Title: Extracting Insight from Noisy Cellular Networks
Abstract: Network biologists attempt to extract meaningful relationships among genes or their products from very noisy data. We argue that what we categorize as noisy data may sometimes reflect noisy biology and therefore may shield a hidden meaning about how networks evolve and how matter is organized in the cell. We present practical solutions, based on existing evolutionary and biophysical concepts, through which our understanding of cell biology can be enormously enriched. Network biologists attempt to extract meaningful relationships among genes or their products from very noisy data. We argue that what we categorize as noisy data may sometimes reflect noisy biology and therefore may shield a hidden meaning about how networks evolve and how matter is organized in the cell. We present practical solutions, based on existing evolutionary and biophysical concepts, through which our understanding of cell biology can be enormously enriched. The spandrels of San Marco is an architectural analogy that Stephen J. Gould and Richard C. Lewontin used to explain the fundamental flaw in systematically ascribing individual traits of an organism to adaptation rather than to a possible coincidental evolution of some other characteristic (Gould and Lewontin, 1979Gould S.J. Lewontin R.C. The spandrels of San Marco and the Panglossian paradigm: a critique of the adaptationist programme.Proc. R. Soc. Lond. B Biol. Sci. 1979; 205: 581-598Crossref PubMed Scopus (4379) Google Scholar). More than three decades on, such adaptationist tendencies remain common to the interpretation of biological data and no less in network biology. The intrinsically beautiful and elegant structure embedded in interpretations, such as functional modularity, can mask important details and understanding of what the data tell us about the organization and evolution of networks. With the accelerating accumulation of data gathered at all layers of the cell, it is useful to return to first principles and ask precisely what we measure and what assumptions we make when analyzing the large-scale data that populate networks. The recent debate around ENCODE regarding how much of the human genome is functional is a clear example of why we need to address these issues (Doolittle, 2013Doolittle W.F. Is junk DNA bunk? A critique of ENCODE.Proc. Natl. Acad. Sci. USA. 2013; 110: 5294-5300Crossref PubMed Scopus (275) Google Scholar, Graur et al., 2013Graur D. Zheng Y. Price N. Azevedo R.B. Zufall R.A. Elhaik E. On the immortality of television sets: “function” in the human genome according to the evolution-free gospel of ENCODE.Genome Biol. Evol. 2013; 5: 578-590Crossref PubMed Scopus (330) Google Scholar, Maher, 2012Maher, B. (2012). Fighting about ENCODE and Junk. http://blogs.nature.com/news/2012/09/fighting-about-encode-and-junk.html.Google Scholar). Interpretations of gene function in a project like ENCODE requires the integration of a number of different types of large- and small-scale experimental data (Gerstein et al., 2012Gerstein M.B. Kundaje A. Hariharan M. Landt S.G. Yan K.K. Cheng C. Mu X.J. Khurana E. Rozowsky J. Alexander R. et al.Architecture of the human regulatory network derived from ENCODE data.Nature. 2012; 489: 91-100Crossref PubMed Scopus (1069) Google Scholar). To discuss all of the issues involved in interpreting these different types of data is beyond the scope of this Perspective. Instead, we take a fresh look at the raw details of one type of data, protein-protein interactions (PPI), and we ask what the experiments upon which they are based measure and take an alternative approach to their interpretation. PPIs constitute the physical link among gene products and thus provide us with essential clues to how biological processes are organized and integrated in cells and organisms (Babu et al., 2012Babu M. Vlasblom J. Pu S. Guo X. Graham C. Bean B.D.M. Burston H.E. Vizeacoumar F.J. Snider J. Phanse S. et al.Interaction landscape of membrane-protein complexes in Saccharomyces cerevisiae.Nature. 2012; 489: 585-589Crossref PubMed Scopus (177) Google Scholar, Gerstein et al., 2012Gerstein M.B. Kundaje A. Hariharan M. Landt S.G. Yan K.K. Cheng C. Mu X.J. Khurana E. Rozowsky J. Alexander R. et al.Architecture of the human regulatory network derived from ENCODE data.Nature. 