The organization of the transcriptional network in specific neuronal classes
Kellen D Winden1,2,3, Michael C Oldham1,2,4, Karoly Mirnics5,6, Philip J Ebert7, Christo H Swan8, Pat Levitt6,7, John L Rubenstein8,9, Steve Horvath4,10 & Daniel H Geschwind1,2,3,4,11
- Interdepartmental Program for Neuroscience, University of California Los Angeles, Los Angeles, CA, USA
- Program in Neurogenetics, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, USA
- Department of Neurology, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, USA
- Department of Human Genetics, University of California Los Angeles, Los Angeles, CA, USA
- Department of Psychiatry, Vanderbilt University, Nashville, TN, USA
- Vanderbilt Kennedy Center for Research on Human Development, Vanderbilt University, Nashville, TN, USA
- Department of Pharmacology, Vanderbilt University, Nashville, TN, USA
- Department of Psychiatry and Langley Porter Psychiatric Institute, University of California, San Francisco, CA, USA
- Nina Ireland Laboratory of Developmental Neurobiology, University of California, San Francisco, CA, USA
- Department of Biostatistics, University of California Los Angeles School of Public Health, Los Angeles, CA, USA
- Semel Institute for Neuroscience and Human Behavior, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, USA
Correspondence to: Daniel H Geschwind1,2,3,4,11 Program in Neurogenetics, Department of Neurology and Semel Institute, David Geffen School of Medicine, Los Angeles, CA 90095, USA. Tel.: +1 310 206 6814; Fax: +1 310 267 2401; Email: dhg@ucla.edu
Received 24 December 2008; Accepted 9 June 2009; Published online 28 July 2009
Article highlights
- The neuronal transcriptome has an underlying network structure that is comprised of robust co-expression relationships and also shows good correspondence to the level of the proteome.
- Network modules, which correspond to functionally related groupings of genes, relate to specific physiological and structural aspects of neuronal phenotype, such as the synapse, that underlie neuronal diversity.
- Network modules can be used to annotate gene function and drive discovery. Here, we show that modules define specific cell types based on developmental programs, and provide a new molecular basis on which to define mitochondrial populations based on sub-cellular localization.
- We show that systems level network relationships based on transcriptional profiling correspond to real in vivo relationships observed at the bench, a first in multi-cellular organisms.
Synopsis
Understanding the molecular basis of neuronal diversity has been aided by the ability to perform genome-wide expression profiling, but achieving broad functional insight remains a considerable challenge. Systems-level analyses that consider relationships between genes permit association of gene expression variation with specific cell phenotypes. Weighted gene co-expression network analysis (WGCNA) groups functionally related genes into modules in an unsupervised manner (Zhang and Horvath, 2005; Horvath et al, 2006; Oldham et al, 2006, 2008), based on the self-organizing properties of complex systems (Barabasi and Albert, 1999; Ravasz et al, 2002). The modularity of the system allows independent analysis of the components, and the identification of relationships between genes facilitates gene annotation based on network position without assumptions about gene function.
Here, we analyzed published microarray data from single neuronal populations (Sugino et al, 2006) using WGCNA to organize the neuronal transcriptome and examine its relationship to cellular function. We identified 13 modules that had characteristic patterns of gene expression and enrichment for specific gene ontology categories. These modules were systematically validated using both transcriptional and proteomic data, showing that the identified co-expression relationships are reproducible and biologically relevant on the protein level. We then show that many modules correspond to specific functions related to neuronal biology, allowing large-scale annotation of function and a new perspective on neuronal diversity.
For example, several modules correspond to developmental programs or origins of neuronal classes. One module corresponds to the subset of interneurons derived from the subpallium and contains all of the Distalless transcription factors that are expressed in the brain (Panganiban and Rubenstein, 2002). A group of genes within this module was specifically regulated in both somatostatin- and parvalbumin-positive interneurons, and one of the most highly connected genes was galectin-1 (Lgals1), which had no known role in these cell populations. Visualization of these genes illustrates that galectin-1 and somatostatin are closely related, and we confirmed this using immunohistochemistry to show that nearly 80% of the galectin-1-positive cells were also somatostatin positive. These data indicate that galectin-1 may serve as a useful marker for this class of cells.
