Relationships between the amyloid precursor protein and its various proteolytic fragments and neuronal systems

Alzheimer's disease (AD) is a progressive neurodegenerative disease and in its familial form is associated with mutations in the amyloid precursor protein (APP) and the presenilins (PSs). Much data regarding the interactions of APP, its proteolytic fragments and PS have been generated, expanding our understanding of the roles of these proteins in mechanisms underlying cognitive function and revealing many complex relationships with wide ranging cellular systems. In this review, we examine the multiple interactions of APP and its proteolytic fragments with other neuronal systems in terms of feedback loops and use these relationships to build a map. We highlight the complexity involved in the APP proteolytic system and discuss alternative perspectives on the roles of APP and its proteolytic fragments in dynamic processes associated with disease progression in AD. We highlight areas where data are missing and suggest potential confounding factors. We suggest that a systems biology approach enhances representations of the data and may be more useful in modelling both normal cognition and disease processes.

Other models suggest diff erent initiating factors, such as cholesterol [14], mitochondria and apoptosis [15], insulin signalling and energy metabolism [16,17], cholinergic insuffi ciency [18], Ca 2+ deregulation [19,20], signalling decoherence [21] or a combination of multiple factors (for example, [22,23]). No single model adequately explains how APP, PS, cholesterol and other factors interact with synaptic plasticity, learning and memory in normal and disease states. Th is suggests that models currently in use in AD research require re-evaluation.
Recent approaches to integrating the data relevant to AD have involved the use of text mining, bioinformatic databases and network analysis for gene co-expression and protein-protein interactions [24][25][26][27][28][29][30][31][32][33]. Studies using automated interrogation of genomic and proteomic databases, such as UNIPROTKB [34], NCBI Entrez Gene [35], and IntAct [36], and text mining in literature databases such as Pubmed [37] have confi rmed the involvement of many previously highlighted neurological processes, extended the connectivity between known molecular targets and revealed new targets for investigation (summarised in Table 1).
Understanding how APP contributes to synaptic plasticity is important in understanding normal cognition and disease progression in AD. Here, we approach the process of integrating the available data relating to the relationships between APP and its proteolytic fragments and neuronal processes based on a systems biology approach and build a map of relationships. We are not aiming to defi ne protein interactions nor fi nd new targets for investigation, but to highlight relationships in the existing literature that illustrate the complexity involved and how this relates to our current understanding of the role of APP in AD. We highlight areas where data are missing and predict relationships that may confound current disease models.

Building the map
Relationships involving APP and each of its proteolytic fragments were identifi ed from literature searches using relevant key words. Each interaction identifi ed was manually entered into a non-hierarchical network using the mapping software Compendium [38]. Th is software allowed each molecule or neuronal system to be represented as a node within the map. Each node in the map could contain further networks of nodes, leading to a multi-dimensional organisation of information. Th ese further nodes represented information relevant to the molecule of the parent node and could include notes describing possible relevance, questions raised and links to relevant entries on various bioinformatic databases. Th e map is not meant to be read as a protein-protein interaction network (PPI), but rather was designed to summarise and collate information relevant to the APP system in a more general way. Physiological outcomes, such as changes in signalling, were included, even though the precise molecular interactions are not clear. Any additional details, such as new biological relationships or new molecular interactions, can be added as they arise. Relationships between molecules are represented by connectors, with annotation if required. Figure 1 shows a limited map built in this way from references cited in the text, with additional references listed in Table 2.
Unlike more automated network studies such as that by Perreau and colleagues [25] where nodes were included in the network only if there was direct evidence of molecular interaction, or various studies [30][31][32][33] where co-expression criteria were applied, no specifi c criteria were imposed and any relationship found was included. Known diff erences in peptide behaviour (for example, [39,40]) necessitated separate nodes for Aβ  and Aβ  . Th is was extended systematically to include all the peptide fragments, including P3  , P3  , secreted APP (sAPP)α and sAPPβ (the extracellular aminoterminal fragments following α-and β-cleavage, respectively), C31, and so on to take account of any potential behavioural diff erences. Th is is in contrast to various network constructions where the various proteolytic fragments are collapsed into an APP parental node, possibly refl ecting the UNIPROTKB identifi er (P05067) being the same for each peptide. Additionally, diff erent aggregation states of each of the peptides (monomers, dimers, oligomers or fi brils) were considered separately. Th ese additional details have been omitted from Figure 1 for clarity. Since there was little or no evidence for many of the Aβ-type peptides [41], these were collapsed into a separate node (not shown).
Th e relationships in Figure 1 were analysed manually by identifying feedback loops. Direct feedback loops were defi ned as relationships between APP, any of its proteolytic fragments and any other molecule or functional change that was represented at more than one point in the map. Possible indirect feedback loops between any neural system, such as neurotransmitter signalling, Ca 2+ homeostasis or the infl ammation cascade, were also investi gated by looking for molecules or functional change in these systems that interact with APP or any of its proteolytic fragments at more than one point in the map.
Following a brief summary of APP cleavage, two examples of the relationships between APP pathways and neuronal processes, looking at the extracellular matrix (ECM) and endocytosis, serve to illustrate the implications of the complexity involved.

