Artificial intelligence and omics-based autoantibody profiling highlights autoimmunity targeting ligand-receptor interaction in dementia

26 Dementia is a neurodegenerative syndrome marked by the accumulation of 27 disease-specific proteins and immune dysregulation, including autoimmune 28 mechanisms involving autoantibodies. Current diagnostic methods are often invasive, 29 time-consuming, or costly. This study explores the use of proteome-wide autoantibody 30 screening (PWAS) for noninvasive dementia diagnosis by analyzing serum samples 31 from Alzheimer's disease (AD), dementia with Lewy bodies (DLB), and age-matched 32 cognitively normal individuals (CNIs). Serum samples from 35 subjects were analyzed 33 utilizing our original wet protein arrays that covers approximately 90% of human 34 transcriptome, revealing elevated gross autoantibody levels in AD and DLB patients 35 compared to CNIs. A total of 229 autoantibodies were differentially elevated in AD 36 and/or DLB, effectively distinguishing between patient groups. Machine learning 37 models showed high accuracy in classifying AD, DLB, and CNIs. Gene ontology 38 analysis highlighted autoantibodies targeting neuroactive ligands/receptors in AD and 39 lipid metabolism proteins in DLB. Notably, autoantibodies targeting neuropeptide B 40 (NPB) and adhesion G protein-coupled receptor F5 (ADGRF5) showed significant 41 correlations with clinical traits including Mini Mental State Examination scores, 42 suggesting a role in dementia pathogenesis. The study demonstrates the potential of 43 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint PWAS and AI integration as a noninvasive diagnostic tool for dementia, uncovering 44 biomarkers that could enhance understanding of disease mechanisms. Limitations 45 include demographic differences, small sample size, and lack of external validation. 46 Future research should involve longitudinal observation in larger, diverse cohorts and 47 functional studies to clarify autoantibodies' roles in dementia pathogenesis and their 48 diagnostic and therapeutic potential. 49 50 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint

51 Dementia is a complex neurodegenerative syndrome affecting millions 52 worldwide. Early diagnosis is crucial for timely intervention, yet many current 53 diagnostic methods are either invasive, time-consuming, or expensive. For instance, 54 psychological assessments require significant time and concern to patients themselves, 55 cerebrospinal fluid examination is invasive, and amyloid positron emission tomography 56 (PET) is costly. Consequently, there is a pressing need for a simpler, noninvasive, and 57 cost-effective diagnostic method for dementia.1–3 58 Pathologically, dementia is marked by the aggregation of disease-specific 59 proteins in the brain. 4 While the pathogenic role of abnormal protein deposition in 60 dementia is well-established, the precise mechanisms behind the initiation and 61 progression of neurodegeneration remain unclear. Meanwhile, emerging evidence has 62 highlighted the role of immune dysregulation in dementia's pathogenesis. Genome-wide 63 association studies have identified common genetic variations in immune system 64 processes that are associated with neurodegenerative diseases such as Alzheimer’s 65 disease (AD), frontotemporal dementia (FTD), and Parkinson’s disease dementia.5–7 66 Autoimmune mechanisms are gaining recognition as a key factor in the 67 pathophysiology of dementia.8–10 Autoantibodies—self-reactive antibodies produced by 68 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint B cells—play a role in immune tolerance and homeostasis. 11 However, due to various 69 genetic and environmental factors, the ability to distinguish “self” from “non-self” 70 deteriorates, leading autoantibodies to trigger and sustain inflammatory processes that 71 cause tissue damage. 12,13 Autoantibodies have been detected in both blood and 72 cerebrospinal fluid of patients with various forms of dementia, including autoimmune 73 dementia and neurodegenerative dementias such as AD, FTD, vascular dementia (VD), 74 and dementia with Lewy bodies (DLB). 14–18 Autoimmune dementia is characterized by 75 progressive cognitive decline with an early onset, atypical clinical presentation, rapid 76 progression, the presence of neural antibodies, cerebrospinal fluid inflammation, brain 77 changes in MRI atypical for neurodegenerative diseases, and a good response to 78 immunotherapy.19 Various neural autoantibodies have been frequently identified in 79 individuals with progressive cognitive decline, targeting cell surface proteins such as 80 the N-methyl-D-aspartate receptor, gamma-aminobutyric acid B receptor, 81 alpha-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor, leucine-rich 82 glioma inactivated protein 1, dipeptidyl-peptidase protein-like 6, and transcobalamin 83 receptor.20–22 However, there is an overlap in the neural autoantibody profiles between 84 autoimmune dementia and neurodegenerative dementias like FTD and DLB, 85 necessitating further research to clarify disease specificity.20 86 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint AD, one of the most well-known forms of dementia, is characterized by the 87 accumulation of amyloid plaques and neurofibrillary tangles in the brain. 23 88 Autoantibodies targeting amyloid- β (A β ), tau, neurotransmitters, and microglia have 89 been reported in AD patients. 24,25 Specifically, autoantibodies against A β are decreased 90 in AD patients, 26,27 suggesting a protective role against A β toxicity, 28,29 in line with 91 clinical efficacy of l ecanemab, a humanized monoclo nal antibody targeting A β soluble 92 protofibrils.30 Additionally, increased levels of autoantibodies against glutamate, 31 93 oxidized low-density lipoproteins, 32 glial markers such as GFAP and S100B, 33 and 94 receptors for advanced glycosylation end products have been observed in AD patients' 95 serum or cerebrospinal fluid. 34 DLB is another progressive neurodegenerative disorder 96 characterized by the presence of Lewy bodies—abnormal aggregates of the protein 97 alpha-synuclein—in the brain. 35 Autoantibodies against alpha-synuclein, A β , myelin 98 oligodendrocyte glycoprotein, myelin basic protein, and S100B, have been identified in 99 some DLB patients. 36,37 Autoantibodies have been detected even in patients with mild 100 cognitive impairment (MCI), indicating a potential role in disease progression. 