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

The pathophysiological mechanisms driving disease progression of frontotemporal lobar degeneration (FTLD) and corresponding biomarkers are not fully understood. 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, MAPT) compared to 39 noncarrier 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 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.


Introduction
Frontotemporal lobar degeneration (FTLD) is among the most common causes of early age of onset dementia (< 65 years), resulting in progressive clinical syndromes that can feature behavioral, language, and/or motor de cits 1,2 .The majority of FTLD is comprised of two major molecular subtypes, de ned by pathological accumulation and cellular inclusions of hyper-phosphorylated tau (FTLD-tau) or transactive response DNA-binding protein 43 kDa (FTLD-TDP) 3 .Current in-vivo molecular biomarkers to aid in diagnosis of sporadic forms of FTLD are limited to nonspeci c indicators of neurodegeneration and axonal loss (e.g., neuro lament light chain [NfL]), as well as markers of Alzheimer's disease (AD) pathology (e.g., Aβ 42/40 , P-tau181) to rule out AD.Although NfL predicts future clinical decline across FTLD clinical syndromes [4][5][6] , a single biomarker does not capture the biological complexity of FTLD and it remains unclear which biological pathways most strongly correlate with disease progression 7 .Molecular signatures that are detectable in-vivo, inform underlying pathophysiology, and contribute to clinical progression are urgently needed to advance understanding of FTLD pathogenesis and guide biomarker development, which will in turn enhance FTLD clinical trials through improved selection of patients and measurement of treatment effects 8 .
Underlying FTLD pathology can be con dently predicted in inherited cases caused by autosomal dominant mutations, which represent approximately 20-30% of all cases 9 .Observational study of FTLD mutation carriers has provided a unique opportunity to chart clinical and biomarker changes across presymptomatic and symptomatic stages of FTLD 5,8,10 .Autosomal dominant forms of FTLD-TDP are most often caused by mutations in chromosome 9 open reading frame 72 (C9orf72 [C9]) or progranulin (GRN).The majority of autosomal dominant FTLD-tau is due to mutations in microtubule-associated protein tau (MAPT) 5,9 .Current understanding of the molecular mechanisms underlying autosomal dominant FTLD has been primarily ascertained from postmortem and non-human studies 3,11 .
Hexanucleotide repeat expansions in the C9orf72 gene may produce both loss-of-function (haploinsu ciency) and toxic gain-of-function (e.g., repeat RNA foci, dipeptide repeat protein aggregates) mechanisms that impair RNA metabolism, nucleocytoplasmic transport, and proteostasis pathways that contribute to nuclear depletion and cytoplasmic aggregation of TDP-43 12,13 .GRN mutations cause haploinsu ciency of progranulin, a lysosomal protein that regulates immune and autophagy pathways implicated in TDP-43 dyshomeostasis 14,15 .MAPT mutations alter tau binding to microtubules and promote tau aggregation, which directly compromises axonal transport, synaptic signaling, cellular metabolism, and the extracellular scaffolds that support neuronal structures [16][17][18] .
Recent large-scale proteomics studies in AD are contributing to discovery of key biological networks and individual proteins re ective of AD pathophysiology, which extend far beyond the classic hallmarks of misfolded amyloid and tau 19,20 .These studies have employed powerful systems biology approaches in brain tissue and bio uids (e.g., cerebrospinal uid [CSF], blood) to identify communities of co-expressed proteins that are reliably altered in clinical cohorts that span the AD disease severity continuum [20][21][22] .Despite the potential to identify novel molecular signatures underlying disease progression, unbiased large-scale analysis of the bio uid proteome in living FTLD patients has lagged.
The present study employed a large-scale aptamer-based proteomic approach (SomaScan) to generate a CSF protein network in 116 participants with autosomal dominant FTLD mutations (C9orf72, GRN, MAPT).Network analysis identi ed dysregulated communities of CSF proteins (modules) involved in RNA splicing, protein degradation, extracellular matrix, and synaptic pathways.These modules were strongly correlated with cross-sectional and longitudinal markers of FTLD clinical progression.We validated the generalizability of our genetic FTLD protein network by demonstrating preservation of network module alterations in CSF of progressive supranuclear palsy (PSP), a sporadic 4-repeat FTLD tauopathy.We further evaluated the disease-speci city and biologic reproducibility of protein network associations by reconstructing modules in a third clinical cohort of sporadic and genetic FTLD and AD using an orthogonal proximity-extension assay based proteomic approach (Olink).Collectively, our approach and results provide new insights regarding in-vivo molecular signatures of central nervous system dysfunction that occur across the disease severity spectrum of FTLD.
