Ethics statement

Input data

Our network-based analysis relies on three inputs: (i) quantitative data, (ii) a set of seed genes, and (iii) a reference interactome.

The GWAS data used as input in the current work were generated in [4]. There the features of the study population were described in detail. For the cases, genotyping data of DNA samples diagnosed with FTD were available from the International FTD-GWAS data set for 634 samples obtained from 8 Italian research centers. After quality check (QC) steps 530 patients diagnosed with bvFTD (n = 418), semantic variant PPA (n = 27), agrammatic variant PPA (n = 61), and FTD-MND (n = 23) survived. All cases were diagnosed according to the Neary criteria [38] and/or the more recent Rascovsky and Gorno-Tempini criteria [1,2]. Age of onset was 64.1 ± 20.7 years (Mean ± SD); range: 29.0–87.0), including 287 (54.2%) women. The cases were collected and genotyped at the University College London by means of Illumina human 660K-Quad Beadchips assayed on the Illumina Infinium platform (Illumina, San Diego, CA, USA).

For the controls, genotyping data were obtained from the HYPERGENES project (European Network for Genetic-Epidemiological Studies; www.hypergenes.eu) [37]. All participants in the sample set (n = 1327; 926 after QC) had no abnormal findings on physical and neurological examination, were unrelated, collected in Italy, and of Caucasian ancestry. Age was 58.2 ± 6.1 years mean (±SD); range, 50.0–97.0), including 349 (37.7%) women. The control samples were genotyped at the University of Milan, using the Illumina 1M-duo array.

Quantitative data was generated by applying supervised Principal Component Analysis (sPCA) [39] over additive-encoded genotypes (0 for the frequent homozygote genotype, 1 for the heterozygote, and 2 for the rare homozygote genotype), and taking the first principal-component score (PC1) for each gene, as reported in our previous study [4]. Briefly, polymorphisms were grouped by gene membership, where the gene region is defined by its locus coordinates ±5000 base pairs. A single gene was then represented as a continuous score by taking the first supervised PC1 on a subset of SNPs selected using outcome (case-control) information, i.e. supervised PC1 was a weighted sum of the SNPs within the gene region that maximize the variance of the gene score, which then varies with the outcome.

Seeds are genes of interest, that can arbitrarily be chosen using existing knowledge (e.g. from databases such as OMIM, DisGeNET, or by text mining), and/or experimental data analysis (e.g. GWAS, expression data, or DNA-binding data). We used the list of FTD-associated genes from our previous SNP-to gene approach [4], applying an FDR threshold < 10%, yielding 280 putatively FTD-associated genes. These genes were referred to as the FTD-seeds.

An interactome is defined as the ensemble of known interactions among biological entities (e.g, genes, proteins, or metabolites) in a given organism, usually represented as a network. The interactome can be retrieved from many different sources, including KEGG [11], Reactome [12], STRING [40], GeneMANIA [41], MINT [42], IntAct [43], BioGRID [44]. We selected KEGG, for different biological and computational reasons, including annotation curation and completeness, and GO biological process mappability for functional annotation. Furthermore, KEGG is a directed interactome, thus causality can be easily interpreted in terms of biological signalling, and can be used for the FTD model validation and extension. However, edge direction is not a necessary requirement in our method. We retrieved KEGG signalling pathways and merged them to obtain a unique network using graphite [45] R software package.

Network edge weighting

The input interactome was converted into a weighted network, endowed with node and edge weights reflecting their perturbation status. Genes (nodes) were weighted as being FTD-seed (weight = 1) and non-seed (weight = 0). Gene-gene interactions (edges) were weighted based on the case/control statistical difference. In our notation, j and k represented any two connected nodes of the network, with j->k being the edge direction. In general, we tested if the total difference between case vs. control groups for gene k through gene j was significant, given data. This implied testing the group change at the same time on gene j, gene k and their direct link j->k. The p-value was yielded by a t-test on the combined difference of the group over the node j, the node k, and their direct connection j->k, fitting a trivariate (X = j-th gene, Y = k-th gene, C = {0, 1}) Structural Equation Model (SEM) via lavaan R package [46].

Then, the edge weights were defined as inverse of negative logarithm of the p-values, w = 1/-log(p-value). In this way, edges with lower p-values had lower weights, on a positive continuous range. Intuitively, this weight can be assumed as the perturbance acting on the relationship between two connected genes in the interactome, due to the genotype difference between groups. The lower the p-value (i.e. the weight), the higher the perturbance. In general, we defined the perturbance over a node k, due to the action of a node j, as the altered status of j and k genes, and their j->k interaction in the diseased sample, comparatively to healthy controls.

The Steiner tree problem

The FTD seeds were mapped to the weighted interactome, and a FTD-related sub-network was constructed by adding new genes to connect FTD genes solving a Steiner tree problem [30–33], minimizing the sum of weights of every edge in the sub-graph. We applied a modified shortest path heuristic (SPH) distance algorithm, from Kou et al. solution [33], implemented in our subnet() R function. Our algorithm selected outgoing shortest paths combing the edge weights by Fisher’s method and testing the statistical significance (p < 0.05) with multiple comparison Bonferroni correction. The resulting Steiner tree, corresponding to the maximum-perturbance sub-graph, preserves the original directed edges. Therefore, we distinguish three types of nodes: “sources” (emitting perturbance) with no incoming connections, “targets” (absorbing perturbance) with no outgoing connections, and “connectors” (transmitting perturbance) with incoming and outgoing connections. We referred to a perturbation route, as a perturbed path originating from a source node, traversing a number of connectors, and terminating in a target node. This tree was used as a backbone for the subsequent augmenting SEM step.

