- Email: [email protected]
Investigation of secreted protein transcripts as early biomarkers for type 1 diabetes in the mouse model
Gene 512 (2013) 161–165
Contents lists available at SciVerse ScienceDirect
Gene journal homepage: www.elsevier.com/locate/gene
Short Communication
Investigation of secreted protein transcripts as early biomarkers for type 1 diabetes in the mouse model Evie Melanitou a,⁎, Fredj Tekaia b, Edouard Yeramian c a b c
Institut Pasteur, Laboratoire Immunophysiologie et Parasitisme, 25, rue du Dr Roux, 75725 Paris, Cedex 15, France Institut Pasteur, Unité de Génétique Moléculaire des Levures (URA 2171 CNRS and UFR927 Univ. P.M. Curie), 25, rue du Dr Roux, 75725 Paris, Cedex 15, France Institut Pasteur, Unité de Bioinformatique Structurale, (URA 2185 CNRS), 25, rue du Dr Roux, 75725 Paris, Cedex 15, France
a r t i c l e
i n f o
Article history: Accepted 14 September 2012 Available online 29 September 2012 Keywords: Diabetes Secretome Biomarkers Autoimmunity Correspondence analysis
a b s t r a c t Type 1 diabetes (T1D) represents a serious health burden in the world, complicated by the fact that disease onset can be preceded by a long time period without evident clinical signs. It would be then of critical importance to detect the disease in its early stages. In this direction, we seek here to identify early preinflammatory markers for autoimmune diabetes, mining our previously reported transcriptome data relevant to distinct early sub-phenotypes in the NOD mouse, associated with early insulin autoantibodies (E-IAA). More specifically we focus on secreted or transmembrane protein transcripts, identifying in this category 71 differentially expressed transcripts which are regulated at the early preinflammatory stages of T1D in the pancreatic lymph nodes (PLN). Following the expression patterns of these 71 transcripts, correspondence analysis (a multivariate analysis method) reveals a clear-cut segregation of the individual samples according to the early subphenotype used. Thus the 71 transcripts coding for secreted proteins constitute a candidate-set of predictive biomarkers for the development of autoimmune damage of the β cells of the pancreas. The majority of these genes have human orthologs and accordingly they represent potential candidate biomarkers for the human disease. In addition, for predictive purposes, the analysis reveals the possibility to reduce significantly the size of the candidate-set in practice, with various genes displaying identical expression profiles. © 2012 Elsevier B.V. All rights reserved.
1. Introduction Type 1 diabetes is characterized by a long asymptomatic period preceding the disease onset in humans and in the NOD mice. For diagnostic purposes, as well as on fundamental grounds, it is then highly desirable to grasp the important changes (environmental, cellular and gene expression) occurring in this critical period of time and leading ultimately to the disease onset. Such a global comprehensive scheme is still elusive at present, but it seems possible, with the increasingly available genomic data and information, to develop appropriate candidate biomarkers for risk assessment. In this direction, autoantibodies are useful in detecting the disease prior to glycemia in the NOD mice and in humans. There are however some limitations for their use notably, the development of transient autoantibodies can be followed by their disappearance months or even years later without the development of diabetes (Abiru et al., Abbreviations: T1D, Type 1 diabetes; NOD, Non obese diabetic; PLN, Pancreatic lymph node; CA, Correspondence analysis; SPG, Secreted protein genes; GEO, Gene expression omnibus. ⁎ Corresponding author at: Immunophysiology and Parasitism laboratory, Department of Parasitology and Mycology, Institut Pasteur, 25-28 rue du Dr Roux, 75015 Paris, France. Tel.: +33 1 40 68 3299; fax: +33 1 45 68 8332. E-mail addresses: [email protected] (E. Melanitou), [email protected] (F. Tekaia), [email protected] (E. Yeramian). 0378-1119/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.gene.2012.09.055
2001; Barker et al., 2004). There also exists the possibility of transplacental transfer of autoantibodies to infants of T1D mothers (Greeley et al., 2002). Despite the identification of several risk factors for type 1 diabetes, early prediction is still missing due to insufficient predictive power of the individual risk factors (Purohit and She, 2008). Accordingly we seek here an alternative solution for the development of candidate biomarkers for early stages of autoimmune diabetes, mining our previously published trancriptome data (Regnault et al., 2009). More specifically we focus on the class of transcripts coding for secreted proteins in the NOD mouse and we demonstrate indeed the potentiality to use this specific category of genes as biomarkers for autoimmune predisposition. The rationale of the focus here on secreted protein transcripts is twofold: (1) On functional grounds, secreted proteins are known to play key roles in various important biological processes such as morphogenesis (Thomas et al., 2011; Tremble et al., 1993), angiogenesis (Onuffer and Horuk, 2002), cellular differentiation (Nalbant et al., 2005; Rosenow et al., 2010), apoptosis (Danielsen and Maihle, 2002) and modulation of immune response (Flavell, 2002; Grandvaux et al., 2002). These proteins are also implicated in disease processes, such as cancer progression (Welsh et al., 2003). Accordingly, this class of proteins has attracted much interest in recent years, with notably the initiation of a concerted effort called “The Secreted Protein Discovery Initiative” (SPDI), aiming to the identification of human secreted and
E. Melanitou et al. / Gene 512 (2013) 161–165
transmembrane proteins with bioinformatic approaches (Clark et al., 2003). (2) A diagnostic potential exists when secreted proteins are found in the serum. Indeed such proteins may be targeted by specific antibodies or small molecules in body fluids and thus constitute potential biomarkers for disease detection. In this general background, we develop here proof of concept for the possibility of early biomarkers for T1D, in the NOD mouse model, based on “secretome” data. More precisely, mining our previously published transcriptome data in this model (Regnault et al., 2009), we identify 71 genes coding for proteins located in the extracellular space. We then exploit the global properties of the expression signatures of these secreted protein genes (SPGs) with multivariate analysis, resorting to correspondence analysis (CA). Such analysis revealed a clear-cut segregation of the individual samples with the various subphenotypes. Examination of the genes involved in the analysis shows that the segregation is not following simple functional characteristics, further highlighting the interest in the global multivariate approach. In addition, such examination also reveals that a significant fraction of the secreted genes share identical expression profiles, allowing the perspective of a significant reduction of the candidate-set size in practice. Finally, with most of the genes in the set having human orthologs, it is expected that it will be possible to extend the developed system towards an efficient test for establishing the association of the presence of the corresponding proteins in the peripheral blood with the early detection of T1D in humans. 2. Materials and methods 2.1. Animals and analysis of microarray data The settings for experimental data collection and microarray analysis have been described previously (Regnault et al., 2009). In brief, NOD/Tac mice were used and monitored for the presence of insulin autoantibodies by radioimmunoassay at one week intervals after weaning (starting at 3 weeks of age) (Melanitou et al., 2004; Yu et al., 2000). Animals were sacrificed at 5 weeks of age as previously described (Melanitou et al., 2004). Two groups of animals were selected (E-IAApos and E-IAAneg) and high quality T-RNAs were prepared (Qiagen) from PLN tissues. T-RNA samples (4.5 μg) were monitored for cDNA preparation and hybridized on the Affymetrix MG_U74A_version 2 arrays (Santa Clara, Ca) containing 12 486 probe sets. The resulting data (in Affymetrix CEL format) were subjected to RMA (Robust Multichip Average) normalization (Bolstad et al., 2003) and transcripts with “Absent” Detection Call in both conditions were removed from further analysis. The RMA-normalized data set was used to identify modulated gene expression by the LPE (Local Pooled Error) statistical test (Jain et al., 2003), following adjustment for multiple comparisons by the Hochberg and Benjamini test (Hochberg and Benjamini, 1990). Box plots metrics of the RMA-normalized data illustrated the robustness of the data for both sets of samples (see supplementary data in Regnault et al., 2009). The data used for the correspondence analyses concerned nine individual mice (PLN samples) and were extracted from the initial data set, deposited in the NCBI's Gene Expression Omnibus (Edgar et al., 2002), GEO: GSE15582. The samples were in two groups, relative to early insulin autoantibodies (E-IAA) subphenotypes (Table S1): 5 negative samples (E-IAAneg) and 4 positive ones (E-IAApos). 2.2. Multivariate analysis The expression patterns of the transcripts for secreted-protein genes were analyzed with correspondence analysis (CA). For this analysis, the expression levels of the various genes were represented by their “expression flags” (A: absent; M: marginal and P: present), after corrections for background noise (as previously described (Regnault et al., 2009)) in the nine individual samples used for the
