Genetic Prioritization of Plasma Proteins Across Dementia Subtypes: A Proteome-Wide Mendelian Randomization and Bayesian Colocalization Analysis

Abstract

Dementia is clinically and biologically heterogeneous, and genetically supported plasma protein associations across dementia subtypes remain incompletely characterized. We aimed to genetically prioritize plasma proteins associated with overall dementia and major dementia subtypes using proteome-wide Mendelian randomization (MR) and Bayesian colocalization. We used the UK Biobank Pharma Proteomics Project (UKB-PPP) as a harmonized source of cis-protein quantitative trait loci (cis-pQTLs) for 2,941 Olink-measured plasma analytes. The European-ancestry discovery cis-pQTL analysis included up to 34,557 participants, whereas 54,219 participants were profiled in the overall UKB-PPP. Outcomes included FinnGen R10 genome-wide association studies (GWAS) for dementia, Alzheimer’s dementia, vascular dementia, frontotemporal dementia, and unspecified dementia, plus a European GWAS for dementia with Lewy bodies. At the nominal threshold of P < 0.05, 134, 132, 102, 78, 101, and 112 exploratory protein–outcome signals were observed for dementia, Alzheimer’s dementia, vascular dementia, frontotemporal dementia, dementia with Lewy bodies, and unspecified dementia, respectively; 23 associations remained significant after outcome-specific Benjamini–Hochberg false discovery rate (FDR) correction. Key FDR-significant proteins included APOE, NECTIN2, PVR, SERPINF2, GRN, TREM2, and ATXN2L. The strongest risk and protective associations were NECTIN2 with Alzheimer’s dementia (odds ratio [OR] = 2.650, 95% confidence interval [CI]: 1.875–3.744, FDR-adjusted P = 1.58 × 10^− 5) and ATXN2L with dementia (OR = 0.408, 95% CI: 0.255–0.654, FDR-adjusted P = 0.0370), respectively. Strong colocalization support was observed for selected SERPINF2 and TREM2 associations. These findings genetically prioritize several plasma proteins as candidate biomarkers and potential therapeutic targets across dementia subtypes, but do not establish that circulating protein abundance directly mediates dementia risk.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s10571-026-01819-2.

Introduction

Dementia is a syndrome characterized by a severe decline in cognitive function affecting approximately 5% of individuals aged > 65 years (Ritchie and Lovestone 2002). It can be classified into various subtypes based on its underlying pathology, including Alzheimer’s dementia (ADD), vascular dementia (VaD), frontotemporal dementia (FTD), dementia with Lewy bodies (DLB), and unspecified dementia (Elahi and Miller 2017). Among these, ADD is the most prevalent and has become one of the most expensive, lethal, and burdensome diseases of this century (Scheltens et al. 2021). Although the exact causes of the different dementia subtypes remain unclear, they are thought to share common neuropathological features (Ritchie and Lovestone 2002). Current therapeutic options for Alzheimer’s disease remain limited, reflecting persistent gaps in understanding and therapeutically targeting its pathogenesis (Loera-Valencia et al. 2019).

With advancements in plasma proteomics and genome-wide association studies (GWAS), new opportunities have emerged for identifying drug targets and understanding comorbid processes among various dementia subtypes. Recent large-scale GWAS have expanded the genetic architecture of Alzheimer’s disease (Jansen et al. 2019; Andrews et al. 2020). Circulating proteins are major regulators of molecular pathways and provide insights into potential therapeutic targets and clinical intervention strategies (Finan et al. 2017). For example, plasma proteomic profiling has identified circulating proteins associated with future dementia risk in initially healthy adults (Guo et al. 2024). However, these observational studies are often subject to reverse causation and confounding factors, making it difficult to determine whether the identified proteins contribute to dementia onset or result from the disease progression (Davies et al. 2018).

Large-scale plasma proteomic studies have identified circulating proteins and biological pathways associated with incident dementia risk, and two-sample Mendelian randomization (MR) analyses have further implicated selected dementia-associated proteins in Alzheimer’s disease (Walker et al. 2021). More recent MR-based proteomic studies have also used genetic instruments to investigate plasma protein associations with Alzheimer’s disease, identifying genetically prioritized protein biomarkers and potential therapeutic targets (Yao et al. 2024; Belbasis et al. 2025; Li et al. 2025). However, most of these studies have focused primarily on Alzheimer’s disease or broader neurodegenerative disease categories, and genetically supported associations between plasma proteins and other dementia subtypes, including VaD, FTD, DLB, and unspecified dementia, remain incompletely explored.

To address this gap, we sought genetic evidence on plasma protein associations across dementia subtypes to support candidate prioritization for future mechanistic and therapeutic studies. MR uses genetic variants as instrumental variables to strengthen causal inference under specific assumptions (Thanassoulis and O’Donnell 2009). Because germline genetic variants are fixed before disease onset, MR is generally less susceptible than conventional observational analyses to confounding and reverse causation; however, it does not by itself establish that the measured plasma protein abundance is the causal mediator. Additionally, Bayesian colocalization can test whether the identified associations are likely to share causal variants rather than reflect linkage disequilibrium (LD) between distinct variants, thereby strengthening the genetic interpretation of selected MR findings. In this study, we conducted a proteome-wide MR analysis using data from the UK Biobank Pharma Proteomics Project (UKB-PPP) and FinnGen project to investigate various types of dementia. Our findings revealed significant associations between several plasma proteins and dementia, with Apolipoprotein E (APOE), Nectin Cell Adhesion Molecule 2 (NECTIN2), Poliovirus receptor (PVR), Serpin family F member 2 (SERPINF2), Granulin (GRN), and Triggering receptor expressed on myeloid cells 2 (TREM2) showing notable association with different dementia subtypes. Bayesian colocalization analysis provided additional support for selected associations. Functional network and druggability analyses further prioritized biologically and pharmacologically relevant candidates. These findings provide a cross-subtype genetic prioritization framework for plasma protein associations and nominate candidates for future mechanistic and therapeutic investigation.

Materials and Methods

Study Design

This study was performed in compliance with the Strengthening the Reporting of Observational Studies in Epidemiology using the Mendelian Randomization (STROBE-MR) guidelines (Skrivankova et al. 2021). A flowchart of the analysis is shown in Fig. 1. We used available GWAS summary statistics to conduct two-sample MR analyses of associations between genetically predicted plasma protein levels and different dementia subtypes. Multiple-testing correction was prespecified for the primary MR estimates and is described in the MR analysis section. For the MR analysis, the following three key assumptions were made: (Amin et al. 2025) the selected instrumental variables (IVs) are strongly associated with the plasma protein exposure; (Andrews et al. 2020) the IVs are independent of confounders of the plasma protein–dementia association; and (Belbasis et al. 2025) the IVs affect dementia outcomes only through the measured plasma protein exposure. These assumptions cannot be verified completely using summary-level data, particularly for associations supported by a single or limited number of instrumental variables. Accordingly, the MR estimates were interpreted as genetically supported associations rather than definitive proof that circulating protein abundance directly causes dementia.

Fig. 1.