Structural Equation Model (SEM) analysis

The sub-graph obtained as a Steiner tree was converted into a SEM [34–36], such that every node in the sub-network corresponds to a variable of the SEM, and every edge is a relationship between variables. In summary, a group node (C = {0, 1}) connected to each gene was added inside the Steiner tree, and the overall sub-graph was converted into a system of linear equations, and then fitted using PC1-trasformed SNP genotype data. The system of linear equations has the form: with a covariance structure: where V(x) and V(y) are, respectively, the sets of the exogenous variables (i.e. source genes) and endogenous variables (i.e, connector plus target genes) in the sub-network. The linear equations define the relationships between the variable Y_j with the group variable C and variables Y_k in the “parents” set, pa (j), quantified by path coefficients β_jC and β_jk, respectively. Unobserved variables U_j represent the variation of each Y_j not explained by its parent nodes. The covariance structure describes the bi-directed relationships between variables in the “siblings” set, sib (j), quantified by covariances ψ_jk, and interpreted as unmeasured common causes for pairs of variables.

SEM analysis allowed to evaluate: (i) the overall goodness of fit of alternative constrains in the models, (ii) the significance of the group difference on every gene, and (iii) how significant the covariance between target nodes (i.e., genes that are not connected by a directed path) was, added to improved SEM goodness of fit. The significant new edges added in (iii), called “ancestral bow-free” covariances in SEM literature closed the directed paths into circuits characterized by the presence of signaling sources and targets.

Minimun AIC (Akaike Information Criterion) score, and a standardized root mean squared residual (SRMR) less than 0.05 were considered for model selection and overall good model fitting, respectively [34]. The statistical significance of the SEM parameters (the regression coefficients and the ancestral bow-free covariances) were estimated by Maximum Likelihood Estimation (MLE) and the beta coefficients were evaluated through t-test with bootstrap standard error (SE) with B = 1000 resamplings, and the significance level established at p < 0.05, two-sided.

After Fisher’s normalizing t-transformation of the covariance, ancestral bow-free covariances, with abs(t) > 2 were selected. Successively, the selected covariances were tested by a SEM with Latent Variables (LVs) [47] using a model in which two target genes are connected through a LV modelling the underlying common unknown cause(s) acting on them. Every LV is subjected to group (C = {case, control}) effect. In this context, a LV is a variable not present in the initial Steiner tree, introduced to capture the differential case/control co-variance between two disconnected target nodes. The goodness of fit of each LV model was evaluated through a Likelihood Ratio Test (LRT), where a good fit corresponds to a p-value > 0.05 (i.e., there is no significant difference between sample and model covariance matrices). The influence of the group C over each LV was evaluated by a t-test (= MLE/bootstrap SE) with p-value < 0.05, two sided. Every step of the SEM analysis was performed calling the lavaan [46] and igraph [48] R packages in our sem2group() R function.

Functional analysis

After model building SEM assessment and model extension with LVs underling covariances between pairs of leaf genes, we proceeded with the biological evaluation. To improve the FTD model interpretation, we initially focused on nodes with specific topological properties (e.g., those connecting different portions of the gene network). We computed various topological indices, including in/out degree, i.e. the number of incoming/outgoing connections for a node, and weighted node betweenness [28], i.e. the number of weighted shortest paths travelling through a node in the network, by means of perturbance. Using these topological indices, we identified four types of nodes: (i) “springs”, i.e. source nodes with high out-degree. (ii) “sinks”, i.e. target nodes with high in-degree; (iii) “hubs”, i.e. connector nodes with high in/out-degree, and (iv) “bottlenecks”, i.e. nodes with high weighted betweenness. We also defined as secondary sources many direct targets of proper sources, with in-degree 1 and out-degree 1, representing ligand-receptor interactions.

We focused on the concept of “essential” node, by combining definitions (iii) and (iv) as critical nodes for the FTD-network integrity, and part of the network backbone. These nodes carry the majority of perturbed information flow through the FTD network, and therefore we expect them to be assumed as suitable novel FTD risk factors.

Annotation enrichment over the main biomedical ontologies was used to functionally validate and extend our biological interpretation of the FTD model. Enrichment analysis was performed using a hypergeometric test with Bonferroni correction (adjusted p-value < 0.05), over GO:BP [10], KEGG [11], Reactome [12], and Disease Ontology (DO) [49]. Network indices and Enrichment analysis were calculated calling the igraph [48], clusterProfiler [50] and DOSE R packages [51].

The FTD sub-network

We first generated the SNPs-to-Genes data set, taking the first principal-component score (PC1) for each gene obtained by applying supervised Principal Component Analysis over additive-encoded genotypes (see Materials and Methods). We then used the 280 FTD-associated genes from our previous study [4] as seed list, applying a False Discovery Rate (FDR) threshold < 10%. We also compiled the reference human interactome by merging KEGG signalling pathways [11]. After mapping PC1 quantitative data (13971 genes) over the KEGG database, our interactome was made of 4073 nodes and 34606 weighted edges (23075 directed and 11530 bidirected), with a median vertex degree of 4 and average directed shortest path distance equal to 5.464. Finally, gene-gene interactions (i.e. edges) were weighted by combining the effects of the group variable (i.e. C = {0, 1}, representing controls and cases, respectively) over pairs of nodes, and their direct connection (see Materials and Methods).

