To better understand the role of biological pathways in the periphery and how they influence AD and its heterogeneous pathophysiology, here we assess associations of approximately 7,000 plasma proteins with multiple AD neuropathologies and endophenotypes, including cognitive function. We also assess how pathways in the periphery are related to those in brain and CSF compartments.

We analyzed 2,645 plasma proteomes from 2,139 individuals across four separate cohorts using the SomaLogic SomaScan 7K aptamer-based platform, which allowed us to quantify the levels of 6,345 proteins using 7,287 aptamer assays without missing values across these samples. Plasma samples were obtained from the Bio-Hermes study, the Emory Goizueta Alzheimer's Disease Research Center (GADRC), the Religious Orders Study and Rush Memory and Aging Project (ROSMAP) as well as a collection of clinical research studies at Emory University (Emory Other) (Fig. 1 and Supplementary Table 1). All cohorts contained participants with and without cognitive impairment and biomarkers of cerebral Aβ. Bio-Hermes had amyloid positron emission tomography (PET) and plasma pTau217 measurements available for nearly all participants to assess the presence and levels of cerebral Aβ. Emory GADRC participants had plasma pTau217 measurements, and Emory Other participants had CSF Aβ and tau measurements to assess levels of cerebral Aβ deposition. All ROSMAP participants underwent autopsy and had direct measurement of Aβ plaque burden. In ROSMAP, we restricted neuropathology correlations with plasma protein measurements to within 3 years of autopsy, which comprised about half of the ROSMAP participants available in our dataset (n = 221 out of 436). We tested whether the associations in the ROSMAP cohort with cerebral Aβ were significantly affected by time to autopsy within 3 years and found that only 2% of proteins were significantly affected, with none of the most strongly associated proteins affected by time between blood draw and autopsy. This finding is consistent with the slow evolution of amyloid pathology that occurs over decades. We first assessed the association of plasma proteins to cerebral Aβ separately in each cohort. Using linear models, we associated each plasma protein measurement with the Aβ PET standardized uptake value ratio (SUVR) in Bio-Hermes, plasma pTau217 in Emory GADRC, CSF total Tau (tTau)/Aβ42 in Emory Other and a composite measure of neuritic and diffuse plaque load in ROSMAP, adjusting for age, sex and race as needed in each cohort based on exploratory analyses (Supplementary Table 2). Four proteins were nominally (without correction for false discovery in each cohort) significantly associated with cerebral β-amyloidosis across all four cohorts: kinetochore protein Spc25 (SPC25), D-3-phosphoglycerate dehydrogenase (PHGDH), vacuolar protein sorting-associated protein 29 (VPS29) and baculoviral IAP repeat-containing protein 2 (BIRC2) (Extended Data Fig. 1a).

To leverage information about associations with β-amyloidosis across cohorts and assessment methods, we performed a meta-analysis of results across all four cohorts. After adjustment for false discovery, 214 proteins were observed to be significantly associated with β-amyloidosis in the meta-analysis, with 131 proteins positively associated and 83 proteins negatively associated (Fig. 2a and Supplementary Table 2). Most of the proteins identified as significantly associated with Aβ plaque using a quantitative linear modeling approach were also observed as significant in a logistic meta-analysis model in which β-amyloidosis was defined using a binary cutoff in each cohort (Extended Data Fig. 1b and Supplementary Table 3). SPC25 was the plasma protein most strongly associated with β-amyloidosis. Gene Ontology analysis showed that plasma proteins positively associated with cerebral β-amyloidosis were enriched in synapse and ECM pathways, whereas proteins negatively associated with β-amyloidosis were enriched in metabolism and proteostasis pathways (Fig. 2d). Because APOE ε4 genotype is strongly positively associated with β-amyloidosis, we assessed which amyloid-associated proteins were dependent upon APOE ε4 (Supplementary Tables 2 and 3). Fifty-six proteins remained significantly associated with β-amyloidosis in the linear model after adjustment for APOE ε4, with complexin 2 (CPLX2) remaining the most strongly associated protein (Fig. 2a,b). One protein -- matrilysin (MMP7) -- gained significance in association after adjustment. Therefore, approximately 75% of proteins were no longer significantly associated with amyloid after APOE ε4 adjustment, including SPC25 (Fig. 2c and Supplementary Table 2). We assessed overlap of the Aβ-associated APOE ε4-dependent proteins in plasma with proteins previously identified as associated with AD and ε4 dependent in serum in the AGES-Reykjavik cohort (Fig. 2e). Approximately half of the ε4-dependent proteins in serum associated with AD overlapped with plasma amyloid-associated ε4-dependent proteins, suggesting that these proteins are robustly specific for the β-amyloidosis AD endophenotype regardless of blood matrix sampling. Aβ plaque-associated proteins after ε4 adjustment reflected similar biological pathways to those without adjustment (Extended Data Fig. 1c). This was also observed for ε4-dependent proteins, although ε4-dependent proteins negatively associated with amyloid were more enriched in neuronal, apoptosis and endosome pathways (Extended Data Fig. 1d). Adjusting for cognitive function in addition to APOE ε4 further reduced the number of proteins associated with cerebral Aβ to 13, highlighting proteins associated with cerebral Aβ regardless of level of cognitive impairment (Extended Data Fig. 1e,f). In summary, we identified plasma proteins robustly associated with cerebral β-amyloidosis across different cohorts and biomarker outcome measurements. These proteins represented primarily synaptic/neuronal, ECM, metabolism, endosome and proteostasis pathways, and many of them were strongly influenced by APOE ε4 genotype.