The amyloid precursor protein
Overview APP is a type I transmembrane protein [42] with a large extracellular amino-terminal domain, a transmembrane domain and an intracellular carboxy-terminal domain that resembles a cell surface receptor [43]. It is expressed in several isoforms, with APP 695 the main isoform expressed by neurons. APP has a high turnover, with a half life ranging from approximately 1 hour [44][45][46] to approximately 4 hours [47][48][49]. In addition to its functions at the cell surface, full length mature APP is also processed via competing pathways that release proteolytic fragments [5,43,50] (recently reviewed in [51]). In eff ect there are four possible routes: APP may remain functionally active at the cell surface; it may be internalised and recycled or degraded; it may be cleaved via the α-pathway, releasing the sAPPα fragment, leaving the membrane bound C83 carboxy-terminal fragment that is further processed to the P3 fragment; or it may be cleaved via the β-pathway releasing sAPPβ, leaving the membrane bound C99 carboxy-terminal fragment that is further processed to the Aβ fragment.
As can be seen in Figure 1, α-and β-cleavage pathways (numbered 1 and 2, respectively) converge on a further intramembrane γ-cleavage (numbered 3) by the PScontaining γ-secretase complex to release the variable length Aβ and P3 peptides and the variable length APP intracellular domain (AICD). Additional cleavages, includ ing caspase cleavage of the carboxyl terminus at residue D664 of APP 695 , producing the alternative length carboxy-terminal fragment C31 [51], and cleavage by BACE2 (beta-site amyloid precursor protein cleaving enzyme 2) within the Aβ sequence [52], have been omitted for clarity.
Unprocessed APP is degraded or recycled via the endosomal or lysosomal pathways and may be recycled back to the membrane and processed within approximately 30 minutes [46], with perhaps a third to a half being processed via the cleavage pathways as measured by secreted sAPPα/β [46]. Regulatory relationships  [25] Evidence of direct interaction between molecules from literature searches Metals, non-protein molecules, poorly characterised proteins, specifi c peptides are included as parent genes Direct PPI involving APP and associated fragments by domain with reference to APP770 isoform but with application to all isoforms; molecular networks with reference to biological processes [33] Twelve causative or susceptibility 'seed' genes previously associated with AD; candidate genes were selected due to close proximity to GWAS-identifi ed gene loci associated with AD Proteins lacking open reading frames; transcription factors, highly glycosylated proteins, extracellular proteins, proteins containing several transmembrane regions; coexpression data were used as a fi lter Identifi cation of genes in AD with reference to direct PPI and biological processes [31] Co-expressed genes that diff er between controls and AD Probe-sets not mapping to any gene or mapping to hypothetical proteins were removed Variations in transcriptomes of AD suggest similarities with molecular networks associated with CVD and diabetes. Cis-regulatory elements identifi ed in several diseases known to co-occur with AD [30] Genes with variable expression between an AD and an aging microarray study Arrays that signifi cantly varied between subjects in each study group; outlier removal, unreliable probe sets defi ned as being present in three or fewer arrays; control probe sets and probes not associated with known genes Transcriptional changes between AD and aging highlight possible contributors to disease pathways. Many biological processes are shared between AD and aging. Many novel associations found, including MAPK pathways and unknown proteins. New functional and disease related association for PSEN1 with glial/neuronal interactions; confi rms and highlights gamma14.3.3 signalling in AD [32] Genes with variable expression between human and mouse from brain sample microarray gene expression data sets Outlier removal, fi ltered (method not specifi ed) to remove datasets with low interspecies expression or low connectivity correlations; only top 5,000 Human and 3,000 Mouse connected genes included, rest removed to reduce 'noise' Mouse and human networks are highly similar with expression levels more preserved than connectivity. Signifi cant species diff erence in: i) co-expression arrays from astroglia and microglia but not neurons; ii) the role of PSEN1 in oligodendrocytes and myelination; and iii) further evidence of species diff erence in glial cells linked to neuroinfl ammation in human AD. Neuronal death is a small part of the biological changes associated with AD in this dataset; new transcription factors associated with AD [27] All AD and related pathways in KEGG database with co-expression data Genes without corresponding data in the assembled PPI network and proteins with no corresponding genes in the co-expression data Crosstalk between pathways involved in AD; close relationships between APP and apoptosis, Notch, Wnt pathways and cytokine-cytokine interactions in this dataset, brain areas vary in specifi c pathway relationships and order of signifi cance promoting expression or cleavage have been collapsed into summary boxes in Figure 1 for simplicity.

