Development and validation of a novel panel of CSF biomarkers for Alzheimer’s disease

Abstract

Background

Using proximity extension proteomic arrays (PEA), we previously identified a nine-biomarker CSF protein panel reflecting different biological processes dysregulated in Alzheimer´s disease (AD) that discriminated AD from non-AD dementias (i.e., dementia with Lewy bodies (DLB) and Frontotemporal dementia (FTD)). We translated this AD-Differential diagnostic panel into immunoassays for large-scale analysis and validated its discriminative performance in an independent cohort.

Method

In-house or commercially available immunoassays for five out of nine proteins (THOP1, ENO2, DDC, MMP7, and ITGB2) were analytically validated on the Ella platform. To verify analytical performance, results in CSF samples (n = 206) were compared between custom-PEA and Ella immunoassays. Clinical validation was performed in patients from the Amsterdam Dementia Cohort, including controls (cognitively unimpaired; n = 55, mean age: 57 ± 4 years, 36% females), amyloid-positive patients with mild cognitive impairment (MCI-Aβ+; n = 39, mean age: 68 ± 4 years, 46% females), AD (n = 47, mean age 66 ± 5 years, 31% females), and a non-AD group including patients with DLB (n = 54, mean age 69 ± 5 years, 17% females) and FTD (n = 50, mean age 64 ± 7 years, 32% females).

Results

Moderate to strong correlations between custom-PEA and Ella immunoassays were observed (Rho’s > 0.646). CSF proteins showed similar though weaker trends compared to custom-PEA results. THOP1 and ENO2 were dysregulated across different AD stages. ENO2 and ITGB2 were increased in AD compared with DLB and FTD, while DDC was specifically increased in DLB. THOP1, ITGB2, and ENO2 correlated strongly with CSF pTau181 and tTau while weak correlations with amyloid were observed. Our CSF panel (THOP1, ENO2, DDC, MMP7, ITGB2) discriminated AD from non-AD dementias with moderate accuracy (AUC = 0.73, 95%CI: 0.64–0.82).

Conclusion

We developed immunoassays for novel AD biomarker candidates, which cover relevant biological processes (i.e., cellular remodeling, neuropeptide degradation, energy metabolic processes, the immune system, and neurotransmitter synthesis) underlying AD pathophysiology. CSF patterns were consistent with our discovery study, albeit diagnostic performance was somewhat lower than observed previously. This highlights the need to further evaluate this panel, which monitors a broader spectrum of biological processes, both in clinical practice and within clinical trial settings.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13195-025-01899-0.

Introduction

Dementia is one of the major causes of disability and dependency within the aging population. Alzheimer’s disease dementia (AD) is the most common form and represents around 70% of dementia cases [1]. This is followed by Frontotemporal dementia (FTD, between 10 and 20%) and dementia with Lewy bodies (DLB, approximately 5%), which both have clinical and neuropathological overlap with AD [2, 3], challenging accurate diagnosis in the clinic. The underlying neuropathology of AD is measured in the cerebrospinal fluid (CSF) where decreased concentrations of amyloid-beta (Aβ), increased concentrations of phosphorylated-tau (pTau), and total-tau (tTau) reflect amyloid and tau pathology and is used in many clinics to support AD diagnosis [4]. However, proteomics and genetics studies suggest that AD pathogenesis is influenced by additional biological processes [5, 6]. Therefore, understanding the complex and heterogeneous biology of AD is essential for the development of tools for biological diagnosis and in clinical trial settings (e.g., patient inclusion or monitoring treatment effects).