arrays. As such, the correspondence analysis concerned 3 × 9 samples, with each mouse individual An (either E-IAAneg or E-IAApos; n indexing variable) associated with three specific expression profiles An.pf(A), An.pf(M) and An.pf(P), according to the expression flags (A, M or P) for the set of genes in the study. Accordingly, with such coding, we are led to a table (not shown) containing 27 samples (9 × 3, corresponding to the various An.pf(X)) versus 71 SPGs. The fold change values for the expression levels (between negative and positive subphenotypes) were not used in the CA analysis. 3. Results 3.1. Identification of secreted protein-genes (SPGs) For type 1 diabetes disease pathogenesis in the NOD mouse, the early pre-inflammatory stages generally occur at the age of 4–6 weeks, after weaning and early priming of T cells reactive to islet β cell antigens takes place in the PLN (Turley et al., 2003). Accordingly, we performed the analyses in this tissue, in mice at 5 weeks of age. In such conditions, considering the E-IAA subphenotype, we identified by microarray analysis 165 transcripts regulated in the PLN of E-IAApos mice (Melanitou et al., 2004; Regnault et al., 2009). Out of the 165 differentially expressed genes identified in the E-IAApos vs E-IAAneg samples, 71 genes (43%), code for secreted proteins, as assigned by Gene Ontology annotations (http://www. geneontology.org/) (Huang et al., 2007) for cellular localization (pb 2.7e−28). Such percentage corresponds to an over-representation of this class of genes, as compared to the overall percentage of genes coding for secreted proteins (a mere 16%) in the Affymetrix arrays (MGU74Av2). Expression differences of the 71 transcripts in the E-IAApos vs E-IAAneg samples varied between 90 and 1.5 fold, with the majority of the genes (66 genes) being up-regulated in the E-IAApos PLN (Fig. 1). Only five transcripts were down-regulated, all of them coding for immunoglobulin-related proteins. With the exception of 3 ESTs, the remaining 68 genes code for known proteins and the majority of them have human orthologs (Table S1). 3.2. Correspondence analysis of expression patterns of SPGs In order to gain a global picture of the -complex- dependencies between expression patterns of the SPGs and the E-IAA subphenotype we resorted to correspondence analysis (CA) a multivariate statistical method (Tekaia and Yeramian, 2006; Tekaia et al., 2002). Our aim was to assess the usage of secreted proteins as biomarkers for autoimmune predisposition in the NOD mouse. 18 16
Mean E-IAApos (Log2)
162
14 12
r=0.65, p≤0.0001
10 8 6 4 2 0 0
2
4
6
8
10
12
14
16
Mean E-IAAneg (Log2) Fig. 1. Relative expression patterns of secreted protein genes (SPGs) in the PLN of E-IAApos and E-IAAneg. Linear regression line is drawn (correlation coefficient r = 0.65 with p ≤0.0001). Down-regulated genes in the E-IAApos samples are represented as triangles.
E. Melanitou et al. / Gene 512 (2013) 161–165
Our analysis was based on the expression levels of the 71 SPG transcripts for the PLN samples from nine individual mice used for the transcriptome analysis (Table S1). For the 71 transcripts, based on expression levels, Absent/Marginal/Present expression flags were calculated (following Affymetrix conventions). With such representation, each mouse individual An (n variable for indexing) was associated with three specific expression profiles An.pf(A), An.pf(M) and An.pf(P), according to the expression flags (A, M or P) for the 71 genes in the given sample. Accordingly, with such coding, we are led to a data table (not shown) containing 27 samples (9 × 3, corresponding to the various An.pf(X)) versus 71 SPGs. Preliminary analyses of this table showed that M profiles were rare and therefore not significant. In addition, four An.pf(A) or An.pf(P) samples appeared as outliers having the same expression flag for the majority of the genes. The An.pf(M) profiles, together with the 4 outliers, were then removed from the final data table used for correspondence analysis. The distribution of the samples (An.pf(X), with X either A or P) and the 71 SPGs on the first factorial plane is displayed in Fig. 2, associated with about 70% of the total information (54.9% and 14.6% following the factorial axes F1 for the F2 respectively leaving 30.5% for the remaining factorial space). This distribution reveals a clearcut segregation between An.pf(P) and An.pf(A) samples (P and A flags, in red and blue respectively), from left to right following the first factorial axis F1 (Fig. 2). In addition, very interestingly, this overall segregation further "separates" the An.pf(P) samples into positive and negative subphenotype categories. More specifically, the separation scheme is visually highlighted in Fig. 2 by two separation lines (dotted lines). In this scheme, all An.pf(A) samples are in the rightside part of the factorial plane, in light blue background, whereas all An.pf(P) samples are in the left-side part of the factorial plane, in red background. This red part of the factorial plane further subdivides into two parts, with the dark-shaded part harboring the positive subphenotype An.pf(P) samples (represented underlined). The only exception to this separation scheme appears to be the E-IAA positive A15.6 sample, for which A15.6pf(A) appears isolated at the very lower right limit of the factorial plane (represented underlined), whilst the A15.6pf(P) clusters together with the other E-IAA positive samples.
163
In summary, based on the information considered here, correspondence analysis leads to a segregation scheme following the F1 axis, associated with the three specific categories in three specific regions (as visually highlighted in Fig. 2 with the dotted lines, from left to right): I) negative subphenotype An.pf(P) samples in the light-shade red part, II) positive subphenotype An.pf(P) samples in the dark-shade red part and finally III) negative subphenotype An.pf(A) samples in the light blue part. Based on this scheme, the positioning of any individual with an unknown subphenotype on the factorial plane, according to the levels of the secreted protein transcripts identified in the periphery (plasma or serum), could in principle help identify an early predisposition to develop autoimmune diabetes. 3.3. Clustering rationale and data size reduction Examination of the genes with CA, in the three categories described above, reveals that the segregation does not follow any simple functional rationale. This feature is simply illustrated, for example, by the kallikrein genes, which appear scattered on the three parts of the factorial plane (Fig. 2). In a more systematic approach, functional annotations (http://david.abcc.ncifcrf.gov/ online annotation tools), were associated with the identified genes (Table S1). Such annotations reveal that the majority of the genes are associated with immune functions, independent of their relative segregation on the factorial plane. It is also worth noting that several serine peptidases cluster with the genes expressed in the E-IAA positive samples, including several genes coding for kallikreins. The latter are proteins found in the plasma, implicated in the regulation of homeostatic functions and also investigated as cancer biomarkers (Diamandis and Yousef, 2002). The functional annotations of the SPGs (Table S1 and Fig. S1) also reveal some interesting features relative to tissue distribution of the identified transcripts: significant association with the pancreas (p= 2.85e-16), plasma (p=1.68e-8), salivary gland (p= 1.75e-8) and the liver (p= 1.947e-7). However some of the transcripts are also expressed in other tissues (Fig. S1). High expression of these transcripts in the E-IAApos PLN might reflect autoimmune-related homeostatic changes, probably due to the presence of dendritic cells (Yin et al.,
Fig. 2. Correspondence analysis of expression patterns of SPG transcripts in nine individual NOD mice: first factorial plane of CA, carried according to the expression levels of the SPG transcripts. Each mouse individual An (n variable for indexing) is associated with specific expression profiles An.pf(A) and An.pf(P) (according to the expression flags ‘A’ or ‘P’: absence or presence of expression, in the corresponding individual, for the 71 genes in the study). E-IAApos subphenotype samples are underlined. Blue dots are associated with genes whose expression patterns are identical for the various individuals. Black dots are associated with genes whose expression patterns are unique. Genes with expression differences (E-IAApos vs E-IAAneg) higher than 25 folds are represented in brown.
