Large-scale network analysis of the cerebrospinal fluid proteome identifies molecular signatures of frontotemporal lobar degeneration

Abstract

The pathophysiological mechanisms driving disease progression of frontotemporal lobar degeneration (FTLD) and corresponding biomarkers are not fully understood. Here we leveraged aptamer-based proteomics (>4,000 proteins) to identify dysregulated communities of co-expressed cerebrospinal fluid proteins in 116 adults carrying autosomal dominant FTLD mutations (C9orf72, GRN and MAPT) compared with 39 non-carrier controls. Network analysis identified 31 protein co-expression modules. Proteomic signatures of genetic FTLD clinical severity included increased abundance of RNA splicing (particularly in C9orf72 and GRN) and extracellular matrix (particularly in MAPT) modules, as well as decreased abundance of synaptic/neuronal and autophagy modules. The generalizability of genetic FTLD proteomic signatures was tested and confirmed in independent cohorts of (1) sporadic progressive supranuclear palsy-Richardson syndrome and (2) frontotemporal dementia spectrum clinical syndromes. Network-based proteomics hold promise for identifying replicable molecular pathways in adults living with FTLD. 'Hub' proteins driving co-expression of affected modules warrant further attention as candidate biomarkers and therapeutic targets.

Fig. 1. Study overview.

CSF was collected in 116 carriers of autosomal dominant mutations for FTLD (47 C9orf72, 32 GRN and 37 MAPT) and 39 non-carrier controls with a family history of genetic FTLD. CSF was analyzed on a modified aptamer-based assay (SomaScan). After data processing, a total of 4,138 proteins were quantified. High-dimensional proteomic data were organized into modules of protein co-expression using WGCNA. CSF protein co-expression modules from the genetic FTLD network were functionally annotated using gene set and cell type enrichment approaches. CSF genetic FTLD modules were examined in relation to cross-sectional (CDR + NACC-FTLD, CSF NfL and bilateral frontotemporal volume) and longitudinal (global cognitive trajectories) indicators of disease severity. On cross-cohort and cross-platform validation analyses, genetic FTLD modules were reconstructed in independent cohorts of sporadic PSP-RS and controls (4RTNI; SomaScan) and FTLD, AD and controls (BioFINDER 2; Olink). Figure created with BioRender.

Fig. 2. Genetic FTLD protein co-expression network.

a, A CSF protein co-expression network was built using WGCNA. The genetic FLTD network consisted of 31 protein co-expression modules. Module relatedness is shown in the dendrogram (right). GO analysis was used to identify the principal biology represented by each module. Within genes, module eigenproteins in symptomatic (Sx) and presymptomatic (PreSx) variant carriers were compared against controls. Increased eigenprotein abundance in FTLD is indicated in green, whereas decreased eigenprotein abundance is indicated in blue. Module eigenproteins were correlated with disease outcomes, including CDR + NACC-FTLD, global cognitive slope, CSF NfL and bilateral frontotemporal volumes (red, positive correlation; blue, negative correlation). The cell type nature of each module was assessed by module protein overlap with cell type-specific marker lists of neurons, oligodendrocytes, astrocytes, microglia and endothelia. The asterisks in the left heat map indicate statistical significance after one-way ANOVA with Tukey’s test (two tailed) ***Tukey P < 0.001, **Tukey P < 0.01 and *Tukey P < 0.05. The asterisks in the middle and right heat maps indicate statistical significance after FDR correction: ***FDR P < 0.001, **FDR P < 0.01 and *FDR P < 0.05. The exact P values are reported in Supplementary Table 4. b, Module eigenprotein levels by case status for six of the most strongly FTLD-associated modules. Individual eigenprotein values are plotted in controls (N = 39) and mutation carriers, which were grouped by Sx and PreSx status (C9orf72: 24 PreSx, 23 Sx; GRN: 12 PreSx, 19 Sx; MAPT: 18 PreSx, 19 Sx). Red P values are statistically significant. Differences in module eigenprotein by case status were assessed by one-way ANOVA with Tukey’s test. Gene-specific P values represent the omnibus significance for gene-stratified comparisons versus controls, with significant effects displayed with red text. The box plots represent the median and 25th and 75th percentiles, and box hinges represent the interquartile range of the two middle quartiles within a group. Min and max data points define the extent of whiskers (error bars). C9, C9orf72; CTL, control.

Fig. 3. Module and hub protein relationships with cognitive trajectory.

The plots display the top five CSF genetic FTLD network modules most strongly associated with global cognitive trajectories in the full sample. Eigenprotein z scores are plotted against the annual rate of global cognitive change during the study period (n = 137). Person-specific cognitive slopes were extracted from linear mixed-effects models that included baseline demographics (age, sex and education) and time (years since baseline). Regression fit lines with 95% confidence intervals are plotted alongside Spearman’s ρ values and two-tailed P values. Information on the association between all network module eigenproteins and cognitive trajectory in the full sample, within each gene and within presymptomatic mutation carriers is provided in Supplementary Table 8. For proteins assigned to each module, an individual protein’s strength of connectivity to the module (x axis) is plotted against the individual protein’s correlation with global cognitive change (y axis). Proteins that exhibited stronger intramodular connectivity also exhibited stronger relationships with cognitive slope. Color-filled triangles represent individual proteins that survived FDR correction in two-tailed proteome-wide differential correlational analyses (n = 646 total proteins; full list in Supplementary Table 8). Proteins in the top 20th percentile of intramodular connectivity are classified as ‘hub’ proteins. Hub proteins that significantly correlated with cognitive trajectory are listed with each plot.