Neurons differ greatly in their characteristic firing activity, and we hypothesized that some modules would be related to this fundamental neuronal phenotype. We tested this by comparing physiological parameters to the gene expression patterns found within the modules. The module that had the highest correlation with firing rate was also enriched for proteins localized to mitochondria and involved in carboxylic acid metabolism. This suggests that the coupling between neuronal activity and oxidative energy production (Kasischke et al, 2004) extends to the transcriptional level.
Neuronal morphology and metabolism are aspects of neuronal phenotypic diversity that we theorized might be reflected at the transcriptional level through variation in organellar composition. We tested this hypothesis by comparing modular organization to proteomic data from a large-scale analysis of subcellular organelles (Foster et al, 2006), and other studies that focused on specific neuronal features (i.e. synaptosome (Schrimpf et al, 2005), postsynaptic density (Collins et al, 2006), presynaptic fraction (Phillips et al, 2005), synaptic vesicles (Morciano et al, 2005)). We observed that two modules showed significant overrepresentation of mitochondrial proteins, but had distinct expression profiles and were related to different aspects of neuronal physiology, leading to the hypothesis that they represented different mitochondrial populations.
Mitochondrial heterogeneity within neurons has been suggested earlier, with one population localized to the cell body and the other to synapses (Lai et al, 1977). We examined the hub genes of these modules and tested their ability to differentiate between mitochondrial populations (Figure 6A). We found that genes in the non-synaptic mitochondrial module (Phb and Fis1/Ttc11) were enriched in the free mitochondrial fraction, whereas the genes in the synaptic mitochondrial module (Vdac2 and Uqcrfs1) were enriched in the synaptosomal fraction (Figure 6B). Using immunocytochemistry, we found that Phb was mainly colocalized with mitochondria in the cell body (Figure 6C), whereas Uqcrfs1 was colocalized with mitochondria in the processes, as well as the cell body (Figure 6D). These data suggest that mitochondrial heterogeneity in neurons is reflected at the transcriptional level.
Figure 6
Submodules within the blue and turquoise modules represent different mitochondrial populations. We examined the expression of genes known to be localized to the mitochondria in submodules of both the blue and turquoise modules subcellular fractionation. (A) Table showing the most highly connected genes in the synaptic and non-synaptic mitochondrial submodules, sorted by intramodular connectivity. The genes that are in bold (Vdac2, Uqcrfs1, Phb, Fis1/Ttc11) denote those genes that we chose to experimentally validate. Although Phb and Fis1/Ttc11 are not within the most highly connected genes, they have a kME>0.75. (B) Representative western blots of the synaptosomal and mitochondrial fractions that show the relative enrichment of specific genes within one fraction that was predicted by the network. Control blots of synaptophysin (34 kDa) and cytochrome (c) (14 kDa) show appropriate enrichment in synaptic and mitochondrial fractions, respectively. (C) Three replicate western blots showing the ratios of synaptosomal or mitochondrial enrichment of each of the proteins (
s.e.m.). The genes in the non-synaptic mitochondrial module, Fis1/Ttc11 (17 kDa) and Phb (30 kDa), were enriched 47- and 6-fold in the free mitochondrial fraction versus the synaptosomal fraction, respectively. The genes in the synaptic mitochondrial module, Uqcrfs1 (25 kDa) and Vdac2 (38 kDa), were enriched 8- and 7-fold in the synaptosomal fraction versus the free mitochondrial fraction, respectively. (D) Primary hippocampal neurons after 3 weeks in vitro. MitoTracker (red) was used to label the mitochondria within a neuron, whereas Map2 (blue) was used to label the neuronal processes. Phb (green) is a hub in the non-synaptic mitochondrial module, and it co-localizes with mitochondria mainly within the cell body (arrows). (E) Uqcrfs1 (Green) is a hub in the synaptic mitochondrial module, and it co-localizes primarily with mitochondria in the neuronal processes (arrows), as well as those in the cell body. These data indicate that genes in the synaptic mitochondrial module are enriched in mitochondria that are localized to neuronal processes. Scale Bar: 20 
m.