Genetic FTLD Protein Co-expression Network.The study design is presented in Fig. 1.Using weighted gene correlational network analysis (WGCNA), we built a protein co-expression network from 4,138 proteins across 155 CSF samples to identify protein communities that are dysregulated in FTLD (Fig. 2A).Network analysis revealed 31 protein co-expression modules across controls and mutation carriers.
Module size ranged from 360 proteins (M1) to 48 proteins (M31).Protein module membership assignments are provided in Supplementary Table 2. Based on gene ontology and cell type enrichment analyses (Supplementary Table 3; Supplementary Fig. 1), 28 out of the 31 modules had a clear primary ontology that was used for annotation, while 3 modules did not evidence clear ontology (labeled "ambiguous").
CSF proteomic signatures of FTLD disease severity.Analyses rst combined FTLD mutation carriers across all 3 genes (N = 116) to maximize sample size.Module eigenproteins, calculated as the rst principal component of module abundance, were compared across symptomatic mutation carriers, presymptomatic mutation carriers, and controls using one-way ANOVA with Tukey's post-hoc correction, adjusting for age and sex.Module eigenproteins were also correlated with continuous indicators of functional severity (CDR®+NACC-FTLD sum of boxes), adjusting for p-values by false discovery rate (FDR) correction according to the Benjamini-Hochberg method.Seven of the 31 modules showed signi cant differences between symptomatic carriers and controls (Supplementary Table 4).Four of these modules exhibited particularly strong relationships with functional severity: M26 spliceosome, M2 presynapse, M28 synapse assembly/axon, and M22 autophagy.M26 spliceosome, highly enriched for proteins involved in mRNA splicing and nuclear transport, was markedly increased in symptomatic carriers vs. controls and positively correlated with functional severity (ρ = 0.41, FDR-p = 2.7e-6).M2 presynapse and M28 synapse assembly/axon, enriched for synaptic/neuronal (M2, M28) and oligodendrocyte (M28) cell-type markers, were decreased in symptomatic carriers vs. controls and negatively correlated with functional severity (M2: ρ=-0.33,FDR-p = 2.6e-4; M28: ρ=-0.35,FDR-p = 1.7e-4).M22 autophagy was also decreased in symptomatic carriers vs. controls and negatively correlated with functional severity (ρ=-0.33,FDR-p = 2.6e-4).CSF proteomic alterations were more subtle and less likely to survive post-hoc correction when comparing presymptomatic mutation carriers vs. controls.However, M9 ion transport, enriched for neuronal cell-type markers involved in ion channel activity and transport, was signi cantly decreased in presymptomatic carriers vs. controls.
We next examined how network modules associated with CSF NfL, an indicator of neuroaxonal degeneration, given that NfL is currently the most validated uid biomarker for disease monitoring in genetic and sporadic FTLD 4,23,24 .Six modules associated with CSF NfL concentrations, measured by Simoa, at an unadjusted p < 0.05, with only M26 spliceosome surviving FDR-correction (ρ = 0.32, FDR-p = 9.3e-3).Upon closer examination, we observed that the SomaScan target for NfL (NEFL) was assigned to the M26 spliceosome module.However, NEFL was not a strong driver of M26 eigenprotein co-expression (intramodular kME = 0.32 [87th out of 88 module members]) and the relationship between M26 spliceosome and CSF NfL (Simoa) persisted when recalculating the M26 spliceosome eigenprotein without NEFL (ρ = 0.31, p = 9.1e-3).
Gene-strati ed FTLD proteomic signatures.We performed gene-strati ed analyses to determine whether proteomic signals observed in the full sample were driven by speci c gene groups.For each gene group, we again compared module eigenprotein levels across symptomatic mutation carriers, presymptomatic mutation carriers, and controls using Tukey's post-hoc correction.We also examined correlations with CDR®+NACC-FTLD sum of boxes and CSF NfL (Supplementary Table 4; Fig. 2B).FDR corrections were not applied to gene-strati ed correlational analyses given the reduced sample sizes.Within each gene group, M26 spliceosome levels were increased and M22 autophagy module levels were decreased in symptomatic carriers vs. controls.However, symptomatic C9orf72 and GRN carriers exhibited more pronounced alterations in M26 and M22 (0.85 to 1 difference in z-score vs. controls) than symptomatic MAPT (0.5 to 0.6 difference in z-score vs. controls).Presymptomatic GRN carriers also showed elevations in M26 spliceosome that were not present in presymptomatic C9orf72 or MAPT mutation carriers.M2 presynapse and M28 synapse assembly/axon were decreased in symptomatic GRN and MAPT, with both modules exhibiting particularly strong relationships with CDR®+NACC-FTLD and CSF NfL in MAPT (M2: CDR®+NACC-FTLD ρ=-0.59,p = 1.0e-4;NfL ρ=-0.54,p = 8.0e-4; M28: CDR®+NACC-FTLD ρ=-0.43,p = 7.4e-3; NfL ρ=-0.46,p = 5.0e-3).M9 ion transport, which showed the strongest presymptomatic signal in full sample analyses, was decreased among both presymptomatic C9orf72 and MAPT carriers vs. controls.