164
E. Melanitou et al. / Gene 512 (2013) 161–165
2012). Indeed these cells are involved in the inflammatory process and have the capacity to migrate to the pancreas-draining lymph nodes where they present self-antigens (Heath et al., 2004; Steinman et al., 2003). From another perspective, it is also interesting to examine the distribution of genes on the factorial plane following fold changes in expression levels. Genes with the largest fold values (with threshold set at 25, for illustration) are represented on the factorial plane (Fig. 2) in brown. It appears that many such genes are in fact clustered in the central part, associated with positive subphenotype An.pf(P) samples. Such observations could potentially suggest the elaboration of more straightforward testing schemes, based for example on changes in expression levels alone. However, the scattering (and mutual positioning) of An.pf(X) individuals and genes on the factorial plane clearly shows that such a scheme could not be effective, with various genes (relative to the threshold 25 in fold changes, noted in brown) also appearing in the left and right parts of the factorial plane. On the other hand, the segregation of An.pf(X) individuals into three well-identified parts in the multivariate analysis does not rely on the fold change information, further highlighting the interest of such approach for the purpose of a reliable discriminative test. Finally, examination of the genes on the factorial plane reveals a series of spots where the positions (of variable number) of genes coincide (in Fig. 2 the corresponding spots are in blue, whereas spots associated with single genes are in black). This observation pinpoints the fact that, at such spots, several genes (with again possibly no direct functional relationship) present identical profiles (in terms of A/M/P flags) for the nine individual samples. As such genes play redundant roles in the discriminatory analysis, it should become then possible – in principle – to reduce significantly the size of the gene set used. However, only further analyses, with increasing number of individuals, could allow determining the ideal compromises between size reduction and enhanced robustness. 4. Discussion The early detection and the possibility of preventing the autoimmune process represent great challenges in T1D. In this respect, it is notably necessary to be able to identify the time of occurrence of this cryptic phase, which includes also insulitis (the β cell-damaging inflammatory state). At present, screening for and detection of autoantibodies against islet autoantigens have a good predictive power for the disease (Orban et al., 2009). It seems however clear that the presence of autoantibodies in the peripheral blood implies that in fact the autoimmune processes per se have already been initiated. In this context, our aim was to address the possibility to identify markers that delineate earlier events, taking place prior to the mechanisms involved in the development of antibodies to islet antigens. We have used the NOD animal model for T1D to address this question. The low penetrance of disease in the NOD mice made it necessary that we proceed to an initial screening for animals that would progress to disease. With this respect, we have used the presence of E-IAA as a T1D subphenotype to select individual mice, and compare the transcription profiles with those of E-IAA negative animals. Such comparisons revealed that a high proportion of the identified transcripts code for secreted proteins, with the majority of them being up-regulated in the E-IAApos PLN. This situation could be related to the recently reported increase of the β cell mass and islet hyperactivity in 6 week old NOD mice (Liang et al., 2011). Taken together with our previous demonstration for the existence of a correlation between the presence of E-IAA and T1D (Melanitou et al., 2004), these results and observations led us to investigate the predictive diagnostic power of the E-IAA PLN “secretome”, as potentially reflecting the early steps preceding the development of islet autoimmunity.
To investigate the validity of this hypothesis, we resorted to correspondence analysis of the expression patterns of secreted proteinencoding genes, in relation to the subphenotype used in our study. For the elaboration of predictive biomarker, it seems natural to use multivariate representations to capture complex dependencies between transcriptome data and various genetic traits such as autoimmune diabetes subphenotypes (E-IAA in this study). Indeed, such methods were used successfully in other instances for the detection of biomarkers based on transcriptome data, such as for the detection of cancer biomarkers (Zhang et al., 2005), with principle component analysis (another variant of a multivariate method). For the transcriptome data, the factorial plane obtained with correspondence analysis (associated with about 70% of the total information) reveals a clear-cut segregation of the samples associated with different E-IAA subphenotypes, following the expression patterns of the identified genes. This representation also highlights the co-localization of (variable number) of genes in several spots of the factorial plane, associated with identical expression profiles. This observation opens the perspective of more or less significant reduction of the number of genes used as predictive biomarkers, at the corresponding spots. Of course, ideally, such reduction could retain only a single gene for each such spot. However, only detailed experimental testing could allow investigating the levels of robustness compatible with reduced sets of genes in the test. In short, no matter the size of the set of genes, the scheme with CA opens the possibility to evaluate the relative risk for a subject to develop T1D, following the positioning in one of the three welldelimited parts of the factorial plane. Finally, the testing scheme described herein can lead to new explorations for the underlying biological features and mechanisms, with the appropriate cautious considerations. Notably, despite the fact that the identified transcripts code for known proteins the question remains of how well transcript levels do actually reflect protein abundance, in particular in the peripheral blood. Indeed regulatory mechanisms at the translation level or posttranslational modifications may influence the effective protein abundance. The links between transcriptomic and proteomic profiles have been investigated by computational methods and generally a poor correlation between mRNA and protein levels has been reported (Griffin et al., 2002; Mootha et al., 2003). More specifically, in these reports, high correlations were found only for components of some molecular pathways, such as the ribosome and cell adhesion complexes (Rogers et al., 2008). On the other hand, contrary to the above observations, a cross comparison of proteomic patterns with mRNA patterns from mouse tissues revealed broad areas of agreement with few inconsistencies (Kislinger et al., 2006). In this respect, concerning our secretome data, ultimately it is only through experimental testing (in the mouse model) or clinical evaluation in human patients that the effectiveness of the diagnostic scheme could be confirmed. In the elaboration of this scheme the interest in resorting to multivariate analysis was highlighted above, as compared to simpler approaches relying for example on the expression level data. With this respect, even though effective discriminatory power involves all 71 transcripts, it is nevertheless clear that a detailed investigation of the six transcripts expressed solely in the E-IAApos animals (and absent in the E-IAAneg ones, Table S1), represents a priority for exploratory studies in the NOD mouse or in human clinical trials. From another point of view, also considering the detailed data, it appears that in our set the only down-regulated transcripts in the E-IAApos samples correspond to genes coding for immunoglobulins, all clustering together on the factorial plane (Fig. 2 and Table S1). Diversity of the antibody repertoire is a prerequisite for efficient recognition of a broad spectrum of invading microorganisms by the immune system. Incorrect or misregulated repertoire can initiate autoimmunity (Ishida et al., 2006; Jankovic and Nussenzweig, 2003). In this context, it is tempting to speculate that down regulation of
