Journal of Bioinformatics, Proteomics and Imaging Analysis | Volume 1 | Issue 3 | 2015-12-14

The Interaction Network of the Presynaptic Active Zone

Martin Wegner

Abstract

The amyloid precursor protein (APP) has been cloned and identified more than 20 years ago as the Precursor protein of amyloid beta (A ) peptide, the main component of senile plaques in brains effected by Alzheimer s disease(AD). APP has recently been allocated to the proteome of the Presynaptic active zone (PAZ). We combined state-of-the-art proteomics and bioinformatics network tools to provide direct evidence for the physiological role of the APP at the PAZ. We constructed a protein-protein interaction network (PPI) based on experimental data in currently available databases, and filtered for experimentally validated physical interactions. The network had 615 proteins and 3,390 physical interactions, and turned out to be scale-free with a degree exponent of 1.3. A visualization of topological centralities gave insights into the functional role of proteins. Community detection algorithms revealed an organization of the interactions into highly connected modules, exerting physiological roles like, e.g., energy metabolism and structural organization. Our experimental data of relative protein abundance in mice with an APP knockout (APP-KO) and with a conditional double knockout (NexCre-cDKO) of APP and amyloid precursor-like protein 2 (APLP2) were both mapped to the modules. The conditional double knockout led to upregulation of proteins associated with energy metabolism, while down regulated proteins were mainly enriched in modules of structural organization of the neurotransmitter release site. The distribution of up-and down regulated proteins in single knockout is more homogeneous than in the double knockout. Our findings of alterations in the proteomic composition in both APP-mutants identified APP as a structural and functional regulator within the hippocampal active zone network.

Keywords: Presynaptic active zone, Protein- protein interaction network, Amyloid precursor protein, Hippo campus

Full text

Dementia is associated with a decline in mental ability, mpairments in memory and changes in personality. The most frequent form of dementia worldwide is the sporadic Alzheimer s disease [1] , and is mainly associated with hyper phosphorylated protein tau and intra- and extracellular deposits of amyloid (A ) oligomers [1] . APP has been cloned and identified more than 20 years ago as the precursor protein of A peptide, the main component of senile plaques in brains affected by AD [2] . The pathophysiology of AD, characterized by a massive loss of synapses, cognitive decline and behavioral changes, has been attributed to the accumulation of A [3] . Within the last decades, much effort has gone into the understanding of the molecular basis of AD progression. However, little is known about the actual physiological function of APP. Allocating APP to the proteome of the structurally and functionally dynamic presynaptic active zone (PAZ) has highlighted APP as a hitherto unknown player with in the setting of the presynapse [4,5] . We combined state-of-the-art proteomics and bioinformatics network tools to provide evidence for the physiological key role of APP in the PAZ. Therefore, we isolated hippo campal neurotransmitter release sites, derived from APP-mutant mice and controls, and determined the proteome by mass spectrometry. Relative protein abundance was measured in APP knockout(APP-KO) experiments and conditional double knockout experiments of APP and the amyloid protein-like protein APLP2 (NexCre-cDKO). A protein-protein interaction (PPI) network was constructed using protein interaction data. The nodes of the networks were the proteins in the PAZ, and the undirected and unweighted edges were experimentally validated physical interactions. We studied various node centralities, topological properties, and the hierarchical organization of the network. Clusters of up- and down regulated proteins in APP-KO mice and NexCre-cDKO mice were identified and analyzed using functional enrichment. Methods Isolation and identification of proteins in the PAZ The experimental setup for sub cellular isolation and immune purification of the hippocampal PAZ was essentially performed as described recently [6] . To compare the proteomic inventory of the PAZ between APP-mutants and controls, samples were subjected to an isobaric labeling (TMT, tandem mass tag) and mass spectrometry. Collection of interaction data PPI data were extracted from protein interaction databases released in February 2015. We decided to collect interaction data as complete as possible from a large set of databases, because the relevance of individual databases for the interactome of the PAZ was unclear in the initial phase of the project. The following databases were used: DIP (Databases of Interacting Proteins) [7] , Int Act [8] , MINT (Molecular Interaction database) [9] , BioGRID (The Biological General Repository for Interaction Datasets) [10] , mentha (mentha: the interactome browser) [11] , iRefIndex [12] , and STRING (Search Tool for the Retrieval of Interacting Genes/Proteins) [13] . Data were stored in the HUPO PSI-MI Tab 2.5 format. Only interactions where both proteins were identified as members of the PAZ were kept for further analysis. UniProt accession numbers were used as unique identifiers for the proteins in the network. Network development We applied the open-source software platform Cytoscape 3.2.1 [14] to create a PPI network. After importing our data collection into Cytoscape, we merged the information into a single network. We applied filters to eliminate all non-physical interactions and those, which were not experimentally validated. The first filter checked the experimental detection method, e.g., yeast two- hybrid, coimmunoprecipitation, and affinity chromatography, among others. The second filter checked the interaction types, e.g., physical association and direct association. The process of filtering reduced the number of 40,796 interactions and 1,457 proteins in our initial collection to 4,990 interactions and 863 proteins. Based on expert knowledge, we identified the largest connected component that was relevant to represent the PPI network of the PAZ. We eliminated singletons, smaller components without connection to the largest connected component, self-loops, and multi-edges. The final network described 3,390 interactions and 615 proteins. The applied filters reduced the network significantly in size and eliminated the drawbacks of the different annotations of the databases. Additional information on proteins as, e.g., localization and function, were collected from UniProt and mapped manually onto the proteins. Finally, our relative abundance data of proteins in knockout mice (APP-KO and NexCre- cDKO) were mapped to the proteins. Network analysis We utilized Cytoscape to generate all visualizations and layouts, analyze the topology, and compute centrality values for all nodes. Centralities of interest were node degree, betweeness, and clustering coefficient. Community detection was performed for our unweighted and undirected network using Radatools 3.2 [15] . Due to its NP-completeness, the computation of the optimal partition into communities is not feasible, so we applied heuristics. A combination of tabu search [16] , extremal optimization [17] , fast heuristics [18] , and spectral optimization [19] was used. A widely accepted criterion for the partition of a network into communities is the optimization of a modularity measure, i.e., the Q-parameter. A Q-value of zero corresponds to a random network without any structure and complex organization. Values of the Q-parameter larger than zero correspond to a well-defined community structure. The theoretical upper bound of Q = 1 is reachable only in the limit of an infinite large network. Functional enrichment analysis for each community was done online using the PANTHER (Protein Analysis Through Evolutionary Relationships) Classification System [20] . Results and Discussion Network (Figure 1A) depicts the final interaction network of the PAZ. Each node denotes one of the 615 proteins, and each edge denotes one of the 3 390 experimentally validated physical interactions. To enable an easy visual comparison, we chose an identical layout for all representations of the network, i.e., the organic layout implemented in Cytoscape. The network is connected and has solely undirected and unweighted edges. It is sparse, with a density of 1.8% of all possible edges. The network diameter was 9, i.e., the maximum length between any two nodes was 9, with an average shortest path length of 3.367, and a network radius, i.e., a minimum eccentricity, of 5. Compared to other PPI networks [21] , the network had a high heterogeneity value of 1.437. The high heterogeneity value of the network was a strong indicator for the existence of hub nodes, large cliques, and a hierarchical organization into highly connected communities. The average number of interaction partners of a protein is 11.024. While most of the proteins have a rather small number of 1 to 5 interactions, a few proteins are highly connected. APP is incorporated into the PAZ interaction network with 75 (9%) interactions. The organic layout algorithm of Cytoscape arranges nodes of the interaction network into a large and highly connected cluster located in the left upper part of (Figure 1A), and nodes that appear to be placed more homogeneously in the layout. The large cluster contains around 54% of all edges. (Figure 1B) gives information about the localization of the proteins. The proteins are grouped into 188 mitochondrial (red), 113 signaling cascade (green), 108 membrane (blue), 81 synaptic vesicle (turquoise), 79 cytoskeletal (yellow), and 46 cytoplasmic (gold) proteins. Interestingly, the organic layout placed red nodes, i.e., the mitochondrial proteins, preferentially together into the large cluster located in the upper left part of (Figure 1B). Grouping mitochondrial proteins preferentially into one cluster by a standard layout algorithm already suggests an underlying topological network structure, that