Several additional module expression patterns emerged in gene-strati ed analyses that were not present when examining modules in all mutation carriers.A cluster of modules adjacent to M26 spliceosome (M24 ubiquitination/translation, M25 protein folding/metabolism, M27 metabolism) were increased and/or positively correlated with disease severity in both symptomatic C9orf72 and GRN.M29 extracellular matrix (ECM), enriched for ECM proteins and microglial cell-type markers, was uniquely elevated in symptomatic MAPT carriers vs. controls.M4 complement/coagulation was closely related to M29 and also selectively elevated in symptomatic MAPT vs. controls.M3 postsynapse/glycosylation did not signi cantly differ between mutation carrier groups and controls, but was one of the strongest correlates of disease severity measures in MAPT (CDR®+NACC-FTLD ρ=-0.53,p = 7.0e-4; NfL ρ=-0.46,p = 5.9e-3).Differential Abundance Analysis.To complement network analyses, we examined differential abundance of all 4,138 proteins to identify individual CSF proteins driving module-level differences in gene-strati ed comparisons of symptomatic carriers, presymptomatic carriers, and controls (Supplementary Table 5, Extended Data Fig. 1).M26 spliceosome had the highest proportion of individual differentially abundant proteins in symptomatic C9orf72 and GRN carriers (Supplementary Table 6), including nuclear proteins TRA2B, TMPO, and HNRNPAB.Proteins assigned to modules with neuronal enrichment were also differentially abundant across multiple gene groups vs. controls, including established synaptic markers NPTX2 (M2 presynapse), CNTNAP2 (M28 synapse assembly/axon), DLG4 (M3 postsynapse/glycoslyation), and 14-3-3 proteins (e.g., YWHAZ; M19 neuron migration).Symptomatic MAPT exhibited the highest proportion of differentially abundant proteins from these neuronal modules, particularly M28 synapse assembly/axon.M26 spliceosome had the highest proportion of differentially abundant protein members in presymptomatic GRN (e.g., XRCC6) and M9 ion transport had the highest proportion of differentially abundant protein members in presymptomatic C9orf72 and MAPT (e.g., KCNE2, GPR6).
GRN mutations are characterized by progranulin protein haploinsu ciency.As expected, CSF progranulin protein (unassigned to a module) exhibited the largest decreased abundance in both pre-and symptomatic GRN carriers vs. controls.C1QTNF1 (unassigned to a module), another lysosomal protein with CSF levels shown to be strongly colinear with GRN levels 25 , exhibited a similar pattern of decreased abundance in GRN carriers.
In differential correlational analyses, 646 proteins were signi cantly correlated with global cognitive trajectories in the full sample (FDR-p < .05;Supplementary Table 6).Over half of those differentially expressed proteins were assigned to modules that also strongly associated with cognitive trajectories (M26, M29, M2, M3, M28).To determine whether individual proteins linked to cognitive trajectories were also of high in uence ('hub proteins') in these target modules, we plotted individual protein correlations with cognitive slope against their intramodular connectivity (intramodular kME).Proteins that were module hubs, de ned as top 20th percentile of intramodular kME 20 , and had FDR-corrected signi cance with cognitive slope are highlighted in the Fig. 3 inset protein lists.Notable 'hub' proteins included the neuronal pentraxins, NPTX2 (M2 presynapse; largest effect on cognitive slope across all proteins) and NPTX1 (M28 synapse assembly/axon), as well as neuroligins NLGN1 and NLGN2 (M3 postsynapse/glycoslyation).Other notable 'hub' proteins linked to cognitive trajectories included transmembrane proteins TMEM106B (M3 postsynapse/glycoslyation) and TMEM132B (M2 presynapse), ECM-linked proteins FSTL1 and TIMP1 (M29 ECM), and nuclear proteins HNRNPA1 and RECQL (M26 spliceosome).