2012; 489: 91-100Crossref PubMed Scopus (1069) Google Scholar, Havugimana et al., 2012Havugimana P.C. Hart G.T. Nepusz T. Yang H. Turinsky A.L. Li Z. Wang P.I. Boutz D.R. Fong V. Phanse S. et al.A census of human soluble protein complexes.Cell. 2012; 150: 1068-1081Abstract Full Text Full Text PDF PubMed Scopus (588) Google Scholar, Zhang et al., 2012Zhang Q.C. Petrey D. Deng L. Qiang L. Shi Y. Thu C.A. Bisikirska B. Lefebvre C. Accili D. Hunter T. et al.Structure-based prediction of protein-protein interactions on a genome-wide scale.Nature. 2012; 490: 556-560Crossref PubMed Scopus (507) Google Scholar). PPI networks are, however, largely difficult to interpret functionally and appear to be both poorly conserved across organisms and immensely large. Statistical strategies for interpreting large data sets have aided greatly in our attempts to understand PPI networks and continue to advance (Collins et al., 2007Collins S.R. Kemmeren P. Zhao X.C. Greenblatt J.F. Spencer F. Holstege F.C. Weissman J.S. Krogan N.J. Toward a comprehensive atlas of the physical interactome of Saccharomyces cerevisiae.Mol. Cell. Proteomics. 2007; 6: 439-450Crossref PubMed Scopus (638) Google Scholar), but for nonspecialists, results of such analyses are abstractions of the physical results that can obscure hidden and important details about how PPIs are organized. As sometimes happens in science, the object of interest becomes the abstract representation itself and not the underlying data. Here, we discuss key problems that may hinder clear understanding of PPIs, PPI networks, and their evolutionary history, and we propose solutions for each of these problems (Box 1) .Box 1Solutions for Common Problems in Interpreting Protein Interaction DataProblem 1: Different methods produce different types of data.Solution 1: Different data need to be conceptualized and assessed differently, and models of reference (ideal PPIs) need to reflect the breadth of methods used to probe PPIs and to reflect the biological diversity of PPIs as well. Reference PPIs should thus be tailored for each method.Problem 2: PPIs often appear as having a poor functional relevance.Solution 2: This observation has a biological explanation—promiscuity. A substantial number of PPIs that we observe serve no discernible function in the cell. Key cellular and chemical parameters, namely protein abundance, complex stoichiometry, and interaction conservation need to be taken into account to single out functional interactions and understand the biology behind networks.Problem 3: Proteins do not always follow rules of organization commonly depicted as molecular modules.Solution 3: An open mind with a combination of the two above-mentioned points. Models of how PPIs are organized should be extracted from the data rather than imposed on the data. Problem 1: Different methods produce different types of data. Solution 1: Different data need to be conceptualized and assessed differently, and models of reference (ideal PPIs) need to reflect the breadth of methods used to probe PPIs and to reflect the biological diversity of PPIs as well. Reference PPIs should thus be tailored for each method. Problem 2: PPIs often appear as having a poor functional relevance. Solution 2: This observation has a biological explanation—promiscuity. A substantial number of PPIs that we observe serve no discernible function in the cell. Key cellular and chemical parameters, namely protein abundance, complex stoichiometry, and interaction conservation need to be taken into account to single out functional interactions and understand the biology behind networks. Problem 3: Proteins do not always follow rules of organization commonly depicted as molecular modules. Solution 3: An open mind with a combination of the two above-mentioned points. Models of how PPIs are organized should be extracted from the data rather than imposed on the data. We begin by asking how, at an essential level, large-scale PPI data are interpreted (Figures 1A and 1B ). For the sake of brevity, we discuss the results of PPI screens for the model eukaryote budding yeast, Saccharomyces cerevisiae, for which there is the greatest amount of data available (Gavin et al., 2002Gavin A.C. Bösche M. Krause R. Grandi P. Marzioch M. Bauer A. Schultz J. Rick J.M. Michon A.M. Cruciat C.M. et al.Functional organization of the yeast proteome by systematic analysis of protein complexes.Nature. 