Identifying Age-Related Protein Mechanisms of Alzheimer's Disease Amyloidosis from Cerebrospinal Fluid Proteomics Using a Novel Machine Learning Approach

Identifying Age-Related Protein Mechanisms of Alzheimer's Disease Amyloidosis from Cerebrospinal Fluid Proteomics Using a Novel Machine Learning Approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Identifying Age-Related Protein Mechanisms of Alzheimer's Disease Amyloidosis from Cerebrospinal Fluid Proteomics Using a Novel Machine Learning Approach Justin Melendez, Leila Mozaffar, Ziqiao Jiao, Kaleigh Roberts, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8791371/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Alzheimer’s disease (AD) is clinically characterized by progressive memory loss and cognitive decline, with aging as the primary risk factor. Early AD pathology includes accumulation of amyloid-beta (Aß) in plaques. This study aims to investigate the molecular mechanisms by which aging contributes to brain amyloidosis using a machine learning-based approach on large proteomic datasets from cerebrospinal fluid (CSF). To accomplish this, we trained a machine learning model to predict CSF Aß42/40, a key biomarker for amyloidosis. Our modified elastic net model using adaptive feature selection achieved robust accuracy (Pearson Correlation of 0.86) predicting CSF Aß42/40 in our validation cohort. Pathway analysis of the model-utilized proteins (and proteins highly correlated to them) revealed age-associated alterations potentially linked to amyloidosis, particularly highlighting dysregulated autophagy and membrane trafficking pathways. These findings suggest that impaired autophagosome-lysosome fusion and endosomal processing may drive the decline in Aß clearance with aging. Our study highlights the power of machine learning in biomarker approximation and biological prediction, enabling insights into multiple diseases. Biological sciences/Neuroscience/Cognitive ageing Biological sciences/Computational biology and bioinformatics/Machine learning Biological sciences/Neuroscience/Diseases of the nervous system/Alzheimer's disease Biological sciences/Developmental biology/Ageing Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Alzheimer’s disease (AD) is a devastating and irreversible neurodegenerative disorder that slowly destroys memory, thinking skills, and, eventually, the ability to carry out daily tasks. Its impact extends far beyond the individual, placing an immense emotional, physical, and financial burden on family members and caregivers. Aging is the single biggest risk factor of sporadic late onset Alzheimer’s disease (LOAD) with the incidence of AD doubling every five years after the age of 65 (Qiu et al., 2009 ). With a rapidly growing older population, AD is an increasing global health concern. Estimates project that the number of Americans aged 65 and older living with AD could grow from approximately 7 million today to nearly 14 million by 2060, underscoring the urgent need for a deeper understanding of the molecular mechanisms by which aging drives the onset of AD (“2025 Alzheimer’s Disease Facts and Figures,” 2025). A critical, early pathological event in AD is the aggregation of amyloid-beta (Aß) protein in amyloid plaques (Ma et al., 2022 ). Aß is generated through the sequential cleavage of the Amyloid Precursor Protein (APP), a transmembrane protein found in neuronal and non-neuronal cells. APP can be processed through two competing pathways: 1) The non-amyloidogenic pathway occurs primarily at the cell membrane surface. Here APP is first cleaved by alpha-secretase within the Aß domain, preventing Aß formation. 2) The amyloidogenic pathway is activated when APP is internalized in a clathrin-dependant manner and trafficked to endosomal compartments. Here, APP is first cleaved by beta-secretase (BACE1), which releases a soluble ectodomain and leaves a C99 membrane-bound fragment. This fragment is then cleaved by the gamma-secretase complex to release the Aß peptide (Campion et al., 2016 ; O’Brien & Wong, 2011 ). However, gamma-secretase cleavage is imprecise, generating Aß peptides of varying lengths, including Aß40 (40 amino acids) and Aß42 (42 amino acids). The Aß42 isoform, which contains two extra hydrophobic amino acids, is significantly more prone to aggregation and plaque formation than the shorter, more abundant Aß40 isoform (Findeis, 2007 ; Haass & Selkoe, 2007 ). Understanding the specific age-related biological factors that drive this initial amyloid-beta (Aß) accumulation may identify novel therapeutic targets capable of delaying or preventing the onset of AD. The ratio of Aß 42 to Aß 40 (Aß42/40) in the cerebrospinal fluid (CSF) is a widely used and important clinical biomarker that reflects the onset of amyloidosis in the brain (Lewczuk et al., 2020 ; Schindler et al., 2019 ). A decline in the CSF Aß42/40 ratio indicates decreased clearance or aggregation of the more aggregation-prone Aß42 isoform and is an early indicator of AD pathology (Amft et al., 2022 ; Lian et al., 2025 ). Previous studies have indicated that the ability to clear central nervous system (CNS) Aß42 and other amyloid-beta species from the brain declines by 400% with aging from the 30s through the 80s, and further Aß42 clearance is specifically altered with amyloidosis, a process strongly correlated with disease progression (Patterson et al., 2015 ). Machine learning is a powerful tool for navigating the high dimensionality and complexity of large biological data sets, such as proteomics, to reveal insightful patterns often imperceptible through traditional statistical methods. This technology has been successfully leveraged to create numerous aging clocks, from large scale omics data, that have increased our understanding of how aging impacts disease (Horvath, 2013 ; Melendez et al., 2024 ; Oh et al., 2023 ; Wang et al., 2026 ). The past decade has seen an explosion of large proteomic data sets generated by high-throughput platforms like SomaLogic, Olink, and Alamar (Bycroft et al., 2018 ; Imam et al., 2025 ). Large-scale studies utilizing the SomaScan platform have identified thousands of proteins associated with AD and other neurodegenerative diseases, distinguishing shared versus disease-specific molecular pathways (Ali, Erabadda, et al., 2025 ). Work using Olink technology has successfully employed multiplex panels to identify high-performance, multi-protein signatures capable of classifying AD from controls with high accuracy (Jiang et al., 2022 ). Our primary goal was to gain biological insights into the age-associated molecular mechanisms underlying amyloidosis. To achieve this, we employed an adaptive elastic net machine learning model with a novel feature selection strategy to predict CSF Aß42/40, a key clinical biomarker of amyloidosis, solely from CSF-derived proteomics (SomaLogic 7k platform). By extracting the features from this model, we were able to better understand the biological basis of the predictions. Furthermore, we hypothesized that the APOE ε4 carrier status, the most significant genetic risk factor for amyloidosis, may modify the progression of Aß accumulation. Therefore, we included APOE ε4 carrier status as a feature in the model. The final model, which integrates chronological age, APOE ε4 status, and proteomics data, accurately estimates CSF Aß42/40 (Fig. 1 ). Finally, by extracting the protein features used by the algorithm, we gained biological insights into the age-associated factors underlying amyloidosis. The proteins identified by the model as most crucial for predicting amyloidosis represent key candidates for further study. We specifically examined the subset of these proteins that were correlated with aging and then performed pathway analysis on this age-associated set (and their highly correlated partners) to identify the age-related biological pathways that may be driving the decline in Aß42 clearance with age and the onset of amyloidosis. Our study illustrates the potential of machine learning not only in biomarker approximation but also in extracting predictive features to reveal biological insights. The