E. Melanitou et al. / Gene 512 (2013) 161–165
immunoglobulin transcripts in the E-IAApos samples might be associated with the initiation of the autoimmune process in these animals. Based on our data and considerations described above, additional investigations combining diagnostic purpose explorations with studies targeting the fundamental early mechanisms in T1D represent a promising route to significant advances in this field. Supplementary data to this article can be found online at http:// dx.doi.org/10.1016/j.gene.2012.09.055. Acknowledgments EM thanks Dr. Geneviève Milon for constant support and encouraging discussions. FT acknowledges financial support from the Institut Pasteur (PTR 370). References Abiru, N., et al., 2001. Transient insulin autoantibody expression independent of development of diabetes: comparison of NOD and NOR strains. J. Autoimmun. 17, 1–6. Barker, J.M., et al., 2004. Prediction of autoantibody positivity and progression to type 1 diabetes: Diabetes Autoimmunity Study in the Young (DAISY). J. Clin. Endocrinol. Metab. 89, 3896–3902. Bolstad, B.M., Irizarry, R.A., Astrand, M., Speed, T.P., 2003. A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics 19, 185–193. Clark, H.F., et al., 2003. The secreted protein discovery initiative (SPDI), a large-scale effort to identify novel human secreted and transmembrane proteins: a bioinformatics assessment. Genome Res. 13, 2265–2270. Danielsen, A.J., Maihle, N.J., 2002. The EGF/ErbB receptor family and apoptosis. Growth Factors 20, 1–15. Diamandis, E.P., Yousef, G.M., 2002. Human tissue kallikreins: a family of new cancer biomarkers. Clin. Chem. 48, 1198–1205. Edgar, R., Domrachev, M., Lash, A.E., 2002. Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res. 30, 207–210. Flavell, R.A., 2002. The relationship of inflammation and initiation of autoimmune disease: role of TNF super family members. Curr. Top. Microbiol. Immunol. 266, 1–9. Grandvaux, N., tenOever, B.R., Servant, M.J., Hiscott, J., 2002. The interferon antiviral response: from viral invasion to evasion. Curr. Opin. Infect. Dis. 15, 259–267. Greeley, S.A., et al., 2002. Elimination of maternally transmitted autoantibodies prevents diabetes in nonobese diabetic mice. Nat. Med. 8, 399–402. Griffin, T.J., et al., 2002. Complementary profiling of gene expression at the transcriptome and proteome levels in Saccharomyces cerevisiae. Mol. Cell. Proteomics 1, 323–333. Heath, W.R., et al., 2004. Cross-presentation, dendritic cell subsets, and the generation of immunity to cellular antigens. Immunol. Rev. 199, 9–26. Hochberg, Y., Benjamini, Y., 1990. More powerful procedures for multiple significance testing. Stat. Med. 9, 811–818. Huang, W., et al., 2007. The DAVID Gene Functional Classification Tool: a novel biological module-centric algorithm to functionally analyze large gene lists. Genome Biol. 8, R183. Ishida, D., et al., 2006. Rap1 signal controls B cell receptor repertoire and generation of self-reactive B1a cells. Immunity 24, 417–427.
165
Jain, N., Thatte, J., Braciale, T., Ley, K., O'Connell, M., Lee, J.K., 2003. Local-pooled-error test for identifying differentially expressed genes with a small number of replicated microarrays. Bioinformatics 19, 1945–1951. Jankovic, M., Nussenzweig, M.C., 2003. OcaB regulates transitional B cell selection. Int. Immunol. 15, 1099–1104. Kislinger, T., et al., 2006. Global survey of organ and organelle protein expression in mouse: combined proteomic and transcriptomic profiling. Cell 125, 173–186. Liang, K., et al., 2011. Contribution of different mechanisms to pancreatic beta-cell hyper-secretion in non-obese diabetic (NOD) mice during pre-diabetes. J. Biol. Chem. 286, 39537–39545. Melanitou, E., Devendra, D., Liu, E., Miao, D., Eisenbarth, G.S., 2004. Early and quantal (by litter) expression of insulin autoantibodies in the nonobese diabetic mice predict early diabetes onset. J. Immunol. 173, 6603–6610. Mootha, V.K., et al., 2003. Integrated analysis of protein composition, tissue diversity, and gene regulation in mouse mitochondria. Cell 115, 629–640. Nalbant, D., et al., 2005. FAM20: an evolutionarily conserved family of secreted proteins expressed in hematopoietic cells. BMC Genomics 6, 11. Onuffer, J.J., Horuk, R., 2002. Chemokines, chemokine receptors and small-molecule antagonists: recent developments. Trends Pharmacol. Sci. 23, 459–467. Orban, T., et al., 2009. Pancreatic islet autoantibodies as predictors of type 1 diabetes in the Diabetes Prevention Trial-Type 1. Diabetes Care 32, 2269–2274. Purohit, S., She, J.X., 2008. Biomarkers for type 1 diabetes. Int. J. Clin. Exp. Med. 1, 98–116. Regnault, B., Osorio, Y.F.J., Miao, D., Eisenbarth, G., Melanitou, E., 2009. Early over expression of messenger RNA for multiple genes, including insulin, in the pancreatic lymph nodes of NOD mice is associated with islet autoimmunity. BMC Med. Genomics 2, 63. Rogers, S., et al., 2008. Investigating the correspondence between transcriptomic and proteomic expression profiles using coupled cluster models. Bioinformatics 24, 2894–2900. Rosenow, A., et al., 2010. Identification of novel human adipocyte secreted proteins by using SGBS cells. J. Proteome Res. 9, 5389–5401. Steinman, R.M., Hawiger, D., Nussenzweig, M.C., 2003. Tolerogenic dendritic cells. Annu. Rev. Immunol. 21, 685–711. Tekaia, F., Yeramian, E., 2006. Evolution of proteomes: fundamental signatures and global trends in amino acid compositions. BMC Genomics 7, 307. Tekaia, F., Yeramian, E., Dujon, B., 2002. Amino acid composition of genomes, lifestyles of organisms, and evolutionary trends: a global picture with correspondence analysis. Gene 297, 51–60. Thomas, S., et al., 2011. Development of secreted protein and acidic and rich in cysteine (SPARC) targeted nanoparticles for the prognostic molecular imaging of metastatic prostate cancer. J. Nanomed. Nanotechnol. 2. Tremble, P.M., Lane, T.F., Sage, E.H., Werb, Z., 1993. SPARC, a secreted protein associated with morphogenesis and tissue remodeling, induces expression of metalloproteinases in fibroblasts through a novel extracellular matrix-dependent pathway. J. Cell Biol. 121, 1433–1444. Turley, S., Poirot, L., Hattori, M., Benoist, C., Mathis, D., 2003. Physiological beta cell death triggers priming of self-reactive T cells by dendritic cells in a type-1 diabetes model. J. Exp. Med. 198, 1527–1537. Welsh, J.B., et al., 2003. Large-scale delineation of secreted protein biomarkers overexpressed in cancer tissue and serum. Proc. Natl. Acad. Sci. U. S. A. 100, 3410–3415. Yin, N., et al., 2012. Functional specialization of islet dendritic cell subsets. J. Immunol. 188, 4921–4930. Yu, L., et al., 2000. Early expression of antiinsulin autoantibodies of humans and the NOD mouse: evidence for early determination of subsequent diabetes. Proc. Natl. Acad. Sci. U. S. A. 97, 1701–1706. Zhang, X.W., Yap, Y.L., Wei, D., Chen, F., Danchin, A., 2005. Molecular diagnosis of human cancer type by gene expression profiles and independent component analysis. Eur. J. Hum. Genet. 13, 1303–1311.