2002; 415: 141-147Crossref PubMed Scopus (3993) Google Scholar, Gavin et al., 2006Gavin A.C. Aloy P. Grandi P. Krause R. Boesche M. Marzioch M. Rau C. Jensen L.J. Bastuck S. Dümpelfeld B. et al.Proteome survey reveals modularity of the yeast cell machinery.Nature. 2006; 440: 631-636Crossref PubMed Scopus (2116) Google Scholar, Ho et al., 2002Ho Y. Gruhler A. Heilbut A. Bader G.D. Moore L. Adams S.L. Millar A. Taylor P. Bennett K. Boutilier K. et al.Systematic identification of protein complexes in Saccharomyces cerevisiae by mass spectrometry.Nature. 2002; 415: 180-183Crossref PubMed Scopus (3072) Google Scholar, Ito et al., 2000Ito T. Tashiro K. Muta S. Ozawa R. Chiba T. Nishizawa M. Yamamoto K. Kuhara S. Sakaki Y. Giaever G. et al.Toward a protein-protein interaction map of the budding yeast: A comprehensive system to examine two-hybrid interactions in all possible combinations between the yeast proteins.Proc. Natl. Acad. Sci. USA. 2000; 97: 1143-1147Crossref PubMed Scopus (662) Google Scholar, Krogan et al., 2006Krogan N.J. Cagney G. Yu H. Zhong G. Guo X. Ignatchenko A. Li J. Pu S. Datta N. Tikuisis A.P. et al.Global landscape of protein complexes in the yeast Saccharomyces cerevisiae.Nature. 2006; 440: 637-643Crossref PubMed Scopus (2333) Google Scholar, Tarassov et al., 2008Tarassov K. Messier V. Landry C.R. Radinovic S. Serna Molina M.M. Shames I. Malitskaya Y. Vogel J. Bussey H. Michnick S.W. An in vivo map of the yeast protein interactome.Science. 2008; 320: 1465-1470Crossref PubMed Scopus (579) Google Scholar, Uetz et al., 2000Uetz P. Giot L. Cagney G. Mansfield T.A. Judson R.S. Knight J.R. Lockshon D. Narayan V. Srinivasan M. Pochart P. et al.A comprehensive analysis of protein-protein interactions in Saccharomyces cerevisiae.Nature. 2000; 403: 623-627Crossref PubMed Scopus (3911) Google Scholar, Yu et al., 2008Yu H. Braun P. Yildirim M.A. Lemmens I. Venkatesan K. Sahalie J. Hirozane-Kishikawa T. Gebreab F. Li N. Simonis N. et al.High-quality binary protein interaction map of the yeast interactome network.Science. 2008; 322: 104-110Crossref PubMed Scopus (1109) Google Scholar). The reader may be surprised to learn that large-scale PPI detection methods do not necessarily detect direct interactions between proteins. Three families of methods have produced the bulk of large-scale PPIs, including first those based on affinity purification followed by mass spectroscopy (AP-MS). This approach provides evidence, largely of stable complexes that can survive conditions of cell lysis and purification (Babu et al., 2012Babu M. Vlasblom J. Pu S. Guo X. Graham C. Bean B.D.M. Burston H.E. Vizeacoumar F.J. Snider J. Phanse S. et al.Interaction landscape of membrane-protein complexes in Saccharomyces cerevisiae.Nature. 2012; 489: 585-589Crossref PubMed Scopus (177) Google Scholar, Gavin et al., 2002Gavin A.C. Bösche M. Krause R. Grandi P. Marzioch M. Bauer A. Schultz J. Rick J.M. Michon A.M. Cruciat C.M. et al.Functional organization of the yeast proteome by systematic analysis of protein complexes.Nature. 2002; 415: 141-147Crossref PubMed Scopus (3993) Google Scholar, Gavin et al., 2006Gavin A.C. Aloy P. Grandi P. Krause R. Boesche M. Marzioch M. Rau C. Jensen L.J. Bastuck S. Dümpelfeld B. et al.Proteome survey reveals modularity of the yeast cell machinery.Nature. 2006; 440: 631-636Crossref PubMed Scopus (2116) Google Scholar, Ho et al., 2002Ho Y. Gruhler A. Heilbut A. Bader G.D. Moore L. Adams S.L. Millar A. Taylor P. Bennett K. Boutilier K. et al.Systematic identification of protein complexes in Saccharomyces cerevisiae by mass spectrometry.Nature. 2002; 415: 180-183Crossref PubMed Scopus (3072) Google Scholar, Krogan et al., 2006Krogan N.J. Cagney G. Yu H. Zhong G. Guo X. Ignatchenko A. Li J. Pu S. Datta N. Tikuisis A.P. et al.Global landscape of protein complexes in the yeast Saccharomyces cerevisiae.Nature. 2006; 440: 637-643Crossref PubMed Scopus (2333) Google Scholar, Zhang et al., 2012Zhang Q.C. Petrey D. Deng L. Qiang L. Shi Y. Thu C.A. Bisikirska B. Lefebvre C. Accili D. Hunter T. et al.Structure-based prediction of protein-protein interactions on a genome-wide scale.Nature. 