38–40 101 Despite the discovery of autoantibodies related to various forms of dementia pathology, 102 further research is needed to assess their potential as diagnostic or prognostic 103 biomarkers and their utility in developing effective immunotherapies for dementia.28 104 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint One promising approach is the use of protein microarrays for autoantibody 105 profiling, which could help identify novel autoantibodies for diagnosing and monitoring 106 MCI and dementia. 39,40 In this pilot study, we utilized a proteome-wide autoantibody 107 screening (PWAS) technique employing wet protein arrays (WPAs) that cover 108 approximately 90% of the human transcriptome. 41,42 This method has previously been 109 used to develop multiplex measurements for disease-related autoantibodies,43,44 identify 110 clinically relevant novel autoantibodies, 45–48 and investigate epitope spreading during 111 disease progression. 49 We have successfully applied this technique to a variety of 112 inflammatory disorders, including systemic sclerosis, 47 and identified autoantibodies to 113 membranous antigens like G protein-coupled receptors (GPCRs) using machine 114 learning approaches.48 In this study, we applied PWAS to serum samples from patients 115 with AD or DLB and age-matched cognitively normal individuals (CNIs) to elucidate 116 the autoantibody landscape in dementia. Our goal was to identify clusters of 117 autoantibodies that may contribute to the pathophysiology of dementia, by integration 118 of artificial intelligence (AI) and omics-based approach. This research aims to uncover 119 novel biomarkers and enhance our understanding of dementia's pathogenesis. 120 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint

121 Demographic and clinical characteristics 122 Serum samples from 35 subjects, including 18 patients with AD, 8 patients 123 with DLB, and 9 CNIs were served for PWAS utilizing WPAs. The baseline 124 demographics across the three groups were similar, except that the proportion of 125 females was highest in the AD group and lowest among CNIs (Extended Table 1). The 126 proportion of females in the AD, DLB, and CNI groups were 82.4%, 62.5%, and 33.3%. 127 The Hasegawa's Dementia Scale-Revised (HDSR) scores for the each group were 128 19.9±5.6, 22.1±5.6, and 27.9 ± 2.0, respectively, while the Mini Mental State 129 Examination (MMSE) scores were 20.2±3.9, 21.1±6.6, and 28.9±1.4. 130 131 Sum of autoantibody levels 132 We defined the sum of autoantibody levels (SAL) as the total serum 133 concentration of all autoantibodies measured in our PWAS. Although not statistically 134 significant, SAL was higher in patients with AD and DLB compared to CNIs ( Figure 135 1A). This trend persisted across all age groups (Extended Figure 1A ) and was 136 relatively higher in females than in males (Extended Figure 1B). 137 138 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Identification of differentially elevated autoantibodies 139 Next, we focused on identifying autoantibodies with serum levels significantly 140 elevated in AD (Figure 1B ) and/or DLB (Figure 1C) compared to CNIs. This analysis 141 revealed 188 autoantibodies elevated in AD and 77 in DLB, with 36 overlapping 142 between the two conditions (Figure 1D), totaling 229 distinct items (Figure 1E). Using 143 these autoantibodies, we performed principal component analysis (PCA), which 144 effectively differentiated AD patients, DLB patients, and CNIs ( Figure 1F), regardless 145 of sex, age, or cognitive impairment severity as measured by HDSR and MMSE 146 (Figure 1G). 147 148 AI-based 2-class classification 149 To identify which of the 229 autoantibodies were most strongly associated with 150 disease status, we employed 14 different machine learning frameworks. Logistic 151 regression with normalization or standardization, along with support vector machines 152 (SVM) under similar conditions, achieved an area under the receiver operating 153 characteristic curve (ROC-AUC) exceeding 0.96, indicating near-perfect accuracy in 154 distinguishing AD patients from others ( Table 1 ). We identified the top 10 features 155 from these four models (Figure 2A), assessed their overlap ( Figure 2B), and analyzed 156 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint the serum levels of 12 autoantibodies highlighted in more than two frameworks (Figure 157 2C). 158 Next, we examined the relationship between these 12 autoantibodies and 159 clinical traits (Figure 2D ). This analysis revealed a significant correlation of serum 160 levels of autoantibodies targeting proteins encoded by TGFB1I1 and KIAA2013 with 161 HDSR scores. However, a database search, utilizing the Human Protein Atlas, 50 162 indicated that these two genes are not specifically expressed in the central nervous 163 system (data not shown). Although serum levels of anti-TGFB1I1 antibodies were 164 significantly associated with sex, trends in the distribution of these autoantibodies 165 among three groups were generally similar between both sex ( Extended Figure 2A ). 166 To evaluate cross-reactivity, we performed a correlation analysis on these 12 167 autoantibodies. Those with moderate to high correlations (Spearman’s r > 0.5) 168 underwent sequence alignment and identity analysis. The correlation matrix ( Extended 169 Figure 2B) revealed five correlated autoantibodies, and sequence analysis showed that 170 all proteins shared less than 25% identity ( Extended Figure 2C ). Additionally, we 171 investigated the prevalence of these highlighted autoantibodies across a broader 172 spectrum of human disorders using the aUToAntiBody Comprehensive Database 173 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint (UT-ABCD).47 Most of these autoantibodies were found to be non-specifically elevated 174 in various pathological conditions (Extended Figure 2D). 175 176 AI-based 3-class classification 177 We also explored multi-class classification among AD, DLB, and CNI by 178 training deep neural networks with five hidden layers using the 249-dimensional 179 autoantibody profiles. The optimal number of epochs was determined based on the 180 accuracy and loss trajectories (Figure 3A ). This approach resulted in high accuracy, 181 with ROC-AUC values reaching up to 0.95 ( Figure 3B), as well as high precision and 182 recall (Figure 3C). 