To determine the relationship between plasma proteins associated with Aβ plaque and brain proteins associated with Aβ plaque, we assessed overlap of plasma proteins associated with Aβ plaque with those directly measured in brain in prior mass spectrometry-based studies. Seventeen proteins out of 425 plaque-enriched or plaque-depleted proteins in brain identified by laser capture microdissection mass spectrometry-based proteomics of Aβ plaques were found in common with plasma proteins associated with Aβ plaque (Fig. 2f and Supplementary Table 4). Proteins such as SPARC-related modular calcium-binding protein 1 (SMOC1), spondin 1 (SPON1) and alpha-1-antichymotrypsin (SERPINA3) demonstrated concordance between brain and plasma, with enrichment in cerebral plaque and plasma levels positively associated with Aβ plaque burden. Others such as APOE, serine protease HTRA1 (HTRA1) and serum amyloid P-component (APCS) were discordant in brain and plasma, with enrichment in brain plaques but plasma levels negatively associated with plaque. CPLX1 and zinc-alpha-2-glycoprotein (AZGP1) were also discordant, with depletion in cerebral plaque but plasma levels positively associated with plaques. We also assessed overlap with plaque-associated brain proteins based on a prior consensus co-expression mass spectrometry-based analysis (Extended Data Fig. 1g). SMOC1, SPON1 and pleiotrophin (PTN) were concordant in association, whereas APOE and HTRA1 were discordant in association, consistent with the laser capture microdissection data. Using previously generated mass spectrometry proteomic data from ROSMAP dorsolateral prefrontal cortex (DLPFC), we also tested the association between plasma protein levels as measured by SomaScan and brain protein levels as measured by mass spectrometry for all overlapping proteins and not those just associated with plaques, restricting the analysis to within 3 years of autopsy and adjusting for time between blood draw and autopsy. We identified 150 plasma proteins positively associated and 87 plasma proteins negatively associated at a nominal statistical threshold with levels in brain (Extended Data Fig. 1h and Supplementary Table 5). Positively associated proteins were enriched in complement, immunoglobulins, reduction-oxidation and high-density lipoprotein pathways, whereas negatively associated proteins were enriched in innate immune response, protein transport, autophagy, actin and Golgi pathways (Extended Data Fig. 1i). Nine proteins overlapped between Aβ-associated and brain-linked proteins in plasma, including acetylcholinesterase (ACHE) (Extended Data Fig. 1j). In summary, we identified cerebral Aβ-associated plasma proteins that were both concordant and discordant with their association with Aβ plaque in brain. A limited number of Aβ-associated plasma proteins were found to be linked to brain levels within individuals in the ROSMAP cohort, suggesting that the relationship between plasma and brain levels for these nine proteins is highly robust across cohorts, proteomic platforms and time to autopsy.