Genetic FTLD modules are preserved in sporadic PSP-RS.We applied the same WGCNA methods to CSF SomaScan data from an independent cohort of individuals with sporadic PSP-RS 26 to determine whether protein co-expression patterns identi ed in genetic FTLD could be reproduced in sporadic FTLD, speci cally a sporadic FTLD tauopathy (Supplementary Table 8B).All 31 modules from the genetic FTLD cohort were highly preserved in the sporadic PSP-RS network (all Z summary scores > 10 [> q = 1 x 10 − 23 ]), indicating strong consistency of the CSF proteomic correlational architecture across genetic and sporadic disease (Fig. 4A).Next, we reconstructed the genetic FTLD network modules in PSP-RS using synthetic eigenproteins to determine whether modules derived from the genetic network could also differentiate PSP-RS from controls.Of the 31 synthetic eigenproteins derived from the genetic FTLD protein module assignments, ve were signi cantly altered in sporadic PSP-RS vs. controls after FDR-correction (Supplementary Table 8A; Fig. 4B; decreased in PSP-RS: M28 synapse assembly/axon, M2 presynapse, M3 postsynapse/glycoslyation, M19 neuron migration; increased in PSP-RS: M29 ECM).Notably, four of these ve modules were those most strongly related to MAPT disease severity (M2, M3, M28, M29), with consistency in directionality of effects (Fig. 4C).These results suggest that, despite differences in initial pathogenesis, genetic and sporadic forms of FTLD-tau exhibit shared proteomic signatures in CSF, characterized by decreases in neuronal cell-type speci c proteins and increases in ECM proteins.
Genetic FTLD modules differentiate FTLD from AD and controls.To assess whether proteomic signatures from the FTLD network were speci c to FTLD or more broadly re ective of neurodegeneration, irrespective of molecular etiology, we constructed synthetic eigenproteins in a second replication cohort (BioFINDER 2; Supplementary Table 9b) comprised of 29 patients with frontotemporal dementia clinical syndromes, 87 AD patients matched on demographics and disease severity, and 248 AD biomarkernegative controls.BioFINDER 2 CSF samples were analyzed using a proximity extension assay platform (Olink).Synthetic eigenproteins were computed based on protein measurements that overlapped between Olink and SomaScan, offering the opportunity to validate proteomic signatures across platforms.CSF proteomic alterations were more pronounced in FTLD than AD.Thirteen synthetic eigenproteins differed between FTLD and controls, 6 differed between FTLD and AD, and only 2 differed between AD and controls (all FDR-p < .05;Supplementary Table 9a; Fig. 4B).Most synthetic eigenproteins that differentiated FTLD from AD and/or controls represented modules that also differed between FTLD mutation carriers and familial controls from ALLFTD (Fig. 4D).These included increased M26 spliceosome (FTLD > AD > control) and decreased M2 presynapse (FTLD < AD, control), M28 synapse assembly/axon (FTLD, AD < control), and M22 autophagy (FTLD < AD, control).These results support the cross-platform and cross-cohort reproducibility of genetic FTLD network-derived modules, which are affected to a greater extent in FTLD than AD.
Module Overlap with CSF AD Networks.The protein co-expression patterns and pathway enrichment present in the CSF genetic FTLD network partially resembled CSF networks previously identi ed in AD 21,27 .We employed module overrepresentation analyses to empirically compare the genetic FTLD network to previously constructed CSF networks in sporadic AD and thus identify CSF modules that overlap across neurodegenerative conditions.These AD networks were built using multi-platform (SomaScan + Olink + TMT-MS; Dammer et al. 21) and TMT-MS (Modeste et.al 27 ) proteomic approaches, which also allowed us to probe the in uence of platform on FTLD and AD network overlap (Fig. 5).Of the 31 genetic FTLD network modules, 14 had signi cant overrepresentation of protein members in at least one corresponding module from the multi-platform AD network and 11 had signi cant overrepresentation in at least one module from the TMT-MS AD network (10 of 14 modules).M4 complement/coagulation was the most strongly overlapping FTLD network module in both AD networks, which also exhibited the strongest preservation across brain tissue, CSF and plasma in the multi-platform study 21 .M29 ECM and neuron/oligodendrocyte-enriched FTLD network modules (M2, M3, M28, M31) also exhibited strong overlap with modules from both CSF AD networks.M26 spliceosome overlapped with a module from the multi-platform network, M30 ribonucleoprotein complex, which did not differ between AD and controls.M26 spliceosome did not overlap with any module from the TMT-MS network, which broadly lacked modules with primary enrichment for RNA binding/splicing pathways.Overall, we observed strong conservation of modules linked to neuronal, oligodendrocyte, ECM, and immune processes in both FTLD and AD CSF samples.