Flowchart of the study design. Plasma protein genome-wide association study (GWAS) summary data were obtained from the UK Biobank Pharma Proteomics Project (UKB-PPP), and dementia outcome GWAS summary data were obtained from FinnGen R10 and the Integrative Epidemiology Unit (IEU) OpenGWAS project. Two-sample Mendelian randomization (MR) was performed using inverse-variance weighted (IVW) and Wald ratio methods as primary methods, with MR-Egger, weighted median, weighted mode, MR robust adjusted profile score (MR-RAPS), contamination mixture (ConMix), Bayesian weighted MR (BWMR), debiased inverse-variance weighted (dIVW), and constrained maximum likelihood with model averaging (cML-MA) methods as supplementary analyses. Sensitivity analyses included Cochran’s Q test, MR-Egger intercept test, and leave-one-out analysis. Bayesian colocalization analysis was then performed, followed by protein-protein interaction network analysis and druggability assessment. ADD Alzheimer’s dementia, VaD vascular dementia, FTD frontotemporal dementia, DLB dementia with Lewy bodies, pQTL protein quantitative trait locus, PP.H4 posterior probability for hypothesis 4

Bayesian colocalization analysis was conducted to assess whether the protein and dementia association signals were compatible with a shared regional causal variant rather than distinct variants correlated through LD. Protein interactions, function, network analyses, and druggability assessments were performed on the identified dementia-related plasma proteins.

The analyses were conducted using the “TwoSampleMR,” “MendelianRandomization,” “LocusCompareR,” and “coloc” packages in R Software 4.3.3 (https://www.R-project.org).

Data Resources

Details of the datasets used in this study are provided in Table 1. Specifically, we prioritized publicly available summary statistics that provided: (Amin et al. 2025) broad proteome coverage for hypothesis-generating proteome-wide MR; (Andrews et al. 2020) internally harmonized measurement and genetic association pipelines to reduce between-study heterogeneity in protein quantification and protein quantitative trait locus (pQTL) derivation; (Belbasis et al. 2025) consistent ancestral background across exposure and outcome datasets to minimize bias from population stratification; (Burgess et al. 2020) large sample size and reliable case ascertainment for dementia phenotypes; and (Burgess et al. 2017) sufficient summary-level information for instrument selection, harmonization, and downstream colocalization analyses. On this basis, we used the UK Biobank Pharma Proteomics Project as the primary exposure resource because it offers large-scale, uniformly processed cis-protein quantitative trait locus (cis-pQTL) data for 2,941 Olink-measured plasma proteins in European participants. For outcomes, we preferentially used FinnGen R10 for overall dementia and most dementia subtypes because it provides a large, recent, and relatively homogeneous European dataset with consistent diagnostic ascertainment across phenotypes. Dementia with Lewy bodies was analyzed using an additional European GWAS because a directly comparable dataset within the same primary outcome framework was not available. We did not combine multiple proteomic resources in the main analysis, as cross-platform differences in protein measurement, normalization, variant processing, and pQTL definition may introduce substantial heterogeneity and reduce the interpretability of subtype-comparative MR results.

Table 1.

Detailed information on the GWAS datasets used in this study

Phenotype	Source	Cases or participants	Controls	Year	Population	PMID
Plasma proteins	UKB-PPP	Up to 34,557	NA	2023	European	37,794,186
Dementia	FinnGen R10	19,157	388,560	2023	Finnish	36,653,562
Alzheimer’s dementia	FinnGen R10	6,145	401,661	2023	Finnish	36,653,562
Vascular dementia	FinnGen R10	2,717	393,024	2023	Finnish	36,653,562
Frontotemporal dementia	FinnGen R10	129	392,463	2023	Finnish	36,653,562
Unspecified dementia	FinnGen R10	4,408	388,560	2023	Finnish	36,653,562
Dementia with Lewy bodies	IEU OpenGWAS	2,591	4,027	2021	European	33,589,841

GWAS genome-wide association study, IEU Integrative Epidemiology Unit, NA not applicable, PMID PubMed identifier, pQTL protein quantitative trait locus, UKB-PPP UK Biobank Pharma Proteomics Project. For plasma proteins, the table reports the maximum European-ancestry discovery pQTL sample size (up to 34,557); the overall UKB-PPP profiled 54,219 participants, and protein-specific effective sample sizes may vary slightly after assay-specific quality control

Notably, plasma proteomic measurements were available for 54,219 participants in the overall UKB-PPP. The European-ancestry discovery pQTL analyses used up to 34,557 randomly selected baseline participants, with the remaining samples used for replication. The corresponding European-ancestry discovery pQTL summary statistics, covering 2,941 plasma proteins measured with the Olink platform, were used for instrument selection. In the discovery analysis, protein phenotypes were inverse-rank normalized before pQTL association testing and were therefore analyzed on an approximately standardized scale; protein-specific effective sample sizes could be slightly lower after assay-specific quality control. These proteins were categorized into eight groups: Cardiometabolism, Cardiometabolism II, Inflammation, Inflammation II, Neurology, Neurology II, Oncology, and Oncology II. The UKB-PPP identified 14,287 major genetic associations, with 81% previously undescribed ones, thus significantly expanding the catalog of genetic tools for downstream analyses (Sun et al. 2023).

Furthermore, GWAS summary data for dementia and its subtypes were obtained from FinnGen R10, comprising 412,181 participants and analyzing up to 21,311,942 variants released in December 2023. The FinnGen outcomes were derived from individuals of Finnish ancestry. Because the exposure and outcome GWAS were obtained from separate resources, substantial participant overlap was not expected; however, it could not be quantified from summary-level data. The sample sizes were as follows: dementia (19,157 cases and 388,560 controls), ADD (6,145 cases and 401,661 controls), VaD (2,717 cases and 393,024 controls), FTD (129 cases and 392,463 controls), and unspecified dementia (4,408 cases and 388,560 controls). All patients were diagnosed in hospitals and genotyped using Illumina and Affymetrix arrays. Further details on data processing and quality control can be found in a previously published study (Kurki et al. 2023). Additionally, data on DLB were obtained from a study by Chia et al., which included 6,618 European participants (2,591 cases and 4,027 controls) (Chia et al. 2021).

Selection of IVs

The pQTLs from UKB-PPP were used to select IVs. Index cis-acting single-nucleotide polymorphisms (SNPs) were defined as SNPs within 1,000 kb of the gene encoding the corresponding protein. Furthermore, SNPs strongly associated with plasma proteins at P < 5 × 10⁻⁸ were selected. Instrument strength was evaluated separately for each variant as F = (β_exposure/SE_exposure)², where β_exposure and SE_exposure denote, respectively, the effect estimate and standard error (SE) for the pQTL association in UKB-PPP. Variants with an F-statistic < 10 were excluded to minimize weak-instrument bias. This direct calculation avoids additional assumptions regarding phenotype variance, allele frequency, and effective per-variant sample size. For the independence of the IVs, LD clumping (r² < 0.001, window size = 10,000 kb) was performed. Palindromic SNPs with intermediate allele frequencies were excluded from the analysis.

LD estimates were calculated using the 1000 Genomes Project Phase 3 European reference panel comprising 503 founder individuals. Because APOE, NECTIN2, and PVR are located within the same extended chromosome 19 region, we assessed exact instrument overlap and pairwise LD among the union of the 15 selected instruments using PLINK v1.9 and the same reference panel. We also queried each selected instrument against all three original UKB-PPP protein measurements and reported variant-level β, SE, and P values for all 45 instrument–protein combinations. Cross-protein associations were evaluated at P < 5 × 10⁻⁸. LD between the APOE instruments and the missense variants rs429358 and rs7412 was also assessed. These marginal pQTL associations were not conditional on rs429358 or rs7412. An r² threshold of 0.8 was used as a conventional criterion for a high-LD proxy, while all exact r² estimates were reported (Carlson et al. 2004).

