Avidity‐mediated virus separation using a hyperthermophilic affinity ligand

Immunoaffinity separation of large multivalent species such as viruses is limited by the stringent elution conditions necessary to overcome their strong and highly avid interaction with immobilized affinity ligands on the capture surface. Here we present an alternate strategy that harnesses the avidity effect to overcome this limitation. Red clover necrotic mosaic virus (RCNMV), a plant virus relevant to drug delivery applications, was chosen as a model target for this study. An RCNMV binding protein (RBP) with modest binding affinity (K_D ～100 nM) was generated through mutagenesis of the Sso7d protein from Sulfolobus solfataricus and used as the affinity ligand. In our separation scheme, RCNMV is captured by a highly avid interaction with RBP immobilized on a nickel surface through a hexahistidine (6xHis) tag. Subsequently, disruption of the multivalent interaction and release of RCNMV is achieved by elution of RBP from the nickel surface. Finally, RCNMV is separated from RBP by exploiting the large difference in their molecular weights (～8 MDa vs. ～10 kDa). Our strategy not only eliminates the need for harsh elution conditions, but also bypasses chemical conjugation of the affinity ligand to the capture surface. Stable non‐antibody affinity ligands to a wide spectrum of targets can be generated through mutagenesis of Sso7d and other hyperthermophilic proteins. Therefore, our approach may be broadly relevant to cases where capture of large multivalent species from complex mixtures and subsequent release without the use of harsh elution conditions is necessary. © 2012 American Institute of Chemical Engineers Biotechnol. Prog., 2013     

Thermodynamic binding studies of bivalent oligosaccharides to galectin-1, galectin-3, and the carbohydrate recognition domain of galectin-3

Galectins are a growing family of animal lectins with common consensus sequences that bind beta-Gal and LacNAc residues. There are at present 14 members of the galectin family; however, certain galectins possess different structures as well as biological properties. Galectin-1 is a dimer of two homologous carbohydrate recognition domains (CRDs) and possesses apoptotic and proinvasive activities. Galectin-3 consists of a C-terminal CRD and an N-terminal nonlectin domain implicated in the oligomerization of the protein and is often associated with antiapoptotic activity. Because many cellular oligosaccharide receptors are multivalent, it is important to characterize the interactions of multivalent carbohydrates with galectins-1 and -3. In the present study, binding of bovine heart galectin-1 and recombinant murine galectin-3 to a series of synthetic analogs containing two LacNAc residues separated by a varying number of methylene groups, as well as biantennary analogs possessing two LacNAc residues, were examined using isothermal titration microcalorimetry (ITC) and hemagglutination inhibition measurements. The thermodynamics of binding of the multivalent carbohydrates to the C-terminal CRD domain of galectin-3 was also investigated. ITC results showed that each bivalent analog bound by both LacNAc residues to the two galectins. However, galectin-1 shows a lack of enhanced affinity for the bivalent straight chain and branched chain analogs, whereas galectin-3 shows enhanced affinity for only lacto-N-hexaose, a naturally occurring branched chain carbohydrate. The CRD domain of galectin-3 was shown to possess similar thermodynamic binding properties as the intact molecule. The results of this study have important implications for the design of carbohydrate inhibitors of the two galectins.     

Binding of phosphorylated peptides and inhibition of their interaction with disease‐relevant human proteins by synthetic metal‐chelate receptors

The modulation of biological signal transduction pathways by masking phosphorylated amino acid residues represents a viable route toward pharmacologic protein regulation. Binding of phosphorylated amino acid residues has been achieved with synthetic metal‐chelate receptors. The affinity and selectivity of such receptors can be enhanced if combined with a second binding site. We demonstrate this principle with a series of synthetic ditopic metal‐chelate receptors, which were synthesized and investigated for their binding affinity to phosphorylated short peptides under conditions of physiological pH. The compounds showing highest affinity were subsequently used to inhibit the interaction of the human STAT1 protein to a peptide derived from the interferon‐γ receptor, and between the checkpoint kinase Chk2 and its preferred binding motif. Two of the investigated ditopic synthetic receptors show a significant increase in inhibition activity. The results show that regulation of protein function by binding to phosphorylated amino acids is possible. The introduction of additional binding sites into the synthetic receptors increases their affinity, but the flexibility of the structures investigated so far prohibited stringent amino acid sequence selectivity in peptide binding. Copyright © 2009 John Wiley & Sons, Ltd.     

