Multiscale causal networks identify VGF as a key regulator of Alzheimer's disease

Abstract

Though discovered over 100 years ago, the molecular foundation of sporadic Alzheimer's disease (AD) remains elusive. To better characterize the complex nature of AD, we constructed multiscale causal networks on a large human AD multi-omics dataset, integrating clinical features of AD, DNA variation, and gene- and protein-expression. These probabilistic causal models enabled detection, prioritization and replication of high-confidence master regulators of AD-associated networks, including the top predicted regulator, VGF. Overexpression of neuropeptide precursor VGF in 5xFAD mice partially rescued beta-amyloid-mediated memory impairment and neuropathology. Molecular validation of network predictions downstream of VGF was also achieved in this AD model, with significant enrichment for homologous genes identified as differentially expressed in 5xFAD brains overexpressing VGF. Our findings support a causal role for VGF in protecting against AD pathogenesis and progression.

Fig. 1. Pipeline overview.

Large-scale, high-dimensional datasets generated in hundreds of subjects serve as the input into our integrative pipeline (a), which comprises a series of steps that first generate the appropriate input features for causal network reconstructions (b, c), then network reconstruction and identification of key driver genes (d), and finally validation via three independent paths: replication (e), human genetic association (f), and experimental disease model (g).

Fig. 2. Characterization of AD traits, and brain gene and protein expression.

a Canonical correlation heatmap of disease traits. The intensity of the red color indicates the strength of correlation between traits; the canonical correlation is indicated in each box. The x- and y-axis represent the traits: clinical dementia rating (CDR), Braak score (bbscore), clinical neuropathology (PATH.Dx), neuropathology category (NP.1), CERAD neuropath criteria (CERJ), and mean neocortical plaque density (number of plaques/mm², PlaqueMean). b, e Breakdown of DE genes (b) and proteins (e). This figure shows the UpsetR plot (“Methods”) of the DE genes or proteins overlapping across tests. The bars represent the size of the DE sets and the points represent the category to which each set belongs. c, f PlaqueMean DE genes (c) and proteins (f). The x-axis is the mean normalized count for each gene or protein and the y-axis is their log fold change. Blue and red genes and proteins correspond to FDR ≥ 0.05 and FDR < 0.05, respectively. The strongest DE genes are highlighted on the plot. d GO term enrichment across all signatures. The heatmap depicted represents the −log 10(FDR) of the top 5 most significant GO terms associated with signatures across all traits. Rows are GO terms and columns signatures. g MsigDB pathway enrichment for all signatures. The barplot represents the union of the top 10 significant MsigDB categories associated to signatures for all traits; x- and y-axis are MsigDB terms and the −log 10(FDR), respectively. Colors represent the individual traits.

Fig. 3. Co-expression network analyses.

a, c Top GO annotations for gene (a) and protein (c) co-expression modules. The x- and y-axis represent the best GO term associated with each module and the −log 10(FDR) of the enrichment, respectively. The color of the bars represents the module names. In bold are the four modules enriched for genes in the union of the DE signatures. b, d Module enrichments for DE genes (b) and proteins (d). The circos plot depicts the enrichment of each module for DE genes or proteins; the “hotness” of the color represents the magnitude of the −log 10(adjusted p value) of the enrichment for the corresponding signature list. The traits are defined as 1 through 12; in bold are the modules enriched for the union of the DE signatures.

Fig. 4. Bayesian causal networks and key drivers.

a Full Bayesian network and an APOE subnetwork. Background visualization of the multiscale AD network described in the main text using an edge-weighted spring embedded layout. The red nodes are proteins and the blue nodes genes. Key driver (KD) genes and proteins are highlighted in yellow. Foreground multiscale subnetwork comprised of genes within a path length of 3 to APOE. Node names and properties are defined in the panel legend. b Density plots of the distribution of pLI scores for genes and proteins by the number of times they appear as global KDs across the three discovery and four replication networks. The yellow dashed line represents the median pLI score for that category. c Distribution of the number of times genes and proteins across all three discovery networks were identified as KDs across all DE signatures. The x- and y-axis depict the different KDs appearing in at least two networks and the number of times they are identified as KDs for DE signatures across all three networks. The colors of the bars are indicative of the network of origin of the KDs. d KD of DE signatures in the multiscale network, as described for (b). The color of the bars is indicative of KDs presence only in the gene expression, protein expression, or in both.

