich, A.; Daura, X. PARS: A Web Server for the Prediction of Protein Allosteric and Regulatory Sites. Bioinformatics 2014, 30, 1314−1315. (9) Mitternacht, S.; Berezovsky, I. N. Binding Leverage as a Molecular Basis for Allosteric Regulation. PLoS Comput. Biol. 2011, 7, e1002148. (10) Goncearenco, A.; Mitternacht, S.; Yong, T.; Eisenhaber, B.; Eisenhaber, F.; Berezovsky, I. N. SPACER: Server for Predicting Allosteric Communication and Effects of Regulation. Nucleic Acids Res. 2013, 41, W266−W272. 2363

DOI: 10.1021/acs.jcim.7b00014 J. Chem. Inf. Model. 2017, 57, 2358−2363