Contents lists available at SciVerse ScienceDirect
Gene journal homepage: www.elsevier.com/locate/gene
Short Communication
Investigation of secreted protein transcripts as early biomarkers for type 1 diabetes in the mouse model Evie Melanitou a,⁎, Fredj Tekaia b, Edouard Yeramian c a b c
Institut Pasteur, Laboratoire Immunophysiologie et Parasitisme, 25, rue du Dr Roux, 75725 Paris, Cedex 15, France Institut Pasteur, Unité de Génétique Moléculaire des Levures (URA 2171 CNRS and UFR927 Univ. P.M. Curie), 25, rue du Dr Roux, 75725 Paris, Cedex 15, France Institut Pasteur, Unité de Bioinformatique Structurale, (URA 2185 CNRS), 25, rue du Dr Roux, 75725 Paris, Cedex 15, France
a r t i c l e
i n f o
Article history: Accepted 14 September 2012 Available online 29 September 2012 Keywords: Diabetes Secretome Biomarkers Autoimmunity Correspondence analysis
a b s t r a c t Type 1 diabetes (T1D) represents a serious health burden in the world, complicated by the fact that disease onset can be preceded by a long time period without evident clinical signs. It would be then of critical importance to detect the disease in its early stages. In this direction, we seek here to identify early preinflammatory markers for autoimmune diabetes, mining our previously reported transcriptome data relevant to distinct early sub-phenotypes in the NOD mouse, associated with early insulin autoantibodies (E-IAA). More specifically we focus on secreted or transmembrane protein transcripts, identifying in this category 71 differentially expressed transcripts which are regulated at the early preinflammatory stages of T1D in the pancreatic lymph nodes (PLN). Following the expression patterns of these 71 transcripts, correspondence analysis (a multivariate analysis method) reveals a clear-cut segregation of the individual samples according to the early subphenotype used. Thus the 71 transcripts coding for secreted proteins constitute a candidate-set of predictive biomarkers for the development of autoimmune damage of the β cells of the pancreas. The majority of these genes have human orthologs and accordingly they represent potential candidate biomarkers for the human disease. In addition, for predictive purposes, the analysis reveals the possibility to reduce significantly the size of the candidate-set in practice, with various genes displaying identical expression profiles. © 2012 Elsevier B.V. All rights reserved.
1. Introduction Type 1 diabetes is characterized by a long asymptomatic period preceding the disease onset in humans and in the NOD mice. For diagnostic purposes, as well as on fundamental grounds, it is then highly desirable to grasp the important changes (environmental, cellular and gene expression) occurring in this critical period of time and leading ultimately to the disease onset. Such a global comprehensive scheme is still elusive at present, but it seems possible, with the increasingly available genomic data and information, to develop appropriate candidate biomarkers for risk assessment. In this direction, autoantibodies are useful in detecting the disease prior to glycemia in the NOD mice and in humans. There are however some limitations for their use notably, the development of transient autoantibodies can be followed by their disappearance months or even years later without the development of diabetes (Abiru et al., Abbreviations: T1D, Type 1 diabetes; NOD, Non obese diabetic; PLN, Pancreatic lymph node; CA, Correspondence analysis; SPG, Secreted protein genes; GEO, Gene expression omnibus. ⁎ Corresponding author at: Immunophysiology and Parasitism laboratory, Department of Parasitology and Mycology, Institut Pasteur, 25-28 rue du Dr Roux, 75015 Paris, France. Tel.: +33 1 40 68 3299; fax: +33 1 45 68 8332. E-mail addresses: [email protected] (E. Melanitou), [email protected] (F. Tekaia), [email protected] (E. Yeramian). 0378-1119/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.gene.2012.09.055
2001; Barker et al., 2004). There also exists the possibility of transplacental transfer of autoantibodies to infants of T1D mothers (Greeley et al., 2002). Despite the identification of several risk factors for type 1 diabetes, early prediction is still missing due to insufficient predictive power of the individual risk factors (Purohit and She, 2008). Accordingly we seek here an alternative solution for the development of candidate biomarkers for early stages of autoimmune diabetes, mining our previously published trancriptome data (Regnault et al., 2009). More specifically we focus on the class of transcripts coding for secreted proteins in the NOD mouse and we demonstrate indeed the potentiality to use this specific category of genes as biomarkers for autoimmune predisposition. The rationale of the focus here on secreted protein transcripts is twofold: (1) On functional grounds, secreted proteins are known to play key roles in various important biological processes such as morphogenesis (Thomas et al., 2011; Tremble et al., 1993), angiogenesis (Onuffer and Horuk, 2002), cellular differentiation (Nalbant et al., 2005; Rosenow et al., 2010), apoptosis (Danielsen and Maihle, 2002) and modulation of immune response (Flavell, 2002; Grandvaux et al., 2002). These proteins are also implicated in disease processes, such as cancer progression (Welsh et al., 2003). Accordingly, this class of proteins has attracted much interest in recent years, with notably the initiation of a concerted effort called “The Secreted Protein Discovery Initiative” (SPDI), aiming to the identification of human secreted and
E. Melanitou et al. / Gene 512 (2013) 161–165
transmembrane proteins with bioinformatic approaches (Clark et al., 2003). (2) A diagnostic potential exists when secreted proteins are found in the serum. Indeed such proteins may be targeted by specific antibodies or small molecules in body fluids and thus constitute potential biomarkers for disease detection. In this general background, we develop here proof of concept for the possibility of early biomarkers for T1D, in the NOD mouse model, based on “secretome” data. More precisely, mining our previously published transcriptome data in this model (Regnault et al., 2009), we identify 71 genes coding for proteins located in the extracellular space. We then exploit the global properties of the expression signatures of these secreted protein genes (SPGs) with multivariate analysis, resorting to correspondence analysis (CA). Such analysis revealed a clear-cut segregation of the individual samples with the various subphenotypes. Examination of the genes involved in the analysis shows that the segregation is not following simple functional characteristics, further highlighting the interest in the global multivariate approach. In addition, such examination also reveals that a significant fraction of the secreted genes share identical expression profiles, allowing the perspective of a significant reduction of the candidate-set size in practice. Finally, with most of the genes in the set having human orthologs, it is expected that it will be possible to extend the developed system towards an efficient test for establishing the association of the presence of the corresponding proteins in the peripheral blood with the early detection of T1D in humans. 2. Materials and methods 2.1. Animals and analysis of microarray data The settings for experimental data collection and microarray analysis have been described previously (Regnault et al., 2009). In brief, NOD/Tac mice were used and monitored for the presence of insulin autoantibodies by radioimmunoassay at one week intervals after weaning (starting at 3 weeks of age) (Melanitou et al., 2004; Yu et al., 2000). Animals were sacrificed at 5 weeks of age as previously described (Melanitou et al., 2004). Two groups of animals were selected (E-IAApos and E-IAAneg) and high quality T-RNAs were prepared (Qiagen) from PLN tissues. T-RNA samples (4.5 μg) were monitored for cDNA preparation and hybridized on the Affymetrix MG_U74A_version 2 arrays (Santa Clara, Ca) containing 12 486 probe sets. The resulting data (in Affymetrix CEL format) were subjected to RMA (Robust Multichip Average) normalization (Bolstad et al., 2003) and transcripts with “Absent” Detection Call in both conditions were removed from further analysis. The RMA-normalized data set was used to identify modulated gene expression by the LPE (Local Pooled Error) statistical test (Jain et al., 2003), following adjustment for multiple comparisons by the Hochberg and Benjamini test (Hochberg and Benjamini, 1990). Box plots metrics of the RMA-normalized data illustrated the robustness of the data for both sets of samples (see supplementary data in Regnault et al., 2009). The data used for the correspondence analyses concerned nine individual mice (PLN samples) and were extracted from the initial data set, deposited in the NCBI's Gene Expression Omnibus (Edgar et al., 2002), GEO: GSE15582. The samples were in two groups, relative to early insulin autoantibodies (E-IAA) subphenotypes (Table S1): 5 negative samples (E-IAAneg) and 4 positive ones (E-IAApos). 2.2. Multivariate analysis The expression patterns of the transcripts for secreted-protein genes were analyzed with correspondence analysis (CA). For this analysis, the expression levels of the various genes were represented by their “expression flags” (A: absent; M: marginal and P: present), after corrections for background noise (as previously described (Regnault et al., 2009)) in the nine individual samples used for the