Complex feedback pathways involving APP and its proteolytic fragments; the extracellular matrix and endocytosis
The extracellular matrix APP interacts with itself and its own proteolytic fragments at many points in the network with various functional consequences. At the cell surface, the large amino-terminal domain of full length APP interacts with many components of the ECM, including heparins and heparin sulphate proteoglycans [42], laminin, collagen [53,54] and β-1-integrin [43,55,56], and contributes to early neuronal development, axon and neurite growth, cell adhesion and synaptic plasticity [43,57]. Many of these interactions are now represented in MatrixDB [58], an ECM interaction database. Th e amino-terminal domain of APP binds heparin in a pH-and zinc-dependant manner and this binding promotes the formation of APP homodimers [42,[59][60][61]. Depending on various factors, including cell compartment and pH, this dimerisation may infl uence APP proteolysis via α-or β-cleavages due to confor mational changes [59], though this requires further investigation.

Details Reference
Akt Aβ increases Akt phosphorylation in the short term via a mechanism involving α7-nAChR and NMDARs, with [99] phosphorylation levels returning to baseline over the long term AMPA receptors AMPA glutamate receptor density is reduced by Aβ oligomers via reduction of CamKII [100,101] ApoE Decreased levels of ApoE lead to increased β-cleavage [102] ApoE promotes polymerisation of Aβ into fi brils and enhances fi brillar Aβ deposition in neuritic plaques.
[103-105] The high affi nity binding of Aβ to ApoE reduces ability of ApoE to bind lipids Ca 2+ sAPPα modulates Ca 2+ signalling by activating high conductance K + channels via a mechanism dependent on cGMP [106] CD74 Interacts with APP and reduces expression of Aβ [107] Cholesterol Interactions of cholesterol and APP may allow APP to react to cholesterol status of the cell [108] Membrane cholesterol correlates with β-secretase activity and inhibition of β-secretase activity leads to [109,110] increased membrane cholesterol levels. Moderate reductions in cholesterol enhance the co-expression of APP and BACE1 and promote the production of Aβ Aβ binds lipids and has high affi nity for cholesterol. Aggregated Aβ  may aff ect lipid transport. [111,112] Aβ binds 24-hydroxycholesterol and aff ects membrane choline carriers

Complement cascade
Aβ activates neuronal complement cascade to induce the membrane attack complex and reduces [113] complement regulatory proteins, increasing complement-mediated cytotoxicity

Furin
Furin enhances cleavage to active forms of ADAM10 and ADAM17, leading to enhanced α-cleavage [122] Glucose/glutamate sAPPα enhances transport of glucose and glutamate in synapses and protects from oxidative stress via a [123] transport mechanism involving cGMP Glutamate signalling sAPPα suppresses NMDA currents rapidly and reversibly at concentrations of approximately 0.011 nM, possibly [124,125] involving cGMP and a protein phosphatase. Reductions in sAPPα lead to reduced tetanically induced NMDA currents while increased sAPPα increased these currents and enhanced LTP G-protein signalling Aβ directly increases TNF-α at high levels and at low levels increases TNF-α release by altering GPCR signalling [126] at early stages of disease progression by indirect eff ects on GPCR kinase 2/5 Full length and processed APP can potentially interact with G proteins via the cytoplasmic tail and this can be [127][128][129] altered by APP mutations around the G protein binding site. This interaction has the potential to alter G-protein signalling with wide ranging eff ects, including Ca 2+ regulation and cell cycle pathways