2012; 490: 556-560Crossref PubMed Scopus (507) Google Scholar). Yeast two-hybrid (Y2H) methods are performed in vivo and may yield direct, binary information albeit in an unnatural compartment for most proteins (the nucleus), and proteins are typically expressed under nonnative promoters (Ito et al., 2000Ito T. Tashiro K. Muta S. Ozawa R. Chiba T. Nishizawa M. Yamamoto K. Kuhara S. Sakaki Y. Giaever G. et al.Toward a protein-protein interaction map of the budding yeast: A comprehensive system to examine two-hybrid interactions in all possible combinations between the yeast proteins.Proc. Natl. Acad. Sci. USA. 2000; 97: 1143-1147Crossref PubMed Scopus (662) Google Scholar, Uetz et al., 2000Uetz P. Giot L. Cagney G. Mansfield T.A. Judson R.S. Knight J.R. Lockshon D. Narayan V. Srinivasan M. Pochart P. et al.A comprehensive analysis of protein-protein interactions in Saccharomyces cerevisiae.Nature. 2000; 403: 623-627Crossref PubMed Scopus (3911) Google Scholar, Yu et al., 2008Yu H. Braun P. Yildirim M.A. Lemmens I. Venkatesan K. Sahalie J. Hirozane-Kishikawa T. Gebreab F. Li N. Simonis N. et al.High-quality binary protein interaction map of the yeast interactome network.Science. 2008; 322: 104-110Crossref PubMed Scopus (1109) Google Scholar). Finally, protein-fragment complementation assays (PCA) are in the middle; they do not provide unambiguous evidence of direct binary PPI but rather provide an indication of spatial proximity between two proteins. An advantage of this method is that proteins are expressed at endogenous levels and within relevant cellular compartments in living cells (Tarassov et al., 2008Tarassov K. Messier V. Landry C.R. Radinovic S. Serna Molina M.M. Shames I. Malitskaya Y. Vogel J. Bussey H. Michnick S.W. An in vivo map of the yeast protein interactome.Science. 2008; 320: 1465-1470Crossref PubMed Scopus (579) Google Scholar). Importantly, applications of criteria to access one type of data can be wholly misleading if applied to another. For instance, a gold standard such as reference protein complexes would include many interactions within complexes that cannot be captured by PCA or Y2H because the proteins are not physically close or in contact. Thus, different standards should be used to assess different data sets. Admittedly, false-positives and biases derived from experimental errors must be eliminated statistically—for instance, based on their reproducibility (Mellacheruvu et al., 2013Mellacheruvu D. Wright Z. Couzens A.L. Lambert J.P. St-Denis N.A. Li T. Miteva Y.V. Hauri S. Sardiu M.E. Low T.Y. et al.The CRAPome: a contaminant repository for affinity purification-mass spectrometry data.Nat. Methods. 2013; 10: 730-736Crossref PubMed Scopus (912) Google Scholar). However, care should be taken to choose appropriate reference PPIs for each particular experimental approach, and ideally, these methods should use information that is orthogonal and based on as many different methods as possible to the PPI detection approach. This would allow for correct assessment of the reliability of the data without biases toward one method or another. A better understanding and consideration of the methods used and their shortcomings may also help explain why so many interactions are not detected. In turn, such understanding could help raise the confidence that a lack of interaction in the data reflects the genuine absence of an interaction in the cell. For instance, some reporters may destabilize the fusion proteins and make interactions impossible to see or may hinder binding interfaces. Some proteins may be unable to work in a particular cell compartment where the reporter is reconstituted. Some screening methods may have a high rate of failure at some point in their procedure. In all cases, better controls on the experimental procedures (e.g., measurement of reproducibility) and of molecular constructs (e.g., confirmation of expression of fusion proteins) may alleviate these shortcomings. Furthermore, the lack of overlap among the current data sets may be relevant to a more fundamental question than those based solely on the reproducibility of a given technique (Bader and Hogue, 2002Bader G.D. Hogue C.W. Analyzing yeast protein-protein interaction data obtained from different sources.Nat. Biotechnol. 