Discussion
We analyzed over 4,000 CSF proteins from patients across the disease severity spectrum of genetic FTLD.We observed that protein co-expression communities most strongly linked to FTLD include proteins involved in RNA processing, synaptic/axonal function, ECM, and protein degradation.These modules, and the hub proteins that drive their co-expression, correlated with cross-sectional markers of clinical severity and axonal degeneration as well as longitudinal cognitive change.We also observed CSF protein alterations in presymptomatic mutation carriers, including decreased M9 ion transport in C9orf72 and MAPT, and increased M26 spliceosome in GRN.CSF protein signatures identi ed in genetic FTLD were also present in PSP-RS patients with autopsy-con rmed PSP (sporadic FTLD-tau), whose protein network alterations closely resembled the neuronal and ECM-linked proteomic signatures of MAPT carriers (genetic FTLD-tau).Genetic FTLD-associated modules also distinguished a cohort comprised of both sporadic and genetic frontotemporal dementia patients from controls and AD, despite cross-cohort differences in proteomic measurement platforms.Our ndings support the utility of large-scale CSF proteomics in identifying biological pathways and candidate biomarkers across the spectrum of FTLD disease severity, with relevance to both genetic and sporadic forms of disease.
Spliceosomal and related nuclear proteins were markedly increased in symptomatic carriers across all three FTLD gene groups.Prior studies report increased RNA splicing protein levels across neurodegenerative disease brain tissues, including FTLD-TDP 28 and tauopathies 29 , implicating RNA splicing dysfunction as a shared mechanism of neurodegeneration 30 .Cross-cohort replication analyses demonstrated a stair-step pattern of increased M26 spliceosome abundance across controls, AD, and FTLD.Although M26 overlapped with RNA pathway modules from the multi-platform AD network 21 , the CSF spliceosome was more strongly dysregulated in FTLD in our study than in prior studies of AD 21,27,31 .M26 was elevated prior to symptom onset and predicted cognitive decline in GRN carriers, highlighting the early stage prognostic utility of spliceosomal proteins in genetic FTLD-TDP.It is possible that increased CSF levels of RNA splicing proteins in FTLD-TDP may re ect compensatory or pathogenic protein network alterations in response to nuclear depletion of TDP-43 and related de cits in RNA metabolism and nucleocytoplasmic transport 32,33 .Nuclear depletion of TDP-43 increases production of HNRNP proteins 34,35 , which were also hub proteins in M26 spliceosome (e.g., HNRNPA1, HNRNPA2B1, HNRNPAB).As efforts continue to develop diagnostic bio uid markers that measure the highly-speci c consequences of TDP-43 loss-of-function 36 , orthogonal development of targeted assays for proteins involved in nuclear processing may be valuable for prognosis, disease monitoring, and measuring therapeutic response.
CSF modules enriched for synaptic and other neuronal proteins, including M2 presynapse, M28 synapse assembly/axon, and M3 postysnapse/glycosylation, were strong predictors of cognitive decline.This neuronal signature of FTLD-related cognitive decline aligns with prior studies demonstrating that synaptic integrity is critical for cognitive function [37][38][39] .A decrease in synaptic protein abundance in FTLD, which was observed across all three study cohorts, may re ect alterations in synaptic protein turnover due to decreased protein production 40 .CSF NPTX2, previously shown to decrease in symptomatic FTLD and other neurodegenerative diseases [41][42][43] , exhibited the strongest association with cognitive decline out of any individual protein.Combining NPTX2 with NfL and additional synaptic markers of neurodegeneration may enhance the molecular phenotyping of neurodegeneration in FTLD and help identify patients at highest risk of cognitive progression.
Decreased abundance of M9 ion transport was a striking feature of both preysmptomatic C9orf72 and MAPT.M9 was enriched for proteins that govern neuronal ion ux, including potassium voltage-gated channel proteins (e.g., KCNIP4, KCNE2) that regulate neuronal excitability and contain genetic risk loci for psychiatric disorders 44,45 .Presymptomatic alterations to M9 may re ect neuronal ion imbalances, which could contribute to neural signaling changes 46,47 that often precede widespread neurodegeneration in FTLD 48 .Decreased M9 ion transport in presymptomatic FTLD may also re ect a longstanding neurodevelopmental molecular signature of genetic dementia, considering that M9 levels did not correlate with clinical severity.M9 hub proteins such as NNAT and FOXG1 are key regulators of early stage neuron differentiation 49,50 .Longitudinal proteomic pro ling will help elucidate the timing and trajectory of changes in these neurodevelopmental proteins, which will inform their suitability as early stage FTLD biomarkers.
Similar to synaptic modules, the microglial-enriched M29 ECM predicted cognitive trajectories, even among presymptomatic mutation carriers.M29 hub proteins included ECM-secreted growth factors (e.g., FSTL1, PDGFD) as well as lysosomal and angiogenic proteins that mediate protein degradation and remodeling of the ECM (e.g., CTSD, TIMP1).In addition to providing structural support for neurons, the ECM also modulates intercellular communication and synaptic plasticity 51,52 .In AD brains, ECM proteins are positively correlated with tau pathology 20 and ECM glycoproteins are elevated in CSF of autosomal dominant AD several decades before symptom onset 53 , suggesting that ECM dysregulation is an early event in AD pathogenesis.M29 was more strongly correlated with symptomatic MAPT and PSP-RS than C9orf72 or GRN, which could re ect interactions between the ECM and extracellular levels of tau that are released during trans-cellular spread 54,55 .