HDL
Aβ binds ApoA-I, ApoA-II, ApoE and ApoJ; binding modulates Aβ solubility [130] Heparins Proteolysis of immature BACE1 to its mature active form is promoted by low concentrations of heparin and [70,131,132]  Aβ binds focal adhesion molecules and integrins and modulates integrin/FA signalling pathways involved in [138,139] cell cycle activation and cell death. The αv integrin subunit is required for Aβ-associated suppression of LTP

Lipids
Binding of Aβ to acidic lipid molecules promotes Aβ aggregation. Aβ binds membrane gangliosides, [140][141][142] sphingolipids and cholesterol, which enhance Aβ aggregation Association of APP and BACE1 with lipid rafts increases Aβ [143] Aβ endocytosis may also involve lipid rafts [144] Sphingolipids enhance α-cleavage via MAPK/ERK signalling [145] Continued overleaf its neuroprotective actions [62][63][64] and may partly explain why sAPPα is approximately 100 times more neuroprotective against excitotoxicity, glucose deprivation and the addition of Aβ in hippocampal cultures than sAPPβ, which lacks the second carboxy-terminal heparin binding site [63]. Dementia status has been associated with both reduced sAPPα levels in cerebrospinal fl uid [65] and an increased half life of sAPPα [49]. Diff ering by only 16 carboxy-terminal residues, sAPPα and sAPPβ share a high degree of sequence homology and have the potential to compete with each other for binding sites. Since both sAPPβ and sAPPα are constantly present in the ECM in dynamically changing ratios, we predict that sAPPβ could act as an antagonist of sAPPα/ APP (shown in Figure 2 as competitive inhibition (line 1)), with the potential to modulate various interactions between APP and sAPPα and the ECM. We predict, therefore, that any functional outcome from the interactions of the APP pathway will depend on the synergy between APP and its amino-terminal fragments. Neurite outgrowth APP enhances neurite outgrowth independently from sAPPα [155] NF-κB May reduce expression of BACE1 [156] NF-κB upregulation by capacitive Ca 2+ entry enhances sAPPα release via mAChR signalling [157] Nicotinic ACh signalling Aβ  and Aβ  reduced Α4β2 nAChR and α7 nAChR currents. Aβ  but not Aβ  increased glutamatergic [40,158] AMPA. Signalling via Α4β2 nAChR is associated with reduced Aβ Aβ has high affi nity for the α7 nAChR and this may be associated with increased Aβ accumulation. Diff erential [39,159,160] eff ects of Aβ  and Aβ  on α7 nAChR as seen by diff erent eff ects on ACh release and Ca 2+ infl ux. Disruption of signalling by α7 nAChR may be associated with Aβ-mediated increases in pre-synaptic Ca 2+ Aβ  has approximately 5,000-fold greater affi nity for α7 nAChR than for Α4β2 nAChR [161] NMDA GluR Aβ promotes endocytosis of NMDARs in cortical neurons with the involvement of protein phosphatase 2B [162]  Increases in TIMP-3 led to decreased surface expression of ADAM10 and APP. The production of Aβ and CTF [171] is increased. TIMP-3 appears to promote endocytosis and β-secretase cleavage
Hunter and Brayne Alzheimer's Research & Therapy 2012, 4:10 http://alzres.com/content/4/2/10 Aβ interacts with heparins in the ECM and at high levels may prevent the catabolism of proteoglycans and promote amyloid formation [66]. Reciprocally, heparins modulate many of the interactions involving Aβ. Heparins enhance both nucleation and elongation processes in the aggregation of Aβ [67] and modulate interactions of Aβ; for example, heparins can limit the neurotoxic and proinfl ammatory activity of Aβ in a dose-dependent manner [68]. Additionally, heparins contribute to the uptake of Aβ by a pathway shared with apolipoprotein E [69] and also contribute to the regulation of the β-secretase BACE in a dose-dependent manner, with low concentrations promoting and high concentrations inhibiting the activation of BACE1 [70].
In terms of direct feedback, Aβ may be part of a negative feedback loop involved in the regulation of APP expression and the promotion of α-cleavage. Fibrillar forms of various proteins, including Aβ, interact with amino-terminal APP and increase APP expression [71] ( Figure 2, line 2) and Aβ  may up-regulate expression of α-secretase via integrin receptors and matrix metalloproteinase-9 activity [72] (Figure 2, line 3). Additionally, Aβ has many indirect negative feedback relationships with α-cleavage -for example, Aβ inhibits protein kinase C [73] and insulin signalling [74], both involved in upregulation of α-cleavage [75] (Figure 2, line 5) -and additionally may negatively feedback on APP expression and its own production via a pathway involving Fe65 and glycogen synthase kinase 3 [76]. Reciprocally, changes to gene expression by sAPPα via the activation of NF-κβ may attenuate the interactions of Aβ [77] and this can be understood as reciprocal indirect feedback from the αpathway on the β-pathway (Figure 2, line 4).
Th e data reviewed above suggest that Aβ may contribute to both direct and indirect feedback loops in the ECM at multiple points, with consequences for the interactions of APP, sAPPα, Aβ and sAPPβ that may ultimately aff ect cell adhesion, neurite outgrowth and synaptic plasticity. Th ese eff ects will depend on the specifi c sequence length, concentration and aggregation state of the Aβ-type peptides, which may dynamically change in response to the neuronal environment.

