Database architecture

The NeuroDNet database was developed in MySQL and is currently hosted on an APACHE http server located in the computer service centre of the Indian Institute of Technology Delhi. The database was built in a three-tier structure - foundation, function and interface (Figure 1). The foundation tier (ndnGNM) of NeuroDNet contains all the genes of the human genome-45020 genes of which 23857 are protein coding. The gene data for ndnGNM was acquired from the NCBI human genome database [ftp://ftp.ncbi.nih.gov/gene/DATA/GENE_INFO/Mammalia/]. Placed above ndnGNM is the table ndnDGN that contains genes associated with neurodegenerative diseases. Disease description and associated information was collected manually from MeSH [http://www.ncbi.nlm.nih.gov/mesh/] and OMIM [http://www.ncbi.nlm.nih.gov/omim/]. The 305 gene populating ndnDGN were the ones associated with patient data reported in the literature. These genes include those reported for sporadic and familial diseases, and in different ethnic groups.

The functional tier consists of six tables - ndnPRO (proteins), ndnINT (human interactome), ndnSNP (SNPs), ndnONT (gene ontology), ndnPHE (phenotype) and ndnPAT (pathway). These tables were populated with data obtained from various public databases. Since data is archived in different formats in these databases, the data acquired was recorded in a format that permitted seamless transition between different databases. The NCBI Gene ID was retained to enable a cross mapping of the NeuroDNet output to other databases. Data for the Protein table (ndnPRO) was obtained from Uniprot/Swiss-Prot database http://www.ebi.ac.uk/uniprot/. The protein-protein interaction data was assembled in two stages. First, the interaction data was downloaded from HPRD, BioGRID, DIP, MINT and Reactome [3, 18, 19]. Next, a PHP script was written to traverse across the downloaded files to capture the gene IDs, protein IDs and protein-protein interactions, and recorded in the ndnINT table. The data consists of 23857 genes possessing 51388 interaction edges. The SNP table was constructed using data obtained from the literature and OMIM database by screening only those that were associated with the NDDs. The gene ontology associated with the ndnDGN was acquired through a MATLAB script from the Gene Ontology Biological Process category and stored in the ndnONT table. Similarly, the phenotype associated with the disease genes was also obtained from published literature and from OMIM through NCBI; the retrieved data was manually verified and then stored in the ndnPHE table. Signaling pathway information obtained from images given in KEGG, Reactome, BIOCARTA, and Cell SnapShots was collated, converted into adjacency matrices and then stored in the ndnPAT table. The collected information was organized into six signal-protein-gene interaction sub-tables, namely, Sig-Sig (ndnPSS), Sig-Prot (ndnPSP), Sig-Gene (ndnPSG), Prot-Prot (ndnPPP), Prot-Gene (ndnPPG) and Gene-Gene (ndnPGG). The information contained in these six tables is used to create signal-protein-gene networks by NeuroDNet when a user accesses the “Network Model” feature of the interface programme. The network is constructed in SBML script assuming first order rates and output as an XML file which is ready for performing simulations in Celldesigner if the user can provide the kinetic parameters.

The user interface consists of two modules. The first module contains the algorithms coded in PHP. It performs computations required to complete the queries requested by the user. The second module, the graphical user interface communicates with the first module to execute a request. It is written in html and it provides the user with the features offered by NeuroDNet database.