Our previous antibody-based CSF proteomics study employed a data-driven modelling approach and identified and validated a panel of nine markers reflecting the biological complexity of AD (e.g., protein phosphorylation, neuropeptide degradation, energy metabolic processes, neurotransmitter synthesis, immune and vascular function). This biomarker panel discriminated AD from a group of non-AD dementias (FTD and DLB) with good performance (AUC’s >0.8) [7]. However, the translation from biomarker discovery to clinical validation can be challenging due to limited availability of specific and sensitive immunoassays [8, 9]. The development of such immunoassays is essential to support their potential clinical implementation. Therefore, in this study, we focus on five out of nine markers (i.e., THOP1, ITGB2, ENO2, DDC, and MMP7). Some of which were detected in previous proteomics studies focusing on AD (ITGB2, THOP1, and ENO2) or Lewy body-related dementias (DDC) [10–18]. In addition, preclinical and pathological studies have shown a relationship between some of these proteins and AD pathophysiology. THOP1 is a neuropeptidase that can degrade only soluble forms of amyloid-beta peptides and was shown to colocalize with Aβ plaques and Tau tangles in AD brain tissue [19, 20]. ITGB2 is a receptor involved in cell-surface mediated signaling and is expressed on many cells including microglia [10, 13]. ENO2, also known as NSE, is primarily expressed in neurons and has been widely investigated as a potential CSF biomarker for AD. A meta-analysis covering 13 studies showed elevated ENO2 concentrations in patients with AD compared to controls [21]. The increased ENO2 concentrations in CSF are probably caused by damage to neuronal bodies or axons [21]. DDC is an enzyme that converts the substrate Levodopa into dopamine and 5-hydroxytryptophan to serotonin [22, 23]. Several studies show increased CSF DDC levels in various Lewy body dementias (LBDs) and atypical parkinsonian disorders, likely reflecting underlying dopaminergic impairment [14–18, 24], a process that is also to some extent impaired in AD pathophysiology [25]. Furthermore, MMP7 has not been associated with AD yet, however, its protein function is related to a pathway dysregulated in AD pathogenesis. MMP7 belongs to a family of matrix metalloproteinases and these proteins have been localized in neuritic senile plaques and Tau tangles in AD brains [26]. Altogether, the relationship of these proteins to diverse biological processes in relation to AD and the observed discriminatory performance highlights their potential as a CSF panel for differential diagnosis to track the multifactorial nature of AD along the disease continuum. Thus, we aimed to translate this five-marker CSF panel to immunoassays on a platform that can be used for large-scale analysis and validate its differential diagnostic potential in an independent clinical cohort.

Materials and methods

Human CSF samples

We included CSF from 55 cognitively unimpaired controls, 39 individuals diagnosed with MCI-Aβ+, 47 AD, and 104 non-AD dementias (54 DLB and 50 FTD) from the Amsterdam Dementia Cohort (ADC) [27, 28]. All individuals visited the ADC for extensive cognitive and neurological assessments. Thereafter, clinical diagnosis was determined in a multidisciplinary consensus meeting, based on diagnostic criteria for each dementia [29–32]. Concentrations of the classical AD-CSF biomarkers Aβ42, tTau, and pTau were used to support AD diagnosis. CSF was collected and stored following consensus biobanking guidelines [33, 34]. CSF concentrations of Aβ42, tTau, and pTau were analyzed, either by commercially available ELISA kits (Innotest Aβ(1–42), Innotest hTAUAg, Innotest phospho-Tau(181P), Fujirebio, Ghent, Belgium) or by the automated Elecsys platform (Roche Diagnostics, Gmbh, Almere, The Netherlands). CSF concentrations measured by ELISA were corrected for the drift in biomarker concentrations that occurred over the years, as described previously [35]. The AD cases included in this study had a positive AD-CSF biomarker profile (ELISA: Aβ42 < 813 pg/mL, pTau >52pg/mL, and tTau >375 pg/mL; Elecsys: Aβ42 < 1000 pg/mL, pTau >19 pg/mL, and tTau >235 pg/mL) [36]. Individuals with subjective cognitive decline had normal clinical and cognitive tests with a negative AD-CSF biomarker profile. Furthermore, controls did not meet the criteria for MCI, dementia, or any other condition causing cognitive decline [29–32]. Neurofilament light (NfL) was measured in CSF from a small subset of the cohort (6 controls, 15 AD patients, and 21 FTD patients). Measurements were performed either by NF-light^® ELISA (Uman Diagnostics, Sweden) or with the single molecule array (Simoa) NF-light™ advantage kit (Quanterix, USA). NfL concentrations measured by ELISA were transformed to Simoa values using Passing-Bablok regression analysis, detailed previously [37].

CSF biomarker analysis

CSF MMP7 concentrations were measured with a commercially available immunoassay (cat. no. SPCKA-PS-000515) following manufacturer’s instructions (ProteinSimple, CA, USA). The concentrations of CSF ENO2, ITGB2, DDC, and THOP1 were measured using four different in-house immunoassays, which were developed on the Ella™ instrument using the 48-Digoxigenin cartridges (ProteinSimple) [24, 38]. Assay details (i.e., antibodies, standard curves, buffers) are provided in Supplementary Table 1. The analytical performance (parallelism, dilution linearity, recovery, protein stability, and intra- and inter-assay coefficient of variation) of all immunoassays were extensively tested for measurements in CSF following international guidelines for immunoassay validation (Supplementary file 1) [39]. The ENO2 protein shares structural similarities with ENO1 and ENO3 and thus cross-reactivity of the antibodies in our immunoassay could occur. Therefore, we additionally evaluated the ENO2 immunoassay for selectivity and specificity parameters. Selectivity was tested by measuring the signal of ENO2 in sample diluent in the presence of different concentrations of ENO1 or ENO3 protein (detailed in Supplementary file 1). Specificity was determined by whether we could measure different concentrations of ENO1 and ENO3 in our ENO2 assay (detailed in Supplementary file 1).