Identification of Novel Contributions to High-affinity Glycoprotein-Receptor Interactions using Engineered Ligands

Engineered receptor fragments and glycoprotein ligands employed in different assay formats have been used to dissect the basis for the dramatic enhancement of binding of two model membrane receptors, dendritic cell-specific intercellular adhesion molecule 3-grabbing nonintegrin (DC-SIGN) and the macrophage galactose lectin, to glycoprotein ligands compared to simple sugars. These approaches make it possible to quantify the importance of two major factors that combine to enhance the affinity of single carbohydrate-recognition domains (CRDs) for glycoprotein ligands by 100-to 300-fold. First, the presence of extended binding sites within a single CRD can enhance interaction with branched glycans, resulting in increases of fivefold to 20-fold in affinity. Second, presentation of glycans on a glycoprotein surface increases affinity by 15-to 20-fold, possibly due to low-specificity interactions with the surface of the protein or restriction in the conformation of the glycans. In contrast, when solution-phase networking is avoided, enhancement due to binding of multiple branches of a glycan to multiple CRDs in the oligomeric forms of these receptors is minimal and binding of a receptor oligomer to multiple glycans on a single glycoprotein makes only a twofold contribution to overall affinity. Thus, in these cases, multivalent interactions of individual glycoproteins with individual receptor oligomers have a limited role in achieving high affinity. These findings, combined with considerations of membrane receptor geometry, are consistent with the idea that further enhancement of the binding to multivalent glycoprotein ligands requires interaction of multiple receptor oligomers with the ligands.     

Development of tacrine clusters as positively cooperative systems for the inhibition of acetylcholinesterase

The synthesis of four tetra-tacrine clusters where the tacrine binding units are attached to a central scaffold via linkers of variable lengths is described. The multivalent inhibition potencies for the tacrine clusters were investigated for the inhibition of acetylcholinesterase. Two of the tacrine clusters displayed a small but significant multivalent inhibition potency in which the binding affinity of each of the tacrine binding units increased up to 3.2 times when they are connected to the central scaffold.     

Harmonic oscillator model of the insulin and IGF1 receptors’ allosteric binding and activation

The insulin and insulin‐like growth factor 1 receptors activate overlapping signalling pathways that are critical for growth, metabolism, survival and longevity. Their mechanism of ligand binding and activation displays complex allosteric properties, which no mathematical model has been able to account for. Modelling these receptors’ binding and activation in terms of interactions between the molecular components is problematical due to many unknown biochemical and structural details. Moreover, substantial combinatorial complexity originating from multivalent ligand binding further complicates the problem. On the basis of the available structural and biochemical information, we develop a physically plausible model of the receptor binding and activation, which is based on the concept of a harmonic oscillator. Modelling a network of interactions among all possible receptor intermediaries arising in the context of the model (35, for the insulin receptor) accurately reproduces for the first time all the kinetic properties of the receptor, and provides unique and robust estimates of the kinetic parameters. The harmonic oscillator model may be adaptable for many other dimeric/dimerizing receptor tyrosine kinases, cytokine receptors and G‐protein‐coupled receptors where ligand crosslinking occurs.     

Solution NMR characterization of WT CXCL8 monomer and dimer binding to CXCR1 N‐terminal domain

Chemokine CXCL8 and its receptor CXCR1 are key mediators in combating infection and have also been implicated in the pathophysiology of various diseases including chronic obstructive pulmonary disease (COPD) and cancer. CXCL8 exists as monomers and dimers but monomer alone binds CXCR1 with high affinity. CXCL8 function involves binding two distinct CXCR1 sites – the N‐terminal domain (Site‐I) and the extracellular/transmembrane domain (Site‐II). Therefore, higher monomer affinity could be due to stronger binding at Site‐I or Site‐II or both. We have now characterized the binding of a human CXCR1 N‐terminal domain peptide (hCXCR1Ndp) to WT CXCL8 under conditions where it exists as both monomers and dimers. We show that the WT monomer binds the CXCR1 N‐domain with much higher affinity and that binding is coupled to dimer dissociation. We also characterized the binding of two CXCL8 monomer variants and a trapped dimer to two different hCXCR1Ndp constructs, and observe that the monomer binds with ～10‐ to 100‐fold higher affinity than the dimer. Our studies also show that the binding constants of monomer and dimer to the receptor peptides, and the dimer dissociation constant, can vary significantly as a function of pH and buffer, and so the ability to observe WT monomer peaks is critically dependent on NMR experimental conditions. We conclude that the monomer is the high affinity CXCR1 agonist, that Site‐I interactions play a dominant role in determining monomer vs. dimer affinity, and that the dimer plays an indirect role in regulating monomer function.     