identified pathways and proteins in this study are potential targets for intervention in research studies to test causality of these pathways. This study also demonstrates approaches identifying potential predictive proteins and pathways across multiple diseases using standardized proteomic data sets. Results We developed a machine learning model to analyze a large proteomic dataset for insights into age associated pathways and proteins predictive of amyloidosis. Our goal was to train models predicting a key AD biomarker and then interrogate the models to understand how their predictions were made. We hypothesized that age associated pathways enriched in model proteins would reveal key markers of the disease. As amyloid plaque formation is one of the earliest stages thought to lead to AD, our model was designed to predict the CSF Aß42/40 ratio, a key biomarker for amyloidosis in Alzheimer's disease pathology. Finally, we extracted protein features from this model and performed pathway analysis to analyze age associated proteomic signatures predictive of amyloidosis. The results from our model, along with our feature selection and age associated protein pathway analyses, are presented below. Machine learning to predict CSF Aß42/40 To investigate age-associated molecular factors linked to amyloidosis, we developed a machine learning approach to predict the CSF Aß42/40 ratio directly from proteomic data. To do this we trained an elastic net machine learning model on 7,008 protein features in CSF from the SomaLogic 7k assay from 856 participants from the Knight Alzheimer’s Disease Research Center (ADRC) to predict Lumipulse measured CSF Aß42/40 ratio, a widely used clinical biomarker used to identify amyloidosis in Alzheimer's disease patients. We then used an algorithm developed in our lab that we call “adaptive feature selection” to simultaneously improve model performance while minimizing protein features used (Fig. 2 ). This created a highly predictive model using a minimal number of protein features (Fig. 3 ). Model performance on the Knight ADRC cohort, used to train the models, was assessed by testing on 100 different train test splits made up of 80% training and 20% testing data held aside for each model. The median performing model had a 0.88 correlation (P < 0.0001) to Lumipulse measured CSF Aß42/40 (Fig. 3 A) and a mean absolute error (MAE) from the participants measured CSF Aß42/40 of 0.009 (Fig. 3 B). The model used just 93 protein features along with the participants age and presence or absence of the APOE ε4 allele to make this prediction (Fig. 3 C) (Model weights are available in Supplementary Materials Table 1). In addition to its predictive performance, we also evaluated the model's utility as a classification tool. This was accomplished by assessing its ability to distinguish between individuals with and without amyloidosis. To do this we used the established CSF Aß42/40 ratio cutoff from the Knight ADRC cohort, where a ratio below 0.0673 indicates amyloidosis positivity and a ratio above this threshold indicates a negative status (Barthélemy et al., 2023 ). When tested against this criterion, the median-performing model demonstrated robust performance, achieving a high accuracy of 94% in classifying amyloidosis positivity versus negativity. Next, we evaluated the pre-trained median Knight ADRC model on an external proteomics data set from ADNI which was locked and not visible to the model during training. The ADNI validation cohort used the Elecsys platform to measure CSF Aß42/40 in contrast to the Knight ADRC training data which used Lumipulse, however measurements form both platforms have been reported to be highly concordant with each other (Dakterzada et al., 2021 ) and interchangeable. We found no harmonization was needed for our Lumipulse trained model to predict Elecsys measured Aß42/40 ratios. Here the model had a Pearson correlation of 0.86 between predicted CSF Aß42/40 ratio and Elecsys measured CSF Aß42/40 ratio in the ADNI validation cohort (48 participants) and mean estimated error of 0.013 from measured values. Additionally, the model was 88% accurate in classifying amyloidosis positive vs negative participants in this validation data (Fig. 3 D). Age association of proteins used in the model Aging is a primary risk factor for Alzheimer's disease (AD) and amyloidosis, with the incidence of AD doubling every five years after the age of 65 (Qiu et al., 2009 ). The body's ability to clear amyloid-beta from the brain declines significantly with age, a process that is strongly correlated with the progression of AD. This age-related decline in clearance is a critical factor in the accumulation of amyloid plaques, a hallmark of AD pathology. As such we sought to examine what proteins in our machine learning model predicting CSF Aß42/40 had strong associations with age in the Knight ADRC cohort. We hypothesize that these proteins and associated pathways may be drivers of declining amyloid beta clearance capacity or compensatory mechanisms to combat the negative effects of this decline. Our model made use of 93 proteins, in addition to chronological age and APOE E4 presence or absence, to predict CSF Aß42/40. As the model was designed to aggressively drop predictively redundant proteins, we looked for proteins that were highly correlated with these model proteins prior to performing pathway analysis (Fig. 4 A). Of the 93 model selected proteins, 728 proteins strongly correlated with them (Pearson correlation > = 0.7) (Supplementary Materials Table 2). Of these 821 proteins 37.5% (308) were correlated with age at a p-value of 0.0001 or lower in the Knight ADRC cohort giving us an estimated false discovery rate (FDR) of 0.027%. We considered a protein age associated if it was significantly (P 0.0673) or post-amyloidosis participants (CSF Aß42/40 < = 0.0673). Pathway analysis of age associated amyloidosis predictive proteins: We next sought to determine which pathways enriched for age associated proteins in our CSF Aß42/40 predictive model. As the set of pre-amyloidosis aging proteins captured nearly all post-amyloidosis aging proteins, and our primary interest was in finding potential age-related drivers of amyloidosis, we focused on the 306 pre-amyloidosis age-related proteins for pathway analysis (Fig. 4 B, C). Pathway analysis was performed using the bioinformatic tool Metascape (metascape.org) on the Reactome pathway knowledgebase using proteins measured in the Somalogic 7k assay as a background (reactome.org) (Fig. 5 ). Our analysis showed signals of dysregulated protein quality control, increased DNA damage repair, increases in cellular stress responses, and dysregulated cell cycle proteins, increased with age and amyloidosis. Top pathways that were enriched included Membrane Trafficking (R-HSA-199997), Autophagy (R-HSA-9612973) and Signaling by Rho GTPases (R-HSA-194315), Formation of Incision Complex in GG-NER (R-HSA-5696395), Attenuation phase (R-HSA-3371568) and Deregulated CDK5 triggers multiple neurodegenerative pathways in Alzheimer's disease models (R-HSA-8862803). Fold change enrichment for each pathway, Reactome pathway identifiers, log10 q-values (p-value adjusted for false discovery rate for each enriched pathway), proteins enriched in each pathway, protein model weights and protein correlations with age in non-amyloidosis participants can be found in Supplementary Materials Table 3. Discussion The onset and progression of Alzheimer’s disease is inextricably linked to aging, with amyloidosis representing one of the earliest detectable pathological stages. Understanding the age-related biological factors that drive the accumulation of Aß is crucial for identifying therapeutic targets. In this study, we leveraged the power of machine learning to process large-scale, high-dimensional proteomic data to approximate a clinical biomarker of amyloidosis, CSF Aß42/40 ratio, and extract novel biological insights related to aging and amyloidosis. Machine Learning Robustly Approximates Clinical Biomarkers from Proteomics A key challenge in utilizing newly available, large proteomic datasets (such