User interface

The home page of NeuroDNet gives a brief description of the database with a link for more information. There are four links given in the welcome panel - Query, Network, Disease Models and Boolean Analysis that enables a user to explore the different features of the website. The quick links are meant to take a revisiting user directly to a feature of interest. Using the first option “Query”, the database may be explored under the following categories - Disease, Gene, Protein, Polymorphism and Pathway. The class “Disease” contains the list of genes grouped by disease. For example, selecting “Alzheimer’s disease” gives the list of all the genes recorded in the database associated with AD. The user is free to pursue a gene of interest in this list by selecting links provided in the table displayed. “Gene ID” automatically connects the user to the corresponding gene page in NCBI, “Protein” to the corresponding Uniprot page for this protein, “Reference” to the research article, and “NDN Gene” to the NeuroDNet database page Gene Description. Gene Description comprises six sections - Gene, Protein, SNPs, Structures from PDB [http://www.rcsb.org/], PPI and Network Model. The database may be explored through links provided in each of these sections depending on the interest of the user. For example, selecting protein-protein interaction (PPI) option will display interacting proteins based on its cellular location. An option is provided to increase the neighbourhood degree for the protein. This enables the user to expand the PPI network to include related proteins in the analysis. At present, the upper limit on the degree is 3. We have also provided the user with the option of filtering the PPI network result by grouping it according to function, phenotype and occurrence in a pathway.

The second option “Network” contains three sub-categories - PPINet, DiseaseNet, and PathwayNet. The PPINet window accepts a list of genes or a text file that are part of a network under study. NeuroDNet processes this information and outputs a table where the connectivity of the network is shown in terms of the binary associations. When the simple example {A2M,APP,APOE} was submitted in the PPINet window, it gave the binary associations between A2M ↔ APOE and APOE ↔ APP. The output text file can be visualized graphically as a network using Cytoscape. The Entrez gene ID and Uniprot/Swiss-Prot protein ID given in the output text file are unique identifiers that may be used for global representation. The purpose of the “DiseaseNet” is to determine if interactions exist between the queried disease and other NDDs [12, 20]. The result is displayed as a text file. “PathwayNet” lists pathways contained in ndnPAT that are associated with NDDs. The user may select any pathway of interest and determine the possible associations with other pathways in the NeuroDNet database. The pathways that possess crosstalk with the one of interest are given in a text file.

“Disease Model” is the third feature offered by NeuroDNet. It lists the collection of disease models that are linked to Celldesigner [http://www.celldesigner.org/] images and the corresponding SBML files stored in the database. These models were redrawn in Celldesigner from their original sources in reported literature. The user can download the SBML file from the database and visualize the model using a suitable network visualization tool. A PubMed reference is included with each model to refer the user to the original publication associated with it. These models may be easily expanded to include additional components for studying the dynamic behaviour of genes or proteins related to the disease.

The fourth option, “Boolean Analysis”, has been widely used to model regulatory networks and signaling pathways. This feature enables the user to determine the dynamic behaviour of a network of interest by assuming that each node of the network possesses a binary state. The Boolean analysis panel accepts two inputs, Nodes and Edges, for converting the network into matrices required for computation. An option is also available for specifying non-regulated nodes (self-degrading nodes). The user must provide the network information in the prescribed format through a window that accepts Boolean functions and select one of the two Boolean rules hosted by the database for determining the steady states of the network. Additional information and illustrative examples are given in the help link and user manual.

Maintenance and update

All database entries were meticulously curated for accuracy. The following procedure was devised for data input, maintenance and update. The data was first acquired automatically by employing programme scripts. This primary information was checked against experimental findings described in the original source reported in the literature. Further, information about the associated pathways, phenotypes, proteins and their interactions were crosschecked and then annotated to the standard syntax. The error-free data was then stored as primary raw data files from which each of the six tables were updated automatically. The current data in NeuroDNet is up-to-date and will be updated quarterly.

Case study 1: Protein neighbourhood network for human angiogenin protein

ALS is a debilitating neurological disease that affects humans across ethnic groups. Missense mutations in the protein angiogenin (ANG) that result in the partial or complete loss of angiogenic functions have been implicated with ALS [21, 22]. Newer therapies for ALS may arise from a better understanding of mutations in ANG. Therefore, there is a need to understand how the loss of angiogenic properties may affect other proteins that are related directly or indirectly through various interactions at the cellular level. Through an exhaustive study of these interactions, it may be possible to elucidate how these mutations or SNPs lead to apoptosis of motor neurons.