Endocytosis
Increases in endocytosis, often associated with synaptic activity, lead to both an increase in Aβ [78] and a reduction in α-cleavage and lower secreted sAPPα levels [79]. Th is may refl ect the regulatory role of compart mentation in APP proteolysis, where α-cleavage occurs at the cell surface and Aβ production is associated with factors that promote endocytosis. No studies could be found investigating the role of endocytosis on P3 production, so this cannot be discussed specifi cally. Endocytosis allows a degree of dissociation between the α-and β-cleavages and γ-cleavage, resulting in levels of sAPPα and sAPPβ that may be independent of Aβ (1-40/42) and P3 levels, respectively. Th is means that in terms of function, APP processing via the β-pathway has the potential to either: modulate actions of sAPPα via competition with sAPPβ; or modulate the actions of sAPPα and initiate actions mediated by Aβ.
P3 corresponds to Aβ (17-40/42) and we predict that this shared sequence could allow P3 to modulate the interactions of Aβ (shown in Figure 2 as competitive inhibition), especially where those interactions involve the carboxy-terminal amino acids. While no studies can be found that address this particular question directly, various amino-terminal truncated forms of Aβ have been associated with diff use plaques [80,81] and the Aβ  fragment, corresponding to P3, has been studied in terms of aggregation [82]. It is clear that amino-truncated Aβ peptides, including P3, have potential for aggregation and can interact with Aβ, but how this aff ects the physiological behaviour of neuronal systems is not clear. In terms of functional fl ow through the APP pathway, processing via the α-pathway to release sAPPα and P3 could either: initiate actions of sAPPα; or initiate actions of sAPPα and modulate actions mediated by Aβ.
In terms of modelling these interactions in functional contexts, PPI networks may require a more detailed approach, where each sequence length and aggregation state is represented as an individual node.