All in-house immunoassays followed the same protocol as previously described in detail [24, 38]. Briefly, antibody pairs targeting the protein of interest were first coupled to digoxigenin (capture antibodies) or biotin (detection antibodies) and were subsequently diluted to assay concentrations (Supplementary Table 1). To analyze clinical samples, CSF was diluted in sample diluent (Supplementary Table 1). Samples, antibodies (35 µL), and wash buffer (1 mL) were dispensed into the corresponding wells on 48-Digoxigenin cartridges (PoteinSimple) and the Ella instrument performed all subsequent analysis steps automatically. Proteins were sequentially measured all having the same f/t cycle and the data obtained was analyzed in triplicate.

Statistical analysis

Data analysis was performed with R version 4.0.3 (packages: emmeans, mcr, and pROC) [40–42]. DDC protein concentrations were determined in two batches of clinical samples. Due to the relatively high inter-assay variation for this novel Ella immunoassays, each round of measurement included 12 bridging samples covering different diagnostic groups. Passing-Bablok regression analysis was performed with bridging samples and protein concentrations from round two were transformed to round one using regression formulas (Method 2 = −10.63 + 0.97 * method 1; Supplementary Fig. 1).

To verify analytical performance, Spearman Rho analysis was used to assess the correlation between our Ella immunoassays and custom-proximity extension assays (PEA) panels developed previously, in overlapping samples (n = 206) [7]. Correlations between platforms are considered as weak when Rho = 0.3, moderate when Rho = 0.3–0.5, or strong when Rho >0.5 [43].

Normal distribution of the data was tested by Shapiro-Wilk test. If protein concentrations did not follow a normal distribution, the protein concentrations were log-transformed. The influence of potential covariates (i.e., age and sex) on protein concentrations was tested with linear regression analysis. We evaluated protein changes along different groups by ANCOVA using log-transformed values, including age and sex as covariates followed by Bonferroni post-hoc correction [44]. We performed an additional analysis where we combined the MCI-Aβ+ and the AD group together as an AD profile group. In addition, we conducted a sensitivity analysis in which we excluded patients with DLB and FTD who exhibited AD co-pathology, as determined by the CSF biomarkers. Receiver operating characteristics curve (ROC) analysis was used to combine the proteins into the differential diagnostic panel and determine the accuracy to discriminate between diagnostic groups.

Spearman Rho correlation analysis was performed to evaluate the correlation between the CSF proteins within the panel and AD pathophysiological features and cognition as measured by the AD-CSF biomarkers and MMSE scores within the total cohort. In addition, principal-component analysis (PCA) was performed to further understand the relationship of the total panel with AD pathophysiology. The first principal component was used as the biomarker panel composite score. This composite score was then correlated to classical AD-CSF biomarkers and MMSE scores. CSF biomarker concentrations measured by the Elecsys were first transformed to the predicted ELISA concentrations, using formulas determined previously [45]. P-values < 0.05 were considered significant.

Results

Cohort characteristics

Demographics and clinical characteristics are presented in Table 1. The control group was younger compared to the other groups. As expected, the DLB group is predominantly male (83%) and MMSE scores were lower in the AD group compared to the other groups. A positive AD CSF biomarker profile was detected in 25% and 8% of the DLB and FTD patients, respectively.

Table 1.

Cohort characteristics

		CON	MCI-Aβ+	AD	DLB	FTD
N		55	39	47	54	50
Sex (%)	F	20 (36.4)	18 (46.2)	31 (66.0)	9 (16.7)	16 (32.0)
	M	35 (63.6)	21 (53.8)	16 (34.0)	45 (83.3)	34 (68.0)
Age (years)		57 (54–61)a, b,c, d	68 (65–71)d, e	65 (62–68)c, e	69 (66–73)b, d,e	64 (59–69)a, c,e
MMSE scores		29 (28–30)a, b,c, d	26 (24–27)b, c,e	21 (19–23)a, c,d, e	24 (21–26)a, b,e	25 (24–27)b, e
Positive CSF AD profile, n (%)		0 (0.0)	35 (92.1)	47 (100.0)	12 (24.5)	4 (8.3)

Continuous data is represented as median ± interquartile range and dichotomous data as the number of cases with a percentage of the total (%)