The interactome contained 83 out of 280 FTD seeds, with a median vertex degree of 7 and average directed shortest path distance equal to 4.396. Notably, a higher vertex degree in combination with a shorter weighted shortest path for the mapped FTD seeds showed that FTD-associated genes are more clustered than the rest of the interactome. To assess the significance of this finding we tested if the mapped FTD-seeds were more connected than a random selection of 83 genes from the interactome (B = 1000 samples) or a node permutation of the interactome (B = 1000 permutations). Results showed that there was a significantly higher degree (p = 0.034 and p = 0.029 for randomization and permutation tests, respectively) and a significantly shorter average shortest path (p < 0.001 for both tests) for the FTD genes.

We applied the Steiner tree algorithm to the mapped FTD seeds, detecting an initial set of 3854 shortest paths between the 83 seeds. The final FTD Steiner tree (Fig 2), including only connector genes traversed by perturbed paths, contained 167 nodes (77 out of 83 FTD seeds) and 166 edges (117 of which significantly perturbed, p < 0.05). The Steiner’s tree topology revealed the presence of: (i) 43 sources, of which 33 FTD seeds; (ii) 49 (38 FTD seeds) targets; and (iii) 75 (6 FTD seeds) connectors. The Steiner’s tree backbone, defining the mainstream perturbation route, was defined by the genes CAMK2A-EP300-TCF7L2, bifurcating to EGFR-CTNNB1 and MAPK8-JUN (Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Steiner tree.

Steiner tree obtained applying a Shortest Path Heuristic (SPH) algorithm. The tree has 167 nodes and 166 edges. The size of each node (i.e. gene) is proportional to its degree (i.e. the number of incoming and outgoing connections). Node colours indicate: perturbed seeds (green), non-perturbed seeds (red), perturbed connectors (blue), and non-perturbed connectors (yellow). Red edges correspond to perturbed interactions, while edge thickness is proportional to their weight (i.e. their perturbance level). A perturbed interaction has a weight w < 0.33 (i.e. p < 0.05 threshold over the nominal p-value). The entire network is characterized by a backbone CAMK2A-TCF7L2- CTNNB1-JUN-MAK8-PRKACG, where CAMK2A and MAPK8-PRKACG are the main perturbed hubs, CTNNB1 and JUN represent the sinks of the entire system, and TCF7L2 is the bottleneck connecting them.

https://doi.org/10.1371/journal.pone.0185797.g002

SEM analysis of the Steiner tree was performed by fitting various models with different constraints on their parameters; the selected model, with fixed beta coefficients, equal variances, and bow-free covariances, yielded the lowest AIC score (AIC = -9324) and good fitting (SRMR = 0.026; see Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. SEM goodness of fit.

Goodness of fit measures for different models, fitted to multivariate data of the extracted Steiner tree. The selected model, indicated by (*), has the lowest Akaike Information Criterion (AIC = -9324.98).

https://doi.org/10.1371/journal.pone.0185797.t001

Non-parametric bootstrap estimate of the SEM beta coefficients (B = 1000 samples with replacement) detected 98 significantly perturbed genes (bootstrapping p-values < 0.05), of which 74 FTD seeds, indicative of association with the disease (S1 Table). After Fisher’s normalizing t-transformation of the covariance, 43 ancestral bow-free covariances, connecting 43 (35 seeds) target genes (S2 Table) with abs(t) > 2 were selected. Adding the latent variables (LVs) underlying covariances between pairs of target genes, we obtained a sub-network of 210 (167 genes + 43 LVs) nodes and 252 (166 + 2*43) edges (S1 Fig). Among the 43 LVs, 20 yielded a good LV-model fit and resulted as significantly perturbed (i.e., LV average values are different in cases respect to controls), having bootstrapping p(C->LV) < 0.05 and p(LRT) > 0.05 (S2 Table).

Topological analysis and essential nodes

To drive and improve our biological interpretation of the FTD network, we initially focused on nodes with specific topological properties, their function, and their possible implication in neurodegeneration. In particular, we searched for those genes in the FTD model that connect different perturbed routes throughout the network. Hereafter, we will use “essential nodes” to designate a class of genes that share a high number of perturbed connections (especially outgoing ones), thus bearing the largest amount of perturbed information spreading through the FTD network. This definition differs from the usual definition of hub, since it takes into account both the number of perturbed connections and perturbed shortest paths traversing a given node (Fig 2 and S1 Fig). Many of these are terminal seed-genes (i.e. peripheral nodes), designed as FTD-associated in our previous analysis [4] and confirmed as perturbed in the present study (i.e. green nodes in Fig 2 and S1 Fig). Besides seeds, we also identified additional FTD-associated genes (blue and red nodes in Fig 2 and S1 Fig), using SEM testing. However, also non-perturbed genes (i.e. not carrying FTD-associated variants) are functionally altered if they are significantly influenced by genes carrying disease-associated variants (i.e. they have perturbed incoming connections). Based on the perturbation routes, we could define topologically-critical FTD-associated genes.

A set of connectors bind sources to targets. These genes can both receive and send perturbed interactions. Steiner’s connectors are genes with generally high centrality. They may have high degree (i.e. hubs, with degree > 3), such as MAPK8, and/or high betweenness (i.e. they are important for connecting network modules), such as TCF7L2. Connectors are key genes revealing the set of critical trait-specific molecular processes for the cell. Among all nodes in the FTD-module, the most central ones are designed as essential. Essential nodes are those genes that cannot be removed without a deep impact on network connectivity. Considering that the FTD network has been generated based on phenotype-associated genotypic variability, essential nodes should be also functionally critical for the FTD phenotype. To point out the most central connectors, we extracted an essential high weighted-betweenness sub-network (Fig 3). Essentiality combines the concepts of biological process perturbation with the importance for network structural integrity. The essential sub-network defines the FTD-network backbone, characterizing the most perturbed interactions. They include: (i) the EGFR-PLCB3 interaction, (ii) the CAMK2A-EP300-TCF7L2 perturbed path, and (iii) the PRKACG-MAPK8 interaction. The essential backbone highlight the presence of nodes with exceptionally high outgoing connectivity (hubs), including MAPK8 and PRKACG, and exceptionally high incoming connectivity (i.e., sink nodes), including CTNNB1 and JUN. Sinks of perturbation represent nodes with no outgoing perturbed interactions, that ideally encompass the model terminal functions, including (post)synaptic plasticity and cell death (see Discussion).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Essential node sub-network.