2002; 20: 991-997Crossref PubMed Scopus (454) Google Scholar, von Mering et al., 2002von Mering C. Krause R. Snel B. Cornell M. Oliver S.G. Fields S. Bork P. Comparative assessment of large-scale data sets of protein-protein interactions.Nature. 2002; 417: 399-403Crossref PubMed Scopus (1921) Google Scholar). This question relates to the incompleteness of our current model of the interactome. If current methods fail to uncover the same relationships among proteins, we expect that many more relationships may have been missed and thus that new technological developments are needed. For instance, approaches that would allow single-cell analysis of interactions based on fluorescent or luminescent reporters or approaches that would allow resolving spatiotemporal dependencies of PPIs could help fill the current gap. Regardless of how data might be validated, an important question is: what are we measuring? In other words, do we understand why a given signal from a given experiment maximizes our ability to predict biologically meaningful PPIs (Balaji et al., 2008Balaji S. Iyer L.M. Babu M.M. Aravind L. Comparison of transcription regulatory interactions inferred from high-throughput methods: what do they reveal?.Trends Genet. 2008; 24: 319-323Abstract Full Text Full Text PDF PubMed Scopus (17) Google Scholar, Jensen and Bork, 2008Jensen L.J. Bork P. Biochemistry. Not comparable, but complementary.Science. 2008; 322: 56-57Crossref PubMed Scopus (47) Google Scholar, Wodak et al., 2013Wodak S.J. Vlasblom J. Turinsky A.L. Pu S. Protein-protein interaction networks: the puzzling riches.Curr. Opin. Struct. Biol. 2013; (Published online September 2, 2013)https://doi.org/10.1016/j.sbi.2013.08.002Crossref PubMed Scopus (68) Google Scholar)? As we describe below, to answer this question, we must first consider how PPIs have evolved. In the past few years, we and others have made several observations suggesting that many PPIs, regardless of whether they are reproducible by different techniques, could have no function in the cell (Landry et al., 2009Landry C.R. Levy E.D. Michnick S.W. Weak functional constraints on phosphoproteomes.Trends Genet. 2009; 25: 193-197Abstract Full Text Full Text PDF PubMed Scopus (216) Google Scholar, Levy et al., 2009Levy E.D. Landry C.R. Michnick S.W. How perfect can protein interactomes be?.Sci. Signal. 2009; 2: pe11Crossref PubMed Scopus (62) Google Scholar). What do we mean by a nonfunctional PPI? From an evolutionary perspective, we mean that such PPIs appeared as the product of evolutionary processes but were not selected for imparting any benefit to the organism in which they arose (or later on) during evolution. Accordingly, such PPIs are not currently maintained by purifying selection. Consequently, disruption of a nonfunctional PPI is not expected to result in any biochemical or phenotypic deleterious consequence. It is also important to distinguish nonfunctional from nonspecific interactions. We usually think of nonspecific interactions as those that arise from, for instance, binding between hydrophobic surfaces of proteins. A nonfunctional interaction, however, could have the hallmarks of a specific interaction, including stereospecificity and shape complementarity, but again, the interaction may impart no beneficial functional consequences to either of the two proteins. This is a difficult idea to fathom, but not without precedent. As Gould and Lewontin, 1979Gould S.J. Lewontin R.C. The spandrels of San Marco and the Panglossian paradigm: a critique of the adaptationist programme.Proc. R. Soc. Lond. B Biol. Sci. 1979; 205: 581-598Crossref PubMed Scopus (4379) Google Scholar argued, not all features of a biological system have evolved because they provide a favorable function to the organism, and the same argument applies to molecular phenotypes. For instance, we know that transcription factors bind to hundreds of sites in a genome, but few are involved in regulating gene expression (Biggin, 2011Biggin M.D. Animal transcription networks as highly connected, quantitative continua.Dev. Cell. 2011; 21: 611-626Abstract Full Text Full Text PDF PubMed Scopus (229) Google Scholar, Euskirchen and Snyder, 2004Euskirchen G. Snyder M. A plethora of sites.Nat. Genet. 