Fig. 4. Cross-cohort validation.

Validation cohorts included 4RTNI, composed of PSP and controls, and BioFINDER 2, composed of patients with FTLD clinical syndromes, biomarker-confirmed AD and controls. The 4RTNI CSF samples were assayed with SomaScan and BioFINDER CSF samples were assayed with Olink. a, WGCNA was applied to 4RTNI SomaScan data to test for module preservation across the genetic FTLD and sporadic PSP-RS networks. Modules that have a z_zsummary score greater than or equal to 1.96 (or q =0.05, dotted blue line) are considered to be preserved, and modules that have a z_summary score greater than or equal to 10 (or q = 1 × 10⁻²³, dotted red line) are considered to be highly preserved. All modules in the genetic FTLD network were highly preserved in the sporadic PSP-RS network. b, Synthetic eigenproteins were reconstructed in 4RTNI and BioFINDER 2 to test for concordance in module relationships with disease groups. The heat map displays average synthetic eigenprotein z score differences between PSP-RS and controls (CTL), FTLD versus CTL, FTLD versus AD and AD versus CTL. The asterisks indicate statistical significance after pairwise two-sided t-test with FDR correction. ***FDR P < 0.001, **FDR P < 0.01 and *FDR P < 0.05. The exact P values are reported in Supplementary Tables 9 and 10. c,d, Synthetic eigenprotein box plots for key modules from 4RTNI (35 CTL and 30 PSP-RS) (c) and BioFINDER 2 (248 CTL, 58 AD and 29 FTLD) (d) analyses. Pairwise differences in module synthetic eigenproteins by case status were assessed by two-sided t-tests with FDR correction. The box plots represent the median and 25th and 75th percentiles, and box hinges represent the interquartile range of the two middle quartiles within a group. Min and max data points define the extent of whiskers (error bars).

Fig. 5. Genetic FTLD CSF network module ORA with AD CSF networks.

Module member ORA of the genetic FTLD CSF network with two CSF protein co-expression networks in AD. Top: a multiplatform network, obtained using SomaScan and TMT-MS (Dammer et al.). Bottom: a single platform network obtained using TMT-MS (Modeste et al.). The box values represent −log₁₀(FDR) value for pairwise module overlap, determined using one-tailed Fisher’s exact test. Bolded AD network modules significantly differed between AD and controls. Modules from the AD networks (y-axis rows) without an overlap value of −log₁₀(FDR) >1 are not included in the heat maps. ***FDR P < 0.005, **FDR P < 0.01 and *FDR P < 0.05. The exact P values are reported in Supplementary Table 12.

Extended Data Fig. 1. Differential Abundance.

Gene-stratified differential abundance analyses examined individual protein differences across symptomatic (Sx) carriers, presymptomatic (PreSx) carriers, and familial non-carrier controls (CTL). Pairwise differential protein abundance is represented by volcano plots of average log₂ difference by negative log₁₀ p-value for each given comparison. Proteins are colored by the module in which they were assigned according to the heatmap shown in Fig. 2a. Annotated proteins represent the top 10 differentially abundant proteins by statistical significance for each direction. Pairwise comparisons were performed using one-way ANOVA with Tukey test. The red horizontal dashed line represents Tukey p < 0.05.

Extended Data Fig. 2. ROC Curves.

Gene-stratified receiver-operating characteristic (ROC) analyses tested the discriminative utility of differentially abundant proteins across symptomatic (Sx) carriers, presymptomatic (PreSx) carriers, and familial non-carrier controls (CTL). ROC curves for the top 5 performing proteins, defined by area under the curve (AUC), are plotted alongside a curve derived from a combined panel of the top 5 proteins.

Extended Data Fig. 3. Correlation of Genetic FTLD-associated Protein Effect Sizes with External AD and PD Datasets.

a-f) Log₂-fold changes of CSF SomaScan proteins with differential abundance in symptomatic FTLD gene carriers versus control (CTL) were correlated against CSF SomaScan protein effect sizes for AD versus CTL (A, C, E; log₂-fold changes reported in Dammer et al.) and PD versus CTL (B, D, F; regression coefficient weights reported in Rutledge et al.). Correlations were performed using Pearson correlation. g) Genetic FTLD network module protein overlap with proteins increaed or decreased in AD or PD datasets. One-tailed Fisher’s exact test was used to determine module-wise enrichment, and results FDR-corrected using the Benjamini-Hochberg method. The exact p-values are reported in Supplementary Table 11.