Node essentiality is determined by considering nodes having both degree centrality and weighted betweenness centrality over the upper-quartile. Essential nodes are placed in non-redundant portions of the network and thus cannot be removed without a deep impact on network connectivity. These genes intercept the network backbone, represented by the axis TCF7L2-JUN-MAK8-PRKACG, carrying the top perturbation levels, especially in proximity of the sources. Nodes and edges are labelled according to the conventions followed in Fig 2.

https://doi.org/10.1371/journal.pone.0185797.g003

Special attention must be given to unobserved variability, represented by LV-mediated interactions (Figs 4 and 5 and S1 Fig). Although latent (i.e. not explicit in the model), LV-associated variability may evidence common causes of perturbation. Considering LVs only, PLCB3, a phospholipase C (PLC) controlled by EGFR in our model, is the sink node with the highest incoming perturbance level (Fig 4). This is particularly interesting as this gene might be responsible of Ca⁺² homeostasis impairments in FTD (see Discussion). Remarkably, PLCB3-specific LVs (LV6-9) connect this gene to four FTD-relevant targets: (i) GABRG3, a perturbed/seed GABA receptor, the major inhibitory neurotransmitter in mammalian brain; (ii) IL12B, an osmotic stress-induced cytokine, targeted by MAPK8; (iii) DGKI, a DAG-kinase interacting with the phospholipase PLCB2; and (iv) ARHGAP10, a GTPase activator involved in apoptosis. Other perturbed LVs connections (Figs 4 and 5) include: CRK, AGK, ATP2A3 (LV1-2); IL1R1, CPT1B (L18); and PPP3CC, VCAM1, PRKD3 (LV14 and LV35). S1 Fig reports the whole gene and LVs connectivity in the FTD-network.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. ncWNT sub-network.

This sub-network focuses on the module characterized by a series of receptors and enzymes regulating Calcium/cAMP homeostasis and involved in the non-canonical Ca⁺²/WNT signaling pathway. Among them, the most perturbed are the EGFR receptor and its target phospholipase PLCB3, and the routes ITPR1-CAMK2A-CALM2-PPP3CC and ITPR1-CAMK2A-EP300-TCF7L2. JUN is a large sink between this module and the MAPK-JNK one. Nodes and edges are labelled according to the conventions followed in Fig 2.

https://doi.org/10.1371/journal.pone.0185797.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. MAPK-JNK sub-network.

This sub-network focuses on the module characterized by a series Serine/Threonine kinases involved in the MAPK-JNK signaling pathway. Among them, MAPK8 is the one having the highest outgoing connectivity of the entire network, and the most perturber incoming interaction carried by PRKACG, another FTD-network hub. Other deeply perturbed interactions include MAPK8-TNF, MAPK8-CRK. JUN is a large sink within this module and the non-canonical WNT pathway (ncWNT) one. Nodes and edges are labelled according to the conventions followed in Fig 2.

https://doi.org/10.1371/journal.pone.0185797.g005

Enrichment analysis

The enrichment analysis was performed on the whole set of nodes of the FTD-network. Enrichment analysis over GO:BP, KEGG, and Reactome showed significant involvement of response to oxygen-containing compounds (ROCC), immune response, and homeostatic processes (S2 Fig), several signaling pathways, including MAPK, RAS, cAMP/cGMP, the WNT pathway, and long-term potentiation (LTP) (S3 Fig), and synaptic transmission-related processes (S4 Fig), including EGFR/FGFR signaling, transmission across chemical synapses, NMDA receptor-associated postsynaptic activation, phospholipase C-mediated cascades, and Calcium pathway. Furthermore, Disease Ontology (DO) data showed enrichment for nervous system and mental health-related terms, including: central nervous system disease, cognitive disorder, psychotic disorder, and neurodegenerative disease (S5 Fig).

The most significant term in GO:BP is the response to oxygen-containing compound (ROCC; GO:1901700; p < 1E-9), involving the majority of source nodes and their targets, suggested that oxygen compounds and reactive oxygen species (ROS) may trigger the whole network perturbation. Particularly, in the FTD-network (Fig 2 and S1 Fig) oxidative stress was impacted due to impaired energetic metabolism [52,53] eliciting the activation of stress-responsive receptors, including Frizzled-4 (FDZ4) and EGFR [54–64], and oxidative stress-inducible Serine/Threonine kinases CAMK2A, PRKACG, PRKACB, and PRKG1 [54–83]. Collaterally, these effectors may further induce neuron membrane potentiation, leading to apoptosis through activation of the MAPK-JNK signaling pathway [73–77] and Ca²⁺/cAMP metabolism deregulation [84–91]. Aberrant Ca²⁺/cAMP metabolism might lead to anomalous neuron potentials and further deficits in neuroprotection, neuron functioning, and survival [69,70,80–100].