MR Analysis

Two-sample MR was performed to estimate associations between genetically predicted plasma protein levels and dementia-related outcomes using 10 methods: inverse-variance weighted (IVW), MR-Egger regression, weighted median, Wald ratio, weighted mode, MR robust adjusted profile score (MR-RAPS), contamination mixture (ConMix), Bayesian weighted MR, debiased IVW, and constrained maximum likelihood with model averaging (cML-MA). The IVW method was used as the primary MR method for associations supported by multiple IVs, whereas the Wald ratio was used for associations supported by a single IV. Other methods were used as supplementary analyses to assess cross-method consistency. Because IVW can only provide an estimate when more than one IV was available, the Wald ratio was used when the number of IVs was only one (Burgess et al. 2017). MR-RAPS provides robust estimation in the presence of weak instruments, ConMix is robust to invalid instruments (Burgess et al. 2020), and cML-MA is designed to provide robust estimation in the presence of correlated and uncorrelated pleiotropic effects (Xue et al. 2021). Associations showing concordant effect directions across applicable supplementary methods were considered to have greater cross-method consistency, but such concordance was not interpreted as definitive causal confirmation.

Sensitivity analyses were performed to evaluate the stability and robustness of the two-sample MR results. Cochran’s Q test was used to assess heterogeneity, and a P value < 0.05 was considered evidence of significant heterogeneity. The MR-Egger intercept test was used to assess directional horizontal pleiotropy, and a P value < 0.05 was considered evidence of significant horizontal pleiotropy. Because these sensitivity analyses require a sufficient number of instrumental variables, Cochran’s Q test was only performed for associations supported by at least two SNPs, whereas the MR-Egger intercept test was only performed for associations supported by at least three SNPs. Leave-one-out analysis was performed for associations supported by at least two SNPs, when applicable, to evaluate whether any single SNP disproportionately drove the observed association.

To assess robustness to instruments with genome-wide-significant cross-protein associations, we repeated the APOE IVW analyses after excluding rs1836278 and the PVR IVW analyses after excluding rs2272021. A protein-specific restricted NECTIN2 analysis could not be estimated because both selected NECTIN2 instruments (rs440277 and rs4452060) were genome-wide significantly associated with APOE; excluding both left no valid instrument. These analyses were treated as nominal sensitivity analyses rather than new outcome-wide FDR tests.

To further assess LD with the APOE missense-variant signals, we repeated the APOE IVW analyses after excluding rs12972156 and rs62117206 one at a time and after excluding both instruments jointly. These analyses were treated as LD-informed exclusion sensitivity analyses, not as formal pQTL analyses conditional on rs429358 and rs7412.

Multiple-testing correction was performed separately within each dementia outcome using the Benjamini–Hochberg false discovery rate (FDR) method across all primary protein–outcome associations tested for that outcome. The primary estimate was the IVW estimate for associations supported by multiple instruments and the Wald ratio estimate for associations supported by a single instrument. A total of 1,925 primary protein–outcome associations were tested for each of dementia, Alzheimer’s dementia, vascular dementia, frontotemporal dementia, and unspecified dementia, whereas 1,817 were tested for dementia with Lewy bodies. Associations with an FDR-adjusted P < 0.05 were considered statistically significant, whereas associations with nominal P < 0.05 but an FDR-adjusted P ≥ 0.05 were treated as exploratory signals.

Colocalization Analysis

Bayesian colocalization analysis was conducted for the FDR-significant associations to assess whether the protein and dementia signals were compatible with a shared regional causal variant rather than distinct variants correlated through LD. The colocalization analysis followed five exclusive hypotheses: (1) no associations with plasma proteins and dementia (H0); (2) associations with only plasma proteins (H1); (3) associations with only dementia (H2); (4) both traits are associated but have distinct causal variants (H3); and (5) both traits share the same causal variant (H4). The posterior probabilities for H3 and H4 were denoted as PP.H3 and PP.H4, respectively. The following default prior probabilities were used: the probability that a SNP was associated with the plasma protein (p1) was 1 × 10⁻⁴, the probability that it was associated with the dementia outcome (p2) was 1 × 10⁻⁴, and the probability that it was associated with both traits (p12) was 1 × 10⁻⁵ (Giambartolomei et al. 2014). The analysis provided strong colocalization evidence if PP.H4 ≥ 0.8 and medium evidence if 0.5 < PP.H4 < 0.8.

Given the complex LD structure of the APOE locus, we additionally determined whether rs429358 and rs7412 were located within each ± 1-megabase (Mb) APOE colocalization window and quantified their pairwise LD with the corresponding regional lead variants. The APOE colocalization findings were interpreted in light of the single-causal-signal assumption of the conventional coloc model.

Function and Network of Identified Proteins and Druggability Assessment

For the cis-genes of different subtypes of dementia-associated plasma proteins identified by the MR method, we constructed a protein-protein interaction (PPI) network based on the STRING database (https://string-db.org/) and GeneMANIA (https://genemania.org/). The STRING database can be used to construct interactions between proteins, including physical and functional associations. GeneMANIA can effectively predict the functions of identified proteins (Warde-Farley et al. 2010; Szklarczyk et al. 2023). Next, to assess whether the identified proteins could serve as potential drug targets for the treatment of different subtypes of dementia, we searched databases, including the Drug-Gene Interaction Database (DGIdb), ChEMBL, and DrugBank, to determine the interactions between the identified proteins and drugs (Wishart et al. 2018; Mendez et al. 2019; Freshour et al. 2021).

Results

Selection of IVs and Overview

The final IVs selected from the UKB-PPP are detailed in Supplementary Table 1, Sheet A. Across 5,683 unique protein–SNP instruments, the recalculated per-variant F-statistics ranged from 29.733 to 15,013.276, and no instrument had an F-statistic ≤ 10. The complete per-variant verification results are provided in Supplementary Table 1, Sheet B (“F-statistic verification”). Using the primary MR estimates, nominal P < 0.05 signals were observed for 134, 132, 102, 78, 101, and 112 protein–outcome associations for dementia, ADD, VaD, FTD, DLB, and unspecified dementia, respectively (Figs. 2 and 3). Given the number of primary tests performed for each outcome, these nominal counts were considered exploratory and were not interpreted as primary findings. Additional MR estimates from the other methods are provided in Supplementary Tables 2, and the sensitivity test results are shown in Supplementary Tables 3, 4, 5. After FDR correction, 23 associations remained significant at an FDR-adjusted P value < 0.05 (Fig. 4). Among these 23 associations, 15 were supported by multiple independent instruments and were estimated using IVW, whereas eight were supported by a single instrument and were estimated using the Wald ratio. Among the 15 multi-instrument associations, 11 had at least three instruments, allowing Cochran’s Q, MR-Egger intercept, and leave-one-out analyses, whereas four had two instruments, allowing Cochran’s Q and leave-one-out analyses but not the MR-Egger intercept test. Conventional multi-instrument sensitivity analyses were not applicable to the eight single-instrument associations. Detailed evidence classification is provided in Supplementary Table 6, Sheet G (“Evidence classification”). Strong colocalization support was observed for three FDR-significant associations. TREM2–dementia was supported by three instruments and strong colocalization, whereas the SERPINF2 associations with dementia and Alzheimer’s dementia showed strong colocalization but were each supported by a single instrument. No association showed moderate colocalization evidence. Bayesian colocalization was conducted for all 23 FDR-significant protein–outcome associations. Druggability searches identified more than 40 candidate drugs or database entries targeting 14 prioritized proteins.