as those generated by SomaLogic, Olink, and Alamar) is the frequent lack of accompanying gold-standard biomarkers (e.g. CSF Aß42/40) necessary for comprehensive disease analysis. To address this issue, we demonstrated that machine learning models could accurately predict the value of a critical biomarker of Alzheimer's Disease (AD) research, CSF Aß42/40, using CSF proteomics data. Our primary goal was to gain insights into aging and amyloidosis progression. Therefore, we needed a model capable of predicting a wide range of amyloidosis states, from non-amyloidosis to amyloidosis positivity. Instead of creating a classifier that determined whether a participant was amyloidosis positive or negative, we chose to predict the actual CSF Aß42/40 biomarker ratio. We hypothesized that a continuous model that could estimate the ratio of these Aß peptides would better capture the progression from non-amyloidosis to amyloidosis. Currently the only way to measure this biomarker ratio accurately is via mass spectrometry or specialized antibody assays such as Lumipulse or Elecsys. This continuous approach would provide more nuanced insights into the biology, compared to classifiers that rely on binary differentiation of amyloidosis positivity versus negativity. We think that the largest differentiators identified by binary classifiers were more likely to represent disease endpoints and miss proteins and pathways important in transitional states. Thus, our aim was to build a machine learning model that was as accurate as possible at estimating an amyloidosis biomarker, the idea being that the better performing models would be more likely to identify proteins and pathways relevant to the underlying biology during the feature extraction stage. To accomplish this, we employed our modified elastic net model, utilizing our novel "adaptive feature selection" approach to reduce features and improve model accuracy. This method successfully estimated CSF Aß42/40, a clinical measure reflecting the onset of amyloidosis. Since CSF Aß42/40 is clinically used to classify amyloidosis positivity vs negativity, though not our primary goal, we also reported how well our model estimated CSF Aß42/40 compared to measured CSF Aß42/40. The model achieved a high correlation of R = 0.88 and 94% classification accuracy on the internal Knight ADRC testing cohort. Crucially, this predictive capability generalized well to the unseen external ADNI validation cohort, maintaining a strong correlation of R = 0.86 and 88% classification accuracy, despite the ADNI cohort using a different but comparable assay platform (Elecsys vs. Lumipulse). This robust cross-cohort validation provides compelling evidence that the protein signatures identified by the model captures fundamental, cohort-independent proteomics changes predictive of Aß accumulation, rather than cohort-specific overfitting. Age-Associated Pathways Driving Amyloidosis: Failure to Complete Autophagy May Lead to Increased Amyloidosis and Neurodegeneration The Autophagy pathway (R-HSA-9612973) showed the highest level of significance for enriched pathways with a log10 q-value of − 3.4 and enrichment fold change over background of 4.8x. Autophagy contributes to cellular homeostasis by degrading damaged organelles and misfolded proteins. A wide range of proteins involved in autophagosome formation and function, including RNASE1, UBB, UBE2N, UBE2V1, GABARAP, GABARAPL2, GABARAPL1, MAP1LC3B, MAP1LC3A, MVB12B, and DYNLL2, were enriched in this pathway (Frake et al., 2015 ; Gong et al., 2016 ; Xie & Klionsky, 2007 ). All model proteins enriched in the pathway had negative model weights indicating that increased amounts of the proteins was associated with amyloidosis. Additionally, these proteins all had positive correlations with chronological age. Overall, we observe that the presence of proteins associated with the initiation of autophagy generally increases with age and predicts a worsening disease state. Initial assessment might suggest this appears contradictory, as an increase in autophagic proteins should result in improved clearance of amyloid beta. However, the literature shows that there is an increase in the number of immature autophagic vacuoles in the neurons of people with AD due to defective autophagosome-lysosome fusion (Boland et al., 2008 ; Hussain et al., 2025 ; Nixon et al., 2005 ). This aligns with our interpretation that the age-related increases in autophagic proteins detected in the CSF may reflect a pathological compensatory response. We hypothesize that age-related increase in these autophagic proteins in the CSF may be a proxy for a pathological compensatory response within cells attempting to initiate and ramp up the autophagic process in response to the mounting burden of misfolded proteins and Aß. However, failure of these autophagic vacuoles to fuse with lysosomes creates a bottleneck for the cellular machinery leading to a buildup of autophagic proteins. It is possible this lack of lysosomal fusion is a driver of amyloid beta clearance deficits leading to increased amyloid plaques with age, however more work will need to be done to see if this is a result or cause of amyloidosis and if what we observe in the CSF truly reflects intracellular and tissue changes. If autophagic impairment is a causal mechanism, an exciting avenue of research may be to correct this fusion failure phenotype and determine if amyloid beta clearance improves, or to target rejuvenation of the lysosomes directly. Interestingly, mutations in Presenilin 1 (PS1), the protein that forms the catalytic core of the gamma-secretase enzyme and the most commonly mutated protein in autosomal dominant Alzheimer's disease (ADAD), have also been implicated in a mechanism independent of Aß production, causing the failure of autophagosomes to fuse with and be degraded by the lysosomes due to impaired lysosomal acidification (Lee et al., 2010 ). Tied with the Autophagy pathway, Membrane Trafficking (R-HSA-199997) was also highly enriched with a score log10 q-value of -3.4 and enrichment fold change over background of 2.6x. Like the Autophagy pathway, most membrane trafficking proteins had negative model weights indicating increased abundance was predictive of worsening pathology, and all model proteins enriched in the pathway were positively correlated with age. Proteins from the 14-3-3 family (YWHAZ, YWHAB, YWHAE, YWHAG, YWHAH) were highly enriched in both the Membrane Trafficking pathway and Signaling by Rho GTPases pathway (R-HSA-194315), the third most highly enriched pathway with a log10 q-value of -3.1 and enrichment fold change over background of 2.5x. These proteins are well-established CSF biomarkers in the Alzheimer's disease continuum, have been identified as being significantly dysregulated in other multi-cohort proteomics studies (Ali, Timsina, et al., 2025 ; Qiang et al., 2023 ) and are recognized as central players at the intersection of aging and neurodegenerative disease (Fan et al., 2019 ). 