Differences between groups were determined using the Kruskal Wallis test with Bonferroni correction

Abbreviations: CON controls, AD Alzheimer’s disease, CSF  cerebrospinal fluid, MCI-Aβ+ mild cognitive impairment with amyoid pathology, DLB dementia with Lewy bodies, FTD Frontotemporal dementia, MMSE mini-mental state examination

^b p < 0.05 compared to AD

^c p < 0.05 compared to DLB

^d p < 0.05 compared to FTD

^e p < 0.05 compared to Controls

Immunoassay validations

As described previously, THOP1 and DDC Ella immunoassays showed good analytical performance (i.e., parallelism, dilution linearity, recovery, and protein stability within the normal range; intra- and inter-assay CV of < 20% and < 20%, respectively) [24, 38]. Our ITGB2 immunoassay showed optimal recovery for the medium ITGB2-spiked samples but not for those spiked with low or high concentrations of the analyte. This suggests that there are some differences between the CSF and calibrator response that might affect the ITGB2 signal, however, clinical samples are not within those extreme ranges but fall within the middle range of the curve. Parallelism and dilutional linearity parameters of the CSF ITGB2 assay were within the acceptance criteria (85% − 115%; Supplementary Fig. 2A-D). All analytical parameters for the MMP7 and ENO2 assays were within acceptable range (Supplementary Fig. 2E-L). Selectivity experiment showed that in the presence of low concentrations of ENO2 (312.5 pg/mL) and ENO3 (312.5 pg/mL), an ENO2 recovery of 125% was detected. The ENO2 recovery remained within the acceptance range when higher concentrations of ENO2 with ENO1 or ENO3 were measured (Supplementary Fig. 3A). In addition, no signal was observed upon measuring different concentrations of ENO1 or ENO3 specifically in our ENO2 immunoassay (Supplementary Fig. 3B). This indicates that there is limited cross-reactivity of the ENO2 antibodies for these structurally similar proteins. The Ella immunoassays showed that all five proteins remained stable up to five f/t cycles (Supplementary Fig. 2) [24, 38]. Intra- and inter-CV% were below 20% for all immunoassays (Supplementary Table 2). We next analyzed the correlations between concentrations measured by our Ella immunoassays with the custom-PEA [7]. We observed that all assays had a moderate to strong positive correlation with the custom-PEA (DDC: Rho = 0.726, THOP1: Rho = 0.707, ENO2: Rho = 0.646, ITGB2: Rho = 0.664, and MMP7: Rho = 0.936; all p < 0.001; Fig. 1).

Fig. 1.

All results generated with Ella immunoassays were significantly correlated with protein levels measured by custom-PEA.

Protein levels were measured by custom-PEA assays (Olink) and with Ella immunoassays. Associations between platforms were determined by Spearman Rho’s correlation analysis, which showed that all assays had a moderate-strong correlation between analytical platforms. Abbreviations: MCI-Aβ+, mild cognitive impairment with amyloid pathology; AD, Alzheimer’s disease; DLB, dementia with Lewy bodies; FTD, Frontotemporal dementia; DDC, dopa decarboxylase; THOP1, thimet oligopeptidase; ENO2, enolase 2; ITGB2, integrin subunit beta 2; MMP7, Matrix metallopeptidase 7; NPX, normalized protein expression. * p < 0.05, ** p < 0.01, and *** p < 0.001

CSF proteins are dysregulated between diagnostic groups

CSF THOP1, ITGB2, ENO2, and DDC concentrations were significantly dysregulated or showed nominal significance across diagnostic groups on the Ella platform (Fig. 2A-D). THOP1 and ENO2 were increased in MCI-Aβ + and AD stages compared to controls, with similar concentrations between MCI-Aβ + and AD (Fig. 2C-D). THOP1, ITGB2 and ENO2 concentrations were higher in AD compared with non-AD dementias (Fig. 2 A and C-D). DDC concentrations were increased specifically in patients having DLB compared with all other groups (Fig. 2B). In addition, MMP7 concentrations did not show any differences between diagnostic groups (Fig. 2E). These trends were weaker though comparable with our custom-PEA assays, developed previously (shown as reference in Fig. 2) [7]. On the custom PEA, CSF ITGB2, THOP1, and ENO2 levels are significantly dysregulated across different AD stages compared with controls and non-AD dementias. DDC levels are significantly increased in patients with DLB specifically, while MMP7 did not show any changes across groups To increase statistical power, MCI-Aβ + and AD groups (i.e., AD profile group) were merged. ITGB2, THOP1, and ENO2 concentrations were dysregulated in patients with AD profile while DDC concentrations were highest in patients with DLB. MMP7 did not show any changes between groups (Fig. 3)

Fig. 2.