Fig. 5. Characterization of AD pathophysiology in wild-type, 5xFAD, and 5xFAD mice overexpressing VGF.

a Immunohistochemical staining of Aβ amyloid plaques and microglial cells in the male mouse cortex of 5xFAD mice overexpressing VGF in the germline. Left panel, green: Aβ (6E10), red: Iba-1, blue: DAPI; right panel, quantification of percent area of Aβ and Iba-1 staining in male and female mice. Quantification of percent area of Aβ and Iba-1 staining in the cerebral cortex, hippocampal CA3, and hilus; data are presented as mean percentage ± SEM of the control group. One-way ANOVA with Newman–Keuls post hoc analysis, cortex (anti-Aβ): F_(3,86) = 30.84, p < 0.0001, CA3 (anti-Aβ): F_(3,86) = 12.44, p < 0.0001, cortex (anti-Iba-1): F_(3,56) = 7.307, p = 0.0003, n = 9, 8, 7, 6 mice per group, 2–3 slices analyzed per animal, *p < 0.05, **p < 0.01, ***p < 0.001; female: n = 7, 6 mice per group, two-sided Student’s t test, p = 0.031 (Aβ), p = 0.0454 (Iba-1). b Doublecortin staining (DCX) of the subgranular zone (SGZ) in the dentate/hilus area of male 5xFAD brains. Upper panel, red: DCX, blue: DAPI; lower panel, average number of DCX-positive cells per subgranular zone. One-way ANOVA with Newman–Keuls post hoc analysis, male: F_{(2, 21)} = 6.652, p = 0.0058, n = 4, 4, 4 mice per group, 2 slices analyzed per animal; female: F_{(2, 21)} = 7.008, p = 0.0047, n = 10, 9, 5 mice per group. **p < 0.01 c Reduced staining of phosphor-Tau and dystrophic neurite clusters in 5xFAD brains with germline VGF overexpression. Upper panel: phosphor-Tau staining; lower panel: quantification results of dystrophic neurite clusters in the hippocampus and cortical area. One-way ANOVA with Newman–Keuls post hoc analysis, cortex: F_{(2, 15)} = 10.92, p = 0.0012, hippocampus: F_{(2, 15)} = 5.549, p = 0.0157, n = 7, 7, 4 male mice/per group. *p < 0.05, ***p < 0.001. d Barnes maze test. Mice were trained daily and WT mice learned the target quarter (TQ) of the hiding zone by increased distance traveled in the TQ (left panel) and increased time spent in the TQ (right panel). 5xFAD mice showed impaired spatial learning on day 4, while germline VGF overexpression (5xFAD,VGF + /Δ) partially restored memory performance. N = 12–14 mice (male + female) per group. Data were analyzed by two-way repeated-measures ANOVA. % of distance spent in TQ: Days (F_(3,108) = 3.215, p < 0.05) and Groups (F_(2,36) = 8.77, p < 0.001), and Days × Groups interaction (F_(6,108) = 1.9, p = 0.0873). % time spent in TQ: Days (F_(3,105) = 2.422, p = 0.07) and Groups (F_(2,35) = 20.01, p < 0.0001), and Days × Groups interaction (F_(6,105) = 4.501, p < 0.001). Tukey’s post hoc test. ^#p < 0.05, **p < 0.01, ****p < 0.0001. All data in b–d are presented as mean percentage ± SEM.

Fig. 6. Characterization of AD pathophysiology in 5xFAD mice with and without AAV5-VGF-driven overexpression of VGF.