14-3-3 proteins are among the most abundant in the brain, functioning as central hub proteins that regulate signal transduction, protein transport, and cytoskeletal dynamics by binding to phosphorylated target proteins (Abdi et al., 2024 ; Pennington et al., 2018 ) and are known to have a protective, chaperone-like role, suppressing the aggregation of tau and other proteins and regulating Aß clearance via the lysosomal pathway (Abdi et al., 2024 ). This fits with a picture of the brain mounting a compensatory response to failed amyloid beta clearance in the brain. Paired with the observation of a failing autophagy system, we hypothesis that this increase in membrane trafficking proteins coupled with worsening amyloidosis could be a symptom of the same lysosomal bottleneck. Normally, APP is recycled from the membrane by internalization into early endosomes along with BACE1 and γ-secretase. This change in pH in the early endosome is conducive of amyloidogenic processing of APP by γ-secretase after which it is either transported to the lysosome for degradation (by fusion of the endosome to a lysosome) or released from the cell. Alternatively, APP is recycled to the membrane via retromers and the trans-Golgi network (Jiang et al., 2014 ; Siegenthaler & Rajendran, 2012 ). We hypothesize that the longer residence time of APP in early endosomes, and the pro-amylogenic environment of those vesicles, leads to APP being more likely to be cleaved into Aß instead of being transported to the Golgi and recycled to the membrane. This could explain how failure of early endosomes to complete autophagy would lead to reduced clearance of Aß leading to upregulation of both membrane trafficking and autophagy machinery in an attempt to overcome this backup of endosome processing. Another key protein in the membrane trafficking pathway is IGF2R (insulin-like growth factor 2 receptor), which has a negative model weight and a positive age correlation. This protein is crucial for transporting lysosomal enzymes from the Golgi to the lysosome (Wang et al., 2020 ). Since Aß is predominantly produced in endosomal and lysosomal compartments, proper trafficking of both APP and the enzymes that cleave it is essential for regulating Aß production and degradation. The dysregulation of a key trafficking protein like IGF2R could compromise the delivery of critical degradative enzymes to the lysosome. This would, in turn, impair the efficiency of the autophagy-lysosome pathway, leading to the accumulation of immature autophagosomes leading to the bottleneck discussed earlier, and reduced cellular clearance of Aß. Conclusion Our findings support the use of machine learning as a powerful tool for biomarker approximation and biological discovery in the era of high throughput 'omics data. Here we have demonstrated that CSF Aß42/40 can be accurately inferred from standard SomaLogic 7k proteomics. This capability offers a path toward a single, integrated biofluid test where routine proteomic screens could simultaneously diagnose, monitor, or test for multiple diseases by simply changing the machine learning model applied to the output. Moving forward, the specific proteins and pathways identified by the model, particularly those highly correlated with age, represent high-priority candidates for targeted functional studies. Future work should focus on validating the causal relationship between the identified age-related proteins and the mechanisms of Aß clearance and production. Specifically, the enrichment of autophagy and membrane trafficking proteins suggests a key pathological bottleneck in the endosomal-lysosomal pathway, characterized by a hypothesized age-related failure of autophagosome-lysosome fusion. This fusion failure may lead to both impaired Aß degradation and a prolonged residence time for APP in pro-amyloidogenic early endosomes, thereby increasing Aß production and driving pathology. Much work needs to be done to vet and validate this hypothesis, however Interventions targeting this fusion failure phenotype as a possible age-related link is a particularly exciting avenue for research. Such work could illuminate the molecular mechanisms by which aging increases AD risk, potentially leading to the development of novel therapeutic interventions aimed at restoring cellular quality control and mitigating age-related stress. Materials and Methods Cohort description CSF proteomics data was obtained for this study via a data request to the Knight Alzheimer Disease Research Center (ADRC) and the Alzheimer's Disease Neuroimaging Initiative (ADNI). The Knight ADRC cohort consisted of 856 participants while the ADNI cohort consisted of 48 participants. Participants in the Knight ADRC were aged 43–99 years old and ranged from cognitively unimpaired (CN, CDR = 0) to those with dementia (CDR > = 1) and covered a wide range of amyloidosis states including those of negative, positive and intermediate status as measured by the CSF Aß42/40 biomarker. Participants in ADNI ranged from ages 60 to 86 and likewise included both CN and CDR > = 1 participants over a similar range of amyloidosis states to the Knight ADRC participants. Participants from the ADRC and ADNI cohorts were selected based on availability of proteomic data and amyloid-β (Aβ) 42/40 ratio measurements obtained within 1 year of CSF collection. CSF 42/40 measurements were quantified using the Lumipulse platform for the Knight ADRC cohort, while they were quantified using the Elecsys immunoassay platform in the ADNI cohort. Proteomic profiling was performed using the SomaLogic SOMAscan 7k assay in both cases. APOE ε4 status information was available for all participants through genetic testing. Adaptive feature selection algorithm Adaptive feature selection is an algorithm we developed that reduces the number of features used in a machine learning model while simultaneously improving predictive performance. We developed this algorithm to solve the frequent problem in biological data sets of having large numbers of features with relatively small sample sizes. It can be applied to any machine learning model where feature impact can be measured. Here we will describe how we applied the algorithm to our elastic net machine learning model of CSF Aß42/40. An elastic net machine learning model was trained on 100 balanced testing and training splits of the CSF proteomics SomaLogic 7k data from the Knight ADRC. This generated 100 differently weighted elastic net models, each predicting the biomarker of interest. In each model the trained elastic net assigned a feature weight to each of the 7,008 proteins. Protein features with a weight of zero were not useful to the model, whereas features with positive weights were considered informative. For each of the 100 models we created a binary array of whether the feature was used (1 for informative, 0 for not used). For each feature we then asked what percentage of the time it was used across all models (Sum of times informative/100 models = percent used). From here we created feature sets based on thresholds of the percentage of features that were used. For instance, if a feature was used in 10% or more of the models it was included in the 10% feature set for that round. If a feature was used in 90% or more models, it was included in the 90% feature set. For simplicity and computational performance feature usage thresholds were set to 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70% 80%, 90%, and 100% or more usage across models. For each feature set, all other features were removed, and the elastic net was run on those features for 100 new train-test splits of the data. From here the median performing model of each feature set was measured via Person correlation of the model’s prediction vs real world measurements of the biomarker. The feature set with the best performing median model was chosen to move on to the next round. If the median model’s performance was better than the prior best median model, the entire adaptive feature selection process was repeated using that model’s features as the starting point. This progressively produced models with fewer features each round and improved performance until the best performing feature set selected no longer outperformed the previous round, at which point the features and weights of the median model would be set and the final, trained model produced. CSF Aß42/40 machine learning model Custom python code was written to train our elastic net-based machine learning algorithm to predict CSF Aß42/40 based on 7008 protein features measured in the SomaLogic 7k data set in human CSF along with chronological age and APOE ε4 presence or absence. The elastic net algorithm was implemented via the scikit-learn machine learning library. Missing data points were imputed using a k nearest neighbor-based approach with K = 2. Prior to training the data was scaled from 0 to 1 using the min-max feature scaler of scikit-learn, which gave the best performance of the data transformation approaches tried. For a single round of elastic-net training the data were randomly split into two class balanced sets: a training set consisting of 80% of the data and a test set consisting of 20% of the data. The training set was then again split into two groups consisting