More specifically, the significant KEGG-enriched (S3 Fig) and Reactome-enriched (S4 Fig) functions sustained by the FTD-network, included: MAKP signaling (KEGG:map04010), RAS signaling (KEGG:map04014), cAMP signaling (KEGG:map04024), cGMP-PKG signaling (KEGG:map04022), Gap Junction (KEGG:map04540), WNT signaling (KEGG:map04310), Insulin signaling (KEGG:04910), Long-Term Potentiation (KEGG:map04720), and FoxO signaling (KEGG:map04068), Signaling by EGFR (Reactome: R-HSA-177929), DAP12 signaling (Reactome: R-HSA-2424491), Transmission across chemical synapses (Reactome: R-HSA-112315), Neurotransmitter receptor binding and downstream transmission in the postsynaptic cell (Reactome: R-HSA-112314), G protein mediated events (Reactome: R-HSA-112040), PLC beta mediated events (Reactome: R-HSA-112043), Ca2+ pathway (Reactome: R-HSA-4086398), and Activation of NMDA receptor upon glutamate binding and postsynaptic events (Reactome: R-HSA-442755).

These processes were supported by oxidative stress-related sources and hubs, including FDZ4, EGFR, THRB, EP300, CAMK2A, PRKACG, PRKACB, and PRKG1. Literature cites their involvement in both Calcium homeostasis [54–61,92–95], LTP [69,70,80,81,91–96], and ROCC [68–83]. Essential sink nodes involved in both ROCC and cell death through MAPK-JNK signaling include JUN, PPP3CC, and the MAPK8 targets TNF, CRK, TP53, and MAPK11 [70,74–78]. Another remarkable perturbed sink is the EGFR [72,73] target phospholipase PLCB3, carrying one of the mostly perturbed receptor-target interactions. This is interesting since the well-known involvement of PLCs in the non-canonical WNT/Ca²⁺ signalling pathway [54]. Topologically, the entire FTD-network can be divided into 2 main modules: the non-canonical WNT/Ca²⁺ (Fig 4), and the PRKACG-MAPK8-JUN (Fig 5) sub-networks. Both have a specific role in neuroprotection and neuron survival, respectively, with a critical interface interaction represented by TCF7L2-JUN-MAPK8 (Fig 2).

In our model perturbance is a measure of edge perturbation significance deriving from a synthesis of genetic variability (PC1) and phenotype (group) effects, considering the whole network architecture (through SEM). To assess the association between perturbance and FTD-related term enrichment, we used existing annotations to map our network onto FTD-associated KEGG and DO annotations. The KEGG database reports 6 pathways related to FTD (KEGG ID: H00078), including: hsa04010 (MAPK signaling pathway), hsa04141 (Protein processing in the endoplasmic reticulum), hsa04144 (Endocytosis), hsa04310 (WNT signaling pathway), hsa04330 (NOTCH signaling pathway), and hsa04722 (Neurotrophin signaling pathway). As shown in S6 Fig, these pathways encompass the fully-perturbed backbone of our network, suggesting that our method can capture the core FTD-associated variability and extend it through the data-driven concept of edge perturbation. Notably, the FTD-network backbone contains the top-perturbed connections. To have a precise measure of the perturbance content we measured the proportion of perturbed connections in the KEGG FTD-specific sub-network comparing it to the total FTD-network, excluding LVs (i.e. the basal perturbance). The basal perturbance proportion is expected to be already high, since the whole disease-network is associated with FTD: 117/166 (= 70%) significantly perturbed edges. In the KEGG FTD-specific sub-network this percentage is even higher: 40/46 (= 87%). We repeated this measure using DO terms included in the Nervous System Disease (DOID:863) and Disease of Mental Health (DOID:150) ontology roots, obtaining the DO Nervous System-specific sub-network (S7 Fig). Also in this case, the sub-network perturbance proportion is higher than the basal one: 40/52 (= 77%).

Biological relevance of the FTD model

In the present study, we investigated whether FTD-associated genetic variability drives the identification of perturbed functional sub-networks, indicating likely impacted biological processes in frontotemporal lobar degeneration. The core structure of the essential FTD-network (Fig 3) highlights the central role of four biological processes in FTD: (i) WNT/Ca²⁺ signaling, (ii) response to oxygen-containing compounds (ROCC), (iii) postsynaptic plasticity and Long Term Potentiation (LTP), and (iv) MAPK-JNK signaling. These functions encompass the entire set of genes and processes defining our theory. As a follow-up to our previous GO-based gene-set analysis [4], several processes appeared replicated, these being neurodevelopment-related processes, Ca²⁺/K⁺ channel regulation and LTP. In addition, in the current study we also found evidence for an implication of oxidative stress and DNA damage, coupled with impairments in the energetic and Ca²⁺/cAMP metabolism, involved in synaptic plasticity and neuroprotection. Particularly, the notion of DNA damage, which is associated with neuronal homeostasis and longevity, appears to gain momentum in the landscape of FTD pathogenesis [81,82,101,102].