Although our study sample is large for a CSF proteomics study in FTLD, it is smaller than CSF proteomics studies from more prevalent neurodegenerative conditions.Our analytic approach emphasized data reduction (module construction) and cross-cohort replication to identify robust biological signals within each gene group while preserving statistical power; however, an increased sample size would support more detailed subgroup comparisons focused on individual proteins, such as candidate biomarkers that discriminate genetic FTLD-TDP vs. FTLD-tau.Another limitation is that our discovery cohort proteomic search was restricted to targets captured by SomaScan, which may introduce bias in pathways identi ed.
SomaScan has been applied in other neurodegenerative diseases 21,25,56,57 and has high coverage for FTLD-relevant markers such as RNA splicing proteins 21 , however platforms such as Olink and TMT-MS may exhibit greater coverage for other molecular pathways (e.g., immune).Importantly, our crosscohort/cross-platform analyses recapitulated SomaScan FTLD signatures with Olink data and demonstrated module overlap between our SomaScan only network and prior networks that included Olink and TMT-MS.
Our analysis of over 4,000 CSF proteins in FTLD, to our knowledge, re ects the largest number of proteins measured in FTLD CSF to date.Our ndings highlight a diverse ensemble of CSF protein network alterations in FTLD that align with preclinical knowledge of FTLD pathophysiology.We demonstrate that FTLD CSF proteomic signatures, including those likely re ecting alterations in RNA processing and synaptic biology, are robust and reproducible across genetic and sporadic FTLD cohorts.Our integrative systems-biology approach also highlights the potential for FTLD proteomics to inform pathophysiological mechanisms that are shared and distinct across proteinopathies, including mixed tauopathies such as AD.Importantly, this FTLD CSF proteomic network provides a data anchor that can be leveraged for multi-layered analyses across matrices and 'omics pathways, thereby offering a rich resource for FTLD and neurodegenerative biomarker research.

Methods
ALLFTD Cohort.Participants included 116 carriers of pathogenic mutations in the C9orf72 (n = 47), GRN (n = 32) or MAPT (n = 37) genes and 39 noncarrier controls from families with a known mutation in one of these genes 58 .Participants were enrolled between 2015 and 2020 in the ARTFL/LEFFTDS Longitudinal Frontotemporal Lobar Degeneration (ALLFTD) consortium (NCT04363684) 10,59 , which includes 23 collaborating centers across the US and Canada.The ALLFTD study was approved through the Trial Innovation Network at Johns Hopkins University.Local ethics committees at each of the sites approved the study, and all participants provided written informed consent or assent with proxy consent.Inclusion in the present study required completion of baseline lumbar puncture for CSF collection.Four participants were excluded from nal analysis due to outlier CSF samples (see below).Clinical stage was determined with the Clinical Dementia Rating Scale (CDR®) plus Behavioral and Language Domains from the National Alzheimer's Coordinating Center (NACC) FTLD module (CDR®+NACC-FTLD) [60][61][62] .Mutation carriers were classi ed as presymptomatic or symptomatic based on a CDR®+NACC-FTLD Global score of 0 (presymptomatic) or ≥0.5 (symptomatic).Clinical and demographic characteristics across the three gene groups and controls are provided in Supplementary Table 1.4RTNI Cohort.A cohort of 35 patients with a clinical diagnosis of PSP-RS were recruited through the UCSF Memory and Aging Center (MAC) and the 4-Repeat Tauopathy Neuroimaging Initiative (4RTNI; NCT01804452).Of these 35 patients, 20 went to autopsy at the UCSF MAC Neurodegenerative Disease Brain Bank and had a con rmatory primary pathological diagnosis of PSP.A comparison group of 39 community-dwelling, cognitively unimpaired controls were recruited through the UCSF Brain Aging Network for Cognitive Health (BrANCH).CSF was collected on all participants and research procedures were approved by the UCSF Institutional Review Board.
CSF collection and processing.CSF was collected at baseline in polypropylene tubes via lumbar puncture in lateral recumbent or sitting positions.CSF samples were centrifuged at 2000 g for 10 minutes at room temperature.Supernatant was aliquoted in 500 microliter polypropylene tubes and stored at -80°C until further analyses.All CSF samples underwent only one freeze/thaw cycle before analysis.CSF ALLFTD samples were analyzed for neuro lament light chain (NfL) using Quanterix Simoa, as previously described 4 .