correlates with the localization of proteins in the PAZ. The mitochondrial proteins appear in a dense cluster and interactions within this cluster are more likely. The non homogeneous distribution of colors in (Figure 1B) reflects an additional more complex structure that, however, is hard to capture by visual inspection. Figure 1: The PPI network of the hippocampal PAZ , 615 proteins, 3,390 interactions. (A) Basic view: green nodes represent proteins, interactions are experimentally reported and physical. (B) Localization: nodes are colored according to the localization of the proteins: 188 mitochondrial (red), 113 signaling cascade (green), 108 plasma membrane (blue), 81 synaptic vesicle (turquoise), 79 cytoskeletal (yellow), and 46 cytoplasmic (gold) proteins. (Figure 2A) gives information about the number of interaction partners, i.e., the node degree of individual nodes. Thefigure shows the node degree by the color and the size of the nodes. A small, green-filled circle indicates the lowest node degree of 1. The gradual color change from green over yellow to red indicates an increasing node degree. Simultaneously the size of the node grows. Nodes with high node degree are distributed over the layout, and may play either the role of a center of a cluster or the role of a communicator between functional subunits. Figure 2: Hubs and bottlenecks in the PPI network of the hippocampal PAZ. (A) Nodes are colored and sized by degree centrality; small size and green color indicate a low degree, yellow color indicates the average number of neighbors 11.024, big size and red color indicate a high degree. (B) Nodes are colored and sized by shortest path betweenness centrality; small size and green color indicate a low betweenness, big size and red color indicate a high betweenness. (Figure 2B) gives information about the betweenness of the individual nodes. Color and size code for the betweenness values are coded similar to (Figure 2A). The large red nodes of high degree in the large cluster of mainly mitochondrial proteins, that are visible in the left upper part of (Figure 2A), appear smaller and less red in the coding for betweenness shown in (Figure 2B). Most likely, these proteins play a central role solely within a functional subnetwork. Proteins with a high number of interaction partners and large betweenness value can be identified in the 2(B). Such proteins as, e.g., the 14-3-3 / adapter protein, APP, and the catalytic subunit A of the vesicle-type proton ATPase are bottlenecks. Our identification of bottleneck proteins by high number of interaction partners and high betweenness is supported also by low clustering coefficients of these proteins (data not shown). (Figure 3A) depicts the node degree distribution, f(k), i.e., the number of proteins that have k interaction partners, which is linear in the log-log plot. We applied the Levenberg-Marquardt algorithm, implemented in gnu plot, to fit a scale-free distribution f(k) ~ k to the data points, and obtained a degree exponent of = 1.286 0.065. Within stochastic fluctuations the scale-free distribution fits the data well. The low value of 0.154 for the centralization of the network excluded an easy explanation of hubs by star-like structures [21] . An additional indicator for a complex topological organization of the PAZ was the high average clustering coefficient of 0.331. Such a high value indicates a small-world topology. (Figure 3B) shows the mean cluster coefficient, C(k), for proteins that have k interaction partners. For the majority of k values, the mean clustering coefficients were in accordance with the overall value of 0.331. The high mean clustering coefficient, for nearly all k, hints at a stable organization of proteins into dense connected subnetworks. Obvious exceptions are the data points in (Figure 3), for which the cluster coefficient drops down to the order of the network density of 1.8%. The data points that exhibit a low cluster coefficient denote the following proteins: 14-3-3 / adapter, APP, and the catalytic subunit A of the vesicle-type proton ATPase. The low cluster coefficient, in combination with a high number of interaction partners, favors the role of a communicator between functional subunits for these nodes. Figure 3: (A) The node degree distribution appears to be linear in a log-log plot. A scale-free distribution f(k) ~ k with = 1.28554 (straight line) fits the asymptotic behavior well. (B) The log-log plot shows the mean cluster coefficient, C(k), of proteins that have k interaction partners. Two horizontal lines show the values of the overall cluster coefficient and the network density, respectively. The 14-3-3 / adapter protein (k = 72), APP (k = 71), and the catalytic subunit A of the vesicle-type proton ATPase (k = 69) have high numbers of interaction partners, and low cluster coefficients in the interaction network of the PAZ. Plots were created with gnuplot 6.4. Functional composition The distribution of centralities in the interaction network of PAZ proteins nominates specific proteins for functional roles like bottleneck communication or center