of 80% of the training set, the “sub-training set,” and 20% of the training set, the “sub-test set,” to find hyperparameters. The sub-training set and sub-test sets were used to train the elastic net and find the optimal tuning parameters (L1 and Alpha), while the original test set was used to gauge performance e for this stage of training. Each model was trained on the Knight ADRC SomaLogic 7k CSF proteomics cohort (N = 856) and validated on the ADNI SomaLogic 7k CSF proteomics cohort (N = 48 for CSF Aß42/40 biomarker measurements). Model performance was trained and assessed on Pearson correlation of the model’s output with clinical biomarker measurements (Lumipulse or Elecsys). Metascape Analysis A pathway enrichment analysis was conducted using the bioinformatic tool Metascape ( https://metascape.org ) to identify biological pathways that were associated with proteins predictive of the CSF Aß42/40 ratio. The original analysis began with 93 protein features identified as predictive of Aß42/40 in the model. Additional proteins that were highly correlated (r > 0.7) with these 93 protein features were included, resulting in 821 total proteins. Of these proteins, 308 were significantly correlated with age in non-amyloidosis participants and were selected for enrichment analysis. Metascape analysis was performed using the Reactome knowledge base with all 7008 possible protein features used as the background protein list. Enrichment significance was evaluated using default Metascape parameters. Among the significantly enriched pathways, SARS-CoV-2 (Reactome: R-HSA-9755779) appeared highly enriched on the list but was chosen to be omitted as it was not deemed relevant to the analysis. Declarations Acknowledgements We thank the entire Bateman lab for their valuable feedback and support. We also thank all the participants and their families, as well as the many institutions involved and their staff. This research was supported by Tracy Family Stable Isotope Labeling Quantitation Center established by the Tracy Family ( https://silqcenter.wustl.edu/ ), Richard Frimel & Gary Werths, GHR Foundation, Pat and Jane Tracy, Anonymous, Anne & Ray Capestrain, Community Foundation Serving West Central Illinois and Northeast Missouri, JTL Family Fund, Payne Family, Mary & Jay Sullivan, Tracy Family Foundation, Catherine & Tom Tracy, Community Foundation for the Land of Lincoln, Jim & Jil Tracy, Joe & Jill Tracy, Sonja & Robert M. Willman, Boniface Foundation, Jean & Michael Buckley, Ann Liberman, Clemence S. Lieber Foundation, Mary Schoolman & Dr. James Hinrichs, and Susan & Scott Stamerjohn brought together by The Foundation for Barnes-Jewish Hospital. Research was also supported by the Healthy Aging and Senile Dementia P01 AG03991, Alzheimer’s Disease Research Center P30 AG066444, Adult Children Study P01 AG026276. Validation cohort data collection and sharing for this project was funded by the Alzheimer's Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U19 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). This work was also supported by access to equipment made possible by the Hope Center for Neurological Disorders, the Neurogenomics and Informatics Center (NGI: https://neurogenomics.wustl.edu/ ) and the Departments of Neurology and Psychiatry at Washington University School of Medicine. Data Availability Statement The data that support the findings of this study are available on request from the corresponding author upon approval from the Knight Alzheimer Disease Research Center ( https://knightadrc.wustl.edu/ ) and the Alzheimer's Disease Neuroimaging Initiative (ADNI). The data are not publicly available due to privacy or ethical restrictions. All trained models and software to run the CSF Aß42/40 model presented in this paper on custom data sets are available upon request from the corresponding author. Code to run the final model is also available through the Bateman lab GitHub at https://github.com/wusm-neurology-batmanlab/CSF_ABeta4240_Proteomics_Clock . References Alzheimer’s disease facts and figures (2025) Alzheimer’s Dement 21(4):e70235. https://doi.org/10.1002/alz.70235 Abdi G, Jain M, Patil N, Upadhyay B, Vyas N, Dwivedi M, Kaushal RS (2024) 14-3-3 proteins—a moonlight protein complex with therapeutic potential in neurological disorder: In-depth review with Alzheimer’s disease. Front Mol Biosci 11:1286536. https://doi.org/10.3389/fmolb.2024.1286536 Ali M, Erabadda B, Chen Y, Xu Y, Gong K, Liu M, Pichet Binette A, Timsina J, Western D, Yang C, Heo G, Vogel JW, Tijms BM, Krish V, Imam F, Hansson O, Winchester L, Cruchaga C (2025) Shared and disease-specific pathways in frontotemporal dementia and Alzheimer’s and Parkinson’s diseases. Nat Med 31(8):2567–2577. https://doi.org/10.1038/s41591-025-03833-1 Ali M, Timsina J, Western D, Liu M, Beric A, Budde J, Do A, Heo G, Wang L, Gentsch J, Schindler SE, Morris JC, Holtzman DM, Ruiz A, Alvarez I, Aguilar M, Pastor P, Rutledge J, Oh H, Cruchaga C (2025) Multi-cohort cerebrospinal fluid proteomics identifies robust molecular signatures across the Alzheimer disease continuum. Neuron 113(9):1363–1379e9. https://doi.org/10.1016/j.neuron.2025.02.014 Amft M, Ortner M, Eichenlaub U, Goldhardt O, Diehl-Schmid J, Hedderich DM, Yakushev I, Grimmer T (2022) The cerebrospinal fluid biomarker ratio Aβ42/40 identifies amyloid positron emission tomography positivity better than Aβ42 alone in a heterogeneous memory clinic cohort. Alzheimer’s Res Therapy 14(1):60. https://doi.org/10.1186/s13195-022-01003-w Barthélemy NR, Saef B, Li Y, Gordon BA, He Y, Horie K, Stomrud E, Salvadó G, Janelidze S, Sato C, Ovod V, Henson RL, Fagan AM, Benzinger TLS, Xiong C, Morris JC, Hansson O, Bateman RJ, Schindler SE (2023) CSF tau phosphorylation occupancies at T217 and T205 represent improved biomarkers of amyloid and tau pathology in Alzheimer’s disease. Nat Aging 3(4):391–401. https://doi.org/10.1038/s43587-023-00380-7 Boland B, Kumar A, Lee S, Platt FM, Wegiel J, Yu WH, Nixon RA (2008) Autophagy Induction and Autophagosome Clearance in Neurons: Relationship to Autophagic Pathology in Alzheimer’s Disease. J Neurosci 28(27):6926–6937. https://doi.org/10.1523/JNEUROSCI.0800-08.2008 Bycroft C, Freeman C, Petkova D, Band G, Elliott LT, Sharp K, Motyer A, Vukcevic D, Delaneau O, O’Connell J, Cortes A, Welsh S, Young A, Effingham M, McVean G, Leslie S, Allen N, Donnelly P, Marchini J (2018) The UK Biobank resource with deep phenotyping and genomic data. Nature 562(7726):203–209. https://doi.org/10.1038/s41586-018-0579-z Campion D, Pottier C, Nicolas G, Le Guennec K, Rovelet-Lecrux A (2016) Alzheimer disease: Modeling an Aβ-centered biological network. Mol Psychiatry 21(7):861–871. https://doi.org/10.1038/mp.2016.38 Dakterzada F, López-Ortega R, Arias A, Riba-Llena I, Ruiz-Julián M, Huerto R, Tahan N, Piñol-Ripoll G (2021) Assessment of the Concordance and Diagnostic Accuracy Between Elecsys and Lumipulse Fully Automated Platforms and Innotest. Front Aging Neurosci 13:604119. https://doi.org/10.3389/fnagi.2021.604119 Fan X, Cui L, Zeng Y, Song W, Gaur U, Yang M (2019) 14-3-3 Proteins Are on the Crossroads of Cancer, Aging, and Age-Related Neurodegenerative Disease. Int J Mol Sci 20(14):3518. https://doi.org/10.3390/ijms20143518 Findeis MA (2007) The role of amyloid β peptide 42 in Alzheimer’s disease. Pharmacol Ther 116(2):266–286. https://doi.org/10.1016/j.pharmthera.2007.06.006 Frake RA, Ricketts T, Menzies FM, Rubinsztein DC (2015) Autophagy and neurodegeneration. J Clin Investig 125(1):65–74. https://doi.org/10.1172/JCI73944 Gong B, Radulovic M, Figueiredo-Pereira ME, Cardozo C (2016) The Ubiquitin-Proteasome System: Potential Therapeutic Targets for Alzheimer’s Disease and Spinal Cord Injury. Front Mol Neurosci 9:4. https://doi.org/10.3389/fnmol.2016.00004 Haass C, Selkoe DJ (2007) Soluble protein oligomers in neurodegeneration: Lessons from the Alzheimer’s amyloid β-peptide. Nat Rev Mol Cell Biol 8(2):101–112. https://doi.org/10.1038/nrm2101 Horvath S (2013) DNA methylation age of human tissues and cell types. Genome Biol 14(10):R115. https://doi.org/10.1186/gb-2013-14-10-r115 Hussain MS, Agrawal N, Ilma