SOMAmer proteomics.CSF samples analyzed using a proprietary version of the SomaScan proteomics platform (SomaLogic, Boulder, CO) that captured 4,138 unique proteins.The SomaScan assay leverages slow off-rate modi ed aptamers (SOMAmers), which are short, single stranded deoxynucleotides that bind to protein targets with high speci city 63 .A volume of 65 microliters of CSF was used to create SOMAmer -protein reactions in 96-well plates.Tagged SOMAmer-protein complexes were captured in a bead-based assay, and levels of SOMAmer bound to sample were quanti ed through a uorescent signal in DNA hybridization microarrays 64 .The reaction signal was detected digitally and expressed as aggregated Agilent relative uorescent units (RFU), which were normalized to scale and subsequently log 2 -transformed for downstream analysis.
Outlier removal and covariate adjustment.Each sample had full protein data availability with no missing values.Outlier samples were identi ed (n = 6 [2 C9orf72, 1 GRN, and 1 control from the genetic cohort, 2 controls from the sporadic cohort]) and removed using a three-fold SD cutoff of Z-transformed sample connectivity, as previously described 20 .Nonparametric bootstrap regression was performed in controls (familial and non-familial) to obtain a median estimated coe cient from 1,000 iterations of tting for the effects of typical aging and sex on each protein.Median estimated coe cients were subsequently multiplied by age and sex and subtracted from participant-speci c protein values to derive adjusted protein values across the entire log 2 (abundance) matrix of cases and controls in ALLFTD and 4RTNI cohorts.
CSF protein co-expression network analysis.In the genetic cohort, the WGCNA algorithm 20,65 was used to generate a CSF protein network from the n = 4,138 log 2 protein abundance × n = 155 case-sample matrix that had undergone covariate correction and network connectivity outlier removal as described above.
Module memberships were then iteratively reassigned to enforce kME table consistency, as previously described 20 .This module membership reassignment procedure increased the smallest module size in the network from 34 to 48 (M31) and reduced the gray (unassigned) protein count for the network from 832 (20.1%) to 199 (4.8%).
Gene ontology and cell enrichment.To characterize protein module biology, we retrieved gene ontology annotations from the Bader Lab's monthly updated .GMT formatted ontology lists downloaded October 2, 2023 66 .A Fisher's exact test for enrichment was performed for each module's protein membership with the background proteome consisting of all proteins measured in the current study.Pruned output with overrepresentation z-scores was visualized using an R script (https://github.com/edammer/GOparallel).Modules were also tested for cell type-speci c enrichment with Fisher's exact one-tailed tests that compared module protein members with an in-house human cell type marker list 20 .
Differential abundance Differentially abundant proteins were identi ed by one-way ANOVA, followed by Tukey's post hoc correction for pairwise comparisons between symptomatic carriers, presymptomatic carriers, and controls (strati ed analysis for each gene).Volcano plots were made using an R script that color-coded individual data points by CSF network module membership (https://github.com/edammer/parANOVA).
Cognitive trajectory analysis.ALLFTD participants completed an annual comprehensive neuropsychological battery, covering episodic memory, executive functions, and language skills (average BioFINDER 2 Cohort.To test whether FTLD CSF proteomic signatures were robust to platform and cohorts, we leveraged CSF proteomic data analyzed on the Olink platform in the Swedish BioFINDER 2 cohort (NCT03174938) 69 .This BioFINDER 2 replication cohort included 29 patients with frontotemporal dementia clinical syndromes, 87 AD-biomarker positive patients with MCI or dementia, and 248 ADbiomarker negative controls 70 .Of the 29 FTLD patients, 10 had genetically con rmed FTLD (8 C9orf72, 1 GRN, 1 MAPT) with a behavioral variant frontotemporal dementia (bvFTD) clinical syndrome.The remaining 19 cases included 13 bvFTD and 6 semantic variant primary progressive aphasia.AD cases were matched to FTLD spectrum cases 3-to-1 for age, sex, disease severity, and mean protein level 69 using k-nearest neighbors.BioFINDER 2 participants were recruited at Skåne University Hospital and the Hospital of Ängelholm, Sweden, diagnosed by multidisciplinary assessment after clinical and neuropsychological examination, brain MRI, and lumbar puncture.The study was approved by the Regional Ethics Committee in Lund, Sweden.All participants gave written informed consent to participate.
Validation analysis: Module preservation.A sporadic PSP-RS (4RTNI) protein co-expression network (4,138 log 2 protein abundance × 74 case-sample matrix) was constructed using the same blockwiseModules parameters and module reassignment procedures as described above.Preservation between the genetic FTLD and sporadic PSP-RS CSF networks was performed using the WGCNA modulePreservation function with 500 permutations.Z summary composite preservation scores for 8 underlying network parameters were obtained using the mutation network as the template and PSP-RS network as the target 20,71 .Statistical signi cance of module preservation (minus log 10 [FDR-adjusted p values]) was visualized as a function of module size.