Fig. 2.

Volcano plots of Mendelian randomization (MR) associations between genetically predicted plasma protein levels and dementia subtypes. The x-axis represents the MR effect estimate β from the primary MR method. For binary dementia outcomes, β corresponds to the natural logarithm of the odds ratio (OR), namely β = ln (OR); therefore, OR = e ^ β, where e is Euler’s number, the base of the natural logarithm. The y-axis represents −log10(P value). Each point represents one plasma protein. Pink points indicate nominal risk associations (β > 0 and P < 0.05), green points indicate nominal protective associations (β < 0 and P < 0.05), and gray points indicate neutral, non-significant associations. The vertical dashed line indicates β = 0, and the horizontal dashed line indicates the nominal significance threshold of P = 0.05. A: dementia. B: Alzheimer’s dementia. C: vascular dementia. D: frontotemporal dementia. E: dementia with Lewy bodies. F: unspecified dementia

Fig. 3.

Circular heatmaps of exploratory nominal Mendelian randomization (MR) signals between genetically predicted plasma protein levels and dementia subtypes. The plots display exploratory protein–outcome signals observed by the primary MR method at P < 0.05. Proteins are arranged according to the P value from the primary MR analysis rather than by odds ratio (OR) magnitude, and the color gradient represents the corresponding P value. These plots are intended to visualize the relative P-value ranking of exploratory nominal signals, whereas the detailed ORs and 95% confidence intervals (CIs) are provided in Supplementary Table 2. A: dementia. B: Alzheimer’s dementia. C: vascular dementia. D: frontotemporal dementia. E: dementia with Lewy bodies. F: unspecified dementia

Fig. 4.

False discovery rate (FDR)-significant Mendelian randomization associations between genetically predicted plasma protein levels and dementia subtypes. Only associations that remained significant after FDR correction are shown. Odds ratios (ORs) greater than 1 indicate risk associations, whereas ORs less than 1 indicate protective associations. Error bars represent 95% confidence intervals (CIs). nsnp indicates the number of single-nucleotide polymorphisms (SNPs) used as instrumental variables. Associations supported by multiple independent SNPs were estimated primarily using the inverse-variance weighted method, whereas associations supported by a single SNP were estimated using the Wald ratio. Conventional multi-instrument sensitivity analyses, including Cochran’s Q, MR-Egger intercept, and leave-one-out analyses, were not applicable to single-instrument findings. Instrument count and evidence classification for all FDR-significant associations are provided in Supplementary Table 6. ACE angiotensin-converting enzyme, APOE apolipoprotein E; ATXN2L ataxin-2-like protein, BLNK B-cell linker protein, CSF3 granulocyte colony-stimulating factor, DPEP1 dipeptidase 1, EPHB4 ephrin type-B receptor 4, FKBPL FK506-binding protein-like, GRN granulin, LILRA5 leukocyte immunoglobulin-like receptor subfamily A member 5, NECTIN2 nectin cell adhesion molecule 2, PVR poliovirus receptor, SERPINF2 serpin family F member 2, TREM2 triggering receptor expressed on myeloid cells 2

Associations Between Plasma Proteins and Dementia/ADD

At nominal P < 0.05, 134 and 132 exploratory protein–outcome signals were observed for dementia and ADD, respectively (Fig. 2A, B; Supplementary Table 2). For dementia, the odds ratio (OR) per standard deviation increase in genetically predicted plasma protein level ranged from 0.408 (95% confidence interval [CI]: 0.255–0.654) for Ataxin-2-like protein to 2.589 (95% CI: 1.571–4.269) for FK506-binding protein-like, based on the primary MR estimates in Supplementary Table 2. For ADD, the OR ranged from 0.101 (95% CI: 0.027–0.373) for HEPACAM family member 2 to 2.650 (95% CI: 1.875–3.744) for NECTIN2, based on the primary MR estimates in Supplementary Table 2. Figure 3A, B visualize these exploratory nominal signals ranked by P value rather than by OR magnitude. After FDR correction, 13 proteins remained significantly associated with dementia and six with ADD (Fig. 4). Among the 13 FDR-significant associations with overall dementia, eight were multi-instrument IVW associations and five were single-instrument Wald ratio associations. Among the six FDR-significant associations with ADD, three were supported by multiple instruments and three by a single instrument. The eight FDR-significant single-instrument associations were GRN–dementia, SERPINF2–dementia, FK506-binding protein-like (FKBPL)–dementia, ataxin-2-like protein (ATXN2L)–dementia, B-cell linker protein (BLNK)–dementia, GRN–Alzheimer’s dementia, SERPINF2–Alzheimer’s dementia, and granulocyte colony-stimulating factor (CSF3)–Alzheimer’s dementia. Single-instrument associations are identified explicitly in Fig. 4 and Supplementary Table 6. Sensitivity analyses requiring multiple SNPs were conducted only when the number of available instrumental variables met the methodological requirements. Specifically, Cochran’s Q test required at least two SNPs, and the MR-Egger intercept test required at least three SNPs; therefore, associations with insufficient SNPs did not generate results in Supplementary Tables 3 and 4. Among associations eligible for these analyses, heterogeneity and horizontal pleiotropy were evaluated using P < 0.05 as the significance threshold (Supplementary Tables 3 and 4). Leave-one-out analyses indicated that most eligible associations were directionally stable. However, the nominal APOE associations with FTD and DLB were sensitive to the exclusion of rs12972156; these APOE-specific findings are described in Sect. 3.5 and Supplementary Table 5.

Associations Between Plasma Proteins and VaD/FTD

At nominal P < 0.05, 102 and 78 exploratory protein–outcome signals were observed for VaD and FTD, respectively (Fig. 2C, D; Supplementary Table 2). For VaD, the OR per standard deviation increase in genetically predicted plasma protein level ranged from 0.101 (95% CI: 0.011–0.922) for placenta growth factor to 7.548 (95% CI: 2.361–24.127) for epigen, based on the primary MR estimates in Supplementary Table 2. For FTD, the OR ranged from 0.002 (95% CI: 4.58 × 10^− 6–0.746) for killer cell lectin-like receptor subfamily B member 1 to 3545.692 (95% CI: 35.355–3.56 × 10⁵) for inhibitor of growth protein 1, based on the primary MR estimates in Supplementary Table 2. Figure 3C, D visualize these exploratory nominal signals according to P value rather than OR magnitude. After outcome-specific FDR correction, APOE was the only protein associated with VaD; this finding was supported by seven instruments and estimated using IVW (OR = 0.670, 95% CI 0.594–0.757; FDR-adjusted P = 1.88 × 10⁻⁷; Fig. 4; Supplementary Table 6, Sheet C). Furthermore, APOE was inversely associated with FTD before FDR correction (OR: 0.720, 95% CI: 0.525–0.987, P = 0.041; Supplementary Table 2). The results of the sensitivity analyses are presented in Supplementary Tables 3, 4, 5.