CSF proteins from the differential diagnostic panel are dysregulated across groups.

Protein concentrations in CSF from (A) ITGB2, (B) DDC, (C) THOP1, (D) ENO2, and (E) MMP7 were sequentially measured either on the custom PEA or the Ella platform. (F) Forest plots display the area under the curve (AUC) of the Receiver operating characteristic analysis from our CSF differential diagnostic panel discriminating different clinical groups. Abbreviations: CON, controls; MCI-Aβ+, mild cognitive impairment with amyloid pathology; AD, Alzheimer’s disease; DLB, dementia with Lewy bodies; FTD, Frontotemporal dementia; ITGB2, integrin subunit beta 2; DDC, dopa decarboxylase; THOP1, thimet oligopeptidase; ENO2, enolase 2; MMP7, Matrix metallopeptidase 7; NPX, normalized protein expression. # p < 0.05 (nominal significance), * q < 0.05, ** q < 0.01, *** q < 0.001 (significant after post-hoc analysis)

Fig. 3.

CSF proteins of the differential diagnostic panel are dysregulated in the AD profile group.

Protein concentrations in CSF from (A) ITGB2, (B) DDC, (C) THOP1, (D) ENO2, and (E) MMP7 were sequentially measured on the Ella platform. Patients having MCI with amyloid pathology and AD were grouped as AD profile. ITGB2, THOP1, and ENO2 were significant or tended to be dysregulated in the AD profile group compared with controls or non-AD dementia. DDC concentrations were increased in the patients with DLB compared with all other groups. Abbreviations: CON, controls; AD Profile, Alzheimer’s disease profile; DLB, dementia with Lewy bodies; FTD, Frontotemporal dementia; ITGB2, integrin subunit beta 2; DDC, dopa decarboxylase; THOP1, thimet oligopeptidase; ENO2, enolase 2; MMP7, Matrix metallopeptidase 7; # p < 0.05 (nominal significance), * q < 0.05, ** q < 0.01, *** q < 0.001 (significant after post-hoc analysis)

Next, we conducted a sensitivity analysis by excluding DLB and FTD patients having AD co-pathology. Trends are consistent with the analysis including DLB and FTD patients with AD co-pathology (Fig. 2). In addition, ENO2 was increased in AD compared to controls and patients with FTD (Supplementary Fig. 4D). Furthermore, in this analysis, MMP7 showed a nominal increase in AD and FTD compared to controls (Supplementary Fig. 4E).

CSF protein correlations with AD-CSF biomarkers and cognitive measurements

We next explored the relationship of the proteins within the differential diagnostic panel with the core AD-CSF biomarkers and MMSE scores, which reflect ongoing AD pathology and cognition. Moderate to strong correlations were observed for THOP1, ITGB2, and ENO2 with pTau and tTau (Rho’s between 0.46 and 0.57, Fig. 4 A). Weak inverse correlations between THOP1, ITGB2, ENO2, and DDC with Aβ42 were observed (Rho’s between − 0.14 and − 0.20; Fig. 4 A). These proteins together with MMP7 had a negative though weak correlation with MMSE scores (Rho’s between − 0.15 and − 0.26; Fig. 4 A). PCA analysis was performed, which showed that principal component 1 explains 32% of the variance (Supplementary Fig. 5). The biomarker composite score of the differential diagnostic panel showed significant though weak correlations with CSF Tau forms (tTau: Rho = 0.17; pTau: Rho = 0.13), CSF Aβ42 (Rho = −0.17), and MMSE score (Rho = −0.27; all p < 0.001; Fig. 4B).

Fig. 4.

Correlation matrix of the differential diagnostic panel proteins with AD-CSF biomarkers and MMSE scores.

Correlation matrix showing the correlations of the individual CSF proteins (A) and the biomarker panel composite score (B) with AD-CSF biomarkers and MMSE scores. The blue color depicts a positive correlation coefficient, and the red color depicts a negative correlation coefficient. Significance between the proteins is shown as * p < 0.05, ** p < 0.01, *** p < 0.001. Abbreviations: tTau, total Tau; pTau, phosphorylated Tau; Aβ42, amyloid-beta 1–42; MMSE, mini-mental state examination; ENO2, enolase 2; ITGB2, integrin subunit beta 2; MMP7, Matrix metallopeptidase 7; DDC, dopa decarboxylase; THOP1, thimet oligopeptidase; PC, Panel score