We then tested basic network predictions in vivo. Network theory predicts that the disruption of a highly connected gene will preferentially affect genes within the same module because of their high degree of co-regulation (Albert et al, 2000; Jeong et al, 2001). To test this hypothesis, we performed microarray expression analysis on the cortex of mice that had a large deletion of Dlx1 and Dlx2 (Anderson et al, 1997b) because Dlx1 was a highly connected gene in our analysis. We observed that the module containing Dlx1 was enriched for differentially expressed genes (Figure 7A), which included two known targets of Dlx1, Arx, and Dlx5 (Zerucha et al, 2000; Zhou et al, 2004; Cobos et al, 2005a). In addition, we observed that genes in the top quartile of connectivity with Dlx1 were more likely to be differentially expressed than genes in lower quartiles, showing a relationship between connectivity and differential expression (Figure 7C). These data show that Dlx1 and Dlx2 have functional roles in regulating genes co-expressed with them, as predicted by their network centrality.
Figure 7
In vivo validation of network model. Validation of the network model using gene expression data from two separate knockout mice. (A) Barplot representing the observed to expected ratio of differentially expressed genes (P<0.01) by module in the Dlx1/2 knockout mice and (B) Rgs4 knockout mice. In both the cases, only the modules containing the deleted gene were significantly enriched in differentially expressed genes (P<0.05). (C) Relationship between a gene's topological overlap or connectedness with a 'hub' gene (i.e. Dlx1/Dlx2 or Rgs4) and differential expression. Genes were ranked by connectivity within the module and the number that was differentially expressed within each quartile was counted and expressed as a percentage of total differentially expressed genes within the module. Error bars show the margin of error for the percentages. In both modules, there is a clear relationship in which genes that are highly connected to the deleted gene are more likely to be differentially expressed than other genes that are not as well connected. Nearly 60% of the genes that were differentially expressed in the enriched modules of either knockout strains were in the top quartile of connectivity with the deleted gene, which is significantly greater than other quartiles (P<0.005).
Full figure and legend (129K)Figures & Tables indexWe then tested the robustness of this relationship by examining the effects of a deletion of Rgs4, which regulates G-protein signaling (Hepler et al, 1997; Ladds et al, 2007) and is associated with schizophrenia (Mirnics et al, 2001; Chowdari et al, 2002). We generated transgenic mice with a deletion of Rgs4 and performed microarrays on their frontal cortex. There was significant enrichment of differentially expressed genes only in the module containing Rgs4 (Figure 7B). In addition, we observed a clear relationship between the intramodular connectivity of a gene with Rgs4 and its chance of being differentially expressed (Figure 7C). These data suggest that relationships identified here reflect real relationships between genes that are present in vivo.
Overall, these analyses provide a framework for understanding the molecular basis of neuronal diversity, which can be used to gain insight into the underlying mechanisms of important biological processes such as neural development, neuronal cell biology, and disease.