Associations Between Plasma Proteins and DLB/Unspecified Dementia

At nominal P < 0.05, 101 and 112 exploratory protein–outcome signals were observed for DLB and unspecified dementia, respectively (Fig. 2E, F; Supplementary Table 2). For DLB, the OR per standard deviation increase in genetically predicted plasma protein level ranged from 0.054 (95% CI: 0.010–0.297) for PMS1 protein homolog 1 to 9.913 (95% CI: 2.254–43.604) for metastasis suppressor 1 (MTSS1), based on the primary MR estimates in Supplementary Table 2. For unspecified dementia, the OR ranged from 0.148 (95% CI: 0.047–0.468) for serine/threonine protein kinase D2 to 6.809 (95% CI: 2.555–18.145) for FK506-binding protein-like, based on the primary MR estimates in Supplementary Table 2. Figure 3E, F visualize these exploratory nominal signals according to P value rather than OR magnitude. Outcome-specific FDR-corrected results based on the primary MR estimates are provided in Supplementary Table 6. Before FDR correction, APOE was inversely associated with DLB (OR = 0.516, 95% CI: 0.370–0.719, P = 9.50 × 10^− 5), but this association did not remain significant after FDR correction (FDR-adjusted P = 0.173; Supplementary Table 6). After outcome-specific FDR correction, APOE remained inversely associated with unspecified dementia (OR = 0.551, 95% CI 0.493–0.617; FDR-adjusted P = 5.86 × 10⁻²²), whereas NECTIN2 and PVR were positively associated with unspecified dementia (NECTIN2: OR = 1.748, 95% CI 1.397–2.188; FDR-adjusted P = 6.75 × 10⁻⁴; PVR: OR = 1.429, 95% CI 1.262–1.618; FDR-adjusted P = 1.86 × 10⁻⁵; Fig. 4; Supplementary Table 6, Sheet F). All three FDR-significant associations with unspecified dementia—APOE, NECTIN2, and PVR—were supported by multiple instruments and estimated using IVW. No DLB association remained significant after FDR correction. The results of the sensitivity analyses are presented in Supplementary Tables 3, 4, 5.

Extended Chromosome 19 Locus Analyses and APOE Sensitivity Analyses

Neither rs429358 nor rs7412 was selected as an instrument, and none of the seven APOE instruments was a protein-coding missense variant. APOE, NECTIN2, and PVR did not share any identical instrumental variants, and pairwise LD among the 15 instruments was low (maximum r² = 0.075461; no pair reached r² ≥ 0.1). However, querying each instrument against all three original UKB-PPP protein measurements identified four genome-wide-significant cross-protein associations among the 30 cross-protein combinations: rs1836278, an APOE instrument, with PVR (β = −0.08695, SE = 0.00956, P = 9.07 × 10⁻²⁰); both NECTIN2 instruments with APOE [rs440277: β = 0.12952, SE = 0.00813, P = 4.23 × 10⁻⁵⁷; rs4452060: β = −0.09429, SE = 0.00764, P = 5.00 × 10⁻³⁵]; and the PVR instrument rs2272021 with APOE (β = 0.13502, SE = 0.01470, P = 4.16 × 10⁻²⁰). The remaining 26 cross-protein associations did not reach P < 5 × 10⁻⁸. Accordingly, the absence of identical instruments and low LD between different selected instruments do not establish protein specificity at this locus (Supplementary Table 7, Sheets A, B, and E).

After exclusion of rs1836278, the four APOE associations that had survived the original outcome-specific FDR correction remained directionally concordant and nominally significant in the sensitivity analyses: overall dementia (OR 0.523, 95% CI 0.460–0.594; P = 2.80 × 10⁻²³), Alzheimer’s dementia (OR 0.470, 95% CI 0.408–0.542; P = 2.14 × 10⁻²⁵), vascular dementia (OR 0.671, 95% CI 0.587–0.766; P = 3.76 × 10⁻⁹), and unspecified dementia (OR 0.551, 95% CI 0.487–0.622; P = 1.17 × 10⁻²¹). The additional APOE association with frontotemporal dementia, which had been nominally significant before FDR correction, also remained directionally concordant and nominally significant after exclusion of rs1836278 (OR 0.720, 95% CI 0.525–0.988; P = 0.0418). The dementia with Lewy bodies estimate was unchanged because rs1836278 was not included in that harmonized instrument set. After exclusion of rs2272021, the three PVR associations that had survived the original outcome-specific FDR correction remained directionally concordant and nominally significant: overall dementia (OR 1.308, 95% CI 1.233–1.386; P = 2.42 × 10⁻¹⁹), Alzheimer’s dementia (OR 1.414, 95% CI 1.317–1.519; P = 1.76 × 10⁻²¹), and unspecified dementia (OR 1.386, 95% CI 1.237–1.552; P = 1.72 × 10⁻⁸). Among the additional outcome-specific PVR estimates, only the association with vascular dementia had been nominally significant before FDR correction and remained directionally concordant and nominally significant after exclusion of rs2272021 (OR 1.244, 95% CI 1.082–1.431; P = 0.00221). The associations with frontotemporal dementia and dementia with Lewy bodies were not nominally significant either before or after exclusion; the corresponding post-exclusion estimates were OR 1.175 (95% CI 0.685–2.015; P = 0.559) and OR 0.890 (95% CI 0.763–1.039; P = 0.140), respectively. A protein-specific restricted NECTIN2 estimate could not be calculated because both selected NECTIN2 instruments were genome-wide significantly associated with APOE, and exclusion of both left no instrument (Supplementary Table 5, Sheet C). These exclusion estimates were nominal sensitivity analyses and were not subjected to a new outcome-wide or proteome-wide FDR correction.

Among the APOE instruments, rs12972156 showed moderate LD with rs429358 (r² = 0.450432), whereas rs62117206 showed weaker but non-negligible LD with rs7412 (r² = 0.106316). In targeted IVW sensitivity analyses, exclusion of rs12972156 preserved the protective direction and nominal P < 0.05 for overall dementia, Alzheimer’s dementia, vascular dementia, and unspecified dementia, whereas the exploratory nominal associations with frontotemporal dementia and dementia with Lewy bodies no longer met nominal P < 0.05. Exclusion of rs62117206 preserved the direction and nominal P < 0.05 for all six APOE estimates (Supplementary Table 5, Sheet B; Supplementary Table 7, Sheet C). The APOE associations with frontotemporal dementia and dementia with Lewy bodies had not survived the original outcome-specific FDR correction.

When rs12972156 and rs62117206 were excluded jointly, nominal P < 0.05 was retained for overall dementia (OR 0.656, 95% CI 0.593–0.726; P = 3.00 × 10⁻¹⁶), Alzheimer’s dementia (OR 0.610, 95% CI 0.525–0.709; P = 1.14 × 10⁻¹⁰), vascular dementia (OR 0.823, 95% CI 0.706–0.959; P = 0.0125), and unspecified dementia (OR 0.669, 95% CI 0.593–0.755; P = 7.39 × 10⁻¹¹), but not for frontotemporal dementia (OR 0.845, 95% CI 0.430–1.661; P = 0.625) or dementia with Lewy bodies (OR 0.895, 95% CI 0.640–1.251; P = 0.516). These joint-exclusion estimates were nominal sensitivity analyses and were not subjected to a new proteome-wide FDR correction.