To determine whether the available plasma proteomic measurements could be used as a diagnostic marker for cerebral Aβ, we tested the ability of a select number of proteins to discriminate Aβ-positive (Aβ) and Aβ-negative (Aβ) participants in each cohort. We searched the top 42 most strongly positively and top 42 most strongly negatively Aβ-associated proteins from the meta-analysis to develop combinatorial protein ratios that best predicted Aβ status, as previously described for proteomic prediction of drug response, with the goal of developing a single value as a predictor of Aβ status that could leverage information from multiple Aβ-associated proteins. The best one-protein, two-protein, three-protein and four-protein ratios (n = 2, 4, 6 and 8 total proteins) are provided in Supplementary Table 6. The discrimination accuracy for the best four-protein ratio (SPC25+CPLX2+SMOC1+ACHE/NEFL+VPS29+PTPRU+NOMO2) reached approximately 80% area under the curve (AUC) in the Bio-Hermes, Emory GADRC and Emory Other cohorts, with lower accuracy in ROSMAP (Extended Data Fig. 2a,b). This was lower than pTau217, which had an AUC of approximately 90% in Bio-Hermes for amyloid positivity based on PET. The accuracy improved slightly if ratios were developed that were specific to each cohort rather than using the meta-analysis results, with the largest increase in accuracy observed in ROSMAP (approximately 64-81%) and the highest accuracy in Emory Other (90%) (Extended Data Fig. 2c). Because many of the top Aβ-associated plasma proteins were dependent upon APOE ε4, and not all Aβ participants carry an ε4 allele, we also explored whether the discrimination accuracy would improve if ratios were developed using amyloid-associated proteins less dependent on ε4 and used in only non-ε4 individuals for amyloid prediction. Discrimination accuracy did not improve using proteins less influenced by ε4 in non-ε4 individuals (Extended Data Fig. 2d and Supplementary Table 6). These results suggest that the proteins measured on the SomaScan 7K platform cannot currently outperform plasma pTau217 as a discriminator of cerebral Aβ positivity.

We performed a similar analysis across cohorts to investigate the associations of plasma proteins with cognitive function, using the Mini-Mental State Examination (MMSE) in Bio-Hermes and Emory GADRC, the Montreal Cognitive Assessment (MoCA) in Emory Other and a composite measure of cognitive function in ROSMAP. One protein, the neuronal pentraxin receptor (NPTXR), was nominally significantly positively associated with cognitive function across all cohorts (Extended Data Fig. 3a and Supplementary Table 7). After meta-analysis, synaptic vesicle membrane protein VAT-1 homolog (VAT1) and amyloid-β precursor-like protein 1 (APLP1) were most strongly positively associated with cognitive function (higher levels associated with better cognitive function), whereas ACHE and SPC25 were most strongly negatively associated with cognitive function (higher levels associated with worse cognitive function) (Fig. 3a and Supplementary Table 7). Pathways associated with better cognitive function included protein translation, cadherins and exosome, whereas pathways associated with worse cognitive function included synapse, immune response and ECM (Extended Data Fig. 3b). Out of a total of 283 proteins associated with cognitive function, 42 were also significantly associated with cerebral β-amyloidosis (Fig. 3b and Supplementary Table 7). After adjusting for Aβ, VAT1 and APLP1 remained significantly positively associated, and ACHE remained significantly negatively associated with cognitive function (Fig. 3a and Supplementary Table 7). The top proteins whose association with cognitive function was dependent upon Aβ levels are shown in Fig. 3c. Proteins positively associated with cognitive function independent of cerebral β-amyloidosis were enriched in protein translation, vesicular transport, cadherin, exosome and mitochondrial pathways, whereas proteins negatively associated with cognitive function independently of amyloid were enriched in inflammation and chemokines, fatty acid metabolism, ECM and endocytosis pathways (Fig. 3d). When comparing proteins associated with cognitive function from the meta-analysis with brain-linked proteins from ROSMAP, we identified 11 proteins that overlapped, including apolipoprotein A-IV (APOA4) and Ras-related protein Rab-1A (RAB1A) that were positively associated with cognitive function and ACHE that was negatively associated with cognitive function (Fig. 3e and Supplementary Table 7). In summary, we identified plasma proteins associated with cognitive function independently of Aβ that were enriched in multiple biological pathways, including mitochondria, immune function, fatty acid metabolism and ECM, highlighting the diversity of biological pathways characterized in plasma related to cognitive function.