An engineered ultra‐high affinity Fab‐Protein G pair enables a modular antibody platform with multifunctional capability

Engineered recombinant antibody‐based reagents are rapidly supplanting traditionally derived antibodies in many cell biological applications. A particularly powerful aspect of these engineered reagents is that other modules having myriad functions can be attached to them either chemically or through molecular fusions. However, these processes can be cumbersome and do not lend themselves to high throughput applications. Consequently, we have endeavored to develop a platform that can introduce multiple functionalities into a class of Fab‐based affinity reagents in a “plug and play” fashion. This platform exploits the ultra‐tight binding interaction between affinity matured variants of a Fab scaffold (Fab^S) and a domain of an immunoglobulin binding protein, protein G (GA1). GA1 is easily genetically manipulatable facilitating the ability to link these modules together like beads on a string with adjustable spacing to produce multivalent and bi‐specific entities. GA1 can also be fused to other proteins or be chemically modified to engage other types of functional components. To demonstrate the utility for the Fab‐GA1 platform, we applied it to a detection proximity assay based on the β‐lactamase (BL) split enzyme system. We also show the bi‐specific capabilities of the module by using it in context of a Bi‐specific T‐cell engager (BiTE), which is a therapeutic assemblage that induces cell killing by crosslinking T‐cells to cancer cells. We show that GA1‐Fab modules are easily engineered into potent cell‐killing BiTE‐like assemblages and have the advantage of interchanging Fabs directed against different cell surface cancer‐related targets in a plug and play fashion.     

General analysis of receptor-mediated viral attachment to cell surfaces.

Viruses are multivalent particles that attach to cells through one or more bonds between viral attachment proteins (VAP) and specific cellular receptors. Three modes of virus binding are presented that can explain the diversity in binding data observed among viruses. They are based on multivalency of attachment and spatial versus receptor saturation effects which are easily distinguished based upon simple criteria. Mode 1 involves only monovalent virus/receptor binding. Modes 2 and 3 involve multivalent bonds between the virus and cell; however, in mode 3 space on the cell surface becomes saturated before receptors. A model is developed for viral attachment that accounts for nonspecific binding, receptor/virus interactions, and spatial saturation effects. The model can describe each mode in different limits and can be applied to virus binding data to extract key physical information such as receptor number and affinity. These values are used to postulate the type of VAP/receptor interaction involved and to predict binding at different parameter values. For the mode 2 binding of Adenovirus 2, the model predicts a receptor number of 4-15 x 10(3) on HeLa cells and an affinity of 2-6 x 10(7) M-1 which closely approximate experimental estimates. For the binding of three, broad-host-range, enveloped viruses, Semliki Forest virus, Vesicular Stomatitis virus, and the baculovirus, Autographa californica nuclear polyhedrosis virus, the model predicts receptor numbers of 10(5) or greater and affinities in the range of 10(4) to 10(5) M-1. These values are indicative of a VAP/oligosaccharide interaction which has been documented for a number of other viruses. Experimental evidence is presented that is the first to demonstrate that baculovirus binding is mediated by a cell surface receptor.     

Quantification of receptor targeting aptamer binding characteristics using single-molecule spectroscopy

This experimental design presents a single molecule approach based on fluorescence correlation spectroscopy (FCS) for the quantification of outer membrane proteins which are receptors to an aptamer specifically designed to target the surface receptors of live Salmonella typhimurium. By using correlation analysis, we also show that it is possible to determine the associated binding kinetics of these aptamers on live single cells. Aptamers are specific oligonucleotides designed to recognize conserved sequences that bind to receptors with high affinity, and therefore can be integrated into selective biosensor platforms. In our experiments, aptamers were constructed to bind to outer membrane proteins of S. typhimurium and were assessed for specificity against Escherichia coli. By fluorescently labeling aptamer probes and applying FCS, we were able to study the diffusion dynamics of bound and unbound aptamers and compare them to determine the dissociation constants and receptor densities of the bacteria for each aptamer at single molecule sensitivity. The dissociation constants for these aptamer probes calculated from autocorrelation data were 0.1285 and 0.3772 nM and the respective receptor densities were 42.27 receptors per µm² and 49.82 receptors per µm². This study provides ample evidence that the number of surface receptors is sufficient for binding and that both aptamers have a high‐binding affinity and can therefore be used in detection processes. The methods developed here are unique and can be generalized to examine surface binding kinetics and receptor quantification in live bacteria at single molecule sensitivity levels. The impact of this study is broad because our approach can provide a methodology for biosensor construction and calculation of live single cell receptor‐ligand kinetics in a variety of environmental and biological applications. Bioeng. 2011; 108:1222–1227. © 2010 Wiley Periodicals, Inc.     