Colocalization Analysis

Bayesian colocalization analysis was conducted to assess whether the FDR-significant protein–dementia associations showed evidence of a shared regional causal signal. The analysis provided strong colocalization evidence for SERPINF2 with both dementia and ADD (dementia: PP.H4 = 0.851; ADD: PP.H4 = 0.905; Fig. 5A, B), and for TREM2 with dementia (PP.H4 = 0.996; Fig. 5C). TREM2–dementia was supported by three instruments and therefore permitted the full set of prespecified multi-instrument sensitivity analyses. In contrast, both strongly colocalized SERPINF2 associations were based on a single instrument and were estimated using the Wald ratio; their strong colocalization support was therefore interpreted as convergent shared-variant evidence rather than definitive causal confirmation.

Fig. 5.

Representative regional association plots from the Bayesian colocalization analysis. The plots show selected protein-dementia loci included in the colocalization analysis. Colors indicate linkage disequilibrium, measured as r², with the lead single-nucleotide polymorphism (SNP). A–C show strong evidence of colocalization based on high posterior probability for hypothesis 4 (PP.H4), whereas D is shown as a representative APOE regional plot characterized by high posterior probability for hypothesis 3 (PP.H3) and low PP.H4. A: SERPINF2 and dementia, with the lead SNP being rs1057335. B: SERPINF2 and Alzheimer’s dementia, with the lead SNP being rs1057335. C: TREM2 and dementia, with the lead SNP being rs580064. D: APOE and dementia, with the lead SNP being rs8106922. Complete colocalization statistics across protein–outcome pairs are provided in Supplementary Table 8

For APOE, high PP.H3 but very low PP.H4 values were observed across several dementia-related outcomes. Under the conventional single-signal coloc model, this posterior pattern did not support a shared leading association signal between plasma APOE levels and the dementia-related traits. Both rs429358 and rs7412 were located within each ± 1-Mb APOE colocalization window. The lead variant rs8106922 showed with rs429358 and with rs7412, whereas rs75627662 showed with rs429358 and with rs7412. These findings indicate substantial local LD complexity and should not be interpreted as definitive evidence that the protein and disease traits are driven by fully distinct causal variants. The relevant LD estimates are provided in Supplementary Table 7, Sheet D (“Coloc lead LD”). The regional plot is shown in Fig. 5D, and complete colocalization statistics are provided in Supplementary Table 8.

Functional Network and Druggability Analysis

Functional network analysis was performed to explore the biological context of the MR-prioritized plasma proteins. The PPI network showed that several FDR-significant proteins were functionally connected rather than isolated signals, with enrichment in pathways related to lipid metabolism, immune regulation, microglial activation, cell adhesion, proteolysis, vascular regulation, and coagulation/fibrinolysis (Supplementary Fig. 1). Among these proteins, APOE, NECTIN2, and PVR were connected with lipid-related and cell-adhesion/immune regulatory processes, whereas TREM2 and GRN were related to microglial and lysosomal pathways. SERPINF2 was linked to coagulation and fibrinolytic regulation, suggesting a potential vascular or inflammatory component in dementia-related protein mechanisms.

Drug database searches further identified more than 40 candidate drugs or drug entries targeting 14 MR-prioritized proteins, including approved drugs and agents in different stages of clinical development (Supplementary Table 9). These results suggest that several genetically prioritized plasma proteins may have pharmacological relevance. However, these druggability findings should be interpreted as target-prioritization evidence rather than direct evidence of therapeutic efficacy in dementia.

Discussion

In this proteome-wide Mendelian randomization study, 23 protein–outcome associations involving 14 plasma proteins remained significant after outcome-specific false discovery rate correction. Fifteen associations were supported by multiple instruments and estimated using inverse-variance weighting, whereas eight were based on a single instrument and estimated using the Wald ratio. The evidence was not uniform across proteins. TREM2–dementia represented the strongest convergent finding because it was supported by three instruments, the applicable heterogeneity, directional-pleiotropy, and leave-one-out analyses, and strong colocalization. The protective SERPINF2 associations with overall dementia and Alzheimer’s dementia also showed strong colocalization, but both were based on single-instrument estimates. APOE showed the broadest cross-subtype pattern, with FDR-significant protective associations across overall dementia, Alzheimer’s dementia, vascular dementia, and unspecified dementia, although its interpretation was complicated by the extended APOE locus. PVR and NECTIN2 showed multi-instrument risk associations across several dementia-related outcomes, while angiotensin-converting enzyme (ACE), ephrin type-B receptor 4 (EPHB4), leukocyte immunoglobulin-like receptor subfamily A member 5 (LILRA5), and dipeptidase 1 (DPEP1) represented additional multi-instrument findings without strong colocalization. GRN, BLNK, CSF3, ATXN2L, and FKBPL were supported by single instruments and therefore require more cautious interpretation. The druggability analysis further nominated candidate protein targets, but these results represent target-prioritization evidence rather than evidence of clinical therapeutic efficacy.

Previous proteome-wide MR studies have principally focused on Alzheimer’s disease or broad neurodegenerative disease categories and have frequently summarized their findings at the pathway level. These studies have implicated immune regulation, microglial and lysosomal function, lipid metabolism, vascular biology, and blood–brain barrier integrity, but pathway overlap alone does not establish whether individual protein findings are replicated, extended to new dementia subtypes, or potentially novel. We therefore performed a structured protein-level comparison of all 14 FDR-significant proteins, considering instrument number, primary estimator, applicability of sensitivity analyses, colocalization evidence, and the closest published genetic or proteomic evidence (Supplementary Table 10). This comparison complements previous Alzheimer’s disease-focused and neurodegenerative-disease-focused proteome-wide MR studies and provides a more proportional basis for interpreting the relative robustness and novelty of the present results (Yao et al. 2024; Belbasis et al. 2025; Salih et al. 2025).

Among the FDR-significant findings, TREM2–dementia showed the strongest convergence of internal and external evidence. The protective association was supported by three independent instruments and strong colocalization (PP.H4 = 0.996), and the available multi-instrument sensitivity analyses did not indicate material heterogeneity, directional pleiotropy, or disproportionate dependence on a single variant. Previous plasma-protein genetic studies similarly reported that higher genetically predicted TREM2 levels were associated with lower Alzheimer’s disease risk, and a recent study across dementia subtypes also identified protective TREM2 associations (Hillary et al. 2022; Salih et al. 2025). TREM2 is a microglial receptor involved in cell survival, phagocytic responses, lipid sensing, and metabolic fitness, providing biological plausibility for its association with neurodegenerative outcomes (Ulland et al. 2017; Yang et al. 2020). Nevertheless, the present exposure represents genetically predicted plasma TREM2 measurement rather than a direct measurement of microglial or cerebrospinal-fluid TREM2. The finding should therefore be regarded as a strongly prioritized protein–dementia association rather than definitive evidence that increasing circulating TREM2 would prevent dementia.