We used the rich neuropathological data from ROSMAP to identify plasma protein associations with multiple different neuropathologies commonly observed in AD cases at autopsy, including Lewy bodies, TDP43, cerebral amyloid angiopathy (CAA) and different types of vascular disease (Fig. 3f and Supplementary Table 8), adjusting for age, sex and race as needed but without adjustment for other pathologies. Biological processes enriched in the plasma protein associations for each neuropathology are provided in Supplementary Fig. 1. We then identified proteins associated with each neuropathology that were also associated with global cognitive function within the ROSMAP cohort. We found that approximately half of the proteins associated with cognitive function in ROSMAP (255 out of 496) were also associated with at least one neuropathology (Fig. 3f and Supplementary Table 8). The other half not associated with measured neuropathologies were enriched for multiple pathways, including T cell biology, glucose metabolism, Golgi and glycosylation, fibronectin and mitochondria (Extended Data Fig. 3c). A large majority of plasma proteins associated with each neuropathology were not significantly correlated with brain levels (Extended Data Fig. 4 and Supplementary Table 8). In ROSMAP, four proteins associated with cognitive function and significantly brain linked were found to be in common with meta-analysis proteins identified in the same analysis and included tumor protein D54 (TPD52L2), alpha-1,3/1,6-mannosyltransferase (ALG2), kininogen 1 (KNG1) and eukaryotic initiation factor 4A-II (EIF4A2) (Fig. 3g). ALG2, KNG1 and EIF4A2 were also not associated with the neuropathologies assessed in ROSMAP (Fig. 3h), indicating their link to other brain processes associated with cognitive function not assessed by neuropathological examination in ROSMAP.

In a separate analysis, we identified cognitively impaired participants who were amyloid negative in Bio-Hermes (254 participants), Emory GADRC (60 participants) and Emory Other (16 participants) cohorts (CI.other) and the plasma proteins associated with this diagnostic state. A total of 211 proteins were associated with CI.other after meta-analysis (Extended Data Fig. 5a and Supplementary Table 9). Proteins positively associated with CI.other were enriched in leukocyte chemotaxis, ECM, heparin binding, cytokine and proteostasis pathways, with immune and ECM pathways found to be in common with pathways negatively associated with cognitive function after adjustment for cerebral Aβ as noted above. Proteins negatively associated with CI.other were enriched in nucleic acid metabolism and peptidase activity pathways, among others (Extended Data Fig. 5b). We observed minimal overlap between proteins associated with CI.other and proteins associated with different neuropathologies in ROSMAP (Extended Data Fig. 5c and Supplementary Table 9), consistent with the findings observed within the ROSMAP-only analysis.

In summary, we identified plasma proteins associated with multiple different neuropathologic endophenotypes commonly observed in AD as well as with cognitive impairment in the absence of cerebral Aβ. Proteins associated with these neuropathologic endophenotypes constituted approximately half of the plasma proteins associated with cognitive function, indicating that a large fraction of plasma protein associations with cognitive function is not accounted for by existing neuropathologic assessments.