183 184 Gene Ontology analysis 185 We aimed to identify autoantibodies with potential pathogenic roles in 186 dementia by conducting gene ontology analysis on the gene lists encoding the 229 187 autoantigens targeted by differentially elevated autoantibodies in AD and/or DLB 188 (Figure 4 ). The analysis highlighted the “neuroactive ligand-receptor interaction” 189 pathway in autoantibodies elevated specifically in AD. We also focused on “regulation 190 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint of lipid metabolic process ” highlighted only in DLB, considering recent advances in 191 understanding the role of lipid metabolism in the pathogenesis of DLB.51–53 192 193 Autoantibodies to neuroactive ligand-receptor interaction-associated proteins 194 There were 12 autoantibodies associated with neuroactive ligand-receptor 195 interaction, and their serum levels are illustrated in Figure 5A . We examined the 196 relationship between these 12 autoantibodies and clinical traits (Figure 5B), revealing a 197 significant association of the serum levels of autoantibodies targeting neuropeptide B, a 198 protein encoded by NPB, with female sex, presence of back pain, and MMSE scores. 199 However, the trend of elevated serum levels of anti-NPB antibody in dementia was 200 observed in both sex ( Extended Figure 3A ). There was no obvious cross-reactivity 201 among the autoantibodies (Extended Figure 3B and 3C ) and showed no disease 202 specificity ( Extended Figure 3D ). To further investigate the potential of anti-NPB 203 antibody to play a role in the pathogenesis of AD, we examined the correlation between 204 serum levels of the autoantibody and all the subscales of MMSE ( Extended Figure 4). 205 As a result, there was statistically significant correlation in memory-related items 206 (“Registration” and “Recall”). In line with this, a database search indicated that NPB is 207 expressed in the CNS ( Extended Figure 5A ), including the hippocampus ( Extended 208 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Figure 5B ). The highest expression was reported in oligodendrocytes ( Extended 209 Figure 5C). 210 211 Autoantibodies to lipid metabolism-associated proteins 212 Finally, we focused on 12 autoantibodies targeting lipid metabolism-associated 213 proteins, whose serum levels are illustrated in Figure 5C. We examined the relationship 214 between these 12 autoantibodies and clinical traits (Figure 5D). This analysis revealed a 215 significant association of the serum levels of autoantibodies targeting Adhesion G 216 Protein-Coupled Receptor F5 (ADGRF5) encoded by ADGRF5 with presence of back 217 pain, lower Comprehensive Geriatric Assessment 7 (CGA7) scores, and lower MMSE 218 scores, especially in “Registration” and “Repetition” subscales (Extended Figure 6 ). 219 There was no big difference between both sex (Extended Figure 7A), cross-reactivity, 220 nor disease specificity. ( Extended Figure 7B, 7C, and 7D ). A database search 221 indicated that the expression of ADGRF5 is ubiquitous across various human tissues 222 (Extended Figure 8A ), including the CNS ( Extended Figure 8B), predominantly in 223 microglial cells (Extended Figure 8C). 224 225 226 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint

227 In this study, we utilized our proprietary PWAS technique to analyze serum 228 samples from patients with AD, DLB, and CNIs. Our results showed an increase in the 229 overall levels of autoantibodies in AD and DLB patients compared to CNIs (Figure 1A). 230 We identified 229 autoantibodies that were differentially elevated in AD and/or DLB 231 (Figure 1D ), effectively distinguishing between AD, DLB, and CNI groups ( Figure 232 1F). Machine learning applied to these 229 autoantibodies demonstrated high accuracy 233 in differentiating AD patients from others ( Table 1 ), and even achieved success in 234 multi-class classification (Figure 3). Gene ontology analysis highlighted autoantibodies 235 targeting neuroactive ligands and receptors in AD, including anti-NPB antibody, as well 236 as lipid metabolism-associated proteins in DLB, such as anti-ADGRF5 antibody 237 (Figure 4 ). Both of anti-NPB and anti-ADGRF5 autoantibodies showed significant 238 correlation with total MMSE scores ( Figure 5B and 5D ) and memory-related subscale 239 scores (Extended Figure 4 and 6 ). Considering the expression of NPB and ADGRF5 240 in the central nervous system ( Extended Figure 5 and 8 ), these findings suggest that 241 autoantibodies targeting NPB or ADGRF5 may contribute to the pathogenesis of 242 dementia. Our results underscore the potential of our systems-based approach in 243 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint developing novel diagnostic tools and propose a new research strategy to explore the 244 autoimmune aspects of dementia. 245 A key highlight of our analysis is the ability of AI integrated with our 246 multiplex autoantibody measurement to achieve near-perfect accuracy in classifying AD 247 versus other groups ( Table 1) and even in multi-class classification tasks (AD, DLB, 248 and CNI; Figure 3). This concept has already been demonstrated in other autoimmune 249 and malignant disorders, 47,48 and is partially available commercially as the 250 Autoantibody Array Assay (A-Cube). 44 Given that blood tests are less invasive than 251 other procedures like cerebrospinal fluid collection and radiological imaging studies and 252 can be conducted without causing undue concern to the patient about suspected 253 cognitive impairment, multiplex measurement of serum autoantibodies using WPAs and 254 AI-based interpretation represents a promising strategy for diagnosing dementia and its 255 subtypes. 256 The NPB gene encodes neuropeptide B, a short biologically active peptide that 257 acts as an agonist for GPCRs known as neuropeptide B/W receptors 1 (NPBWR1) and 2 258 (NPBWR2).54 Neuropeptide B is believed to play roles in regulating feeding, the 259 neuroendocrine system, memory, learning, and the pain pathway. 55 Research by 260 Nagata-Kuroiwa R et al. on NPBWR1 knockout mice revealed increased autonomic and 261 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint neuroendocrine responses to physical stress and abnormalities in contextual fear 262 conditioning, suggesting a role for NPBWR1 in stress vulnerability and fear memory. 56 263 Histological and electrophysiological studies indicate that NPBWR1 acts as an 264 inhibitory regulator on a subpopulation of GABAergic neurons in the lateral division of 265 the central nucleus of the amygdala, terminating stress responses. Additionally, 266 Watanabe N et al. demonstrated that a single nucleotide polymorphism in NPBWR1, 267 associated with impaired molecular function, affected valence evaluation and 268 dominance ratings in response to seeing angry faces in humans, suggesting NPBWR1's 269 involvement in social interaction. 57 These insights highlight the potential role of 270 autoantibodies affecting the NPB-NPBWR1 signaling system in social behavior, 271 suggesting its potential contribution to the clinical manifestations of AD, particularly its 272 behavioral and psychological symptoms. 