2004; 36: 325-326Crossref PubMed Scopus (15) Google Scholar, Hahn et al., 2003Hahn M.W. Stajich J.E. Wray G.A. The effects of selection against spurious transcription factor binding sites.Mol. Biol. Evol. 2003; 20: 901-906Crossref PubMed Scopus (72) Google Scholar). Formal predictions of nonfunctional transcription factor binding have been explored, and we have extended the same analyses to PPIs (Hahn et al., 2003Hahn M.W. Stajich J.E. Wray G.A. The effects of selection against spurious transcription factor binding sites.Mol. Biol. Evol. 2003; 20: 901-906Crossref PubMed Scopus (72) Google Scholar, Levy et al., 2009Levy E.D. Landry C.R. Michnick S.W. How perfect can protein interactomes be?.Sci. Signal. 2009; 2: pe11Crossref PubMed Scopus (62) Google Scholar). For one type of PPI, that of protein kinases with their substrates, we estimate that nonfunctional PPIs may compose higher than 50% of observables (Landry et al., 2009Landry C.R. Levy E.D. Michnick S.W. Weak functional constraints on phosphoproteomes.Trends Genet. 2009; 25: 193-197Abstract Full Text Full Text PDF PubMed Scopus (216) Google Scholar). Following these studies, investigators have used evolutionary and structural information as a means of prioritizing posttranslational modifications for functional studies, for instance for modifications that regulate PPIs (Beltrao et al., 2012Beltrao P. Albanèse V. Kenner L.R. Swaney D.L. Burlingame A. Villén J. Lim W.A. Fraser J.S. Frydman J. Krogan N.J. Systematic functional prioritization of protein posttranslational modifications.Cell. 2012; 150: 413-425Abstract Full Text Full Text PDF PubMed Scopus (298) Google Scholar). One might expect that nonfunctional interactions should be eliminated by natural selection. This, however, is likely to occur only if nonfunctional PPIs are deleterious to the cell and if the appropriate mutational and population genetics requirements are met (Figures 1C and 1D) (Fernández and Lynch, 2011Fernández A. Lynch M. Non-adaptive origins of interactome complexity.Nature. 2011; 474: 502-505Crossref PubMed Scopus (94) Google Scholar, Levy et al., 2009Levy E.D. Landry C.R. Michnick S.W. How perfect can protein interactomes be?.Sci. Signal. 2009; 2: pe11Crossref PubMed Scopus (62) Google Scholar). A PPI could arise from point mutations (Grueninger et al., 2008Grueninger D. Treiber N. Ziegler M.O. Koetter J.W. Schulze M.S. Schulz G.E. Designed protein-protein association.Science. 2008; 319: 206-209Crossref PubMed Scopus (119) Google Scholar)—or perhaps even due to a change in expression level (Gagnon-Arsenault et al., 2013Gagnon-Arsenault I. Marois Blanchet F.C. Rochette S. Diss G. Dube A.K. Landry C.R. Transcriptional divergence plays a role in the rewiring of protein interaction networks after gene duplication.J. Proteomics. 2013; 81: 112-125Crossref PubMed Scopus (21) Google Scholar) or subcellular localization of one of the partners (Kuriyan and Eisenberg, 2007Kuriyan J. Eisenberg D. The origin of protein interactions and allostery in colocalization.Nature. 2007; 450: 983-990Crossref PubMed Scopus (298) Google Scholar)—but have no functional consequence. Furthermore, a particular PPI may be an inevitable consequence of a function of another PPI in which one of the partners is involved. What we observe then may be a tradeoff between the specificity of PPIs and the ability of proteins to perform specific functions (Pechmann et al., 2009Pechmann S. Levy E.D. Tartaglia G.G. Vendruscolo M. Physicochemical principles that regulate the competition between functional and dysfunctional association of proteins.Proc. Natl. Acad. Sci. USA. 2009; 106: 10159-10164Crossref PubMed Scopus (130) Google Scholar). Nonfunctional PPIs may appear to be an obstacle to our understanding of how the cell works. We argue the opposite: understanding nonfunctional PPIs provide a window into the past, the present, and the future of evolving PPI networks. For instance, the birth and death of PPIs may contribute to the evolution of biochemical networks and to speciation of organisms (Tawfik, 2010Tawfik D.S. Messy biology and the origins of evolutionary innovations.Nat. Chem. Biol. 2010; 6: 692-696PubMed Google Scholar). What will be functional in the future is impossible to