Acknowledgements
We thank Sugino et al for their generous sharing of the data and Sacha Nelson for his valuable comments on an earlier draft of this manuscript. We thank members of the Geschwind lab for their helpful input and Ezra Rosen specifically for testing the network explorer on this data set. We thank Dr Kelsey Martin and members of the Martin lab, Dr Carrie Heusner, and Rachel Jefferey, for valuable discussions and technical expertise. We thank Dr Linda Baum for generous gift of the galectin-1 antibody and the Animal Core of CURE, Digestive Diseases Division, UCLA for the generous gift of the somatostatin antibody (CURE.S6). For her excellent work on generating the microarray data for the Rubenstein lab, we thank Winnie Liang at the NIH Neuroscience Microarray, Translational Genomics Research Institute, Neurogenomics Division. We acknowledge support from NIH grants NIMH #R37 MH60233-06A1 (DHG, KW, MO), NINDS #U24 NS52108 (DHG, SH), NIGMS GM08042 (KW), Neurobehavioral Genetics Training Grant T32MH073526-01A1 (KW). Medical Scientist Training Program UCLA (KW), Aesculapians Fund of the UCLA School of Medicine (KW), the Miriam and Sheldon Adelson Program in Neural Repair and Rehabilitation (DHG, SH), and the JLRR from Nina Ireland, NIMH RO1 MH49428-01, RO1, and K05 MH065670 (JR, CS).
References
- Albert R, Jeong H, Barabasi A (2000) Error and attack tolerance of complex networks. Nature 406: 378–382 | Article | PubMed | ISI | ADS | ChemPort |
- Anderson SA, Qiu M, Bulfone A, Eisenstat DD, Meneses J, Pedersen R, Rubenstein JL (1997b) Mutations of the homeobox genes dlx-1 and dlx-2 disrupt the striatal subventricular zone and differentiation of late born striatal neurons. Neuron 19: 27–37 | Article | PubMed | ISI | ChemPort |
- Barabasi A, Albert R (1999) Emergence of scaling in random networks. Science 286: 509–512 | Article | PubMed | ISI
- Chowdari KV, Mirnics K, Semwal P, Wood J, Lawrence E, Bhatia T, Deshpande SN, Thelma BK, Ferrell RE, Middleton FA, Devlin B, Levitt P, Lewis DA, Nimgaonkar VL (2002) Association and linkage analyses of rgs4 polymorphisms in schizophrenia. Hum Mol Genet 11: 1373–1380 | Article | PubMed | ISI | ChemPort |
- Cobos I, Broccoli V, Rubenstein JLR (2005a) The vertebrate ortholog of aristaless is regulated by dlx genes in the developing forebrain. J Comp Neurol 483: 292–303 | Article | PubMed | ChemPort |
- Collins MO, Husi H, Yu L, Brandon JM, Anderson CN, Blackstock WP, Choudhary JS, Grant SG (2006) Molecular characterization and comparison of the components and multiprotein complexes in the postsynaptic proteome. J Neurochem 97(Suppl 1): 16–23
- Foster LJ, de Hoog CL, Zhang Y, Zhang Y, Xie X, Mootha VK, Mann M (2006) A mammalian organelle map by protein correlation profiling. Cell 125: 187–199 | Article | PubMed | ISI | ChemPort |
- Hepler JR, Berman DM, Gilman AG, Kozasa T (1997) Rgs4 and gaip are gtpase-activating proteins for gq alpha and block activation of phospholipase c beta by gamma-thio-gtp-gq alpha. Proc Natl Acad Sci USA 94: 428–432 | Article | PubMed | ADS | ChemPort |
- Horvath S, Dong J (2008) Geometric interpretation of gene co-expression network analysis. PloS Computational Biology 4: e1000117 | Article | PubMed | ChemPort |
- Jahr CE, Stevens CF (1987) Glutamate activates multiple single channel conductances in hippocampal neurons. Nature 325: 522–525 | Article | PubMed | ISI | ADS | ChemPort |
- Jansen R, Greenbaum D, Gerstein M (2002) Relating whole-genome expression data with protein-protein interactions. Genome Res 12: 37–46 | Article | PubMed | ISI | ChemPort |