Longitudinal data were available in the Emory GADRC and ROSMAP cohorts. Approximately 30 participants in the Emory GADRC cohort were cognitively normal (control) at the first plasma collection but converted to mild cognitive impairment or AD at a later visit, and, therefore, because of the low conversion numbers and the fact that the Emory GADRC data did not meet statistical assumptions for Cox proportional hazard analysis, we instead focused on the ROSMAP cohort for risk analyses in which there were 172 participants who became cognitively impaired as determined by a change in diagnostic status out of a total of 310 selected for analysis. In the risk analyses, cognitive decline did not have to occur within 3 years of autopsy. Although no proteins survived correction for multiple comparisons, proteins whose levels at baseline were nominally associated with increased risk of all-cause cognitive decline were enriched in ECM, heparin binding, cytokines and immune cells (Fig. 4a,d and Supplementary Table 10), regardless of pathology at autopsy in converters and non-converters. When converters were restricted to those who had a sufficiently high degree of AD pathology confirmed on autopsy (n = 131), proteins that nominally increased risk for cognitive decline were associated with heparin binding, ECM, fibronectin and muscle (Fig. 4b,e and Supplementary Table 10). Further restricting the analysis to include only those non-converters with low levels of Aβ and tau pathology on autopsy (total n = 187 with n = 131 converters), we identified Erk signaling, axon regeneration, myelination, T cell, cytokine and SNARE biology nominally associated with increased risk of cognitive decline (Fig. 4c,f and Supplementary Table 10). Proteins in common among the three analytical approaches that either increased or decreased risk of cognitive decline were enriched in cell recognition, T cell, cytokine/immune, glycosaminoglycan binding, ECM and post-synapse pathways (Fig. 4g, Extended Data Fig. 6a and Supplementary Table 10). We assessed for overlap of proteins in ROSMAP associated with risk of cognitive decline at a nominally significant level (all converters) with proteins associated with risk of cognitive decline in the Atherosclerosis Risk in Communities (ARIC) cohort, and we identified CPLX1, ephrin type-A receptor 2 (EPHA2) and collagen alpha-1(X) chain (COL10A1) that were commonly associated with increased risk of cognitive decline (Fig. 4h). We performed the same analysis with the AGES-Reykjavik cohort in which conversions were defined as AD based on clinical evaluation and observed SMOC1, tissue factor pathway inhibitor (TFPI), myomesin 2 (MYOM2) and cyclic GMP-AMP phosphodiesterase SMPDL3A (SMPDL3A) as associated with increased risk of conversion and linker for activation of T cells family member 2 (LAT2) and anaphase-promoting complex subunit 10 (ANAPC10) as associated with decreased risk of conversion (Fig. 4i). COL10A1, EPHA2, MYOM2 and LAT2 survived five-fold cross-validation (Extended Data Fig. 6b,c). In summary, we identified proteins associated with risk of cognitive decline in ROSMAP that were enriched in pathways also associated with cognitive function, including immune, ECM and synapse, indicating that alterations in these pathways as assessed in plasma can precede cognitive impairment.

As a separate approach to identify the biological pathways associated with neuropathological and cognitive AD endophenotypes in plasma, we applied a systems biology technique called protein co-expression analysis to reduce the dimensionality of the plasma proteomic data into clusters, or 'modules', of proteins defined by their co-abundance across cases. These clusters represent biological processes identified in a data-dependent fashion based on the underlying plasma proteomic measurements with which we could then assess for enrichment in proteins identified to be associated with each AD endophenotype. A co-expression network built from all individuals across the four cohorts identified 27 modules representing diverse biological pathways and processes (Fig. 5, Extended Data Fig. 8 and Supplementary Table 11), with most modules having clear primary ontologies. Comparison of our network to a previously published serum network from the AGES-Reykjavik cohort showed excellent overlap, particularly in lipoprotein, complement and neuronal modules (Extended Data Fig. 7 and Supplementary Table 12). Proteins associated with cerebral β-amyloidosis were most strongly enriched in the M16 cholesterol/lipid/APOE module, as expected, but also showed strong enrichment in M19 insulin growth factor binding, M17 small molecule metabolism, M4 ECM/TNF/neuron development and M5 neuron development modules. Using proteins associated with different neuropathologies as assessed in the ROSMAP-specific analyses to test for module enrichment, M4 ECM/TNF/neuron development also showed strong enrichment for proteins associated with Aβ plaques, whereas M12 CXCR chemokine/neutrophil chemotaxis, M16 cholesterol/lipid/APOE and M3 mitosis regulation/transcription modules showed the strongest enrichment for proteins associated with tau tangles. Proteins associated with TDP43 deposits were enriched in M3 mitosis regulation/transcription and M2 transcription factor/DNA binding/receptor ligand activity modules, whereas proteins associated with cerebral atherosclerosis were associated with M16 cholesterol/lipid/APOE, among other modules. Proteins associated with Lewy bodies did not show significant enrichment in network modules except for M25, which did not have a clear primary ontology. Other than the large M1 ambiguous module that was likely unrelated to a coherent biological pathway and more likely related to technical factors based on our prior experience with SomaScan networks in biofluids, the modules most strongly enriched in proteins associated with cognitive function from the meta-analysis were M7 protein folding/chaperone and M13 ubiquitin ligase/protein secretion, highlighting the importance of the relationship between proteostasis and cognitive function. Modules that showed the strongest enrichment for proteins involved in risk for cognitive decline included M18 coagulation/complement/wound healing, M3 mitosis regulation/transcription, M21 skeletal muscle, M4 ECM/TNF/neuron development and M5 neuron development. Additional information for network module enrichments with endophenotypes is provided in Extended Data Fig. 8 and with risk-associated proteins in Extended Data Fig. 9, with individual proteins provided in Supplementary Table 13. In summary, protein co-expression analysis highlighted biological pathways consistent with those identified by ontology analysis of linear modeling results, such as neuronal/synapse and ECM associations with cerebral β-amyloidosis, but also provided additional insights into other associations, such as cognitive function with proteostasis and complement and skeletal muscle proteins with risk of cognitive decline, among others.