The CSF differential diagnostic panel discriminated between groups

Next, we investigated the discriminative performance of our CSF differential diagnostic panel. The CSF panel discriminated controls from MCI-Aβ+  (AUC = 0.771, 95% CI: 0.673–0.868, Fig. 2 F) and AD (AUC = 0.832, 95% CI: 0.752–0.911, Fig. 2 F) with moderate to good performance but performance was less upon discriminating AD from MCI-Aβ+ (AUC = 0.615, 95% CI: 0.490–0.740, Fig. 2 F). In addition, we observed that the CSF panel discriminated AD from non-AD dementia with moderate performance (AUC = 0.734, 95% CI: 0.646–0.822, Fig. 2 F). However, when the non-AD dementia group was stratified for clinical diagnosis, a moderate discriminative performance between AD and FTD was observed (AUC: 0.687, 95% CI: 0.575–0.798) while the discriminative performance was higher between AD and DLB (AUC: 0.831, 95% CI: 0.745–0.916).

We additionally compared the AD differential diagnostic panel to CSF NfL in a small subset of the cohort. NfL concentrations are increased in FTD and AD compared to controls, with significantly higher concentrations in FTD compared to AD (Supplementary Fig. 6A). The performance of the AD differential diagnostic panel was comparable to CSF NfL for discriminating AD from FTD (AUC CSF Panel: 0.811, 95% CI: 0.659–0.962; AUC CSF NfL: 0.767, 95%CI: 0.589–0.944; Supplementary Fig. 6B).

Discussion

In this study, we developed and analytically validated different immunoassays targeting five proteins from the differential diagnostic panel identified in our previous CSF proteomic study [7]. These proteins cover different biological pathways affected in AD pathology (i.e., cellular remodeling, neuropeptide degradation, energy metabolic processes, the immune system, and neurotransmitter synthesis). CSF proteins were dysregulated along different AD stages (THOP1, ENO2, and ITGB2) and between AD and non-AD dementias (THOP1, ITGB2, ENO2, and DDC). Most of the proteins correlated with pTau and tTau, highlighting the association with AD Tau pathology. The combination of THOP1, ITGB2, ENO2, DDC, and MMP7 discriminated between AD and controls with good performance, while between AD and non-AD groups moderate to good performances were observed.