arrays. As such, the correspondence analysis concerned 3 × 9 samples, with each mouse individual An (either E-IAAneg or E-IAApos; n indexing variable) associated with three specific expression profiles An.pf(A), An.pf(M) and An.pf(P), according to the expression flags (A, M or P) for the set of genes in the study. Accordingly, with such coding, we are led to a table (not shown) containing 27 samples (9 × 3, corresponding to the various An.pf(X)) versus 71 SPGs. The fold change values for the expression levels (between negative and positive subphenotypes) were not used in the CA analysis. 3. Results 3.1. Identification of secreted protein-genes (SPGs) For type 1 diabetes disease pathogenesis in the NOD mouse, the early pre-inflammatory stages generally occur at the age of 4–6 weeks, after weaning and early priming of T cells reactive to islet β cell antigens takes place in the PLN (Turley et al., 2003). Accordingly, we performed the analyses in this tissue, in mice at 5 weeks of age. In such conditions, considering the E-IAA subphenotype, we identified by microarray analysis 165 transcripts regulated in the PLN of E-IAApos mice (Melanitou et al., 2004; Regnault et al., 2009). Out of the 165 differentially expressed genes identified in the E-IAApos vs E-IAAneg samples, 71 genes (43%), code for secreted proteins, as assigned by Gene Ontology annotations (http://www. geneontology.org/) (Huang et al., 2007) for cellular localization (pb 2.7e−28). Such percentage corresponds to an over-representation of this class of genes, as compared to the overall percentage of genes coding for secreted proteins (a mere 16%) in the Affymetrix arrays (MGU74Av2). Expression differences of the 71 transcripts in the E-IAApos vs E-IAAneg samples varied between 90 and 1.5 fold, with the majority of the genes (66 genes) being up-regulated in the E-IAApos PLN (Fig. 1). Only five transcripts were down-regulated, all of them coding for immunoglobulin-related proteins. With the exception of 3 ESTs, the remaining 68 genes code for known proteins and the majority of them have human orthologs (Table S1). 3.2. Correspondence analysis of expression patterns of SPGs In order to gain a global picture of the -complex- dependencies between expression patterns of the SPGs and the E-IAA subphenotype we resorted to correspondence analysis (CA) a multivariate statistical method (Tekaia and Yeramian, 2006; Tekaia et al., 2002). Our aim was to assess the usage of secreted proteins as biomarkers for autoimmune predisposition in the NOD mouse. 18 16
Mean E-IAApos (Log2)
162
14 12
r=0.65, p≤0.0001
10 8 6 4 2 0 0
2
4
6
8
10
12
14
16
Mean E-IAAneg (Log2) Fig. 1. Relative expression patterns of secreted protein genes (SPGs) in the PLN of E-IAApos and E-IAAneg. Linear regression line is drawn (correlation coefficient r = 0.65 with p ≤0.0001). Down-regulated genes in the E-IAApos samples are represented as triangles.
E. Melanitou et al. / Gene 512 (2013) 161–165
Our analysis was based on the expression levels of the 71 SPG transcripts for the PLN samples from nine individual mice used for the transcriptome analysis (Table S1). For the 71 transcripts, based on expression levels, Absent/Marginal/Present expression flags were calculated (following Affymetrix conventions). With such representation, each mouse individual An (n variable for indexing) was associated with three specific expression profiles An.pf(A), An.pf(M) and An.pf(P), according to the expression flags (A, M or P) for the 71 genes in the given sample. Accordingly, with such coding, we are led to a data table (not shown) containing 27 samples (9 × 3, corresponding to the various An.pf(X)) versus 71 SPGs. Preliminary analyses of this table showed that M profiles were rare and therefore not significant. In addition, four An.pf(A) or An.pf(P) samples appeared as outliers having the same expression flag for the majority of the genes. The An.pf(M) profiles, together with the 4 outliers, were then removed from the final data table used for correspondence analysis. The distribution of the samples (An.pf(X), with X either A or P) and the 71 SPGs on the first factorial plane is displayed in Fig. 2, associated with about 70% of the total information (54.9% and 14.6% following the factorial axes F1 for the F2 respectively leaving 30.5% for the remaining factorial space). This distribution reveals a clearcut segregation between An.pf(P) and An.pf(A) samples (P and A flags, in red and blue respectively), from left to right following the first factorial axis F1 (Fig. 2). In addition, very interestingly, this overall segregation further "separates" the An.pf(P) samples into positive and negative subphenotype categories. More specifically, the separation scheme is visually highlighted in Fig. 2 by two separation lines (dotted lines). In this scheme, all An.pf(A) samples are in the rightside part of the factorial plane, in light blue background, whereas all An.pf(P) samples are in the left-side part of the factorial plane, in red background. This red part of the factorial plane further subdivides into two parts, with the dark-shaded part harboring the positive subphenotype An.pf(P) samples (represented underlined). The only exception to this separation scheme appears to be the E-IAA positive A15.6 sample, for which A15.6pf(A) appears isolated at the very lower right limit of the factorial plane (represented underlined), whilst the A15.6pf(P) clusters together with the other E-IAA positive samples.
163
In summary, based on the information considered here, correspondence analysis leads to a segregation scheme following the F1 axis, associated with the three specific categories in three specific regions (as visually highlighted in Fig. 2 with the dotted lines, from left to right): I) negative subphenotype An.pf(P) samples in the light-shade red part, II) positive subphenotype An.pf(P) samples in the dark-shade red part and finally III) negative subphenotype An.pf(A) samples in the light blue part. Based on this scheme, the positioning of any individual with an unknown subphenotype on the factorial plane, according to the levels of the secreted protein transcripts identified in the periphery (plasma or serum), could in principle help identify an early predisposition to develop autoimmune diabetes. 3.3. Clustering rationale and data size reduction Examination of the genes with CA, in the three categories described above, reveals that the segregation does not follow any simple functional rationale. This feature is simply illustrated, for example, by the kallikrein genes, which appear scattered on the three parts of the factorial plane (Fig. 2). In a more systematic approach, functional annotations (http://david.abcc.ncifcrf.gov/ online annotation tools), were associated with the identified genes (Table S1). Such annotations reveal that the majority of the genes are associated with immune functions, independent of their relative segregation on the factorial plane. It is also worth noting that several serine peptidases cluster with the genes expressed in the E-IAA positive samples, including several genes coding for kallikreins. The latter are proteins found in the plasma, implicated in the regulation of homeostatic functions and also investigated as cancer biomarkers (Diamandis and Yousef, 2002). The functional annotations of the SPGs (Table S1 and Fig. S1) also reveal some interesting features relative to tissue distribution of the identified transcripts: significant association with the pancreas (p= 2.85e-16), plasma (p=1.68e-8), salivary gland (p= 1.75e-8) and the liver (p= 1.947e-7). However some of the transcripts are also expressed in other tissues (Fig. S1). High expression of these transcripts in the E-IAApos PLN might reflect autoimmune-related homeostatic changes, probably due to the presence of dendritic cells (Yin et al.,
Fig. 2. Correspondence analysis of expression patterns of SPG transcripts in nine individual NOD mice: first factorial plane of CA, carried according to the expression levels of the SPG transcripts. Each mouse individual An (n variable for indexing) is associated with specific expression profiles An.pf(A) and An.pf(P) (according to the expression flags ‘A’ or ‘P’: absence or presence of expression, in the corresponding individual, for the 71 genes in the study). E-IAApos subphenotype samples are underlined. Blue dots are associated with genes whose expression patterns are identical for the various individuals. Black dots are associated with genes whose expression patterns are unique. Genes with expression differences (E-IAApos vs E-IAAneg) higher than 25 folds are represented in brown.