SERPINF2 constituted the second group of findings with strong shared-variant support. Genetically predicted higher SERPINF2 was protectively associated with both overall dementia and Alzheimer’s dementia, with PP.H4 values of 0.851 and 0.905, respectively. These directions were consistent with recent plasma-protein MR studies that reported protective SERPINF2 associations with Alzheimer’s disease (Sun et al. 2025; Zhan et al. 2025). SERPINF2, also known as α2-antiplasmin, is a principal inhibitor of plasmin and a key regulator of fibrinolysis (Wu et al. 2019). Given the importance of blood–brain barrier dysfunction in neurodegeneration and dementia (Noe et al. 2020), the SERPINF2 association warrants further investigation in relation to neurovascular integrity. However, both SERPINF2 estimates were based on a single instrument. Strong colocalization provides convergent evidence that the protein and disease association signals share a regional genetic signal under the conventional model, but it does not permit conventional heterogeneity, MR-Egger, or leave-one-out analyses and does not by itself prove that circulating SERPINF2 abundance is the causal mediator.

APOE displayed the broadest cross-subtype pattern and was supported by seven instruments. Higher genetically predicted plasma APOE was protectively associated with overall dementia, Alzheimer’s dementia, vascular dementia, and unspecified dementia after FDR correction. This direction is consistent with a previous large Mendelian randomization study in which genetically lower plasma apoE was associated with higher risks of Alzheimer’s disease and all dementia (Rasmussen et al. 2018). APOE nevertheless requires a different interpretation from TREM2. The common apoE isoforms are defined by the missense variants rs429358 and rs7412, which alter amino acids 112 and 158 and have structural and functional consequences extending beyond protein abundance (Martens et al. 2022; Raulin et al. 2022). In the present study, rs12972156 showed moderate LD with rs429358. Exclusion of rs12972156 preserved the protective associations with overall dementia, Alzheimer’s dementia, vascular dementia, and unspecified dementia, whereas the nominal associations with frontotemporal dementia and dementia with Lewy bodies became non-significant; the latter two findings had already failed to survive the original FDR correction. Joint exclusion of rs12972156 and rs62117206 yielded the same qualitative pattern: the four associations that had survived the original FDR correction retained nominal P < 0.05, whereas the FTD and DLB associations did not. The four FDR-significant APOE associations are therefore comparatively more stable, but they cannot automatically be attributed exclusively to circulating APOE concentration.

The APOE colocalization results further support cautious interpretation. High PP.H3 and very low PP.H4 values did not provide evidence for a shared leading association signal under the conventional single-signal coloc model. This should not be interpreted as proof that the protein and dementia traits are driven by completely distinct causal variants, because the extended APOE locus has complex LD and may contain multiple association signals. Previous analyses have shown that associations assigned to neighboring genes, including NECTIN2, can be markedly attenuated after adjustment for the APOE ε2/ε3/ε4-defining variants (Curtis 2021). Consequently, the APOE results provide broad and reproducible genetic prioritization evidence, but the specific contribution of circulating apoE abundance, isoform-dependent structural effects, and other local regulatory pathways cannot be separated fully without valid conditional pQTL analyses and multiple-signal colocalization.

PVR and NECTIN2 also showed multi-instrument risk associations across more than one dementia-related outcome. A recent PubMed-indexed preprint integrating plasma and brain proteomics reported PVR as a candidate protein in both Alzheimer’s disease and vascular dementia, providing emerging protein-level support for the present findings; however, its preprint status and the absence of strong colocalization in our analyses preclude treating it as definitive external replication (Amin et al. 2025). Previous circulating-protein and proteome-wide MR studies implicated NECTIN2 in cognitive and Alzheimer’s disease–related phenotypes, including a higher-risk direction for Alzheimer’s disease (Tin et al. 2023). Proteogenomic work has also identified NECTIN2 as a modulator of soluble TREM2 biology (Wang et al. 2024). These data provide biological and genetic context for the current risk associations with overall dementia, Alzheimer’s dementia, and unspecified dementia. However, NECTIN2 lies within the extended APOE region, conventional colocalization did not support a shared leading signal, and regional disease associations may partly reflect APOE-related LD. The PVR and NECTIN2 findings should therefore be interpreted as multi-instrument prioritization signals with external support of varying maturity, rather than as independently established protein-mediated effects.

Variant-level cross-protein analyses showed that both selected NECTIN2 instruments were genome-wide significantly associated with APOE, whereas rs2272021, a PVR instrument, was associated with APOE and rs1836278, an APOE instrument, was associated with PVR. After exclusion of rs1836278 and rs2272021, respectively, the APOE and PVR associations that had survived the original outcome-specific FDR correction remained directionally concordant with nominal P < 0.05 in the sensitivity analyses. Nevertheless, these exclusions do not identify the specific protein mediator, and a protein-specific restricted NECTIN2 estimate was not identifiable because removing both cross-associated instruments left no valid instrument. The APOE-, NECTIN2-, and PVR-related findings should therefore be interpreted as locus-level genetic-prioritization signals that may reflect shared cis regulation or cross-protein pleiotropy rather than independently established protein-mediated effects.

Among the remaining multi-instrument findings, higher genetically predicted ACE was protectively associated with overall dementia. This direction is compatible with previous analyses in which genetically proxied ACE inhibition, corresponding to lower ACE activity or expression, was associated with higher risks of selected dementia outcomes (Nassan et al. 2023). The present result extends this evidence to an overall dementia phenotype, although the lack of strong colocalization limits mechanistic attribution. EPHB4 and LILRA5 had emerging cross-tissue, genetic, or proteomic support, but the available studies were not direct peer-reviewed replications of the present plasma-protein associations, as summarized in Supplementary Table 10. DPEP1 was supported by eight instruments, making it internally more informative than the single-instrument findings, but no directly comparable PubMed-indexed dementia plasma-protein MR study was identified in our targeted comparison. DPEP1 may therefore represent a potentially novel prioritization signal within the reviewed literature, but this designation does not establish first discovery, and replication and shared-signal validation remain necessary.

GRN and BLNK illustrate a different evidence pattern: both were supported by direct prior protein-level studies, but the present associations were based on single instruments and lacked strong colocalization. Previous proteome-wide MR and multi-omics studies reported higher GRN as protective and higher BLNK as associated with greater Alzheimer’s disease risk, consistent with the directions observed here (Belbasis et al. 2025; Zhan et al. 2025). Additional work integrating proteomic MR with neuropathological evidence has prioritized GRN-related microglial and lysosomal networks (Yuan et al. 2025). GRN is biologically relevant to lysosomal homeostasis, neuronal survival, and microglial regulation, and reduced progranulin function is implicated in several neurodegenerative disorders (Mendsaikhan et al. 2019; Rhinn et al. 2022). External consistency strengthens the prioritization of GRN and BLNK, but it does not overcome the inability to assess heterogeneity, directional pleiotropy, or single-variant influence in the present Wald ratio estimates. They should therefore be considered externally supported but internally less robust than the multi-instrument findings.

The remaining single-instrument findings should be viewed as hypothesis-generating. The protective CSF3–Alzheimer’s dementia estimate was not directionally consistent with all published genetic evidence; a recent MR study reported that higher genetically predicted granulocyte colony-stimulating factor was associated with higher Alzheimer’s disease risk across several datasets (Guo et al. 2026). This discrepancy may reflect differences in protein assay, pQTL architecture, instrument selection, disease definition, or residual pleiotropy and argues against strong mechanistic interpretation. No directly comparable PubMed-indexed dementia plasma-protein MR study was identified for ATXN2L or FKBPL in the targeted comparison. Although these associations survived outcome-specific FDR correction, their single-instrument basis, absence of strong colocalization, and limited external replication require independent validation before they are considered biomarkers or therapeutic targets.