Lock and chop: A novel method for the generation of a PICK1 PDZ domain and piperidine‐based inhibitor co‐crystal structure

The membrane protein interacting with kinase C1 (PICK1) plays a trafficking role in the internalization of neuron receptors such as the amino‐3‐hydroxyl‐5‐methyl‐4‐isoxazole‐propionate (AMPA) receptor. Reduction of surface AMPA type receptors on neurons reduces synaptic communication leading to cognitive impairment in progressive neurodegenerative diseases such as Alzheimer disease. The internalization of AMPA receptors is mediated by the PDZ domain of PICK1 which binds to the GluA2 subunit of AMPA receptors and targets the receptor for internalization through endocytosis, reducing synaptic communication. We planned to block the PICK1‐GluA2 protein–protein interaction with a small molecule inhibitor to stabilize surface AMPA receptors as a therapeutic possibility for neurodegenerative diseases. Using a fluorescence polarization assay, we identified compound BIO124 as a modest inhibitor of the PICK1‐GluA2 interaction. We further tried to improve the binding affinity of BIO124 using structure‐aided drug design but were unsuccessful in producing a co‐crystal structure using previously reported crystallography methods for PICK1. Here, we present a novel method through which we generated a co‐crystal structure of the PDZ domain of PICK1 bound to BIO124.     

The bulky and the sweet: How neutralizing antibodies and glycan receptors compete for virus binding

Numerous viruses rely on glycan receptor binding as the initial step in host cell infection. Engagement of specific glycan receptors such as sialylated carbohydrates, glycosaminoglycans, or histo‐blood group antigens can determine host range, tissue tropism, and pathogenicity. Glycan receptor‐binding sites are typically located in exposed regions on viral surfaces—sites that are also generally prone to binding of neutralizing antibodies that directly interfere with virus‐glycan receptor interactions. In this review, we examine the locations and architecture of the glycan‐ and antibody‐binding sites in four different viruses with stalk‐like attachment proteins (reovirus, influenza virus, norovirus, and coronavirus) and investigate the mechanisms by which antibodies block glycan recognition. Those viruses exemplify that direct molecular mimicking of glycan receptors by antibodies is rare and further demonstrate that antibodies often partly overlap or bind sufficiently close to the receptor‐binding region to hinder access to this site, achieving neutralization partially because of the epitope location and partly due to their sheer size.     

Cholesterol‐β1AR interaction versus cholesterol‐β2AR interaction

Two 8‐µs all‐atom molecular dynamics simulations have been performed on the two highly homologous G protein‐coupled receptor (GPCR) subtypes, β₁‐ and β₂‐adrenergic receptors, which were embedded in a lipid bilayer with randomly dispersed cholesterol molecules. During the simulations, cholesterol molecules accumulate to different surface regions of the two receptors, suggesting the subtype specificity of cholesterol–β‐adrenergic receptor interaction and providing some clues to the physiological difference of the two subtypes. Meanwhile, comparison between the two receptors in interacting with cholesterols shed some new light on general determinants of cholesterol binding to GPCRs. Our results indicate that although the concave surface, charged residues and aromatic residues are important, neither of these stabilizing factors is indispensable for a cholesterol interaction site. Different combinations of these factors lead to the diversified binding modes of cholesterol binding to the receptors. Our long‐time simulations, for the first time, revealed the pathway of a cholesterol molecule entering the consensus cholesterol motif (CCM) site, and the binding process of cholesterol to CCM is accompanied by a side chain flipping of the conserved Trp4.50. Moreover, the simulation results suggest that the I‐/V‐/L‐rich region on the extracellular parts of helix 6 might be an alternatively conserved cholesterol‐binding site for the class‐A GPCRs. Proteins 2014; 82:760–770. © 2013 Wiley Periodicals, Inc.     