Topologically, the FTD-network can be divided into in two main modules: (i) the non-canonical WNT/Ca2+ signaling (ncWNT)-associated module (Fig 4), and (ii) the MAPK/JNK-associated module (Fig 5), including PRKACG-MAPK8, AKT3, and CREB3L2, and their sink JUN. The interface between the two sub-modules is represented by the TCF7L2-JUN-MAPK8 interaction (Figs 2 and 3), depicting an overall disruption of the DAG-IP3/Ca2+/cAMP homeostasis and energetic metabolism, leading to the induction of several oxidative stress-responsive Serine/Threonine (Ser/Thr) kinases and their related targets, including members of the CREB family, protein phosphatases, junction proteins, K⁺ cannels, and members of the FoxO signalling pathway. These critical Ser/Thr kinases include: (a) the Ca²⁺-dependent CAMK2A, PRKCG, PRKACG, PRKACB, and PRKG1; (b) the MAP-kinases MAPK8 and MAPK11; and (c) the AKT-kinase AKT3. Every Ser/Thr kinase has a specific influence over a group of functionally-related nodes, converging towards common target nodes. Specifically, we may define two high-degree sinks (Fig 3): CTNNB1 and JUN. Given these premises, we hypothesized that oxidative stress and DNA damage generated in response to energetic metabolism and Ca²⁺/K⁺ homeostasis impairments, could be among the major potential molecular underpinnings of FTD pathogenesis.

The non-canonical WNT/Ca²⁺ module

A large portion of our model involves the ncWNT pathway, converging to the sink CTNNB1-CREBBP, and targeted by a series of perturbed receptors: ITPR1-CAMK2A, WNT3-FZD4, FGF11-EGFR, and THRB. In the landscape of WNT signalling, the WNT3-FZD4 interaction might be a novel susceptible functional element in FTD. FZDs are G Protein-Coupled Receptors (GPCRs), involved in downstream DAG-IP3/Ca²⁺/cAMP signalling. WNT/FZD-mediated Phospholipases C (PLCs) activation leads to diacylglycerol (DAG) and inositol triphosphate (IP3) production, stimulating intracellular Ca²⁺ release. Calcium release activates CAMK2A, inhibiting the GSK3 kinase and causing nuclear accumulation of beta-catenin (CTNNB1) and TCF-mediated gene regulation [54–92]. Our model suggests that the disruption of this mechanism may involve CAMK2A directly, as a perturbed (i.e. disease-variant carrying) gene, and through its interaction with EP300-TCF7L2-CTNNB1. Beta-catenin has been implicated in many forms of neurological and cognitive impairments, including the Autism Spectrum Disorder, cell adhesion impairments, dendritic branching, and Long-Term Potentiation (LTP) [103]. Therefore, mutations in the CTNNB1 gene could lead to altered transcriptional activity and impaired synaptic plasticity that may result in brain malformation, intellectual disability, and neuronal loss [104].

TCF7L2 (a perturbed/seed node) is a beta catenin-interacting transcription factor [63,92] whose involvement in WNT signalling is well documented, although its physiological role in adult brain is still unclear [63]. This supports previous reports about calcium-dependent activation of CAMK2A and its implication in neurodevelopment, transcriptional regulation, cell fate determination [54,57,93–96], and neurodegeneration [70]. Furthermore, both WNT3-FZD4 and FGF11-EGFR interactions reveal a key role of the PLCs PLCB1, PLCB2, and PLCB3 within our FTD-network (Fig 4). FTD-associated variants in PLCs may lead to improper intracellular Ca²⁺ levels and thus dysregulation of the Ser/Thr kinases controlling the downstream MAPK-JNK signaling. Previous studies on PLCB1 showed an association between its depletion and early-onset epileptic encephalopathy [88], suggesting its involvement in learning impairments and neurodegeneration. PLCs activity was recently associated with the activation of the repair factor PARP1. Neuronal DNA is constantly exposed to ROS due to the intense mitochondrial activity in response to the high energy demand required by the Central Nervous System (CNS). This massive exposure to ROS causes the accumulation of single-strand breaks (SSBs) that remain unresolved, despite the presence of repair mechanisms [105]. In particular, aged cerebral neurons showed a group of genes targeted of PARP1-bound SSBs that are implicated in synaptic plasticity and long-term memory.

CAMK2A also controls other, peripheral, FTD-relevant targets. In our model, perturbance propagates from CAMK2A to calcineurin catalytic subunit PPP3CC through calmodulin CALM2 (Fig 2). Notably, reactive oxygen species (ROS) enhance LTP through calcineurin suppression, inducing cellular and DNA damage, and leading to cell death in neurons [70,80], with constitutively (as per genetic association) impaired metabolism. Together with active cell damage, loss of neuroprotection may contribute to neural cell death. CAMK2A is connected to three different perturbed routes, involving cAMP and fatty acids metabolism [84–91], and including several mitochondrial proteins. The first route involves the adenylate cyclase ADCY3, the phosphodiesterases PDE1A and PDE10A, and the GMP-deaminase GMPR (Fig 2). Several studies highlighted the importance of intracellular levels of cAMP and cGMP for neuroprotection and neuron survival [84–91], and CAMK2-dependent ADCY3 activation has been documented in mouse and rat brain [87]. Specifically, neuron regeneration, survival and synaptic plasticity seem to be enhanced by high intracellular levels of cAMP and cGMP [84–87], and antagonized by phosphodiesterases (PDEs) and myelin-associated glycoproteins (MAGs) [86–88]. In a recent study [88] PDE-dependent cAMP/cGMP control has been documented in Alzheimer’s disease (AD), depression and multiple sclerosis (MS), but not in FTD, to date.

The second CAMK2-secific route includes EP300, CPT1B, and ACSBG1 (Fig 2). These nodes are involved in fatty acid metabolism and all of them can be found in mitochondria [106–108]. Moreover, CPT1B has been recently associated with behavioural disorders characterizing post-traumatic stress both in human and rodent models [107], and ACSBG1 participate in myelinogenesis [108].