The structured protein-level comparison clarifies that the 14 FDR-significant proteins do not represent a single evidential category (Supplementary Table 10). TREM2–dementia was the most strongly convergent result because multi-instrument MR, applicable sensitivity analyses, strong colocalization, and prior plasma-protein genetic evidence were aligned. SERPINF2 replicated previously reported protective Alzheimer’s disease associations and extended them to overall dementia with strong colocalization, but remained limited by single-instrument estimation. APOE, NECTIN2, PVR, and ACE had direct or emerging prior protein-level evidence, while their colocalization and locus-specific interpretability varied. GRN and BLNK were directionally consistent with previous reports but were supported by single instruments in the present analysis. DPEP1 was potentially novel within the targeted literature and supported by multiple instruments, whereas EPHB4 and LILRA5 had emerging but not directly comparable external evidence as summarized in Supplementary Table 10. CSF3 showed inconsistent published direction, and ATXN2L and FKBPL lacked directly comparable evidence. Accordingly, the replicated, multi-instrument, and strongly colocalized findings were prioritized, whereas single-instrument, non-colocalized, or externally unreplicated signals were interpreted more cautiously.

This study has several strengths. First, the use of a harmonized UKB-PPP exposure resource and a largely consistent FinnGen outcome framework reduced heterogeneity in protein measurement, variant processing, ancestry, and dementia ascertainment across subtype analyses. Second, primary MR findings were subjected to outcome-specific FDR correction, and multi-instrument IVW estimates were distinguished explicitly from single-instrument Wald ratio estimates. Third, the applicability of Cochran’s Q, MR-Egger intercept, and leave-one-out analyses was reported according to instrument number rather than treating missing sensitivity results as evidence of stability. Fourth, Bayesian colocalization was used to identify findings with additional shared-variant support, while the single-signal assumption and locus-specific limitations were acknowledged. Fifth, the additional chromosome 19 LD, cross-protein, and exclusion sensitivity analyses improved the transparency of the APOE, NECTIN2, and PVR interpretation. Finally, the protein-level literature comparison, functional networks, and druggability analyses placed the statistical findings in a broader biological and translational context without treating predicted druggability as clinical efficacy.

Several limitations should also be considered. First, one harmonized proteomic resource was deliberately used to maximize internal comparability. This strategy reduced cross-platform heterogeneity but limited formal replication in independent pQTL resources and may have missed proteins detectable on other platforms. The outcomes were also predominantly obtained from one European framework, and the findings may not generalize to other ancestries. Second, statistical power differed substantially across dementia subtypes, particularly for frontotemporal dementia and dementia with Lewy bodies. Therefore, similar numbers of nominal P < 0.05 signals across outcomes should not be interpreted as evidence of comparable statistical support across dementia subtypes. The absence of FDR-significant findings for these rarer outcomes should not be interpreted as evidence that relevant protein associations do not exist. Colocalization was restricted to FDR-significant associations, improving specificity but potentially missing shared signals among suggestive associations.

A further limitation concerns the biological interpretation of plasma protein measurements in a disorder whose principal pathology occurs in the brain. The MR estimates reflect genetically predicted circulating protein measurements and should not be assumed to represent the corresponding protein abundance or activity within brain tissue. Tissue-specific expression, blood–brain barrier transport, protein processing, isoform composition, and assay-dependent epitope recognition may separate a plasma pQTL association from the relevant central nervous system mechanism. Eight of the 23 FDR-significant associations were supported by only one instrument, precluding conventional assessments of heterogeneity, directional pleiotropy, and single-variant influence. Even for multi-instrument associations, the exclusion-restriction assumption cannot be verified completely using summary-level data. The results therefore represent genetic prioritization rather than definitive proof that circulating protein abundance directly causes or mediates dementia risk.

The extended APOE locus constitutes an additional specific limitation. Although APOE, NECTIN2, and PVR did not share identical instruments and pairwise LD among the selected instruments was low, four of the 30 cross-protein associations reached genome-wide significance, including APOE associations for both selected NECTIN2 instruments. Thus, non-overlapping, low-LD instrument sets did not ensure protein specificity, and a protein-specific restricted NECTIN2 estimate was not identifiable. Selected APOE instruments and colocalization lead variants also showed non-negligible LD with rs429358 and rs7412. The separate and joint LD-informed exclusion analyses reduced but could not eliminate the possibility that the APOE estimates partly reflect isoform-dependent structural effects or other local pathways. Accordingly, the locus-level MR estimates cannot be uniquely attributed to a single measured plasma protein. A formal pQTL analysis conditional on rs429358 and rs7412 was not performed because the individual-level genotype–protein data and valid conditional pQTL summary statistics required to re-estimate the SNP–protein associations were unavailable.

Moreover, conventional coloc assumes at most one causal association signal per trait within the analyzed region. This assumption may be overly restrictive at the extended APOE locus and at other genetically complex regions. Signal-specific conditional or Sum of Single Effects–based colocalization was not performed because the complete harmonized regional summary statistics and LD inputs required for a valid multiple-signal analysis were unavailable. Secondary shared signals cannot be excluded. More generally, MR and colocalization cannot completely eliminate residual horizontal pleiotropy, phenotype misclassification, tissue-specific discordance, or platform-specific measurement bias. The druggability results likewise identify genetically prioritized candidates rather than clinically validated therapeutic targets. Independent pQTL replication, tissue-specific proteogenomic analyses, multiple-signal colocalization, and experimental studies are required to determine which prioritized proteins participate directly in dementia biology and are suitable for therapeutic modulation.

Conclusion

This proteome-wide MR study genetically prioritized plasma proteins associated with multiple dementia-related outcomes. Fifteen of the 23 FDR-significant associations were supported by multiple instruments, whereas eight were based on a single instrument and therefore require more cautious interpretation. TREM2–dementia was supported by both multi-instrument MR and strong colocalization, while the strongly colocalized SERPINF2 associations were based on single-instrument estimates. The APOE-, NECTIN2-, and PVR-related findings should be interpreted as locus-level genetic prioritization rather than evidence that the circulating abundance of any one of these proteins independently mediates dementia risk. These findings nominate candidate biomarkers and therapeutic targets but do not establish that circulating protein abundance directly causes or mediates dementia risk. Future tissue-specific, functional, and experimental studies are required to clarify the underlying biological mechanisms and therapeutic relevance.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

Xinyang Yan and Longxiao Zhang: Conceptualization, Methodology, Writing-original draft, Data curation, Formal analysis, Project administration, Software, Validation and Visualization. Jiaxi Li and Jinning Song: Funding acquisition, Writing-review and editing.

Funding

This work was supported by the National Natural Science Foundation of China [Grant Number 82102185].

Data Availability

No datasets were generated or analysed during the current study.

Declarations

Competing Interests

The authors declare no competing interests.

Informed Consent and Ethical Approval

All summary-level GWAS datasets used in this study were obtained from publicly available databases or previously published studies, with detailed data sources listed in Table 1 and the corresponding cited publications. This study was based exclusively on publicly available, deidentified, summary-level genetic association data. Ethical approval and informed consent had been obtained in the original studies according to their respective protocols. No new human participants, animal experiments, biological samples, individual-level data, or identifiable personal information were involved; therefore, additional institutional ethical approval was not required. The study was conducted in accordance with the principles of the Declaration of Helsinki. All the authors have approved the final article for publication.