Through the example of ANG, we illustrate how the database may be used to generate protein neighbourhood networks for visualization and analysis. ALS and ANG were selected from the Disease and Gene options from the Query panel respectively to enter the Gene Description page. The graphical view of PPI showed the first neighbourhood proteins {ACTC1, ACTN2, ATP6AP1, PTEN, RNH1, TDGF1, TNFSF8} and their locations in a cellular environment. The interactions may be visualized using the Cellular View or Graphical View; the result may also be downloaded and visualized in Cytoscape [http://www.cytoscape.org/] (Additional file 1: Figure S1).

Increasing the neighbourhood degree increases the number of interactions and complexity of the interaction network. ANG is an extracellular protein; when transported to the nucleus from the cytoplasm, it interacts with seven proteins (Additional file 1: Figure S1a) that participate in regulation of angiogenesis (RNH1), cardiac muscle development (ACTC1, TDGF1, PTEN), vasculature development (ANG, TDGF1, PTEN) and apoptosis (TDGF1, PTEN, ACTN2, TNFSF8, ATP6AP1). The second neighbours arising from these seven proteins include 119 proteins possessing 124 edges of interactions (Additional file 1: Figure S1b). Exploring the network further to include third neighbours recruited an additional 2580 proteins with 4424 edges of interactions (Additional file 1: Figure S1c). Network expansion to include higher degree neighbours is useful in identifying proteins engaging in crosstalk with other signalling pathways through direct or indirect interactions. Ahn et al.[23] had observed that ANG causes up regulation of Rb1, RBX1, NBN, CDK4, CDC34, SMARCA2, STAT3, CDK1, and MAp2; whereas down regulation of IκB-alpha, p55cdc, p35, BRCA, KAP, PCNA proteins in human umbilical vein endothelial cells (HUVECs). When the network was expanded to the third degree neighbour, we observed that all these proteins are included in the PPI network result. PPI networks, such as the one generated for ANG (Figure 2), may be created for other proteins and correlated with observed experimental observations.

Case study 2: PSEN1 network model and effect on calcium homeostasis

Presenilin-1 and presenilin-2 (PSEN1 and PSEN2), are transmembrane protein modulators associated with early onset familial Alzheimer’s disease (FAD) [24]. It is a conserved polytopic transmembrane protein predominantly localised in endoplasmic/sarcoplasmic reticulum (ER/SR) and participates in the cleavage of the amyloid precursor protein (APP) which results in the formation of amyloid-β-peptides [25]. The aggregation of the “sticky” form, amyloid-β42, into fibrillar structures has been established as one of the mechanisms leading to all known clinical symptoms of Alzheimer’s disease. Consequently, we decided to investigate the role of PSEN1 in neuronal cell death in AD.

NeuroDNet tool under “Gene” was employed to construct the wiring diagram. This interactive tool allows the user to select a gene of interest from a disease category and retrieve interactions from the information contained in the six tables of “Pathway” to assemble the signal-gene-protein interactions. The information is retrieved and stored as an Excel spreadsheet, which is then converted into the SBML format using a PHP script. The XML file is displayed as a network using Celldesigner (Additional file 1: Figure S2).

PSEN1 is part of the γ-secretase complex involved in APP cleavage, notch signaling pathway and calcium homeostasis. The “Network Model” feature of NeuroDNet revealed that PSEN1 protein is an important vertex that regulates the cell proliferation, differentiation and apoptosis [26–28]. The network gives details of known interactions of PSEN1 mediated through the vertices Notch, APP and Ca²⁺. The information obtained may be used in verifying experimental results, building global systems model or examining specific phenomena. We illustrate how the interaction of the PSEN1 vertex with its neighbours may be used to understand how calcium leak from intracellular stores promotes neurodegeneration.

We asked the question whether the loss of channel function or its permanent opening caused by PSEN1 mutation disrupted calcium homeostasis. To answer this question we first studied the PSEN1 interactions. The SBML network generated using NeuroDNet showed how Ca²⁺ interacts with channels and regulatory proteins. The differential equation model described by Marhl et al.[29] was obtained from the SBML repository http://www.ebi.ac.uk/biomodels/. This model explained complex intracellular Ca²⁺ oscillations using Ca²⁺ kinetics to describe fluxes from the ER and mitochondria to the cytoplasm, and the role of Ca²⁺ binding proteins in regulating the homeostasis.