273 Our study also revealed a strong association between serum anti-NPB antibody 274 levels and the presence of back pain, likely due to the role of NPB-NPBWR1 signaling 275 in pain transmission. NPB knockout mice exhibit different responses to pain; they show 276 hyperalgesia to acute inflammatory pain but not to thermal or chemical pain. 58 277 Intrathecal administration of NPB reduced mechanical allodynia via activation of 278 NPBWR1 receptors without affecting thermal hyperalgesia. 59 These effects were not 279 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint inhibited by naloxone, an opioid receptor antagonist, indicating the involvement of a 280 non-opioid analgesic pathway, possibly related to myelin-forming Schwann cells, which 281 express low levels of NPBWR1 under physiological conditions but much higher levels 282 in patients with inflammatory neuropathies. Thus, anti-NPB antibodies may play a role 283 in modulating nociceptive transmission. 284 ADGRF5, a member of the adhesion GPCR (aGPCR) family, which is the 285 second largest GPCR subfamily, has recently garnered attention for its biological 286 functions, disease relevance, and potential as a drug target. 60 Predominantly expressed 287 in the lung and kidney, ADGRF5 may play a crucial role in regulating surfactant protein 288 synthesis acid-base balance in these organs. 61–63 DiBlasi et al. identified a single 289 nucleotide polymorphism in the ADGRF5 gene linked to an increased risk of suicide, 64 290 suggesting its psychiatric role. Additionally, Kaur et al. found that plasma levels of 291 ADGRF5 are associated with the APOE genotype, 65 a known risk factor for DLB and 292 AD.52,53 Elevated levels of anti-ADGRF5 antibodies correlated with global geriatric 293 function scores assessed by CGA7 ( Figure 5D), and the fact that ADGRF5 expression 294 is not exclusive to the CNS ( Extended Figure 8), may reflect systemic aspects of DLB 295 affecting multiple organs.66 296 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint It is important to note that not all patients had anti-NPB nor anti-ADGRF5 297 antibodies, and their serum levels in AD were not specific to the condition ( Extended 298 Figure 3D and 7D). This suggests that while the presence of these autoantibodies may 299 not explain the entire pathogenesis of dementia, they could influence disease 300 manifestation and progression as bystanders. Further investigation is needed to clarify 301 the role of anti-NPB and anti-ADGRF5 antibodies in the pathophysiology, including 302 functional assays to assess the effects of these antibodies on neurons or glial cells, 303 passive immune challenge in AD animal models by administering anti-NPB or 304 anti-ADGRF5 antibodies, and active immunization of animals with NPB or ADGRF5 305 antigens. 306 Our study has several strengths. First, by including multiple types of dementia 307 (AD and DLB), as well as CNIs, we were able to identify autoantibodies that are 308 differentially elevated in each condition and develop machine learning methodologies 309 for distinguishing different types of dementia in a relatively non-invasive way. Second, 310 the use of a wheat-germ in vitro protein synthesis system and the manipulation 311 technique for WPAs allowed for high-throughput expression of a wide range of human 312 proteins, including soluble proteins, on a single platform. 41,42,67 This enabled our 313 autoantibody measurement to cover an almost proteome-wide range of antigens, 314 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint allowing the application of omics-based bioinformatics approaches to interpret the data. 315 Third, integration of AI and omics-based approach allowed us to conduct an unbiased 316 and holistic investigation, resulting in novel discoveries. 317 A major limitation of our study is the demographic differences among the 318 human subjects, particularly in terms of sex (Extended Table 1). Moreover, the sample 319 size was modest, lacked external validation, and was cross-sectional. Future studies 320 should target larger, more demographically balanced patient groups with a wider range 321 of dementia types, such as VD and FTD. Recruiting longitudinal specimens and data 322 from elderly individuals before and after the onset of MCI in prospective 323 population-based cohorts would be a valuable challenge to explore the causal 324 relationship between autoantibodies and dementia pathogenesis. 325 326 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint

327 1. Umeda-Kameyama, Y. et al. Development of a novel convenient Alzheimer’s 328 disease assessment scale, the ABC Dementia Scale, using item response theory. 329 Geriatr. Gerontol. Int. 19, 18–23 (2019). 330 2. Wada-Isoe, K. et al. ABC Dementia Scale Classifies Alzheimer’s Disease 331 Patients into Subgroups Characterized by Activities of Daily Living, Behavioral 332 and Psychological Symptoms of Dementia, and Cognitive Function. J. 333 Alzheimer’s Dis. 73, 383–392 (2020). 334 3. Umeda-Kameyama, Y. et al. Screening of Alzheimer’s disease by facial 335 complexion using artificial intelligence. Aging (Albany. NY). 13, 1765–1772 336 (2021). 337 4. Wilson, D. M. et al. Hallmarks of neurodegenerative diseases. Cell 186, 693–714 338 (2023). 339 5. Yokoyama, J. S. et al. Association Between Genetic Traits for Immune-Mediated 340 Diseases and Alzheimer Disease. JAMA Neurol. 73, 691–697 (2016). 341 6. Witoelar, A. et al. Genome-wide Pleiotropy Between Parkinson Disease and 342 Autoimmune Diseases. JAMA Neurol. 74, 780–792 (2017). 343 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 7. Ferrari, R. et al. Frontotemporal dementia and its subtypes: a genome-wide 344 association study. Lancet. Neurol. 13, 686–699 (2014). 345 8. Archelos, J. J. & Hartung, H. P. Pathogenetic role of autoantibodies in 346 neurological diseases. Trends Neurosci. 23, 317–327 (2000). 347 9. D’Andrea, M. R. Add Alzheimer’s disease to the list of autoimmune diseases. 348 Med. Hypotheses 64, 458–463 (2005). 349 10. Sardi, F. et al. Alzheimer’s disease, autoimmunity and inflammation. The good, 350 the bad and the ugly. Autoimmun. Rev. 11, 149–153 (2011). 351 11. Elkon, K. & Casali, P. Nature and functions of autoantibodies. Nat. Clin. Pract. 352 Rheumatol. 4, 491–498 (2008). 353 12. Chatanaka, M. K., Sohaei, D., Diamandis, E. P. & Prassas, I. Beyond the amyloid 354 hypothesis: how current research implicates autoimmunity in Alzheimer’s 355 disease pathogenesis. Crit. Rev. Clin. Lab. Sci. 60, 398–426 (2023). 356 13. Wu, J. & Li, L. Autoantibodies in Alzheimer’s disease: potential biomarkers, 357 pathogenic roles, and therapeutic implications. J. Biomed. Res. 30, 361–372 358 (2016). 