As described previously, following the fluid biomarker development workflow [8, 9], we aimed to translate our novel differential diagnostic panel identified by our previous antibody-based discovery study to immunoassays on an accessible technology that could be used for wide-scale analysis and applied in clinical routine or trials [46]. We encountered technical difficulties in the development of immunoassays for four out of nine proteins previously identified (i.e., ABL1, VEGFR-3, PTK7, and GZMB). It is worth noting that PTK7 and GZMB were not detectable with the custom-PEA assays, while ABL1 and VEGFR-3 exhibited concentrations close to the lower limit of detection. Thus, technologies with a higher sensitivity such as Simoa may help to optimally measure ABL1 and VEGFR-3 protein concentrations in CSF. Nonetheless, we successfully developed four in-house immunoassays (i.e., THOP1, DDC, ITGB2, and ENO2) and used one commercial immunoassay for the measurement of MMP7. When comparing these protein concentrations to protein levels previously obtained with our custom-PEA assays, we observed moderate to strong correlations between platforms (Rho’s between 0.64 and 0.93), which indicates that we are likely measuring similar protein isoforms across these technologies [7]. It should be noted that MMP7 showed the highest correlation between technologies. As this is a commercially available assay, it has likely undergone extensive validation and optimization by the manufacturer, which may account for its higher analytical performance. This efficient translation of the proteins from the discovery phase to analytical and clinical validation was also observed in other antibody-based proteomics studies [10, 38, 47], emphasizing the use of the workflow employed in this study for the effective development of immunoassays for novel fluid biomarkers. Noteworthy, these novel immunoassays could be useful to quantify CSF protein concentrations along the AD continuum and also in other matrices and cellular models to further understand the involvement of these proteins in AD pathogenesis. The proteins from the CSF differential diagnostic panel for which we had optimal immunoassays are associated with different biological pathways such as immune system (ITGB2), neuropeptide degradation (THOP1), energy metabolic processes (ENO2), neurotransmitter synthesis (DDC), and cellular remodeling (MMP7). The protein changes of CSF ITGB2, THOP1, and ENO2 observed in this study are in agreement with previous findings [10, 21, 38], and underscore their potential as AD-specific biomarkers reflecting a dysregulation of these pathways. In addition, the elevated DDC concentrations in patients with DLB align with previous research suggesting DDC as a potential biomarker for LBD [14–18, 24, 48]. Noteworthy, when patients with AD co-pathology were excluded from the non-AD groups, stronger effects of CSF ITGB2, THOP1, and ENO2 were observed, which underscores the associations of these proteins with ongoing AD pathophysiology. In addition, we show that all the CSF proteins analyzed correlated with different pathophysiological features associated with AD or cognitive decline. Some proteins (ITGB2, ENO2, and DDC) showed weak correlations with CSF amyloid, while the correlations of other proteins (THOP1, ITGB2, ENO2, and MMP7) with CSF pTau and tTau were stronger. The effect and direction of these correlations align with previous findings, highlighting the association for some of these proteins with Tau pathology and neurodegenerative processes [7, 14, 38, 49]. Noteworthy, although THOP1 is functionally linked to amyloid processing, we observed stronger correlations with pTau and tTau, suggesting a closer association with tau pathology and neurodegenerative processes that warrants further investigation. In addition, a correlation between THOP1 and Aβ42 was observed within the control group [38], which suggests that THOP1’s association with amyloid may be disease-dependent and is not consistently detectable when different diagnostic groups are combined. In line with our previous findings, we observed negative though weak correlations of CSF ITGB2, ENO2, MMP7, and DDC with MMSE scores [7]. When we combined the proteins to a biomarker panel composite score, we observed a stronger correlation with MMSE scores, which suggests an association of the CSF differential diagnostic panel with cognition and possibly clinical dementia progression. Given that MMSE reflects global cognition, it would be valuable to explore the correlation between this panel and additional neuropsychological tests, such as those assessing memory, language, attention, and executive functioning, over an extended period. This could help to determine the prognostic value of this CSF biomarker panel. We next investigated the discriminative performance of the CSF differential diagnostic panel with five out of the nine proteins originally selected within the discovery study [7]. The combination of CSF THOP1, ITGB2, ENO2, DDC, and MMP7 could discriminate AD from non-AD dementia with moderate accuracy (AUC: 0.72), which was lower compared to our previous findings (AUC’s >0.8) [7]. The difference between this study and previous discovery results could be explained by the heterogeneity of the non-AD group, which includes both FTD and DLB individuals. These types of dementia are highly heterogeneous with limited biological phenotyping compared to AD, and thus differences within these groups across studies may partially explain the lower performance [50, 51]. Furthermore, ABL1, a tyrosine kinase expressed in neurons and involved in Tau phosphorylation, was one of the most significantly dysregulated proteins between AD and non-AD dementias [52]. We could not develop an Ella immunoassay for ABL1 and was therefore not included in the current panel. The exclusion of ABL1 could to a great extent explain the lower performance observed in this study [7]. However, we observed that the five-biomarker CSF panel performed well in discriminating AD from DLB patients (AUC: 0.83), but this is likely due to the inclusion of DDC, a novel biomarker associated with DLB pathophysiology [14–18, 24, 48]. The development of more sensitive immunoassays (e.g., Simoa) for the remaining biomarker candidates (e.g., ABL1 and VEGFR-3) could help to increase and ultimately establish the optimal performance of our CSF panel. In addition, hierarchical clustering analyses based on the CSF panel would be highly informative. Unfortunately, the present study lacks sufficient statistical power to reliably perform such analyses, but this approach should be explored in future studies. Furthermore, we compared the performance of the CSF panel to CSF NfL in a subset of the cohort, which showed similar discriminative performance. Though results should be interpreted with caution considering the large confidence intervals, probably due to the limited sample size. Nonetheless, the proteins in this panel show added value since they encompass a wide range of biological pathways, including those related to the immune system, neuropeptide degradation, energy metabolic processes, neurotransmitter synthesis, and cellular remodeling. The incorporation of such diverse pathways in a panel would enrich our understanding of various processes associated with AD in a clinical setting. It would also enable the monitoring of these processes in clinical trials, allowing for the assessment of their responsiveness to treatments [53].

This study is not without limitations. The added value of this biomarker panel in comparison to the classical AD-CSF biomarkers could not be evaluated since the patients with AD were selected based on these criteria. Furthermore, our cohort is relatively small, and only patients from our memory clinic were included. In addition, we cannot exclude misdiagnosis in some of the FTD and DLB from the non-AD group. However, diagnoses were made during a multidisciplinary consensus meeting in a specialized memory clinic, likely decreasing the chance of misdiagnosis. It would be of interest to further validate this panel in external independent clinical and neuropathological cohorts. One of the strengths of our study is the workflow we applied. We started the biomarker discovery process with an antibody-based technology, which resulted in a swift translation to clinical validation using independent immunoassays, potentially facilitating the implementation of novel fluid biomarkers.