To assess overlap of plasma proteins with brain and CSF proteins, we first used the protein co-expression network approach to identify co-expression modules in common between the compartments. At the level of co-expression, few plasma modules had strong overlap with a previously reported brain network, with the strongest overlap being complement and acute phase proteins measured in both compartments (Fig. 6a and Supplementary Table 14). More overlap was observed between CSF and plasma compartments (Fig. 6b and Supplementary Table 15). In addition to complement/coagulation, modules that were strongly associated with APOE in CSF as previously reported (M33 oxidant detoxification/MAPK signaling, M26 neddylation and M34 mitochondrion) overlapped with M16 cholesterol/lipid/APOE in plasma. In addition, multiple neuronal CSF modules overlapped with M4 ECM/TNF/neuron development and M5 neuron development modules in plasma, suggesting that the co-expression relationships among the proteins in these modules are preserved between CSF and plasma compartments. We tested for overlap between proteins associated with cerebral β-amyloidosis and cognitive function in plasma with brain and CSF networks (Fig. 6c,d). Plasma proteins associated with cerebral β-amyloidosis were most strongly enriched in the brain M26 complement/acute phase and M42 matrisome modules, whereas plasma proteins associated with cognitive function were enriched in multiple modules, including M7 MAPK/metabolism, M4 synapse/neuron, M2 mitochondria, M37 endosome, M30 proteasome, M38 heat shock/folding, M14 protein folding, M25 sugar metabolism, M26 complement/acute phase, M27 ECM, M42 matrisome and M11 cell-ECM interaction modules, consistent with many ontologies identified through ontology analysis of plasma protein associations with these traits (Fig. 3d and Extended Data Fig. 3b). More overlap for cerebral β-amyloidosis-associated plasma proteins was observed with CSF modules compared to brain modules (Fig. 6d). M33 oxidant detoxification/MAPK signaling and M26 neddylation, both APOE-associated CSF modules, were strongly enriched in plasma proteins negatively associated with cerebral β-amyloidosis. Proteins positively associated with cerebral β-amyloidosis were enriched in multiple neuronal CSF modules along with M2 complement/coagulation and other modules such as M10 golgi/glycosylation. For cognitive function, M14 translation was the most strongly enriched CSF module for proteins positively associated with cognitive function, consistent with one of the primary ontologies individually identified from linear modeling for this set of plasma proteins. Immune (M2 complement/coagulation) and proteostasis (M4 autophagy/ubiquitination) pathways in CSF were enriched in plasma proteins negatively associated with cognitive function.

We observed in prior studies that some proteins appear to be anticorrelated in their levels between CSF and plasma compartments. To further explore this phenomenon, we assessed overlap of plasma proteins positively and negatively associated with cerebral amyloidosis and cognitive function with CSF proteins measured on the same SomaScan platform positively and negatively correlated with these traits in CSF (Fig. 7 and Supplementary Table 16). Immune, ECM and general metabolism pathways had a concordant direction of association with cerebral amyloidosis between CSF and plasma compartments, whereas synaptic pathways had notable discordant direction of association (increased in plasma, decreased in CSF) (Fig. 7a). For cognitive function, metabolism, integrin binding and ECM pathways were also altered in concordant fashion in CSF and plasma, and transferase, ATP binding and synapse pathways were discordant (Fig. 7b).

In summary, we observed more overlap in co-expression between plasma and CSF compartments than plasma and brain compartments. More plasma proteins associated with cerebral β-amyloidosis had overlap with CSF modules compared to brain modules, but plasma proteins associated with cognitive function had similar degree of overlap with both brain and CSF modules. At both the co-expression and trait association overlap levels, we observed that complement/acute phase proteins had good overlap among brain, CSF and plasma compartments. However, the direction of association for certain types of proteins, such as some synaptic proteins, was opposite in plasma and CSF compartments, illustrating the importance of the compartment when studying disease pathways and associations.