a Immunohistochemical staining of Aβ amyloid plaques and VGF in the 5xFAD mouse brain 4 months after AAV5-VGF or AAV5-GFP infusion into the dorsal hippocampus. Left panel, red: VGF, cyan: Aβ, green: GFP; right panel, quantification of percent area of Aβ amyloid plaque in different brain areas. N = 4, 5 male mice per group. Data are presented as mean percentage ± SEM (of the control group, two-sided Student’s t test. **p = 0.004, *p = 0.0121. b Doublecortin staining (DCX) in the dentate/hilus area. Upper panel, red: DCX, blue: DAPI; lower panel, average number of DCX-positive cells per subgranular zone. N = 4–5 male mice per group. Data were analyzed by one-way ANOVA with Newman–Keuls post hoc analysis, F_{(2, 23)} = 6.574, p = 0.0055, n = 4, 4, 5 male mice per group, 2 slices analyzed per animal. *p < 0.05, **p < 0.01. c Reduced staining of phosphor-Tau and reduction of dystrophic neurite cluster number and diameter in 5xFAD brains with AAV5-VGF overexpression. Upper panel, phosphor-Tau staining; lower panel, quantification results of dystrophic neurite cluster number and diameter in the dorsal hippocampus. N = 4, 5 male mice per group. Data were analyzed by two-sided Student’s t test. *p = 0.0162, **p = 0.0021. d Barnes maze test. Mice were trained daily and on Day 4 WT mice learned the target quarter (TQ) of the hiding zone, as revealed by increased distance traveled in the TQ (left panel), and increased time spent in the TQ (right panel). 5xFAD mice with AAV5-GFP showed impaired spatial learning on day 4, while in 5xFAD with AAV5-VGF overexpression, memory performance was significantly rescued. N = 12, 9, 10, 7 mice (male + female) per group. Data were analyzed by two-way repeated-measures ANOVA. % of distance spent in TQ: Days (F_(3,102) = 5.000, p < 0.01) and Groups (F_(3,34) = 5.997, p < 0.01), and Days × Groups interaction (F_(9,102) = 2.371, p < 0.05). % time spent in TQ: Days (F_(3,102) = 11.39, p < 0.0001) and Groups (F_(3,34) = 13.62, p < 0.0001), and Days × Groups interaction (F_(9,102) = 3.824, p < 0.001). Tukey’s post hoc test. ^#p < 0.05, **p < 0.01, ^$$$p < 0.001, ****^,$$$$p < 0.0001. e Impaired DHPG-mediated long-term depression (LTD) in 5xFAD mice is partially restored by AAV-VGF expression in the dHc. N: WT (AAV-GFP) = 8 slices from seven mice; 5xFAD (AAV-GFP) = 12 slices from six mice; WT (AAV-VGF) = 8 slices from six mice; 5xFAD (AAV-VGF) = 9 slices from five mice. f Summary graph of data from e indicating the average fEPSP slope [mV/ms (% of baseline)] during the last 5 min of recording. Data were analyzed by two-way ANOVA. Slope mV/ms (% of baseline): Genotype (F_(1,34) = 9.396, p < 0.001) and Groups (AAV-VGF and AAV-GFP) (F_(1,34) = 0.3282, p = 0.5705) and Genotype × Groups interaction (F_(1,34) = 5.045, p < 0.01). Newman–Keuls post hoc test. *p < 0.01, **p < 0.001. All data in b–f are presented as mean ± SEM.

Fig. 7. Chronic i.c.v. administration of TLQP-62 peptide ameliorated pathophysiological changes in the 5xFAD mouse brain.

a Immunohistochemical staining of Aβ amyloid plaques and microglial cells in the male 5xFAD mouse cortex and dentate gyrus after 28-day i.c.v. administration of TLQP-62 peptide or vehicle control (aCSF). Red: Aβ (6E10), green: Iba-1. b Quantification of percent area of Aβ and Iba-1 staining in both peptide-treated male and female 5xFAD mouse brains. Data are presented as mean percentage ± SEM. Results of Aβ(6E10) staining were analyzed by two-sided Student’s t test. Male: cortex, p = 0.0015; DG p = 0.0316; female: cortex, p = 0.0101; CA1, p = 0.022. Iba-1 staining were analyzed by one-way ANOVA with Newman–Keuls post hoc analysis, male cortex: F_(2,36) = 8.449, p = 0.001, n = 4, 5, 4 mice per group, 3 slices analyzed per animal; female cortex: F_(2,14) = 12.53, p = 0.0008, n = 5, 7, 5 mice per group, *p < 0.05, **p < 0.01, ***p < 0.001. c Reduced staining of Lamp1-immunoreactive dystrophic neurite cluster number in 5xFAD brains after 28-day TLQP-62 i.c.v. infusion. Red: Lamp1, green: 6E10, blue: DAPI. N = 6, 6 male mice per group. Data are presented as mean ± SEM and analyzed by two-sided Student’s t test. *p = 0.024.

Fig. 8. Molecular validation of VGF.

a Consensus subnetwork within a path length of 2 of VGF. The consensus subnetworks around VGF, two steps away from VGF, across all three networks are depicted. The blue and red nodes are genes and proteins, respectively. The blue edges originate from the gene-only network, the red edges from the protein only network, and the purple edges from the multiscale network. VGF and its known partners are in bold in the plot. b Density plot of the distribution of differential expression nominal p values for genes downstream and not downstream (causally independent of the expression levels) of VGF in the gene-only network for mouse DE genes (5xFAD, AAV5-GFP versus 5xFAD, AAV5-VGF brains). The x-axis is the −log 10(p value) for differential expression, and the y-axis represents the densities at the different −log10(p value). The red and blue curves are for genes downstream and not downstream of VGF in the network, respectively. c Summary of DE results of VGF network genes in the 5xFAD, AAV5-GFP versus 5xFAD, AAV5-VGF brains overlaid on the VGF gene-only subnetwork. The nodes are colored by log fold change from green (negative) to orange (positive). The size of the node represents the DE FDR. Gray genes names are not significantly DE and white nodes have no orthologous genes in mice.