B, Nayak MMR, Kaur PP, Khachi M, Goyal A, Rekha K, Gupta A, Gupta S, G., Dua K (2025) Autophagy and Cellular Senescence in Alzheimer’s Disease: Key Drivers of Neurodegeneration. CNS Neurosci Ther 31(7):e70503. https://doi.org/10.1111/cns.70503 Imam F, Saloner R, Vogel JW, Krish V, Abdel-Azim G, Ali M, An L, Anastasi F, Bennett D, Pichet Binette A, Boxer AL, Bringmann M, Burns JM, Cruchaga C, Dage JL, Farinas A, Ferrucci L, Finney CA, Frasier M, Lovestone S (2025) The Global Neurodegeneration Proteomics Consortium: Biomarker and drug target discovery for common neurodegenerative diseases and aging. Nat Med 1–11. https://doi.org/10.1038/s41591-025-03834-0 Jiang S, Li Y, Zhang X, Bu G, Xu H, Zhang Y (2014) Trafficking regulation of proteins in Alzheimer’s disease. Mol Neurodegeneration 9:6. https://doi.org/10.1186/1750-1326-9-6 Jiang Y, Zhou X, Ip FC, Chan P, Chen Y, Lai NCH, Cheung K, Lo RMN, Tong EPS, Wong BWY, Chan ALT, Mok VCT, Kwok TCY, Mok KY, Hardy J, Zetterberg H, Fu AKY, Ip NY (2022) Large-scale plasma proteomic profiling identifies a high-performance biomarker panel for Alzheimer’s disease screening and staging. Alzheimer’s Dement 18(1):88–102. https://doi.org/10.1002/alz.12369 Lee J-H, Yu WH, Kumar A, Lee S, Mohan PS, Peterhoff CM, Wolfe DM, Martinez-Vicente M, Massey AC, Sovak G, Uchiyama Y, Westaway D, Sisodia SS, Cuervo AM, Nixon RA (2010) Lysosomal Proteolysis and Autophagy Require Presenilin 1 and Are Disrupted by Alzheimer-Related PS1 Mutations. Cell 141(7):1146–1158. https://doi.org/10.1016/j.cell.2010.05.008 Lewczuk P, Łukaszewicz-Zając M, Mroczko P, Kornhuber J (2020) Clinical significance of fluid biomarkers in Alzheimer’s Disease. Pharmacol Rep 72(3):528–542. https://doi.org/10.1007/s43440-020-00107-0 Lian P, Guo Y, Yu J (2025) Biomarkers in Alzheimer’s disease: Emerging trends and clinical implications. Chin Med J 138(9):1009–1012. https://doi.org/10.1097/CM9.0000000000003579 Ma C, Hong F, Yang S (2022) Amyloidosis in Alzheimer’s Disease: Pathogeny, Etiology, and Related Therapeutic Directions. Molecules 27(4):1210. https://doi.org/10.3390/molecules27041210 Melendez J, Sung YJ, Orr M, Yoo A, Schindler S, Cruchaga C, Bateman R (2024) An interpretable machine learning-based cerebrospinal fluid proteomics clock for predicting age reveals novel insights into brain aging. Aging Cell 23(9):e14230. https://doi.org/10.1111/acel.14230 Nixon RA, Wegiel J, Kumar A, Yu WH, Peterhoff C, Cataldo A, Cuervo AM (2005) Extensive Involvement of Autophagy in Alzheimer Disease: An Immuno-Electron Microscopy Study. J Neuropathology Experimental Neurol 64(2):113–122. https://doi.org/10.1093/jnen/64.2.113 O’Brien RJ, Wong PC (2011) Amyloid Precursor Protein Processing and Alzheimer’s Disease. Annu Rev Neurosci 34:185–204. https://doi.org/10.1146/annurev-neuro-061010-113613 Oh HS-H, Rutledge J, Nachun D, Pálovics R, Abiose O, Moran-Losada P, Channappa D, Urey DY, Kim K, Sung YJ, Wang L, Timsina J, Western D, Liu M, Kohlfeld P, Budde J, Wilson EN, Guen Y, Maurer TM, Wyss-Coray T (2023) Organ aging signatures in the plasma proteome track health and disease. Nature 624(7990):164–172. https://doi.org/10.1038/s41586-023-06802-1 Patterson BW, Elbert DL, Mawuenyega KG, Kasten T, Ovod V, Ma S, Xiong C, Chott R, Yarasheski K, Sigurdson W, Zhang L, Goate A, Benzinger T, Morris JC, Holtzman D, Bateman RJ (2015) Age and amyloid effects on human central nervous system amyloid-beta kinetics. Ann Neurol 78(3):439–453. https://doi.org/10.1002/ana.24454 Pennington KL, Chan TY, Torres MP, Andersen JL (2018) The dynamic and stress-adaptive signaling hub of 14-3-3: Emerging mechanisms of regulation and context-dependent protein–protein interactions. Oncogene 37(42):5587–5604. https://doi.org/10.1038/s41388-018-0348-3 Qiang Q, Skudder-Hill L, Toyota T, Huang Z, Wei W, Adachi H (2023) CSF 14-3-3β is associated with progressive cognitive decline in Alzheimer’s disease. Brain Commun 5(6):fcad312. https://doi.org/10.1093/braincomms/fcad312 Qiu C, Kivipelto M, von Strauss E (2009) Epidemiology of Alzheimer’s disease: Occurrence, determinants, and strategies toward intervention. Dialog Clin Neurosci 11(2):111–128. https://doi.org/10.31887/DCNS.2009.11.2/cqiu Schindler SE, Bollinger JG, Ovod V, Mawuenyega KG, Li Y, Gordon BA, Holtzman DM, Morris JC, Benzinger TLS, Xiong C, Fagan AM, Bateman RJ (2019) High-precision plasma β-amyloid 42/40 predicts current and future brain amyloidosis. Neurology 93(17):e1647–e1659. https://doi.org/10.1212/WNL.0000000000008081 Siegenthaler B, Rajendran L (2012) Retromers in Alzheimer’s disease. Neurodegenerative Dis 10(1–4) Article 1–4. https://doi.org/10.1159/000335910 Wang R, Qi X, Schmiege P, Coutavas E, Li X (2020) Marked structural rearrangement of mannose 6-phosphate/IGF2 receptor at different pH environments. Sci Adv 6(7):eaaz1466. https://doi.org/10.1126/sciadv.aaz1466 Wang Y, Xiao S, Liu B, Jiang R, Liu Y, Hang Y, Chen L, Chen R, Vitiello MV, Bennett D, Wang B, Lv J, Yu C, Haslam DE, Zheng Q, Gerszten RE, Bao Y, Shi J, Xie J, Chan AT (2026) Organ-specific proteomic aging clocks predict disease and longevity across diverse populations. Nat Aging 6(1):162–180. https://doi.org/10.1038/s43587-025-01016-8 Xie Z, Klionsky DJ (2007) Autophagosome formation: Core machinery and adaptations. Nat Cell Biol 9(10):1102–1109. https://doi.org/10.1038/ncb1007-1102 Additional Declarations There is NO Competing Interest. Supplementary Files SupplementaryMaterials.xlsx Supplementary Materials Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8791371","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":594546258,"identity":"2ea9776e-8274-4225-aa25-6ec4441e96be","order_by":0,"name":"Justin Melendez","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDUlEQVRIiWNgGAWjYBACxgYgkQBisQMxDwODHFj4AVQGhxbGBrAWZogWY7BgAh4tDHA5qJbEBkJamGekP3/wcAdDND8z87EHbyrq0jfcAIokMNjIbjiAw4oZOYYNiWcYcmc2s6UbzjnDlrvhBlAkgSHNGI8WxobENobcDYd5zKR523hyt93IATnscCJuLekPwVr2H+b/Js37TyLd7AZQJIHhPx4tCYYQW5h52KR5GwwSzG4kgBx2ALeWnjeGMxLbJHJnHGYzk5xzLMFw/xmgSIJBsvFMHFoM29MffPzZZpPb3978TOJNTZ28JFDkw4cKO9k+XFoawJQEurgBduUgII9bahSMglEwCkYBFAAAD4Jju5zfNiQAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0003-3629-3274","institution":"Washington University in St Louis School of Medicine","correspondingAuthor":true,"prefix":"","firstName":"Justin","middleName":"","lastName":"Melendez","suffix":""},{"id":594546259,"identity":"def323b7-a783-47aa-8181-9bcdc3bc5ead","order_by":1,"name":"Leila Mozaffar","email":"","orcid":"","institution":"Washington University in St Louis School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Leila","middleName":"","lastName":"Mozaffar","suffix":""},{"id":594546260,"identity":"3fff9b9a-79ec-4b2b-8896-982ba3de23b5","order_by":2,"name":"Ziqiao Jiao","email":"","orcid":"","institution":"Washington University in St Louis School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Ziqiao","middleName":"","lastName":"Jiao","suffix":""},{"id":594546261,"identity":"0824157d-0237-4932-a137-2574d42d38ba","order_by":3,"name":"Kaleigh Roberts","email":"","orcid":"","institution":"Washington University in St Louis School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Kaleigh","middleName":"","lastName":"Roberts","suffix":""},{"id":594546262,"identity":"1c83f641-9eb2-46c7-a1c6-27be7f86a995","order_by":4,"name":"Carlos Cruchaga","email":"","orcid":"https://orcid.org/0000-0002-0276-2899","institution":"Washington University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Carlos","middleName":"","lastName":"Cruchaga","suffix":""},{"id":594546263,"identity":"443c047a-1790-4667-88c1-b4c3e3d4b6b3","order_by":5,"name":"Miranda (E) Orr","email":"","orcid":"https://orcid.org/0000-0002-0418-2724","institution":"Washington University at St. Louis","correspondingAuthor":false,"prefix":"","firstName":"Miranda","middleName":"(E)","lastName":"Orr","suffix":""},{"id":594546264,"identity":"0a371143-f94c-46a9-bf07-c26c50176239","order_by":6,"name":"Andrew Yoo","email":"","orcid":"https://orcid.org/0000-0002-0304-3247","institution":"Washington University at St