359 14. Gibson, L. L. et al. Neuronal surface autoantibodies in dementia: a systematic 360 review and meta-analysis. J. Neurol. 268, 2769–2779 (2021). 361 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 15. Terryberry, J. W., Thor, G. & Peter, J. B. Autoantibodies in Neurodegenerative 362 Diseases: Antigen-Specific Frequencies and Intrathecal Analysis. Neurobiol. 363 Aging 19, 205–216 (1998). 364 16. Arshad, F. et al. Role of Autoantibodies in Neurodegenerative Dementia: An 365 Emerging Association. Dement. Geriatr. Cogn. Disord. 50, 153–160 (2021). 366 17. Nagele, E., Han, M., Demarshall, C., Belinka, B. & Nagele, R. Diagnosis of 367 Alzheimer’s disease based on disease-specific autoantibody profiles in human 368 sera. PLoS One 6, e23112 (2011). 369 18. Reddy, M. M. et al. Identification of candidate IgG biomarkers for Alzheimer’s 370 disease via combinatorial library screening. Cell 144, 132–142 (2011). 371 19. Flanagan, E. P. et al. Autoimmune dementia: clinical course and predictors of 372 immunotherapy response. Mayo Clin. Proc. 85, 881–897 (2010). 373 20. Hansen, N. Current Nosology of Neural Autoantibody-Associated Dementia. 374 Front. Aging Neurosci. 13, 711195 (2021). 375 21. Banks, S. A., Sechi, E. & Flanagan, E. P. Autoimmune encephalopathies 376 presenting as dementia of subacute onset and rapid progression. Ther. Adv. 377 Neurol. Disord. 14, 1756286421998906 (2021). 378 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 22. Pluvinage, J. V. et al. Transcobalamin receptor antibodies in autoimmune 379 vitamin B12 central deficiency. Sci. Transl. Med. 16, (2024). 380 23. Scheltens, P. et al. Alzheimer’s disease. Lancet (London, England) 388, 505–517 381 (2016). 382 24. Henstridge, C. M., Hyman, B. T. & Spires-Jones, T. L. Beyond the 383 neuron-cellular interactions early in Alzheimer disease pathogenesis. Nat. Rev. 384 Neurosci. 20, 94–108 (2019). 385 25. Colasanti, T., Barbati, C., Rosano, G., Malorni, W. & Ortona, E. Autoantibodies 386 in patients with Alzheimer’s disease: pathogenetic role and potential use as 387 biomarkers of disease progression. Autoimmun. Rev. 9, 807–811 (2010). 388 26. Du, Y. et al. Reduced levels of amyloid beta-peptide antibody in Alzheimer 389 disease. Neurology 57, 801–805 (2001). 390 27. Weksler, M. E. et al. Patients with Alzheimer disease have lower levels of serum 391 anti-amyloid peptide antibodies than healthy elderly individuals. Exp. Gerontol. 392 37, 943–948 (2002). 393 28. Britschgi, M. et al. Neuroprotective natural antibodies to assemblies of 394 amyloidogenic peptides decrease with normal aging and advancing 395 Alzheimer’s disease. Proc. Natl. Acad. Sci. U. S. A. 106, 12145–12150 (2009). 396 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 29. Paul, S., Planque, S. & Nishiyama, Y. Immunological origin and functional 397 properties of catalytic autoantibodies to amyloid beta peptide. J. Clin. Immunol. 398 30 Suppl 1, S43-9 (2010). 399 30. van Dyck, C. H. et al. Lecanemab in Early Alzheimer’s Disease. N. Engl. J. Med. 400 388, 9–21 (2023). 401 31. Davydova, T. V, Voskresenskaya, N. I., Fomina, V. G., Vetrile, L. A. & 402 Doronina, O. A. Induction of autoantibodies to glutamate in patients with 403 Alzheimer’s disease. Bull. Exp. Biol. Med. 143, 182–183 (2007). 404 32. Kankaanpää, J., Turunen, S. P., Moilanen, V., Hörkkö, S. & Remes, A. M. 405 Cerebrospinal fluid antibodies to oxidized LDL are increased in Alzheimer’s 406 disease. Neurobiol. Dis. 33, 467–472 (2009). 407 33. Mecocci, P. et al. Serum anti-GFAP and anti-S100 autoantibodies in brain aging, 408 Alzheimer’s disease and vascular dementia. J. Neuroimmunol. 57, 165–170 409 (1995). 410 34. Mruthinti, S. et al. Autoimmunity in Alzheimer’s disease: increased levels of 411 circulating IgGs binding Abeta and RAGE peptides. Neurobiol. Aging 25, 412 1023–1032 (2004). 413 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 35. McKeith, I. G. et al. Diagnosis and management of dementia with Lewy bodies: 414 Fourth consensus report of the DLB Consortium. Neurology 89, 88–100 415 (2017). 416 36. Maetzler, W. et al. Autoantibodies against amyloid and glial-derived antigens are 417 increased in serum and cerebrospinal fluid of Lewy body-associated dementias. 418 J. Alzheimers. Dis. 26, 171–179 (2011). 419 37. Marchionatti, A., Woodhall, M., Waters, P. J. & Sato, D. K. Detection of 420 MOG-IgG by cell-based assay: moving from discovery to clinical practice. 421 Neurol. Sci. Off. J. Ital. Neurol. Soc. Ital. Soc. Clin. Neurophysiol. 42, 73–80 422 (2021). 423 38. Di Domenico, F. et al. Autoantibodies Profile in Matching CSF and Serum from 424 AD and aMCI patients: Potential Pathogenic Role and Link to Oxidative 425 Damage. Curr. Alzheimer Res. 13, 112–122 (2016). 426 39. Ehtewish, H. et al. Profiling the autoantibody repertoire reveals autoantibodies 427 associated with mild cognitive impairment and dementia. Front. Neurol. 14, 428 1256745 (2023). 429 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 40. Shim, S.-M., Koh, Y. H., Kim, J.-H. & Jeon, J.-P. A combination of multiple 430 autoantibodies is associated with the risk of Alzheimer’s disease and cognitive 431 impairment. Sci. Rep. 12, 1312 (2022). 432 41. Goshima, N. et al. Human protein factory for converting the transcriptome into 433 an in vitro-expressed proteome. Nat. Methods 5, 1011–1017 (2008). 434 42. Fukuda, E. et al. Identification and characterization of the antigen recognized by 435 the germ cell mAb TRA98 using a human comprehensive wet protein array. 436 Genes to Cells 26, 180–189 (2021). 437 43. Matsuda, K. M. et al. Autoantibody Landscape Revealed by Wet Protein 438 Array/i2 : Sum of Autoantibody Levels Re fl ects Disease Status. Front. Immunol. 439 13, 1–14 (2022). 440 44. Kuzumi, A. et al. Comprehensive autoantibody profiling in systemic 441 autoimmunity by a highly-sensitive multiplex protein array. Front. Immunol. 14, 442 (2023). 443 45. Matsuda, K. M., Kotani, H., Yamaguchi, K., Okumura, T. & Fukuda, E. 444 Significance of anti-transcobalamin receptor antibodies in cutaneous arteritis 445 revealed by proteome-wide autoantibody screening. J. Autoimmun. 135, 102995 446 (2023). 447 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 46. Matsuda, K. M. et al. Autoantibodies to nuclear valosin-containing protein-like 448 protein: systemic sclerosis-specific antibodies revealed by in vitro human 449 proteome. Rheumatology (Oxford). (2024) doi:10.1093/rheumatology/keae063. 450 47. Matsuda, K. M. et al. Proteome-wide autoantibody screening and holistic 451 autoantigenomic analysis unveil COVID-19 signature of autoantibody landscape. 452 medRxiv 2024.06.07.24308592 (2024). 