The eff ects of concentration, sequence length aggregation state and affi nity
Aβ is produced in a range of sequence lengths [41] and can form monomers, dimers, oligomers and fi brils. Th e amount of Aβ produced is regulated and can dynamically change in response to many factors, including increased cholesterol [83], increased synaptic activity [84], heparins via activation of BACE1 [70], reduced acetylcholine (Ach) signalling via muscarinic receptors [70], sustained increased cytosolic Ca 2+ [85], and hypoxia [86]. Changes in concentration are a well recognised mechanism of regulation in cellular processes; for example, Aβ  and Aβ  both promote angiogenesis at nanomolar concentrations but inhibit it at higher micromolar concentrations [87].
Diff erent sequence lengths show diff erent behaviours; for example, Aβ  is more prone to aggregation than Aβ  due to a more rigid carboxyl terminus [88]. Aβ monomers appear to have structured and unstructured regions, with diff erences in aqueous conformations between Aβ  and Aβ   [89]. While both Aβ  and Aβ  reduced currents via α7 and α4β2 ACh receptors (AChRs), only Aβ  increased glutamatergic signalling via AMPA receptors [40]. Another study found that the Figure 2. Feedback between amyloid precursor protein and its proteolytic fragments. Inhibition is shown in red, enhancement in green; numbers are referred to in the text. Functionally diff erent aggregation states for the Aβ-type peptides are collapsed into P3 and Aβ nodes for clarity. Aβ, amyloid beta protein; AICD, APP intracellular domain; APP, amyloid precursor protein; PKC, protein kinase C; sAPP, secreted amyloid precursor protein. eff ects of Aβ  and Aβ  on α7 AChRs, as measured by ACh release and Ca 2+ infl ux, were diff erent [39]. Additionally, Aβ  may protect neurons from the eff ects of Aβ  by disrupting the aggregation of Aβ   [90] and this raises the possibility that P3-type peptides may act in similar ways. Th e functional eff ects of sequence length are only just beginning to be elucidated and should be expanded to include all the Aβ-type peptides, including P3. In addition to diff erences in peptide sequence and concentration, aggregation state also modulates function. Aβ oligomers appear to be spherical [91,92] and Aβ fi brils appear as fi ne fi bres [92]. Aβ oligomers, but not monomers or fi brils, enhance lipid release [93] and at low concentrations enhance neuronal diff erentiation [92] whereas Aβ fi brils, but not monomers or oligomers, may interact with APP and increase its expression [71]. While both soluble and aggregated forms of Aβ  inhibit longterm potentiation, the fi brillar forms may additionally aff ect neurotransmitter release [94]. Within a specifi c cellular compartment, whether one interaction is more likely to occur than another depends on the relative affi nities of the reactants involved. While dissociation constants have been measured for some reactions in rodent models (for example, K(i) values of Aβ  for α7 nicotinic AChRs from rat and guinea pig are 4.1 and 5.0 pM, respectively, and the K(i) of Aβ  with α4β2 AChRs is approximately 5,000-fold lower at 30 and 23 nM, respectively [95]), we are missing dissociation constants for most interactions in humans.
Taken together, the data discussed above suggest that the interactions of APP and its proteolytic fragments with synaptic systems are complex and involve subtle changes in the relative ratios between APP and all its fragments. Given the complexity of the APP system, multiple neuronal systems have the potential to regulate the expression and proteolysis of APP and could also initiate imbalance within this system, leading to multiple possible disease pathways.
Th is complexity has the potential to confound purely empirical studies into the behaviour of the peptides in model systems. Over-expressing Aβ peptides in, for example, mouse models without knowing their concentrations, sequences and aggregation states and the background concentrations, sequences and aggregation states of the other relevant peptides in the experimental system that is being studied means that we do not know if the experimental results obtained are due to the properties of the Aβ peptides themselves or refl ect a disrupted balance in a complex system over time, or perhaps both. While this distinction may seem subtle, given the self-organising and self-referencing properties of the human brain, an approach based on the principles in systems biology may better represent the roles of the APP pathway in functional contexts.