- Kitamura K, Yanazawa M, Sugiyama N, Miura H, Iizuka-Kogo A, Kusaka M, Omichi K, Suzuki R, Kato-Fukui Y, Kamiirisa K, Matsuo M, Kamijo S, Kasahara M, Yoshioka H, Ogata T, Fukuda T, Kondo I, Kato M, Dobyns WB, Yokoyama M et al (2002) Mutation of arx causes abnormal development of forebrain and testes in mice and x-linked lissencephaly with abnormal genitalia in humans. Nat Genet 32: 359–369 | Article | PubMed | ISI | ChemPort |
- Kokoszka JE, Waymire KG, Levy SE, Sligh JE, Cai J, Jones DP, MacGregor GR, Wallace DC (2004) The adp/atp translocator is not essential for the mitochondrial permeability transition pore. Nature 427: 461–465 | Article | PubMed | ISI | ADS | ChemPort |
- McKay RD, Hockfield SJ (1982) Monoclonal antibodies distinguish antigenically discrete neuronal types in the vertebrate central nervous system. Proc Natl Acad Sci USA 79: 6747–6751 | Article | PubMed | ADS | ChemPort |
- Molyneaux BJ, Arlotta P, Hirata T, Hibi M, Macklis JD (2005) Fezl is required for the birth and specification of corticospinal motor neurons. Neuron 47: 817–831 | Article | PubMed | ISI | ChemPort |
- Morciano M, Burré J, Corvey C, Karas M, Zimmermann H, Volknandt W (2005) Immunoisolation of two synaptic vesicle pools from synaptosomes: a proteomics analysis. J Neurochem 95: 1732–1745 | Article | PubMed | ChemPort |
- Naga KK, Sullivan PG, Geddes JW (2007) High cyclophilin d content of synaptic mitochondria results in increased vulnerability to permeability transition. J Neurosci 27: 7469–7475 | Article | PubMed | ChemPort |
- Oldham MC, Konopka G, Iwamoto K, Langfelder P, Kato T, Horvath S, Geschwind DH (2008) Functional organization of the transcriptome in human brain. Nat Neurosci 11: 1271–1282 | Article | PubMed | ChemPort |
- Panganiban G, Rubenstein JLR (2002) Developmental functions of the distal-less/dlx homeobox genes. Development 129: 4371–4386 | PubMed | ISI | ChemPort |
- Polleux F (2005) Genetic mechanisms specifying cortical connectivity: let's make some projections together. Neuron 46: 395–400 | Article | PubMed | ISI | ChemPort |
- Qiu M, Bulfone A, Ghattas I, Meneses JJ, Christensen L, Sharpe PT, Presley R, Pedersen RA, Rubenstein JL (1997) Role of the dlx homeobox genes in proximodistal patterning of the branchial arches: mutations of dlx-1, dlx-2, and dlx-1 and -2 alter morphogenesis of proximal skeletal and soft tissue structures derived from the first and second arches. Dev Biol 185: 165–184 | Article | PubMed | ISI | ChemPort |
- Shawlot W, Behringer RR (1995) Requirement for lim1 in head-organizer function. Nature 374: 425–430 | Article | PubMed | ISI | ADS | ChemPort |
- Yoshimura Y, Callaway EM (2005) Fine-scale specificity of cortical networks depends on inhibitory cell type and connectivity. Nat Neurosci 8: 1552–1559 | Article | PubMed | ISI | ChemPort |
- Zapala MA, Hovatta I, Ellison JA, Wodicka L, Del Rio JA, Tennant R, Tynan W, Broide RS, Helton R, Stoveken BS, Winrow C, Lockhart DJ, Reilly JF, Young WG, Bloom FE, Lockhart DJ, Barlow C (2005) Adult mouse brain gene expression patterns bear an embryologic imprint. Proc Natl Acad Sci USA 102: 10357–10362 | Article | PubMed | ChemPort |
- Zerucha T, Stühmer T, Hatch G, Park BK, Long Q, Yu G, Gambarotta A, Schultz JR, Rubenstein JL, Ekker M (2000) A highly conserved enhancer in the dlx5/dlx6 intergenic region is the site of cross-regulatory interactions between dlx genes in the embryonic forebrain. J Neurosci 20: 709–721 | PubMed | ISI | ChemPort |