Conclusion

Overall, we have developed a CSF differential diagnostic panel covering different biological pathways dysregulated in with AD (i.e., cellular remodeling, neuropeptide degradation, energy metabolic processes, the immune system, and neurotransmitter synthesis) using immunoassays that are optimal for wide-scale analysis. While our CSF panel showed lower discriminative performance than our original discovery study, the differences and trends observed with all proteins are consistent. Thus, this CSF panel could still be useful to capture the complexity of processes involved in AD pathophysiology. Further validation studies with other relevant markers (e.g., ABL1) in external independent cohorts, including those from clinical trials, are needed to establish the relevance and potential context-of-use for this CSF panel.

Authors’ contributions

YSH contributed to the study design and was responsible for acquisition, experiments, statistical analysis, interpretation of the data, and drafting of the manuscript. BB analytically validated the ITGB2 immunoassay and contributed to data acquisition. KB developed the DDC immunoassay and performed measurements of DDC in clinical samples. LB supervised the analytical validation of the MMP7 immunoassay. DR developed and analytically validated the THOP1 and ENO2 immunoassay. AL, WF, and YP provided patient material. MC and CT conducted the study concept and design, data interpretation, and critically revised the manuscript and supervised the study. All authors read and approved the manuscript.

Funding

This project was funded by the memorable project PRIDE (Project No. WE.03-2018-05), supported by Alzheimer Nederland. MC is supported by the attraction talent fellowship from Comunidad de Madrid (2018-T2/BMD-11885).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

All procedures involving human subjects were performed following the Helsinki Declaration of 1975. All participants or their authorized representatives provided written informed consent for the use of medical data and biomaterials for research purposes. The study was approved by the Amsterdam UMC VUmc medical ethical committee.

Consent for publication

Not applicable.

Competing interests

Y.S.H., B.B, K.B., L.B., D.R., A.L., and M.C. report no conflicts of interest. Y.P. received funding from Stichting Dioraphte, although this funding source had no role in the design, practice, or analysis of this study. As of 1-11-2025, W.F is executive director at Alzheimer Nederland, Amersfoort, the Netherlands. Before 1-11-2025, research programs of W.F. have been funded by ZonMW, NWO, EU-FP7, EU‐JPND, Alzheimer Nederland, Hersenstichting CardioVascular Onderzoek Nederland, Health∼Holland, Topsector Life Sciences & Health, stichting Dioraphte, Gieskes‐Strijbis fonds, stichting Equilibrio, Edwin Bouw fonds, Pasman stichting, stichting Alzheimer & Neuropsychiatrie Foundation, Philips, Biogen MA Inc, Novartis‐NL, Life‐MI, AVID, Roche BV, Fujifilm, Eisai, and Combinostics. W.F. holds the Pasman chair. W.F. is recipient of ABOARD, which is a public‐private partnership receiving funding from ZonMW (#73305095007) and Health∼Holland, Topsector Life Sciences & Health (PPP‐allowance; #LSHM20106). Before 1-11-2025, W.F. has been an invited speaker at Biogen MA Inc, Danone, Eisai, WebMD Neurology (Medscape), NovoNordisk, Springer Healthcare, NovoNordisk, and European Brain Council. W.F. is a consultant to Oxford Health Policy Forum CIC, Roche, and Biogen MA Inc. W.F. participated in advisory boards of Biogen MA Inc, Roche, and Eli Lilly. 2024-2025, W.F has been member of the steering committee of phase 3 EVOKE/EVOKE+ studies (NovoNordisk). In 2025, W.F has been member of the steering committee op phase 3 Trontinemab study (Roche). All funding has been paid to Amsterdam UMC. W.F. was associate editor of Alzheimer’s Research & Therapy in 2020/2021. W.F. was associate editor at Brain 2021-2025. W.F is chair of the Scientific Leadership Group of InRAD.W.F is member of Supervisory Board (Raad van Toezicht) Trimbos Instituut. C.T. has a collaboration contract with ADx Neurosciences, Quanterix, and Eli Lilly, and performed contract research or received grants from AC‐Immune, Axon Neurosciences, Bioconnect, Bioorchestra, Brainstorm Therapeutics, Celgene, EIP Pharma, Eisai, Grifols, Novo Nordisk, PeopleBio, Roche, Toyama, and Vivoryon. C.T. serves on editorial boards of Medidact Neurologie/Springer, Alzheimer’s Research & Therapy, and Neurology: Neuroimmunology & Neuroinflammation and is editor of a Neuromethods book (Springer). She had speaker contracts for Roche, Grifols, and Novo Nordisk.