The third CAMK2-secific perturbed route is represented by the central network axis CAMK2A-EP300-TCF7L2, showing a bifurcation through CTNNB1 and JUN (Fig 2). We also identified several important junction proteins, such as Ca²⁺, IP3, and cAMP transporters, highlighting the perturbed status of both ncWNT and MAPK signalling pathways. The interaction PRKCG-GJD2-TJP1 is one of the perturbation sources acting on CTNNB1. PRKCG is a Ca²⁺-activated Ser/Thr protein-kinase C (PKC), which mutations are known to be associated with spinocerebellar ataxia, characterized by cognitive impairment, tremor, and sensory loss [109]. Enrichment analysis showed how this PKC is involved in ROCC and LTP. Our model highlights GJD2 and TJP1 as possible perturbed PRKCG interactors. Former evidences demonstrated a strong association of GJD2 and TJP1 with schizophrenia [110].

Stress-induced Ser/Thr-kinases and neural cell death

The largest sink in our model is JUN, and its largest incoming hub is MAPK8 (aka JNK1). The second module of our FTD-network is centred on these two components of the c-Jun N-terminal kinase (JNK) pathway (Fig 5). The two mostly perturbed interactions acting on JUN in the entire FTD-network are PRKG1-CREB3L2 (aka BBF2H7), and PRKACG-MAPK8; together with the ncWNT-related route CAMK2A-EP300-TCF7L2.

The cGMP-activated Ser/Thr-kinase PRKG1 is a key mediator of the nitric oxide (NO)/cGMP signalling pathway, involved in LTP and neuron branching [111]. CREB3L2, a PRKG1 target in the cGMP signalling pathway, is perturbed in our FTD-model. Remarkably, it has been demonstrated that CREB3L2, whose expression is induced by endoplasmic reticulum (ER) stress, is involved in preventing the accumulation of unfolded proteins in normal damaged neurons, and protecting neuroblastoma cells from ER-stress induced cell death [112].

On the other hand, PRKACG and MAPK8 are the FTD-connectors with the largest perturbed outgoing connectivity, and sharing one of the mostly perturbed interaction, indicating that they may play a critical role in FTD aetiology. MAPK8 is a key Ser/Thr kinase capable of phosphorylating JUN and prompting stress-induced cell apoptosis by phosphorylating other transcription factors, including TP53 [74–78], a direct MAPK8 target. Moreover, our model evidenced a large variety of potentially FTD-associated pro-apoptotic stress-responsive MAPK8 perturbed interactions and targets. The main target is the disease-variant carrying cytokine TNF, sharing with MAPK8 the second-largest perturbed interaction of the model (after MAPK8-PRKACG). Other important targets include: the Ser/Thr kinase MAPK11, the transcription factor FOXO1, the adapter protein CRK, the interleukin IL12B, the Ca²⁺-dependent phospholipase PLA2G4B, and the cytidylyltransferase PCYT1A. Conversely, MAPK8 is influenced by few perturbed interactions, among which the most important is PRKACG. The latter is the most degree-central Ser/Thr kinase in the FTD-model after MAPK8, and part of the perturbed route MAPK8-PRKACG-GRIN2B-PRKACB (Fig 5). PRKACG is a protein kinase A (PKA) involved in lipid and glucose metabolism, immune system response, and G1-checkpoint response during cell cycle. It can be dysregulated in response to different kind of stresses, including metabolic stress, DNA damage and cancer [20,78,93,100]. Alongside PRKACG, we found other protein kinases, NMDA receptors, and GPCRs that are involved in stress response, (neuronal) apoptosis, and critical neurological functions, such as LTP, learning, and behaviour [69,70,80,81,91–95,99,100]. They include: PRKACB, PRKCG, PRKG1, AKT3, PIK3R1, PIK3CA, GRIN2B, together with the EGFR, CAMK2A, and FZD4. PRKACG and PRKG1 also regulate the activity of the potassium channels KCNMB4 and KCNMA1, respectively, through perturbed interactions. These channels are responsible for membrane excitation and sensitivity to Ca²⁺ levels [113].

MAPK and NMDA receptor signalling, oxidative stress, and neurodegeneration in FTD

The roots of oxidative damage studies in biology date back to Harman’s free radical theory of aging, in 1954 [68]. The response to oxidative damage involves several different pathways regarding cell fate. We here found that the most ligand-receptors triggering perturbed interactions, control members of the MAPK-JNK signalling pathway. Oxygen radicals, present in ROS, induce phosphorylation of MAPK-related GPCRs or NMDA receptors (NMDARs), activating their coupled PLCs (i.e. PLCB1, PLCB2, and PLCB3), which in turn activate downstream cGMP-dependent kinases, such as PRKG1, the Ser/Thr kinases PKA (i.e. PRKACG, PRKACB), PKC (i.e. PRKCG), PKII (i.e. CAMK2A), or members of the PI-3K family (i.e. AKT3 and PIK3CA) [57,70,76]. A crucial NMDAR that emerged from our FTD model is GRIN2B, part of the NR2 subunit, and acting as glutamate agonist. It is the major excitatory receptor in the mammalian brain. Information exchange through the post-synaptic excitatory CNS neurons occurs mainly through the ionotropic glutamate NMDARs. Synaptic plasticity (i.e. the modulation of the activity of glutamatergic synapses) may persist over long periods of time, leading to the so called LTP. LTP leads to NMDARs activation and influx of Ca²⁺ ions in the post-synaptic membrane. These processes deeply affect key cellular processes for learning and memory [57,70,75].