Binding parameters of antibodies: pseudo-affinity and other misconceptions

In order to evaluate and compare monoclonal antibodies (Abs), functional affinities are generally determined. While the equations that define affinity relate to monovalent interactions, it has been considered that the binding of Abs to multivalent antigens such as the cell surface could be described by an apparent, or functional affinity. We demonstrate here that this concept is incorrect, since the binding interactions that occur cannot be described in terms of a functional affinity, and the values that are obtained serve only to obscure the true interactions. Bivalent Ab binding must be considered to be an irreversible reaction, in most cases. A correct understanding of Ab binding will be useful in the further development of Abs for therapeutic purposes.     

A solid-phase method for synthesis of dimeric and trimeric ligands: Identification of potent bivalent ligands of 14-3-3σ

Multivalent protein-protein interactions including bivalent and trivalent interactions play a critical role in mediating a wide range of biological processes. Hence, there is a significant interest in developing molecules that can modulate those signaling pathways mediated by multivalent interactions. For example, multimeric molecules capable of binding to a receptor protein through a multivalent interaction could serve as modulators of such interactions. However, it is challenging to efficiently generate such multimeric ligands. Here, we have developed a facile solid-phase method that allows for the rapid generation of (homo- and hetero-) dimeric and trimeric protein ligands. The feasibility of this strategy was demonstrated by efficiently synthesizing fluorescently-labeled dimeric peptide ligands, which led to dramatically increased binding affinities (~400-fold improvement) relative to a monomeric 14-3-3σ protein ligand.     

Ligand-receptor interactions: detailed evaluation of occupancy-dependent affinity

The exact nature of the curvilinearity of Scatchard plots derived from hormonal and nonhormonal binding systems has not been definitively resolved. Such plots are compatible with heterogeneous receptors with different but fixed affinities and with negatively interacting binding sites resulting in occupancy-dependent affinity. In the current study we examined in detail the effect of receptor occupancy by the ligand on receptor affinity under a variety of experimental conditions. We chose the human lymphocyte-leukoagglutinin (LPHA) system, which closely mimics the IM9-insulin model. Reliable estimates of total binding capacity (728 ng/10(6) cells) essential to our report were calculated from a wide database by the least-squares model. At occupancies greater than or equal to 0.085, receptors are associated with low and fixed affinity (1.5 X 10(6) M-1), whereas at occupancies less than or equal to 0.085, affinity is high and fixed (1.8 X 10(8) M-1) or high but variable (1 X 10(7) M-1 to 1.5 X 10(6) M-1) depending on whether the binding is assumed to be noncooperative or cooperative, respectively. Calculation of receptor-ligand complex dissociation velocity over a wide range of occupancies (0.01-0.40) suggested that occupancy exerts an inversely proportional effect on affinity that is rapid and sustained. Cell activation (DNA synthesis) is initiated at receptor occupancy of approximately equal to 0.004 and is magnified as ligand binding to high affinity receptors increases up to approximately equal to 0.07 occupancy (functional sites), beyond which point further binding (to low affinity sites) becomes increasingly ineffective and cytotoxic (redundant sites). These findings suggest that occupancy influences affinity as postulated by the hypothesis of negative cooperativity. Through this effect occupancy may play a significant role in regulating ligand-induced cell responses.     

Human interleukin‐23 receptor antagonists derived from an albumin‐binding domain scaffold inhibit IL‐23‐dependent ex vivo expansion of IL‐17‐producing T‐cells

Engineered combinatorial libraries derived from small protein scaffolds represent a powerful tool for generating novel binders with high affinity, required specificity and designed inhibitory function. This work was aimed to generate a collection of recombinant binders of human interleukin‐23 receptor (IL‐23R), which is a key element of proinflammatory IL‐23‐mediated signaling. A library of variants derived from the three‐helix bundle scaffold of the albumin‐binding domain (ABD) of streptococcal protein G and ribosome display were used to select for high‐affinity binders of recombinant extracellular IL‐23R. A collection of 34 IL‐23R‐binding proteins (called REX binders), corresponding to 18 different sequence variants, was used to identify a group of ligands that inhibited binding of the recombinant p19 subunit of IL‐23, or the biologically active human IL‐23 cytokine, to the recombinant IL‐23R or soluble IL‐23R‐IgG chimera. The strongest competitors for IL‐23R binding in ELISA were confirmed to recognize human IL‐23R‐IgG in surface plasmon resonance experiments, estimating the binding affinity in the sub‐ to nanomolar range. We further demonstrated that several REX variants bind to human leukemic cell lines K‐562, THP‐1 and Jurkat, and this binding correlated with IL‐23R cell‐surface expression. The REX125, REX009 and REX128 variants competed with the p19 protein for binding to THP‐1 cells. Moreover, the presence of REX125, REX009 and REX115 variants significantly inhibited the IL‐23‐driven expansion of IL‐17‐producing primary human CD4⁺ T‐cells. Thus, we conclude that unique IL‐23R antagonists derived from the ABD scaffold were generated that might be useful in designing novel anti‐inflammatory biologicals. Proteins 2014; 82:975–989. © 2013 The Authors. Proteins: Structure, Function, and Bioinformatics Published by Wiley Periodicals, Inc.     