Louis","correspondingAuthor":false,"prefix":"","firstName":"Andrew","middleName":"","lastName":"Yoo","suffix":""},{"id":594546265,"identity":"e0b17e8b-0e0a-4382-8017-d5d9d62c231f","order_by":7,"name":"Suzanne Schindler","email":"","orcid":"https://orcid.org/0000-0002-1680-1465","institution":"Washington University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Suzanne","middleName":"","lastName":"Schindler","suffix":""},{"id":594546266,"identity":"0462efbd-7c63-40e6-96d2-8cdacc5e2b4f","order_by":8,"name":"John Budde","email":"","orcid":"https://orcid.org/0000-0001-7778-7276","institution":"Washington University in St. Louis","correspondingAuthor":false,"prefix":"","firstName":"John","middleName":"","lastName":"Budde","suffix":""},{"id":594546267,"identity":"366ca078-e9e0-40fc-a2a6-013063e26632","order_by":9,"name":"Jigyasha Timsina","email":"","orcid":"https://orcid.org/0000-0003-4794-9187","institution":"Department of Psychiatry, Washington University School of Medicine, St. Louis, MO, USA","correspondingAuthor":false,"prefix":"","firstName":"Jigyasha","middleName":"","lastName":"Timsina","suffix":""},{"id":594546268,"identity":"a5eed1b0-8152-4903-8dba-08563d0e5b76","order_by":10,"name":"Menghan Liu","email":"","orcid":"https://orcid.org/0009-0008-9902-5983","institution":"Washington University in St. Louis","correspondingAuthor":false,"prefix":"","firstName":"Menghan","middleName":"","lastName":"Liu","suffix":""},{"id":594546269,"identity":"bd57baf6-e46b-49ed-9288-52ed73cd83f4","order_by":11,"name":"Randall Bateman","email":"","orcid":"https://orcid.org/0000-0002-7729-1702","institution":"Department of Neurology, Washington University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Randall","middleName":"","lastName":"Bateman","suffix":""}],"badges":[],"createdAt":"2026-02-05 02:36:46","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8791371/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8791371/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103505772,"identity":"4a5049ab-85b3-4cff-a146-d6b44acf1eac","added_by":"auto","created_at":"2026-02-26 13:32:58","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":396720,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eA graphical overview outlining the CSF Aß42/40 machine learning model and age associated pathway analysis.\u003c/strong\u003e \u003cstrong\u003e1.\u003c/strong\u003e CSF proteomics from the SomaLogic 7k platform was obtained from the Knight Alzheimer’s Disease Research Center (ADRC). This data contains measures of 7,008 protein analytes on the CSF of 856 participants. \u003cstrong\u003e2. \u003c/strong\u003eAn adaptive elastic net machine learning model was developed to predict CSF Aß42/40 from APOE E4 status, age and the proteomics data. Adaptive elastic net is a variant of elastic net we developed that makes use of our adaptive feature selection algorithm described in the methods. \u003cstrong\u003e3. \u003c/strong\u003eProteins used by our model to predict CSF Aß42/40, and those that highly correlated with them, were extracted and filtered for proteins that correlated with aging prior to amyloidosis. \u003cstrong\u003e4. \u003c/strong\u003ePathway enrichment was performed on age associated model proteins to gain insight into the age-related predictors of amyloidosis.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8791371/v1/929c4784139dabd2e6ee4616.png"},{"id":103505706,"identity":"4636ffac-3dfa-434d-813f-ee71ae2b5ef4","added_by":"auto","created_at":"2026-02-26 13:32:43","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":103065,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eProteins features needed to predict CSF Aß42/40 are reduced while model performance improves by applying our adaptive feature selection algorithm to the elastic net model.\u003c/strong\u003e We applied a feature selection algorithm developed in our lab which we’ve dubbed “adaptive feature selection” to the elastic net model, which converges on a set of features while improving predictive model performance of CSF Aß42/40. At round 0, we start with 7,008 features presented to the model with no training. Round 1 represents a classic elastic net which used 1,777 of the features and achieved a median Pearson correlation of 0.8 between CSF Aß42/40 predicted by the model and CSF Aß42/40 measured via Lumipulse in our internal test cohort over 100 different train-test splits of the data. Our adaptive feature selection method was then applied for several rounds until median performance over the 100 train test splits leveled off at 0.88 correlation with measured data using 93 features, at which point no subsequent rounds would improve performance.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8791371/v1/c63d5f2772bbcd8bf574b292.png"},{"id":103506207,"identity":"b96e9b72-b56f-4131-b4af-02924cab0c48","added_by":"auto","created_at":"2026-02-26 13:34:35","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":264156,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eA modified elastic net-based machine learning algorithm estimates CSF Aß42/40. \u003c/strong\u003eA modified elastic net model was trained to predict CSF Aß42/40 from SomaLogic 7K proteomics measurements of 7,008 protein features. Training was done on 846 participants from the Knight ADRC over a wide range of ages and amyloidosis states. Adaptive feature selection, a process developed in our lab was employed to automatically reduce features while improving performance during training. \u003cstrong\u003eA, B)\u003c/strong\u003e After converging on a set of features, cross-testing was performed on 100 different splits of the training data. Histograms of model performance based on Pearson correlation of model predicted vs measured CSF Aß42/40 (A) and mean absolute error of model predicted vs measured CSF Aß42/40 are shown (B). Highlighted yellow bars denote histogram segments containing models of median performance when provided with the selected set of features. Final weights and features were selected from the closest median performing model based on Pearson correlation and MAE. \u003cstrong\u003eC)\u003c/strong\u003eThis selected median model made use of 93 proteins as well as chronological age and the presence or absence of the APOE E4 allele to make predictions. The model was able to predict CSF Aß42/40 ratio with a 0.88 correlation to Lumipulse measured CSF Aß42/40 ratio, the standard measurement used by the Knight ADRC, and a mean absolute error (MAE) of 0.009 from measured ratio values on Knight ADRC testing data. Blue dotted lines display the positivity cutoff used by the Knight ADRC for determining amyloidosis positivity via CSF Aß42/40 ratio (below 0.0673). Using this cutoff the model classified amyloidosis positivity with a 94% accuracy. \u003cstrong\u003eD) \u003c/strong\u003eThis model was then validated on an entirely external data set from ADNI (not seen during training). The validation data consisted of 48 participants including 20 positive amyloidosis and 28 amyloidosis negative individuals. On this validation data, our model achieved an 88% classification accuracy, Pearson correlation of 0.86 between predicted CSF Aß42/40 ratio and Elecsys measured CSF Aß42/40 ratio and mean estimated error of 0.013 from measured values.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8791371/v1/4c0c75ab2d4018e218cab884.png"},{"id":103285211,"identity":"66876175-cee3-4ba4-858f-e249dd62283c","added_by":"auto","created_at":"2026-02-24 04:31:43","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":226938,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAge-Correlated Proteins Extracted from the CSF Aß42/40 Proteomics Model.