The calcium leak was attributed to the zymogen form of PSEN1 and the parameter associated with the corresponding equation, the rate constant k _leak ^PSEN1, was perturbed (from normal to aberrant) over 0.05 - 0.3 s^-1 range. We observed that the amplitude of cytoplasmic Ca²⁺ oscillations decreased while the frequency increased (Figure 3a). Absence of the “burst” signal of Ca²⁺ in the cytoplasm indicated a severe dysfunction. When the ER store of Ca²⁺ was depleted below 0.72 μM, the cytoplasmic control was lost completely (Figure 3b). Similar results were observed by Zampese et al.[30] in PSEN2 mutant SH-SY5Y and HeLa cells. The mitochondrial regulation of the ER was also examined through a phase diagram. The limit cycle observed for normal parameter values disappeared when k _leak ^PSEN1 increased to 0.3 s^-1, suggesting breakdown of ER-mitochondria Ca²⁺ control (Figure 3c). Our results suggest that PSEN1 malfunction increased Ca²⁺ leak frequency, which in turn was responsible for the loss of ER-mitochondria Ca²⁺ regulation. Loss of this regulation effectively nullified the mitochondrial Ca²⁺ induced Ca²⁺ release (mCICR). It was shown that changes in spatiotemporal concentration of mCICR caused a permanent opening of the permeability transition pore (PTP) which ultimately triggered apoptosis [31–33].

An alternative explanation emerges when the ability of presenilins to form ion channels is considered. The zymogen form of presenilins localized in the ER/SR function as calcium channels and is responsible for the “leak” calcium currents involved in calcium homeostasis [34, 35]. The mechanism of calcium leak into the cytosol has remained largely unexplained. Hofer et al.[36] observed a decrease in free calcium in the ER lumen in studies where the SERCA (sarco/endoplasmic reticulum calcium ATP-ases) was inhibited by the addition of thapsigargin. This leak remained unaffected even when the RyRs and IP3Rs were inhibited indicating that the basal leak rate of calcium was dependent on other unknown mechanisms. However, evidence that is more recent suggests that RyRs and IP3Rs may participate in the calcium leak under certain pathological conditions and in apoptotic phenotypes. Structural studies have shown that the cytoplasmic domain becomes dysfunctional creating a “channel only” behaviour [37]. Under normal physiological conditions, basal calcium leak has been attributed to natural ionophores, golgi, endosomes, and the potential difference between ER and cytoplasm [38]. Although the literature presents conflicting views on how the basal calcium leak is regulated, it is clear that calcium imbalance in the cytosol may push the cell into apoptotic state. Aberration in the PSEN1 network has been implicated in AD and its role in calcium homeostasis deserves further experimental investigation.

Case study 3: Mammalian cell cycle - NeuroDNet Boolean analysis tool

Biological networks comprise the intricate circuitry of signals, genes, proteins and metabolites that brings about the kaleidoscope of changes within the cell. The interactions between the different molecular species, results in the observed complexity in cellular functions. Usually, computationally manageable sub-networks are constructed and analyzed to obtain insight into specific aspects of cellular behaviour. The dynamics of these sub-networks have been described with ordinary differential equations (ODE) but the ODE formalisms is constrained by the size of the sub-network and the dearth of kinetic and parametric information [39]. In such circumstances, it is more desirable to use Boolean networks, which are based on logical formalisms.