of a subnetwork. The centralities are, however, insufficient to reveal the organizational structure of the interaction network. Applying the combination of several algorithms offered by Radatools (see Material and Methods), we yielded a modularity measure of Q = 0.530853 for a partition into 7 communities. The high value of Q = 0.530853 is sufficient to prove the existence of a highly organized topological structure in the interaction network of the PAZ. We applied the PANTHER tool for a GO enrichment analysis for each of the modules. Examples of enriched GO-terms in the individual modules were e.g., synaptic transmission and cellular component organization (275 proteins), protein/ vesicle-mediated transport and receptor-mediated endocytosis, (96 proteins), and high number of energy metabolism-associated proteins mainly localized to the mitochondrion. Due to space restrictions, we abstain to discuss the functional composition of the modules in detail. Examples of assigned biological processes for each community are given in (Table 1). It is worth mentioning that APP and its mammalian homologues, APLP1 and APLP2, are members of the largest community, which contains 275 proteins. Table 1: The communities of the interaction network of the PAZ. The first column denotes the number of the community, size is the number of nodes in the community, the third row gives examples of associated PANTHER GO-Slim biological process terms. Community Size GO-Slim Biological Process 1 275 cell communication, vesicle-mediated transport, cellular component organization, synaptic transmission 2 96 localization, transport, endocytosis 3 82 oxidative phosphorylation, respiratory electron chain, generation of precursor metabolites and energy 4 71 carbohydrate metabolic process, glycolysis, tricarboxic acid cycle, cellular amino acid process 5 51 fatty acid metabolic process, fatty acid beta-oxidation, lipid metabolic process 6 35 protein folding, protein complex assembly and biogenesis, response to stress 7 5 cytokinesis, mitosis, cell cycle APP Knockout and Conditional double knockout: The question of how the deletion of specific proteins alters the functional organization at the PAZ is of utmost importance to understand the development of AD. We mapped our experimental data of relative protein abundance in APP knockout mice (APP-KO), and in mice with a conditional double knockout of APP and APLP2 (NexCre-cDKO), to the network. (Figure 4A) shows the degree of up regulation or down regulation of each protein in APP-KO. Green nodes denote down regulated proteins. The color changes gradually over yellow (no effect) to magenta for maximal up regulated proteins. The size of each node correlates with the absolute value of alteration; proteins that are maximally up- or down regulated are visually highlighted by their big sizes, whereas small yellow nodes denote no alteration. A visual inspection of (Figure 4A) shows a major impact of the knockout of APP. A high number of down regulated proteins, i.e., green nodes, are distributed over the entire layout. Consequently, a knockout of APP has no specific local effect but an impact on several functional subunits in the interaction network of the PAZ. (Figure 4B) shows the degree of upregulation or downregulation of each protein in NexCre-cDKO. (Figure 4B) applies a color and size code for the up-regulation or down- regulation similar to (Figure 4A). Only a small subset of proteins of the total were up-regulation or down-regulated. Please note that, only a subset of 615 proteins from the complete set of proteins that were identified by mass spectrometry and considered for normalization were displayed in figure 4A. The higher degree of down-regulation results from the selection of proteins which play a role in the PAZ, i.e., the largest connected component of the network. Applying a different normalization in a non-standard way, e.g., by normalizing the protein ratios to the median of the selected subset may result in an equal number of up-regulated and down-regulated proteins, but would not reflect the real regulatory effects observed in the knockout experiments. Interestingly, the double knockout leads, compared to a single APP-KO, to an increased number of up-regulated proteins, i.e., large pink nodes. Most of the up-regulated proteins are located in the upper left part of the layout, where a large number of highly connected mitochondrial proteins cluster together (supplementary figure 1). We postpone a detailed discussion of the effect of APP knockout and conditional double knockout of APP and APLP2 on the functional modules, bottleneck proteins, and cluster centers to forthcoming work. Figure 4: The distribution of up- and downregulated proteins in the PPI network of the hippocampal PAZ. (A) Up- and downregulated proteins in APP-KO. Upregulated proteins are more concentrated in the dense cluster of mitochondrial proteins in the upper part of the layout, downregulated proteins are distributed across the network. (B) Up- and downregulated proteins in NexCre-cDKO. Most of the upregulated proteins are clustered in the dense