Validation analysis: Synthetic modules.validation was also assessed in the 4RTNI and BioFINDER 2 cohorts by calculating synthetic eigenproteins 20,22 , which captures the variance of all protein module members present in the target cohorts.We leveraged every protein module membership assignment from the genetic FTLD network to construct synthetic eigenproteins in sporadic PSP-RS, given that CSF from both cohorts were analyzed on the same platform.Synthetic eigenprotein calculations in the BioFINDER 2 Olink dataset were restricted to Olink proteins that overlapped with SomaScan proteins from the genetic FTLD network and had missing frequency < 75% (856 total proteins).Boxplots visualized synthetic eigenprotein differences across cases and controls in each target cohort.
Validation analysis: Module overrepresentation.To determine the degree to which CSF modules in the genetic FTLD network overlapped with CSF modules in AD, we performed over-representation analysis (ORA) of module gene symbols between the CSF genetic FTLD network and two CSF AD protein coexpression networks that differed in depth of proteomic coverage and sample size.Speci cally, the CSF genetic FTLD network was compared to a 1) deep, three-platform CSF network (Somalogic, Olink, tandem mass tag-based mass spectrometry [TMT-MS]; n = 7,158 proteins, 38 modules) constructed in a small cohort of AD (n = 18) and controls (n = 18); and 2) a shallower TMT-MS CSF network (n = 1,840 proteins, 14 modules) constructed in a larger and more racially diverse (~ 50% African American) cohort of AD (n = 98) and controls (n = 105).Module overlap was determined using one-tailed Fisher's exact test, followed by correction of p-values for multiple testing using the Benjamini-Hochberg method.Overlap of module gene symbols between networks was visualized using a custom in-house script.
Other statistics.Statistical analyses were performed in R (v4.3.1).Boxplots represent the median, 25th, and 75th percentile extremes; thus, hinges of a box the interquartile range of the two middle quartiles of data within a group.Minimum and maximum data points de ne the extent of whiskers (error bars).Correlations were performed using Spearman's rho (ρ) coe cients.Comparisons between two groups were performed by a two-sided t test.Comparisons among three or more groups were performed with ANOVA with Tukey's pairwise comparison of signi cance.P values were adjusted for multiple comparisons by FDR correction according to the Benjamini-Hochberg method where indicated.Figure 3 applied to 4RTNI SomaScan data to test for module preservation across the genetic FTLD and sporadic PSP-Richardson syndrome networks.Modules that have a Z summary score greater than or equal to 1.96 (or q = 0.05, blue dotted 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 , red dotted 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.Heatmap displays average synthentic eigenprotein z-score differences between PSP-RS and controls (CTL), FTLD vs. CTL, FTLD vs. AD, and AD vs.

[
range] visits = 3.4[1-7])10,67  .Raw test scores were converted to z-scores based on the score distribution in the larger cognitively unimpaired BrANCH cohort 68 , which were then averaged into domain-based composite scores.Domain-based composite scores were averaged together to create a global cognitive composite score.The rate of cognitive change over time for each participant was determined via a linear mixed-effects model that examined global cognitive scores as a function of time (years since baseline), entered as a xed and random effect, adjusting for baseline age, sex and years of education.Personspeci c random slopes were extracted to represent the annual rate of cognitive change per participant, which was the primary outcome in downstream analyses that examined differential correlation of modules and individual proteins with cognitive trajectory.
Overview.Cerebrospinal uid (CSF) was collected in 116 carriers of autosomal dominant mutations for frontotemporal lobar degeneration (FTLD; 47 C9orf72 [C9], 32 GRN, 37 MAPT) and 39 noncarrier controls with a family history of genetic FTLD.CSF was analyzed on a modi ed aptamer-based assay (SomaScan).After data processing, a total of 4,138 proteins were quanti ed.High-dimensional proteomic data were organized into modules of protein co-expression using weighted gene co-expression network analysis (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 neuro lament light [NfL]) and longitudinal (global cognitive trajectories) indicators of disease severity.In cross-cohort and cross-platform validation analyses, genetic FTLD modules were reconstructed in independent cohorts of sporadic PSP-Richardson syndrome (PSP-RS) and controls (4RTNI; SomaScan) and FTLD, AD, and controls (BioFINDER 2; Olink).

Figure 2 Genetic
Figure 2 CTL. Asterisks indicate statistical signi cance after false discovery rate correction C, D) Synthetic eigenprotein boxplots for key modules from 4RTNI (C) and BioFINDER 2 (D) analyses.Pairwise differences in module synthetic eigenproteins by case status were assessed by two-sided t-test with FDR-correction.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 de ne the extent of whiskers (error bars).