Effects of clustered epitopes in multivalent ligand-receptor interactions

Many biological ligands are composed of clustered binding epitopes. However, the effects of clustered epitopes on the affinity of ligand-receptor interactions in many cases are not well understood. Clustered carbohydrate epitopes are present in naturally occurring multivalent carbohydrates and glycoproteins, which are receptors on the surface of cells. Recent studies have provided evidence that the enhanced affinities of lectins, which are carbohydrate binding proteins, for multivalent carbohydrates and glycoproteins are due to internal diffusion of lectin molecules from epitope to epitope in these multivalent ligands before dissociation. Indeed, binding of lectins to mucins, which are large linear glycoproteins, appears to be similar to the internal diffusion mechanism(s) of protein ligands binding to DNA, which have been termed the "bind and slide" or "bind and hop" mechanisms. The observed increasing negative cooperativity and gradient of decreasing microaffinity constants of a lectin binding to multivalent carbohydrates and glycoproteins result in an initial fraction of lectin molecules that bind with very high affinity and dynamic motion. These findings have important implications for the mechanisms of binding of lectins to mucins, and for other ligand-biopolymer interactions and clustered ligand-receptor systems in general.     

Glioblastoma targeting via integrins is concentration dependent

A novel approach to treat cancer more selectively is achieved by targeting drugs to cells via conjugating the drug or imaging agent to an antibody or ligand for a cell surface receptor that is over‐expressed by the target cell population. Previous work by us has suggested that enhanced specificity can be obtained by multivalency of binding moieties. In this study we investigated the binding specificity of a multivalent construct including three peptides segments (TWYKIAFQRNRK), which bind the α₆β₁‐integrin, linked by poly(ethylene glycol) spacers. The binding specificity of the constructs was calculated by quantifying their binding to target cells (glioma cells, SF 767) relative to non‐targeted cells (normal human astrocytes, NHA). Dodecapeptide constructs (monovalent) exhibit specificity equal to the ratio of receptor expression at all concentrations. However, trivalent constructs demonstrated a sharp increase in specificity at concentrations less than the affinity of the receptor–ligand bond (4.28 µM). These experiments (conducted at 4°C) were consistent with the theoretical prediction and indicate that the biophysical model captures the basic trend of the data in the absence of receptor internalization, although the concentration at which increased specificity is observed is greater than predicted. The biophysical model does not predict the results of 37°C experiments, and this is shown to be due to internalization which occurs at 37°C but not at 4°C. Biotechnol. Bioeng. 2009; 104: 408–417 © 2009 Wiley Periodicals, Inc.     

Feature selection and classification of protein–protein complexes based on their binding affinities using machine learning approaches

Protein–protein interactions are intrinsic to virtually every cellular process. Predicting the binding affinity of protein–protein complexes is one of the challenging problems in computational and molecular biology. In this work, we related sequence features of protein–protein complexes with their binding affinities using machine learning approaches. We set up a database of 185 protein–protein complexes for which the interacting pairs are heterodimers and their experimental binding affinities are available. On the other hand, we have developed a set of 610 features from the sequences of protein complexes and utilized Ranker search method, which is the combination of Attribute evaluator and Ranker method for selecting specific features. We have analyzed several machine learning algorithms to discriminate protein‐protein complexes into high and low affinity groups based on their K_d values. Our results showed a 10‐fold cross‐validation accuracy of 76.1% with the combination of nine features using support vector machines. Further, we observed accuracy of 83.3% on an independent test set of 30 complexes. We suggest that our method would serve as an effective tool for identifying the interacting partners in protein–protein interaction networks and human–pathogen interactions based on the strength of interactions. Proteins 2014; 82:2088–2096. © 2014 Wiley Periodicals, Inc.