zone were mitochondrial proteins are located, the downregulated proteins appear all over the network. Discussion Within this study, we presented a PPI network of the hippocampal PAZ, consisting of proteins that were significantly identified for all animals analyzed. We combined state-of-the-art proteomics and bioinformatics tools to identify the core proteome, and highlighted the important physiological role of APP within the PAZ. Identifying the core proteome was a prerequisite to construct the PPI network. The filtering process reduced the network size significantly and unified the annotations of the data bases. The network turned out to have topological properties as scale-freeness and small-world that are well known for other PPI. Within the network, we can identify important proteins that play key roles like hubs, bottlenecks, and cluster centers. Standard algorithms, based on a modularity measure, i.e., the Q-value, were capable to partition the network into biological meaningful communities. Drawbacks of algorithms based on the Q-value are their resolution limit and the unique assignment of each protein to exactly one community. The interactions of proteins at the PAZ most likely reflect a complex hierarchical organization of functions. Key proteins influence a broad range of functional subunits. Consequently, the applied algorithms of community detection might have limited ability to resolve the functional organization of the PAZ. These limits drive the development of alternative concepts [22] . Our motivation is to address the question which functional role APP is executing at the PAZ. Deletion studies on members of the amyloid protein family, i.e., APP and its mammalian homologues APLP1 and APLP2, could be directly correlated to pronounced alterations in protein abundance of numerous PAZ constituents [4,5] . Interestingly, lack of APP induces an age-dependent phenotype. While younger mice reveal only a mild phenotype, e.g., reduced body weight and grip strength, this condition is changed in the elderly mice, e.g. impairments in learning and memory and reduced dendritic spine density. Since changes at protein level already occurred in younger adults, these changes point to a yet unknown link incorporating APP into the physiological network of the PAZ. Consistent with our data, loss of APLP2 increases the impact of APP deletion at the hippocampal PAZ. Impairments in the induction and maintenance of long term potentiation (LTP) in hippocampal neurons, paralleled by reduced performance in memory-related and behavioral tasks, could be attributed only to aged APP-KO mice. This points to compensatory mechanisms, mainly by APLP2 [23] . Over long periods, the impact of APP/APLP2 deletion on neuronal physiology was complicated by postnatal lethality of DKO mice. However, the design of a conditional APP/APLP2 double knockout in excitatory forebrain neurons under the Nex Cre promotor (Nex- Cre-cDKO) circumvented postnatal lethality [24] . These mice display severe impairments in memory-related and behavioral tasks associated with reduced LTP already in younger animals [24] . Recently, APP was suggested as a G-protein coupled receptor (GPCR) mediating presynaptic signaling and neurotransmitter release by activation of Ca 2+ -channels [25] . This is in line with the down-regulation of G-protein subunits G(o)/G(i) and Ca 2+ -binding proteins at neurotransmitter release sites in both mutants. Furthermore, dysregulation of Ca 2+ -mediated pathways points to impairments in synaptic plasticity at the hippocampal PAZ and may account for the reported deficits for NexCre-cDKO [24] . Moreover, -synuclein, an important mediator of coordinated neurotransmitter release, was severely decreased at the presynaptic active zone derived from NexCre-cDKO. -synuclein was unraveled as a promotor for the assembly of SNARE proteins into the SNARE-complex during fusion [26] . The analysis of synuclein triple knockout mice demonstrated that SNARE-complex assembly was drastically decreased, despite the fact that the basal SNARE-complex protein levels for the proteins VAMP2, syntaxin-1 and SNAP25 were unaltered [26] . This was in accordance with our data, which showed no alterations in protein abundance for these three proteins. Since the LTP-deficits have occurred only in the aged APP-KO mice, a compensatory effect mediated by APLP2 has been suggested [23] . Consistent with our data, a pronounced dysregulation of PAZ proteins involved in transmission, plasticity and signaling, the loss of APLP2 increases the impact of APP deletion at the presynaptic active zone. Conclusion Our findings of alterations in the proteomic composition in both APP-mutants identified APP as a structural and functional mediator within the hippocampal active zone network. Future studies have to unravel how deletion of APP affects the spatiotemporal organization of the interactome of the PAZ and leads to impairments in learning and memory. Alterations induced by knockout of APP may provide a molecular link to the pathogenesis of Alzheimer s disease and open new strategies to identify therapeutic