Boolean networks have been widely used to model gene regulatory networks and signaling pathways [40–42]. Each element of the system has a binary state ({0, 1} = {on, off}) and is therefore discrete, deterministic and parameter free [43]. The Boolean network is a graph, where the node (or vertex) v denotes in general, a gene or protein, and the edge e defines the nature of the interaction between two nodes. A network of N nodes has 2 ^N possible states. Time evolution of the network may be synchronous or asynchronous and the eventual steady states are called “attractor” states. The attractors, represented by the on-off state of genes in the network, correspond to important physiological states of the cells. For example, the state may indicate growth, differentiation or apoptosis [44]. Boolean analyses are particularly useful in cases where a qualitative insight is sought and for a better understanding of the network structure without invoking computationally intensive procedures.

More recently, the research has focused on developing generalized logical formalisms [45], examining robustness of networks [46], using interaction graph representations [47], investigating scalability across systems [48] and designing new efficient algorithms [49]. A number of systems such as, A. thaliana morphogenesis, yeast cell cycle, T-cell signaling and T-lymphocyte survival signaling, have been studied using Boolean networks to gain intuitive understanding [50–53].

We illustrate the use of the Boolean analysis tool featured in NeuroDNet using the mammalian cell cycle described by Fauré et al.[54] (Figure 4a). Emerging evidence suggests that a strong relationship exists between the expression of cell cycle proteins and neurodegenerative diseases [55]. We describe the steps used to study this mammalian cell cycle to predict cellular fates (Figure 4a). The Boolean network comprising 10 nodes and 49 edges explained the emergence of G0 quiescent phase and non-spurious dynamical cycle. The Fauré cycle was reconstructed using NeuroDNet’s “Boolean Network” feature. The network properties {G = (10, 49)} were first keyed in. Next, the syntax for defining the logical operations of the network was written in the designated text window (Table 1). In the equations, the i ^th node of a network was represented by $node i, (i = 1, 2, …, 10). The Boolean rules and algorithm for computation were then selected and executed to obtain the attractor states (steady states). The sign associated with edge of the network, qualified the kind of interaction existing between two nodes - activation (+1) or inhibition (−1). At each node this information was processed using logic gates such as AND, OR and NOT. The dynamic status of each node was updated synchronously. The state of a node {0 or 1} was assigned the result of the logical operation at each instant, that is

$S_i\left(t+1\right)=\{\begin{array}{c}\hfill 1\leftrightarrow {\displaystyle \sum _j{a}_{\mathit{ij}}{S}_j\left(t\right)\ge 1}\hfill \\ \hfill 0\leftrightarrow {\displaystyle \sum _j{a}_{\mathit{ij}}{S}_j\left(t\right)<1}\hfill \end{array}$

(1)

Where S _i (t) and S _i (t + 1) represent the successive states of the i ^th node, S _j (t), a _ij  = ±1 represents interacting nodes depending on whether the edge is activating or inhibiting. The network details submitted by the user are processed by an algorithm written in PHP script residing in the user interface module. The algorithm outputs a text file containing the 2 ^N Boolean outcomes and the list of singleton attractors and cyclic attractors which signify the states that the network can acquire. The output text files can be visualized for determining dynamic trajectories in a suitable network visualization tool such as Cytoscape (Figure 4b, c). Each attractor state represented as the binary sequence shown in (b) and (c), corresponds to the state of nodes taken in the order given in Table 1.

Cyclin D (CycD) controls the expression of retinoblastoma (Rb), a key tumor suppressor. During the G1 to S phase transition, E2F transcription factor (E2F) activates transcription of Cyclin E (CycE) and Cyclin A (CycA), which in turn controls the anaphase-promoting complex (APC) in cyclic fashion. We obtained one singleton attractor state that represented the quiescent G0 phase {0100010100} where each Boolean state corresponds to node (i = 1, 2, …, 10) (Table 1). The other stable states represent cyclic attractors consisting of seven successive states describing dynamical cycle consistent with those reported by Fauré et al.[54]. Mutational studies were also performed using this cell cycle network. The Boolean network for Rb mutant was executed and it was observed that the network lost its singleton attractor. Instead, a cyclic attractor was created that depicted lack of restriction point as shown by Novak and Tyson [56]. The results of this example showed that Boolean modeling tool correctly predicted the outcomes of the network and its dynamic behaviour qualitatively.