453 48. Matsuda, K. M. et al. Deciphering autoantibody landscape of systemic sclerosis 454 through systems-based approach: insights from a B-cell depletion clinical trial. 455 medRxiv (2024) doi:10.1101/2024.07.30.24311212. 456 49. Kotani, H. et al. Diversity and epitope spreading of anti-RNA polymerase /i2 457 antibodies in systemic sclerosis: a potential biomarker for skin and lung 458 involvement. Arthritis Rheumatol. in press, (2024). 459 50. Thul, P. J. et al. A subcellular map of the human proteome. Science (80-. ). 356, 460 (2017). 461 51. García-Sanz, P., M.F.G. Aerts, J. & Moratalla, R. The Role of Cholesterol in 462 α -Synuclein and Lewy Body Pathology in Parkinson’s Disease. Mov. Disord. 36, 463 1070–1085 (2021). 464 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 52. Guerreiro, R. et al. Investigating the genetic architecture of dementia with Lewy 465 bodies: a two-stage genome-wide association study. Lancet. Neurol. 17, 64–74 466 (2018). 467 53. Tsuang, D. et al. APOE ε 4 increases risk for dementia in pure synucleinopathies. 468 JAMA Neurol. 70, 223–228 (2013). 469 54. Singh, G. & Davenport, A. P. Neuropeptide B and W: neurotransmitters in an 470 emerging G-protein-coupled receptor system. Br. J. Pharmacol. 148, 471 1033–1041 (2006). 472 55. Wojciechowicz, T. et al. The Role of Neuropeptide B and Its Receptors in 473 Controlling Appetite, Metabolism, and Energy Homeostasis. Int. J. Mol. Sci. 22, 474 (2021). 475 56. Nagata-Kuroiwa, R. et al. Critical Role of Neuropeptides B/W Receptor 1 476 Signaling in Social Behavior and Fear Memory. PLoS One 6, e16972 (2011). 477 57. Watanabe, N. et al. A single nucleotide polymorphism of the neuropeptide B/W 478 receptor-1 gene influences the evaluation of facial expressions. PLoS One 7, 479 e35390 (2012). 480 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint 58. Kelly, M. A. et al. Neuropeptide B-deficient mice demonstrate hyperalgesia in 481 response to inflammatory pain. Proc. Natl. Acad. Sci. U. S. A. 102, 9942–9947 482 (2005). 483 59. Yamamoto, T., Saito, O., Shono, K. & Tanabe, S. Anti-hyperalgesic effects of 484 intrathecally administered neuropeptide W-23, and neuropeptide B, in tests of 485 inflammatory pain in rats. Brain Res. 1045, 97–106 (2005). 486 60. Einspahr, J. M. & Tilley, D. G. Pathophysiological impact of the adhesion G 487 protein-coupled receptor family. Am. J. Physiol. Physiol. 323, C640–C647 488 (2022). 489 61. Bridges, J. P. et al. Orphan G protein-coupled receptor GPR116 regulates 490 pulmonary surfactant pool size. Am. J. Respir. Cell Mol. Biol. 49, 348–357 491 (2013). 492 62. Yang, M. Y. et al. Essential regulation of lung surfactant homeostasis by the 493 orphan G protein-coupled receptor GPR116. Cell Rep. 3, 1457–1464 (2013). 494 63. Abe, J., Suzuki, H., Notoya, M., Yamamoto, T. & Hirose, S. Ig-Hepta, a Novel 495 Member of the G Protein-coupled Hepta-helical Receptor (GPCR) Family That 496 Has Immunoglobulin-like Repeats in a Long N-terminal Extracellular Domain 497 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint and Defines a New Subfamily of GPCRs *. J. Biol. Chem. 274, 19957–19964 498 (1999). 499 64. DiBlasi, E. et al. Rare protein-coding variants implicate genes involved in risk of 500 suicide death. Am. J. Med. Genet. Part B Neuropsychiatr. Genet. 186, 508–520 501 (2021). 502 65. Kaur, G., Poljak, A., Masters, C. L., Fowler, C. & Sachdev, P. Impact of APOE 503 ε 3 and ε 4 genotypes on plasma proteome signatures in Alzheimer’s disease. 504 bioRxiv 2022.01.29.478291 (2022) doi:10.1101/2022.01.29.478291. 505 66. Beach, T. G. et al. Multi-organ distribution of phosphorylated alpha-synuclein 506 histopathology in subjects with Lewy body disorders. Acta Neuropathol. 119, 507 689–702 (2010). 508 67. Sawasaki, T., Ogasawara, T., Morishita, R. & Endo, Y. A cell-free protein 509 synthesis system for high-throughput proteomics. Proc. Natl. Acad. Sci. U. S. A. 510 99, 14652–14657 (2002). 511 68. Zhou, Y. et al. Metascape provides a biologist-oriented resource for the analysis 512 of systems-level datasets. Nat. Commun. 10, 1523 (2019). 513 514 515 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint

516 We thank Ms. Maiko Enomoto and her colleagues for their secretarial work. 517 We appreciate K. Yamaguchi, T. Okumura, C. Ono, A. Sato, A. Miya, and N. Goshima 518 from ProteoBridge Corporation for preparing the WPAs. We also acknowledge R. 519 Uchino, Y. Murakami, and H. Matsunaka from TOKIWA Pharmaceuticals Co. Ltd. for 520 providing technical assistance with autoantibody measurement. 521 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Author Contributions 522 KM Matsuda primarily engaged in autoantibody measurement, data analysis, 523 visualization, and writing the first draft of the manuscript. Y Umeda-Kameyama 524 oversaw clinical sample and data collection and was involved in revising the manuscript. 525 K Iwadoh participated in machine learning analysis. M Miyawaki, M Yakabe, S Ogawa, 526 and M Akishita participated in clinical sample and information collection. S Sato 527 conceptualized and supervised the study. A Yoshizaki conceptualized, launched, and 528 supervised this study, and was involved in revising the manuscript. 529 530 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Conflict-of-interest statement 531 A Yoshizaki belongs to the Social Cooperation Program, Department of 532 Clinical Cannabinoid Research, The University of Tokyo Graduate School of Medicine, 533 Tokyo, Japan, supported by Japan Cosmetic Association and Japan Federation of 534 Medium and Small Enterprise Organizations. The remaining authors declare that the 535 research was conducted in the absence of any commercial or financial relationships that 536 could be construed as a potential conflict of interest. 537 538 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint

539 Participants 540 We enrolled 26 dementia participants who were admitted to the Department of 541 Geriatric Medicine, The University of Tokyo Hospital, Tokyo, Japan, for evaluation of 542 cognitive impairment. All participants were diagnosed by experienced geriatricians 543 using DSM-IV criteria for AD (n=18), and Revised 2017 Clinical Diagnostic Criteria 544 for Dementia with Lewy Bodies for DLB (n=8). 35 Nine participants were NCIs who 545 admitted to the Department of Geriatric Medicine, The University of Tokyo Hospital, 546 for other reasons, except acute illness and autoimmune disease. Patients with malignant 547 disorders were excluded. We made precise diagnoses using psychological tests, 548 information from family, laboratory data, brain structural imaging (X-ray computed 549 tomography or nuclear magnetic resonance imaging). We also used 550 N-isopropyl-p-iodoamphetamine brain perfusion single photon emission computed 551 tomography (SPECT), metaiodobenzylguanidine, ioflupane dopamine transporter 552 SPECT. Clinical metrics included number of comorbidities, Charlson’s Comorbidity 553 Index, Comprehensive Geriatric Assessment-short version (CGA7), MMSE, HSDR, 554 Barthel Index, Lowton’s Instrumental Activities of Daily Living (IADL) scores, 555 Geriatric Depression Scale 15 (GDS15), and Vitality Index. All procedures were 556 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint approved by the Ethical Review Board at The University of Tokyo Hospital and The 557 University of Tokyo (approval number 2797). The clinical study guidelines of the 558 University of Tokyo, which conform to the Declaration of Helsinki, were strictly 559 adhered to CNIs, dementia patients and their families. They were provided with detailed 560 information about the study, and all provided written informed consent to participate. 561 562 Autoantibody measurement 563 WPAs were arranged as previously described. 43 First, proteins were 564 synthesized in vitro utilizing a wheat germ cell-free system from 13,455 clones of the 565 HuPEX.41 Second, synthesized proteins were plotted onto glass plates (Matsunami 566 Glass, Osaka, Japan) in an array format by the affinity between the GST-tag added to 567 the N-terminus of each protein and glutathione modified on the plates. The WPAs were 568 treated with human serum diluted by 3:1000 in the reaction buffer containing 1x 569 Synthetic block (Invitrogen), phosphate-buffered saline (PBS), and 0.1% Tween 20. 570 Next, the WPAs were washed, and goat anti-Human IgG (H+L) Alexa Flour 647 571 conjugate (Thermo Fisher Scientific, San Jose, CA, USA) diluted 1000-fold was added 572 to the WPAs and reacted for 1 hour at room temperature. Finally, the WPAs were 573 washed, air-dried, and fluorescent images were acquired using a fluorescence imager 574 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint (Typhoon FLA 9500, Cytiva, Marlborough, MA, USA). Fluorescence images were 575 analyzed to quantify serum levels of autoantibodies targeting each antigen, following 576 the formula shown below: 577 578 /g1827/g1873/g1872/g1867/g1853/g1866/g1872/g1861/g1854/g1867/g1856/g1877 /g1864/g1857/g1874/g1857/g1864 /g4670 /g1827/g1847 /g4671 /g3404 /g1832 /g3028/g3048/g3047/g3042/g3028/g3041/g3047/g3036/g3034/g3032/g3041 /g3398 /g1832 /g3041/g3032/g3034/g3028/g3047/g3036/g3049/g3032 /g3030/g3042/g3041/g3047/g3045/g3042/g3039 /g1832 /g3043/g3042/g3046/g3036/g3047/g3036/g3049/g3032 /g3030/g3042/g3041/g3047/g3045/g3042/g3039 /g3398 /g1832 /g3041/g3032/g3034/g3028/g3047/g3036/g3049/g3032 /g3030/g3042/g3041/g3047/g3045/g3042/g3039 /g3400 100 AU : arbitrary unit 579 F autoantigen : fluorescent intensity of autoantigen spot 580 F negative control : fluorescent intensity of negative control spot 581 F positive control : fluorescent intensity of positive control spot 582 583 AI-based analysis 584 We applied supervised machine learning techniques using Python (v3.10.12) 585 with libraries from Scikit-learn and the PyTorch framework to construct classifiers for 586 the diagnosis of dementia based on the autoantibody measurement data. The 587 performance of the classifiers was evaluated with 3-fold cross validation, using the 588 metrics of area under the receiver operating characteristics curve (AUC), area under the 589 precision-recall curve, accuracy, precision, recall, and F1-score, with the higher score 590 indicating the better classification performance. Machine learning models from 591 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Scikit-learn included simple linear regression, Lasso regression, Ridge regression, 592 logistic regression, support vector machine (SVM), random forest, XGBoost, 593 LightGBM, CatBoost, decision trees, gradient boosting machines and naïve Bayes to 594 conduct binary classification. Additionally, we also used deep neural networks with five 595 hidden layers in PyTorch to classify between three types of dementia. 596 597 Statistical analysis 598 Fisher’s exact test was performed to compare categorical variables. 599 Mann-Whitney U test was performed to compare continuous variables. Spearman 600 correlation test was used for correlation analysis. P values of < 0.05 were considered 601 statistically significant. Data analyses were conducted using R (v4.2.1). 602 603 Protein functional enrichment analysis 604 Gene Ontology Analysis using web-based tools targeted the list of the entry 605 clones coding the differentially highlighted autoantigens was performed for gene-list 606 enrichment analysis, gene-disease association analysis, and transcriptional regulatory 607 network analysis with Metascape.68 608 609 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Sequence identity analysis 610 To assess cross-reactivity among proteins that express similar antigen epitopes 611 and are highly correlated, we checked the correlation of the differentially expressed 612 autoantibodies. The corresponding proteins of the highly correlated autoantibodies 613 (Spearman’s r > 0.5) were then aligned with the highly correlated proteins using the 614 Uniprot alignment tool. 615 616 Data visualization 617 Box plots, scatter plots, hierarchical clustering, and correlation matrix were 618 visualized by using R (v4.2.1). Box plots were defined as follows: the middle line 619 corresponds to the median; the lower and upper hinges correspond to the first and third 620 quartiles; the upper whisker extends from the hinge to the largest value no further than 621 1.5 times the interquartile range (IQR) from the hinge; and the lower whisker extends 622 from the hinge to the smallest value at most 1.5 times the IQR of the hinge. 623 624 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Figure legends Figure 1. Autoantibodies differentially elevated in dementia. (A) The SAL in AD, DLB, and CNI. (B) Volcano plot that shows autoantibodies differentially elevated in AD compared to NCI. The vertical dash line indicates P = 0.05. The horizontal dash line indicates fold change = ± 2. (C) Volcano plot that shows autoantibodies differentially elevated in DLB compared to NCIs. The vertical dash line indicates P = 0.05. The horizontal dash line indicates fold change = ± 2. (D) Venn diagram that illustrates the inclusion relationship between autoantibodies differentially elevated in AD and/or DLB. The vertical dash line indicates P = 0.05. The horizontal dash line indicates fold change = ± 2. (E) Heat map that shows the serum levels of 229 autoantibodies differentially elevated in AD and/or DLB. (F) PCA of 229 autoantibodies differentially elevated in AD and/or DLB. In the scatter plot, individual subjects as points. (G) PCA plots colored by sex, age, HDSR, and MMSE. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Figure 2. Autoantibodies highlighted in 2-class classification tasks by AI. (A) Autoantibodies that were mostly highlighted according to feature importance by Logistic regression and SVM with standardization or normalization. (B) UpSet plot shows the inclusion relationship of autoantibodies highlighted by the four machine learning frameworks. (C) Box plots describe the serum levels of autoantibodies highlighted by more than two frameworks in AD, DLB, and CNI. (D) Heatmap illustrates correlation between autoantibodies highlighted in machine learning analysis and demographic and clinical characteristics of dementia. *: P < 0.05, **: P < 0.01. P values were calculated by Spearman’s correlation test. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Figure 3. Performance of deep neural network for 3-class classification by AI. (A) Learning curves of the deep neural network model in 3-fold cross validation. (B) ROC curves of the deep neural network model in 3-fold cross validation. Class 1: CNI, class 2: AD, class 3: DLB. (C) Precision-recall curves of the deep neural network model in 3-fold cross validation. Class 1: CNI, class 2: AD, class 3: DLB. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Figure 4. Autoantibodies to neuroactive ligand-receptor interaction-associated proteins. Gene ontology analysis encompassing the genes coding proteins targeted by autoantibodies differentially elevated in AD and/or DLB. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Figure 5. Correlation between autoantibodies highlighted in gene ontology analysis and clinical traits of dementia. (A) Box plots describe the serum levels of autoantibodies to neuroactive ligand-receptor interaction-associated proteins. (B) Heatmap illustrates correlation between autoantibodies to neuroactive ligand-receptor interaction-associated proteins and demographic and clinical characteristics of dementia. (C) Box plots describe the serum levels of autoantibodies to regulation of lipid metabolic process-associated proteins. (D) Heatmap illustrates correlation between autoantibodies to regulation of lipid metabolic process-associated proteins and demographic and clinical characteristics of dementia. *: P < 0.05, **: P < 0.01. P values were calculated by Spearman’s correlation test. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Extended Figure 1. Sum of autoantibody levels by age and sex. (A) Box plots show SAL by age groups. (B) Box plots show SAL by sex. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Extended Figure 2. Additional information for autoantibodies highlighted in 2-class classification tasks by AI. (A) Box plots describe the serum levels of autoantibodies highlighted in 2-class classification tasks by sex. (B) A correlation matrix of the autoantibodies highlighted in 2-class classification tasks using Spearman’s correlation. Only statistically significant pairs (P 0.5). (D) Box plots describe the serum levels of autoantibodies highlighted in 2-class classification tasks in COVID-19, atopic dermatitis, anti-neutrophil cytoplasmic antibody-associated vasculitis, systemic lupus erythematosus, systemic sclerosis, and healthy controls. The data derives from the UT-ABCD. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Extended Figure 3. Additional information for autoantibodies to neuroactive ligand-receptor interaction-associated proteins. (A) Box plots describe the serum levels of autoantibodies to neuroactive ligand-receptor interaction-associated proteins by sex. (B) A correlation matrix of the autoantibodies to neuroactive ligand-receptor interaction-associated proteins using Spearman’s correlation. Only statistically significant pairs (P 0.5). (D) Box plots describe the serum levels of autoantibodies to neuroactive ligand-receptor interaction-associated proteins in COVID-19, atopic dermatitis, anti-neutrophil cytoplasmic antibody-associated vasculitis, systemic lupus erythematosus, systemic sclerosis, and healthy controls. The data derives from the UT-ABCD. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Extended Figure 4. Correlation between serum levels oof anti-NPB antibodies and MMSE subscales. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Extended Figure 5. Expression of the NPB gene in human tissues and single cells. (A) Expression of NPB in multiple human tissues measured by bulk RNA-sequencing from the Human Protein Atlas. (B) Expression of NPB in the CNS from the Human Protein Atlas. (C) Expression of NPB in the CNS evaluated by single-cell RNA-sequencing from the Human Protein Atlas. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Extended Figure 6. Correlation between serum levels oof anti-ADGRF5 antibodies and MMSE subscales. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Extended Figure 7. Additional information for autoantibodies to regulation of lipid metabolic process-associated proteins. (A) Box plots describe the serum levels of autoantibodies to regulation of lipid metabolic process-associated proteins by sex. (B) A correlation matrix of the autoantibodies to regulation of lipid metabolic process-associated proteins using Spearman’s correlation. Only statistically significant pairs (P 0.5). (D) Box plots describe the serum levels of autoantibodies to regulation of lipid metabolic process-associated proteins in COVID-19, atopic dermatitis, anti-neutrophil cytoplasmic antibody-associated vasculitis, systemic lupus erythematosus, systemic sclerosis, and healthy controls. The data derives from the UT-ABCD. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Extended Figure 8. Expression of the ADGRF5 gene in human tissues and single cells. (A) Expression of ADGRF5 in multiple human tissues measured by bulk RNA-sequencing from the Human Protein Atlas. (B) Expression of ADGRF5 in the CNS from the Human Protein Atlas. (C) Expression of ADGRF5 in the CNS evaluated by single-cell RNA-sequencing from the Human Protein Atlas. . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint A B E F G Matsuda KM et al. Figure 1 C D 152 36 41 CNI vs AD CNI vs DLB P = 0.16 P = 0.48 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Matsuda KM et al. Figure 2 A B C D . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Matsuda KM et al. Figure 3 A B C . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint Matsuda KM et al. Figure 4 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint A Matsuda KM et al. Figure 5 C D B . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted September 23, 2024. ; https://doi.org/10.1101/2024.09.20.24313547doi: medRxiv preprint