164
E. Melanitou et al. / Gene 512 (2013) 161–165
2012). Indeed these cells are involved in the inflammatory process and have the capacity to migrate to the pancreas-draining lymph nodes where they present self-antigens (Heath et al., 2004; Steinman et al., 2003). From another perspective, it is also interesting to examine the distribution of genes on the factorial plane following fold changes in expression levels. Genes with the largest fold values (with threshold set at 25, for illustration) are represented on the factorial plane (Fig. 2) in brown. It appears that many such genes are in fact clustered in the central part, associated with positive subphenotype An.pf(P) samples. Such observations could potentially suggest the elaboration of more straightforward testing schemes, based for example on changes in expression levels alone. However, the scattering (and mutual positioning) of An.pf(X) individuals and genes on the factorial plane clearly shows that such a scheme could not be effective, with various genes (relative to the threshold 25 in fold changes, noted in brown) also appearing in the left and right parts of the factorial plane. On the other hand, the segregation of An.pf(X) individuals into three well-identified parts in the multivariate analysis does not rely on the fold change information, further highlighting the interest of such approach for the purpose of a reliable discriminative test. Finally, examination of the genes on the factorial plane reveals a series of spots where the positions (of variable number) of genes coincide (in Fig. 2 the corresponding spots are in blue, whereas spots associated with single genes are in black). This observation pinpoints the fact that, at such spots, several genes (with again possibly no direct functional relationship) present identical profiles (in terms of A/M/P flags) for the nine individual samples. As such genes play redundant roles in the discriminatory analysis, it should become then possible – in principle – to reduce significantly the size of the gene set used. However, only further analyses, with increasing number of individuals, could allow determining the ideal compromises between size reduction and enhanced robustness. 4. Discussion The early detection and the possibility of preventing the autoimmune process represent great challenges in T1D. In this respect, it is notably necessary to be able to identify the time of occurrence of this cryptic phase, which includes also insulitis (the β cell-damaging inflammatory state). At present, screening for and detection of autoantibodies against islet autoantigens have a good predictive power for the disease (Orban et al., 2009). It seems however clear that the presence of autoantibodies in the peripheral blood implies that in fact the autoimmune processes per se have already been initiated. In this context, our aim was to address the possibility to identify markers that delineate earlier events, taking place prior to the mechanisms involved in the development of antibodies to islet antigens. We have used the NOD animal model for T1D to address this question. The low penetrance of disease in the NOD mice made it necessary that we proceed to an initial screening for animals that would progress to disease. With this respect, we have used the presence of E-IAA as a T1D subphenotype to select individual mice, and compare the transcription profiles with those of E-IAA negative animals. Such comparisons revealed that a high proportion of the identified transcripts code for secreted proteins, with the majority of them being up-regulated in the E-IAApos PLN. This situation could be related to the recently reported increase of the β cell mass and islet hyperactivity in 6 week old NOD mice (Liang et al., 2011). Taken together with our previous demonstration for the existence of a correlation between the presence of E-IAA and T1D (Melanitou et al., 2004), these results and observations led us to investigate the predictive diagnostic power of the E-IAA PLN “secretome”, as potentially reflecting the early steps preceding the development of islet autoimmunity.
To investigate the validity of this hypothesis, we resorted to correspondence analysis of the expression patterns of secreted proteinencoding genes, in relation to the subphenotype used in our study. For the elaboration of predictive biomarker, it seems natural to use multivariate representations to capture complex dependencies between transcriptome data and various genetic traits such as autoimmune diabetes subphenotypes (E-IAA in this study). Indeed, such methods were used successfully in other instances for the detection of biomarkers based on transcriptome data, such as for the detection of cancer biomarkers (Zhang et al., 2005), with principle component analysis (another variant of a multivariate method). For the transcriptome data, the factorial plane obtained with correspondence analysis (associated with about 70% of the total information) reveals a clear-cut segregation of the samples associated with different E-IAA subphenotypes, following the expression patterns of the identified genes. This representation also highlights the co-localization of (variable number) of genes in several spots of the factorial plane, associated with identical expression profiles. This observation opens the perspective of more or less significant reduction of the number of genes used as predictive biomarkers, at the corresponding spots. Of course, ideally, such reduction could retain only a single gene for each such spot. However, only detailed experimental testing could allow investigating the levels of robustness compatible with reduced sets of genes in the test. In short, no matter the size of the set of genes, the scheme with CA opens the possibility to evaluate the relative risk for a subject to develop T1D, following the positioning in one of the three welldelimited parts of the factorial plane. Finally, the testing scheme described herein can lead to new explorations for the underlying biological features and mechanisms, with the appropriate cautious considerations. Notably, despite the fact that the identified transcripts code for known proteins the question remains of how well transcript levels do actually reflect protein abundance, in particular in the peripheral blood. Indeed regulatory mechanisms at the translation level or posttranslational modifications may influence the effective protein abundance. The links between transcriptomic and proteomic profiles have been investigated by computational methods and generally a poor correlation between mRNA and protein levels has been reported (Griffin et al., 2002; Mootha et al., 2003). More specifically, in these reports, high correlations were found only for components of some molecular pathways, such as the ribosome and cell adhesion complexes (Rogers et al., 2008). On the other hand, contrary to the above observations, a cross comparison of proteomic patterns with mRNA patterns from mouse tissues revealed broad areas of agreement with few inconsistencies (Kislinger et al., 2006). In this respect, concerning our secretome data, ultimately it is only through experimental testing (in the mouse model) or clinical evaluation in human patients that the effectiveness of the diagnostic scheme could be confirmed. In the elaboration of this scheme the interest in resorting to multivariate analysis was highlighted above, as compared to simpler approaches relying for example on the expression level data. With this respect, even though effective discriminatory power involves all 71 transcripts, it is nevertheless clear that a detailed investigation of the six transcripts expressed solely in the E-IAApos animals (and absent in the E-IAAneg ones, Table S1), represents a priority for exploratory studies in the NOD mouse or in human clinical trials. From another point of view, also considering the detailed data, it appears that in our set the only down-regulated transcripts in the E-IAApos samples correspond to genes coding for immunoglobulins, all clustering together on the factorial plane (Fig. 2 and Table S1). Diversity of the antibody repertoire is a prerequisite for efficient recognition of a broad spectrum of invading microorganisms by the immune system. Incorrect or misregulated repertoire can initiate autoimmunity (Ishida et al., 2006; Jankovic and Nussenzweig, 2003). In this context, it is tempting to speculate that down regulation of