predict. However, one could argue that nonfunctional interactions may provide templates for the accumulation of beneficial mutations in the future. Accordingly, the wandering of PPI networks in the nonfunctional space may allow cells to explore configurations not directly available to beneficial mutations or modify the functional space so as to affect the neutrality of future mutations (Doolittle, 2013Doolittle W.F. Is junk DNA bunk? A critique of ENCODE.Proc. Natl. Acad. Sci. USA. 2013; 110: 5294-5300Crossref PubMed Scopus (275) Google Scholar). Thus, on the one hand, nonfunctional PPIs can be a source of annoyance to those trying to understand PPI data, but on the other, they may represent a feature of ongoing evolution of cellular networks (Levy et al., 2009Levy E.D. Landry C.R. Michnick S.W. How perfect can protein interactomes be?.Sci. Signal. 2009; 2: pe11Crossref PubMed Scopus (62) Google Scholar, Levy et al., 2010Levy E.D. Landry C.R. Michnick S.W. Cell signaling. Signaling through cooperation.Science. 2010; 328: 983-984Crossref PubMed Scopus (44) Google Scholar, Lynch, 2007aLynch M. The evolution of genetic networks by non-adaptive processes.Nat. Rev. Genet. 2007; 8: 803-813Crossref PubMed Scopus (230) Google Scholar, Lynch, 2007bLynch M. The frailty of adaptive hypotheses for the origins of organismal complexity.Proc. Natl. Acad. Sci. USA. 2007; 104: 8597-8604Crossref PubMed Scopus (505) Google Scholar, Zhang et al., 2008Zhang J. Maslov S. Shakhnovich E.I. Constraints imposed by non-functional protein-protein interactions on gene expression and proteome size.Mol. Syst. Biol. 2008; (Published online August 5, 2008)https://doi.org/10.1038/msb.2008.48Crossref Scopus (80) Google Scholar). Furthermore, functional and nonfunctional PPIs can be separated based on simple biophysical and evolutionary concepts. Chemical principles of PPI may provide important clues of functionality (Figure 1D) (Schreiber and Keating, 2011Schreiber G. Keating A.E. Protein binding specificity versus promiscuity.Curr. Opin. Struct. Biol. 2011; 21: 50-61Crossref PubMed Scopus (183) Google Scholar). Existing PPI data analyses implicitly test chemical parameters. For instance, intensity or frequency of an observable can be thought of as measuring the affinities or rates of association or dissociations of complexes. Recently, we have demonstrated that the proportion of protein phosphorylation on specific residues (or stoichiometry) can provide meaningful predictions of their functionality (Landry et al., 2009Landry C.R. Levy E.D. Michnick S.W. Weak functional constraints on phosphoproteomes.Trends Genet. 2009; 25: 193-197Abstract Full Text Full Text PDF PubMed Scopus (216) Google Scholar, Levy et al., 2012bLevy E.D. Michnick S.W. Landry C.R. Protein abundance is key to distinguish promiscuous from functional phosphorylation based on evolutionary information.Philos. Trans. R. Soc. Lond. B Biol. Sci. 2012; 367: 2594-2606Crossref PubMed Scopus (73) Google Scholar). Whether this principle applies to PPIs or other types of biomolecular interactions remains to be explored but, if true, could provide strong evidence of functionality. In addition, the thermodynamics of PPIs could be even more useful to distinguishing functional versus nonfunctional PPIs. For instance, it has been recently demonstrated that functional transcription factor binding in a genome could be distinguished from nonfunctional interactions by virtue that transcription factors exchange slower at functional sites, where transcription occurs, than at sites where no transcription is initiated (Lickwar et al., 2012Lickwar C.R. Mueller F. Hanlon S.E. McNally J.G. Lieb J.D. Genome-wide protein-DNA binding dynamics suggest a molecular clutch for transcription factor function.Nature. 2012; 484: 251-255Crossref PubMed Scopus (175) Google Scholar). Similar conclusions have been reached regarding the occupancy of transcription factors during development (Fisher et al., 2012Fisher W.W. Li J.J. Hammonds A.S. Brown J.B. Pfeiffer B.D. Weiszmann R. MacArthur S. Thomas S. Stamatoyannopoulos J.A. Eisen M.B. et al.DNA regions bound at low occupancy by transcription factors do not drive patterned reporter gene expression in Drosophila.Proc. Natl. Acad. Sci. USA. 2012; 109: 21330-21335Crossref PubMed