targets. In this context, molecular network studies can help to assess potential side effects of targeting APP in pharmaceutical approaches by concepts of systems biology. However, a PPI network generates a rather static view on the complex variety of actions, reactions, and interactions that take place at the PAZ. Integrating temporal data might enable insights into the dynamic behavior of the constituents of the PAZ. Here, especially the identification of specific pathways altered by gene knockouts would give valuable information about the physiological function of APP. In summary, unraveling the dynamics of the molecular functions of APP at neurotransmitter release sites will provide novel therapeutic strategies targeting Alzheimer s disease. Acknowledgment We are grateful to Michael Karas, Benjamin M ller, Marion Rohmer and Tabiwang Array for performing mass spectrometry analysis. We thank Ulrike M ller and Meike Hick for providing mutant mice. Conflict of interest The authors have declared no conflicts of interest. References 1. Leuner, K., Muller, W.E., Reichert, A.S. From mitochondrial dysfunction to amyloid beta formation: novel insights into the pathogenesis of Alzheimers disease. (2012) Mol Neurobiol 46(1): 186-193. 2. Glenner, G.G., Wong, C.W. Alzheimers disease and Downs syndrome: sharing of a unique cerebrovascular amyloid fibril protein. (1984) Biochem Biophys Res Commun 122(3): 1131-1135. 3. Jacobsen, K.T., Iverfeldt, K. Amyloid precursor protein and its homologues: a family of proteolysis-dependent receptors. (2009) Cell Mol Life Sci 66(14): 2299-2318. 4. Lasek, M., Weingarten, J., Einsfelder, U., et al. Amyloid precursor proteins are constituents of the presynaptic active zone. (2013) J Neurochem 127(1): 48-56. 5. Lasek, M., Weingarten, J., Acker-Palmer, A., et al. Amyloid precursor protein knockout diminishes synaptic vesicle proteins at the presynaptic active zone in mouse brain. (2014) Curr Alzheimer Res 11(10): 971-980. 6. Weingarten, J., Lasek, M., Muller, B.F., et al. Regional specializations of the PAZ proteomes derived from mouse hippocampus, olfactory bulb and cerebellum. (2015) Proteomes 3(2): 74-88. 7. Xenarios, I., Rice, D.W., Salwinski, L., et al. DIP: the database of interacting proteins. (2000) Nucleic Acids Res 28(1): 289 291. 8. Hermjakob, H., Montecchi-Palazzi, L., Lewington, C., et al. IntAct: an open source molecular interaction database. (2004) Nucleic Acids Res 32: D452-D455. 9. Zanzoni, A., Montecchi-Palazzi, L., Quondam, M., et al. MINT: a Molecular interaction database. (2002) FEBS Lett 513(1): 135-140. 10. Stark, C., Breitkreutz, B.J., Reguly, T., et al. BioGRID: a general repository for interaction datasets. (2006) Nucleic Acids Res 34: D535-D539. 11. Calderone, A., Castagnoli, L., Cesareni, G. Mentha: a resource for browsing integrated protein-interaction networks. (2013) Nat Methods 10(8): 690-691. 12. Razick, S., Magklaras, G., Donaldson, I.M. iRefIndex: a consolidated protein interaction database with provenance. (2008) BMC Bioinformatics 9(1): 405. 13. Franceschini, A., Szklarczyk, D., Frankild, S., et al. STRING v9. 1: protein-protein interaction networks, with increased coverage and integration. (2013) Nucleic Acids Res 91(D1): D808-D815. 14. Shannon, P., Markiel, A., Ozier, O., et al. Cytoscape: a software environment for integrated models of biomolecular interaction networks. (2003) Genome Res 13(11): 2498-2504. 15. Gomez, S., Fernandez, A., Borge-Holthoefer, J., et al. (2011). Radatools 3.2 16. Arenas, A., Fernandez, A., Gomez, S. Analysis of the structure of complex networks at different resolution levels. (2008) New Journal Phys 10(5): 053039 17. Duch, J. Arenas, A. Community detection in complex networks using extremal optimization. (2005) Phys Rev EStat Nonlin Soft Matter Phys 72(2): 027104. 18. Newman, M.E. Fast algorithm for detecting community structure in networks. (2004) Phys Rev E Stat Nonlin Soft Matter Phys 69(6): 066133. 19. Newman, M. E. Modularity and community structure in networks. (2006) Proc Nati Acad of Sci usa 103(23): 8577-8582. 20. Mi, H., Muruganujan, A., Casagrande, J.T., et al. Large-scale gene function analysis with the PANTHER classification system. (2013) Nature Protocols 8(8): 1551-1566. 21. Dong, J., Horvath, S. Understanding network concepts in modules. (2007) BMC Syst Biol 1: 24. 22. Fortunato, S. Community detection in graphs. (2010) Physics Reports, 486(3):75-174. 23. Weyer, S.W., Klevanski, M., Delekate, A., et al. APP and APLP2 are essential at PNS and CNS synapses for transmission, spatial learning and LTP. (2011) EMBO J 30(11): 2266-2280 24. Hick, M., Herrmann, U., Weyer, S.W., et al. Acute function of secreted amyloid precursor protein fragment APPs in synaptic plasticity. (2015) Acta Neuropathol 129(1): 21-37 25. Fogel, H., Frere, S., Segev, O., et al. APP homodimers transduce an amyloid- -mediated increase in release probability at excitatory synapses. (2014) Cell Reports 7(5): 1560-1576. 26. Burre, J., Sharma, M., Tsetsenis, T., et al. -synuclein promotes SNARE-complex assembly in vivo and in vitro . (2010) Science 329(5999): 1663-1667.