E. Melanitou et al. / Gene 512 (2013) 161–165
immunoglobulin transcripts in the E-IAApos samples might be associated with the initiation of the autoimmune process in these animals. Based on our data and considerations described above, additional investigations combining diagnostic purpose explorations with studies targeting the fundamental early mechanisms in T1D represent a promising route to significant advances in this field. Supplementary data to this article can be found online at http:// dx.doi.org/10.1016/j.gene.2012.09.055. Acknowledgments EM thanks Dr. Geneviève Milon for constant support and encouraging discussions. FT acknowledges financial support from the Institut Pasteur (PTR 370). References Abiru, N., et al., 2001. Transient insulin autoantibody expression independent of development of diabetes: comparison of NOD and NOR strains. J. Autoimmun. 17, 1–6. Barker, J.M., et al., 2004. Prediction of autoantibody positivity and progression to type 1 diabetes: Diabetes Autoimmunity Study in the Young (DAISY). J. Clin. Endocrinol. Metab. 89, 3896–3902. Bolstad, B.M., Irizarry, R.A., Astrand, M., Speed, T.P., 2003. A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics 19, 185–193. Clark, H.F., et al., 2003. The secreted protein discovery initiative (SPDI), a large-scale effort to identify novel human secreted and transmembrane proteins: a bioinformatics assessment. Genome Res. 13, 2265–2270. Danielsen, A.J., Maihle, N.J., 2002. The EGF/ErbB receptor family and apoptosis. Growth Factors 20, 1–15. Diamandis, E.P., Yousef, G.M., 2002. Human tissue kallikreins: a family of new cancer biomarkers. Clin. Chem. 48, 1198–1205. Edgar, R., Domrachev, M., Lash, A.E., 2002. Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res. 30, 207–210. Flavell, R.A., 2002. The relationship of inflammation and initiation of autoimmune disease: role of TNF super family members. Curr. Top. Microbiol. Immunol. 266, 1–9. Grandvaux, N., tenOever, B.R., Servant, M.J., Hiscott, J., 2002. The interferon antiviral response: from viral invasion to evasion. Curr. Opin. Infect. Dis. 15, 259–267. Greeley, S.A., et al., 2002. Elimination of maternally transmitted autoantibodies prevents diabetes in nonobese diabetic mice. Nat. Med. 8, 399–402. Griffin, T.J., et al., 2002. Complementary profiling of gene expression at the transcriptome and proteome levels in Saccharomyces cerevisiae. Mol. Cell. Proteomics 1, 323–333. Heath, W.R., et al., 2004. Cross-presentation, dendritic cell subsets, and the generation of immunity to cellular antigens. Immunol. Rev. 199, 9–26. Hochberg, Y., Benjamini, Y., 1990. More powerful procedures for multiple significance testing. Stat. Med. 9, 811–818. Huang, W., et al., 2007. The DAVID Gene Functional Classification Tool: a novel biological module-centric algorithm to functionally analyze large gene lists. Genome Biol. 8, R183. Ishida, D., et al., 2006. Rap1 signal controls B cell receptor repertoire and generation of self-reactive B1a cells. Immunity 24, 417–427.
165
Jain, N., Thatte, J., Braciale, T., Ley, K., O'Connell, M., Lee, J.K., 2003. Local-pooled-error test for identifying differentially expressed genes with a small number of replicated microarrays. Bioinformatics 19, 1945–1951. Jankovic, M., Nussenzweig, M.C., 2003. OcaB regulates transitional B cell selection. Int. Immunol. 15, 1099–1104. Kislinger, T., et al., 2006. Global survey of organ and organelle protein expression in mouse: combined proteomic and transcriptomic profiling. Cell 125, 173–186. Liang, K., et al., 2011. Contribution of different mechanisms to pancreatic beta-cell hyper-secretion in non-obese diabetic (NOD) mice during pre-diabetes. J. Biol. Chem. 286, 39537–39545. Melanitou, E., Devendra, D., Liu, E., Miao, D., Eisenbarth, G.S., 2004. Early and quantal (by litter) expression of insulin autoantibodies in the nonobese diabetic mice predict early diabetes onset. J. Immunol. 173, 6603–6610. Mootha, V.K., et al., 2003. Integrated analysis of protein composition, tissue diversity, and gene regulation in mouse mitochondria. Cell 115, 629–640. Nalbant, D., et al., 2005. FAM20: an evolutionarily conserved family of secreted proteins expressed in hematopoietic cells. BMC Genomics 6, 11. Onuffer, J.J., Horuk, R., 2002. Chemokines, chemokine receptors and small-molecule antagonists: recent developments. Trends Pharmacol. Sci. 23, 459–467. Orban, T., et al., 2009. Pancreatic islet autoantibodies as predictors of type 1 diabetes in the Diabetes Prevention Trial-Type 1. Diabetes Care 32, 2269–2274. Purohit, S., She, J.X., 2008. Biomarkers for type 1 diabetes. Int. J. Clin. Exp. Med. 1, 98–116. Regnault, B., Osorio, Y.F.J., Miao, D., Eisenbarth, G., Melanitou, E., 2009. Early over expression of messenger RNA for multiple genes, including insulin, in the pancreatic lymph nodes of NOD mice is associated with islet autoimmunity. BMC Med. Genomics 2, 63. Rogers, S., et al., 2008. Investigating the correspondence between transcriptomic and proteomic expression profiles using coupled cluster models. Bioinformatics 24, 2894–2900. Rosenow, A., et al., 2010. Identification of novel human adipocyte secreted proteins by using SGBS cells. J. Proteome Res. 9, 5389–5401. Steinman, R.M., Hawiger, D., Nussenzweig, M.C., 2003. Tolerogenic dendritic cells. Annu. Rev. Immunol. 21, 685–711. Tekaia, F., Yeramian, E., 2006. Evolution of proteomes: fundamental signatures and global trends in amino acid compositions. BMC Genomics 7, 307. Tekaia, F., Yeramian, E., Dujon, B., 2002. Amino acid composition of genomes, lifestyles of organisms, and evolutionary trends: a global picture with correspondence analysis. Gene 297, 51–60. Thomas, S., et al., 2011. Development of secreted protein and acidic and rich in cysteine (SPARC) targeted nanoparticles for the prognostic molecular imaging of metastatic prostate cancer. J. Nanomed. Nanotechnol. 2. Tremble, P.M., Lane, T.F., Sage, E.H., Werb, Z., 1993. SPARC, a secreted protein associated with morphogenesis and tissue remodeling, induces expression of metalloproteinases in fibroblasts through a novel extracellular matrix-dependent pathway. J. Cell Biol. 121, 1433–1444. Turley, S., Poirot, L., Hattori, M., Benoist, C., Mathis, D., 2003. Physiological beta cell death triggers priming of self-reactive T cells by dendritic cells in a type-1 diabetes model. J. Exp. Med. 198, 1527–1537. Welsh, J.B., et al., 2003. Large-scale delineation of secreted protein biomarkers overexpressed in cancer tissue and serum. Proc. Natl. Acad. Sci. U. S. A. 100, 3410–3415. Yin, N., et al., 2012. Functional specialization of islet dendritic cell subsets. J. Immunol. 188, 4921–4930. Yu, L., et al., 2000. Early expression of antiinsulin autoantibodies of humans and the NOD mouse: evidence for early determination of subsequent diabetes. Proc. Natl. Acad. Sci. U. S. A. 97, 1701–1706. Zhang, X.W., Yap, Y.L., Wei, D., Chen, F., Danchin, A., 2005. Molecular diagnosis of human cancer type by gene expression profiles and independent component analysis. Eur. J. Hum. Genet. 13, 1303–1311.