According to our model, the oxidative stress-responsive Ser/Thr kinase PRKACG also influences MAPK8. MAPK8 interacts, in turn, with MAPK11 (aka p38-beta), a member of p38-MAPK family [75], involved in neuro-inflammatory processes preluding neurodegeneration. This is also supported by the highly perturbation of the osmotic stress-induced cytokine TNF, target of MAPK8 (Fig 5). Furthermore, MAPK8 may elicit neuronal cell death through activation of the perturbed target TP53 [73–76]. Another interesting MAPK8 target is FOXO1, involved in oxidative stress response and neuronal cell death [77]. In this respect, our current study supports previous work highlighting a role for oxidative damage and cell degeneration neurological disorders [68–83], including FTD [52,53].

Notably, our findings support neurodegeneration-associated mechanisms, particularly in astrocytes. Astrocytes are glial cells implicated in various neurodegenerative diseases, including Amyotrophic Lateral Sclerosis (ALS), Alzheimer’s Disease (AD), Huntington’s Disease (HD) and Parkinson’s Disease (PD), although not much is known to date in FTD [114]. In general, Astrocyte-mediated onset and progression of these disorders seems to be a consequence of loss of homeostatic functions (e.g., Ca²⁺/cAMP homeostasis, implicated in neuroprotection), and gain of disruptive functions (e.g. abnormal LTP and ROS accumulation) [115,116]. Interestingly, astroglial cells are known to modulate neuronal activity through glutamate release, causing an NMDAR-mediated increase in Ca²⁺ [117,118]. An excess of neuronal metabolic activity, accompanied by an impaired calcium homeostasis, due to astroglial-associated disorders, could make CNS synapses more sensitive to oxidative damage. As reported in the results section, oxygen radicals present in ROS induce phosphorylation of MAPK-associated NMDARs, activating their coupled PLCs, and consequently cGMP-dependent oxygen-responsive kinases (including PRKG1, PRKACG, PRKACB, PRKCG, and CAMK2A) [57,70,76]. The activity of these oxygen-responsive kinases could be the cause of abnormal LTP, neuroinflammatory response, and neuronal cell death [52,53,68–83].

Methodological and theoretical aspects

The core of our method consists in reducing the initial network to a minimum set nodes and edges bearing the disease-associated information content of the interactome (KEGG, in this case), filtered to exclude misleading interactions and noise. Although the initial interactome is given, it is then transformed and cross-connected, respectively, by perturbance and covariance-driven LVs inclusion. The last passage enables to also evaluate possible unobserved sources of co-variation. Several disease-relevant aspects apply to our ‘seed-and-expand’ method.

First, our approach assumes that genetic variability influences both single genes and their interaction. We expect that phenotype-associated genetic variants influence proteins’ chemical-physical properties and thus their functions as well as interactions. In this perspective, perturbance is a measure of deviation from the given (i.e. KEGG) interactome. Our supervised-PC1 scores of SNP-to-gene quantification reflects genetic variability at a given locus, therefore the concept of perturbance can be explained in terms of phenotype-associated (i.e. FTD-associated) genetic variation over gene-gene interactions. More in general, the key entity comprised in the definition of perturbance is the information transmitted through such interactions, from the source of perturbation to the final target. If the genotype is altered, also the biochemical properties of the effectors are influenced. Therefore, the genetic alterations that underlie edge perturbations are not limited to deep structural rearrangements, but my also include mild phenotypes that cause small changes in protein functionality and connectivity, resulting in functional impairments [18].

Second, given the scale-free nature of biological regulatory networks [28] one natural class of relevant nodes is represented by hubs, many of which are not perturbed, including FZD4, TJP1, EP300, JUN, and PRKACG (Figs 2, 3 and 4). This supports the hypothesis according to which hubs are essential for embryonic development and thus are less likely to accumulate disease-related variants that usually turn to be lethal [28,29]. Nevertheless, their functional involvement in the disease phenotype is critical since they send and receive most perturbed connections. However, our method recovered also perturbed hubs, especially those connecting different functional sub-modules, including TCF7L2 and MAPK8. Furthermore, seeds and novel perturbed genes are not necessarily the closest to hubs, but rather those which are mostly influenced by disease-related genetic variation. In our model this is shown through paths maximizing edge perturbation, allowing to avoid the classical and often misleading guilt-by-association rule and favouring disease-relevant (or in general phenotype-relevant) variation. In this scenario, the discovery of novel disease-associated genes is at the same time dependent on gene perturbation, local gene context (i.e. edge perturbation), and the whole disease-network topology, through the overall FTD-model fitting.

The final FTD-network revealed several useful notions: (i) the presence of significantly relevant routes, sources, connectors (including hubs), and targets of perturbance; (ii) the presence of latent “common cause(s)” underling hidden gene-circuits, which connect perturbed routes; (iii) the detection of novel FTD-biomarkers based on functional perturbed connectors. Of note, other than for effect of genetic data, the pipeline is also suited for assessing any kind of quantitative data, including eQTL, expression microarrays, and RNA-seq, in the context of complex disorders.

In conclusion, the current study allowed to expand our previous work [4], revealing the potential functional impact of the genetic variation associated with FTD in the Italian population, using a newly developed network model. Specifically, our study suggests that the genetics underpinning (Italian) FTD is associated with metabolic stress involving possible oxygen compound excess and Calcium/cAMP homeostasis deregulation. This happens through different complex molecular mechanisms, including Adenylate Cyclases, PDEs, stress-induced Ser/Thr-kinases, and WNT/EGFR and NMDA receptors malfunctioning. According to our hypothesis, these impairments lead to neuronal damage, loss of neuroprotection, and eventually to cell death. Moreover, by indicating these complex biological processes we also indicate their key functional players, i.e. those genes that we called novel potential FTD-biomarkers, which appear to be perturbed, and having synergic and small effect size.