Discussion
It is not possible to include the full extent of the known interaction of APP and its proteolytic fragments here and further interactions may yet be discovered. Data relating to the roles of APP and its proteolytic fragments in synaptic plasticity and AD have accumulated from many diverse fi elds and these need to be placed into context. One way of doing this is to investigate the relationships between APP, its proteolytic fragments and wider neuronal systems by mapping networks of interactions.
It is interesting that the networks generated in the various studies reviewed in Table 1 do not always overlap and diff erent studies highlight diff erent pathways or biological processes, for example, Fe 2+ [28], apoptosis [96], or cardiovascular disease/diabetes [31]. Each study has diff erent starting points, inclusion/exclusion criteria and network construction methods, so this lack of agreement is no surprise. It is diffi cult to assess the degree to which the various starting points, criteria and network construction methods bias results towards an outcome.
Th e study by Soler-Lopez and colleagues [33] may not model the interactions of full length APP in the membrane adequately, as many of the ECM proteins that might be expected to interact are excluded due to diffi culties involved in expressing them in the experimental microarray used. Th is may shift the focus of their network more towards intracellular interactions. Given the importance of the various interactions of APP with components of the ECM, any study excluding such proteins and proteoglycans could be seriously confounded.
While the study by Perreau and colleagues [25] maps PPIs to specifi c domains of full length APP, they do not distinguish between the diff erent proteolytic peptides. Additionally, transient interactions and biologically essential post-translational modifi cations, such as glycosylation and phosphorylation, cannot yet be fully represented in PPI networks [27], severely limiting the modelling of regulation and control in these models.
Studies comparing mouse and human gene expression networks [32] or aging and AD pathways [30] have revealed interesting results. Perhaps most striking is the study revealing a human-specifi c network for PS function in oligodendrocytes and myelination, with the potential to confound the current approach to modelling AD in rodents [32]. Miller and colleagues [30,32] also show important contributions from both neuronal and glial pathways, perhaps refl ecting the contributions of glia to disease pathways found in humans [97,98]. Th e diff erences found in networks between glia and neurons could be extended to diff erentiate between specifi c neuronal types and future network models may usefully investigate why cholinergic neurons in AD appear more susceptible to degeneration than others.
We do not yet have totally inclusive, gene co-expression or PPI models of APP and related pathways to the level of detail required and this places limitations on their interpretation. Given the limitations, however, network studies do show the usefulness of the systems biology approach in integrating huge volumes of data and a simple search of patents in Google with the terms 'protein interaction network Alzheimer disease' retrieves more than 700 applications based on this approach. Network studies highlight the idea that multiple processes are involved in AD disease pathways, reveal new targets for further investigation and have generated new hypotheses. Network models have the potential to avoid biases inherent in other approaches, such as the emphasis on Aβ in animal models based on the more reductionist amyloid cascade hypothesis, and allow multiple disease pathways to be represented in a single model, perhaps better relating to the processes involved in human disease.
We have suggested that APP and its proteolytic fragments appear to dynamically modulate each other by multiple mechanisms with evidence for both direct and indirect feedback loops between Aβ, sAPPα, sAPPβ and full length APP. Additionally, the fragments derived from γ-cleavage, including the variable length fragments, P3 and Aβ, have the potential to interact with each other and modulate multiple neuronal processes in subtle and dynamic ways. Unfortunately, evidence for the role(s) of P3 is almost entirely missing from the literature and that for sAPPβ is limited. Without this information, we cannot have a full picture of the contributions of the APP proteolytic system to normal neuronal processes and how these change during disease progression.
Given the number and range of reciprocal feedback relationships that APP and its proteolytic fragments are involved in, the eff ects of perturbing the APP proteolytic system may be better understood as dynamic, synergistic actions involving multiple players. A wide range of neuronal processes have the potential to interact with and regulate APP processing, allowing APP to monitor the current state of the cell. Th e multiple proteolytic pathways allow APP to assimilate signals from wideranging cellular systems and pass these signals on via the ratios of APP fragments. Th e constant turnover of APP contributes to the ability of neurons to sense the current state of the cell and use this information to set in motion future responses. Th e potential for subtly diff erent signals both from the α-and β-pathways via alternative peptides and the complex feedback relationships between them described above fi ts well with the fi ne control required for integrating neuronal systems.
Th e predicted synergy between APP and all its proteolytic fragments implies that the ratios of full length APP and all the various fragments, not just Aβ  and Aβ  , are more important than absolute levels in coordinating neuronal responses. Th e complexity of the relationships described here has the potential to confound purely empirical investigations into the activities of any one proteolytic fragment studied in isolation. To avoid these predicted confounding eff ects, experimental approaches should be extended to account for the diff erent concentrations, affi nities and aggregation states of each peptide length, and experimental systems should be followed over time for each biological process in a systematic manner. With over 40 Aβ-type peptides [41], this presents a challenge.
Generating useful network models will require a more detailed and comprehensive approach to experimental design, involving a range of experimental controls to account for all the confounding factors. Th ese data will be essential for generating network models that are truly relevant to our understanding of normal and abnormal processes in the context of the human population.

Competing interests
The authors declare that they have no competing interests.