\u003c/strong\u003e \u003cstrong\u003eA)\u003c/strong\u003eA Sankey diagram categorizing proteins used to predict CSF Aß42/40 extracted from our machine learning model. Out of 7,008 possible proteins the model selected a minimum set of 93 to calculate CSF Aß42/40 for participants. Since the modified elastic net model was designed to be highly aggressive at eliminating informatively redundant proteins, we restored any proteins highly correlated to model derived proteins (R \u0026gt;= 0.7) giving us a total set of 821 protein features. Of these proteins, 308 (38%) were significantly correlated with age (P\u0026lt;0.0001) in the Knight ADRC proteomics cohort in either the amyloidosis negative (N=480) or amyloidosis positive (N=366) groups\u003cstrong\u003e. B) \u003c/strong\u003eA Venn diagram depicting the overlap of age correlated proteins in non-amyloidosis participants vs amyloidosis participants. All but two proteins that correlated with age in the amyloidosis group also correlated with age in non-amyloidosis participants. \u003cstrong\u003eC)\u003c/strong\u003e A volcano plot of non-amyloidosis age correlated proteins derived from model proteins and proteins highly correlated to model proteins. The x-axis is the degree of correlation to aging and direction. Positively correlated proteins go up with age while negatively correlated proteins go down with age. The y-axis is the significance of the correlation (p-value) on a –log10 scale. Each point represents a protein and top negatively correlated and positively correlated proteins are labeled. Proteins are color coded by model weight with blue proteins having negative model weights and red proteins having positive model weights. Negative model weights indicate that increasing amounts of the protein was predicative of lower CSF Aß42/40 scores (pro-amyloidosis) whereas positive model weights indicate increasing amounts of the protein predicted higher CSF Aß42/40 scores (less amyloidosis).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8791371/v1/866832b35b57d84e99aee50d.png"},{"id":103285207,"identity":"16d0a5b5-45b0-40d3-9944-8b2fe806f35b","added_by":"auto","created_at":"2026-02-24 04:31:43","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":361029,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePathway enrichment analysis of age associated proteins used in our model or highly correlated with model proteins. \u003c/strong\u003ePathway enrichment analysis using the Metascape tool (metascape.org) was performed on model\u003cstrong\u003e \u003c/strong\u003eproteins, and those proteins highly correlated with them (R\u0026lt;=0.7, P\u0026lt;=0.0001), that were also correlated with age in non-amyloidosis participants (P\u0026lt;0.0001, FDR 0.027%). Proteins measured by all 7,008 possible aptamer analytes (SomaLogic 7k) were used as the background and pathway enrichment analysis was performed on the Reactome knowledgebase (reactome.org). \u003cstrong\u003eA.\u003c/strong\u003e Pathways are listed in order of statistical significance on a -Log10(P-value) scale. Log10 q-values (adjusted p-value after accounting for false discovery rate) for each pathway can be found in the supplementary materials, table 3. The color represents the mean correlation (Pearson correlation) with age in non-amyloidosis participants of proteins enriched in each pathway. B. Proteins enriched in pathway analysis are plotted by their correlations with age in non-amyloidosis individuals vs their model weights in the CSF Aß42/40 model. Negative model weights mean increased amount of protein predict worsening amyloidosis, whereas positive model weights mean increasing amount of protein predicts less amyloidosis. Model weights of proteins that were not directly in the model, but instead highly correlated to model proteins (as previously described), were inferred from the model proteins they correlated with. Proteins from each pathway are denoted by color and shape.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8791371/v1/acfb84d1d6e0d33354f288d6.png"},{"id":103509625,"identity":"e786cb2c-ff0d-4c2a-9ed2-deea42a9acdd","added_by":"auto","created_at":"2026-02-26 14:00:07","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2297794,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8791371/v1/8160db4a-b685-4490-b61b-1644dad3322a.pdf"},{"id":103285210,"identity":"6d4ce535-e8b4-4aa5-9daf-1c5963e86b01","added_by":"auto","created_at":"2026-02-24 04:31:43","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1828752,"visible":true,"origin":"","legend":"Supplementary Materials","description":"","filename":"SupplementaryMaterials.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8791371/v1/2baec2a838064feb911741dd.xlsx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Identifying Age-Related Protein Mechanisms of Alzheimer's Disease Amyloidosis from Cerebrospinal Fluid Proteomics Using a Novel Machine Learning Approach","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAlzheimer\u0026rsquo;s disease (AD) is a devastating and irreversible neurodegenerative disorder that slowly destroys memory, thinking skills, and, eventually, the ability to carry out daily tasks. Its impact extends far beyond the individual, placing an immense emotional, physical, and financial burden on family members and caregivers. Aging is the single biggest risk factor of sporadic late onset Alzheimer\u0026rsquo;s disease (LOAD) with the incidence of AD doubling every five years after the age of 65 (Qiu et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). With a rapidly growing older population, AD is an increasing global health concern. Estimates project that the number of Americans aged 65 and older living with AD could grow from approximately 7\u0026nbsp;million today to nearly 14\u0026nbsp;million by 2060, underscoring the urgent need for a deeper understanding of the molecular mechanisms by which aging drives the onset of AD (\u0026ldquo;2025 Alzheimer\u0026rsquo;s Disease Facts and Figures,\u0026rdquo; 2025).\u003c/p\u003e \u003cp\u003eA critical, early pathological event in AD is the aggregation of amyloid-beta (A\u0026szlig;) protein in amyloid plaques (Ma et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). A\u0026szlig; is generated through the sequential cleavage of the Amyloid Precursor Protein (APP), a transmembrane protein found in neuronal and non-neuronal cells. APP can be processed through two competing pathways: 1) The non-amyloidogenic pathway occurs primarily at the cell membrane surface. Here APP is first cleaved by alpha-secretase within the A\u0026szlig; domain, preventing A\u0026szlig; formation. 2) The amyloidogenic pathway is activated when APP is internalized in a clathrin-dependant manner and trafficked to endosomal compartments. Here, APP is first cleaved by beta-secretase (BACE1), which releases a soluble ectodomain and leaves a C99 membrane-bound fragment. This fragment is then cleaved by the gamma-secretase complex to release the A\u0026szlig; peptide (Campion et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; O\u0026rsquo;Brien \u0026amp; Wong, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). However, gamma-secretase cleavage is imprecise, generating A\u0026szlig; peptides of varying lengths, including A\u0026szlig;40 (40 amino acids) and A\u0026szlig;42 (42 amino acids). The A\u0026szlig;42 isoform, which contains two extra hydrophobic amino acids, is significantly more prone to aggregation and plaque formation than the shorter, more abundant A\u0026szlig;40 isoform (Findeis, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2007\u003c/span\u003e; Haass \u0026amp; Selkoe, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2007\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eUnderstanding the specific age-related biological factors that drive this initial amyloid-beta (A\u0026szlig;) accumulation may identify novel therapeutic targets capable of delaying or preventing the onset of AD. The ratio of A\u0026szlig; 42 to A\u0026szlig; 40 (A\u0026szlig;42/40) in the cerebrospinal fluid (CSF) is a widely used and important clinical biomarker that reflects the onset of amyloidosis in the brain (Lewczuk et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Schindler et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). A decline in the CSF A\u0026szlig;42/40 ratio indicates decreased clearance or aggregation of the more aggregation-prone A\u0026szlig;42 isoform and is an early indicator of AD pathology (Amft et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Lian et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Previous studies have indicated that the ability to clear central nervous system (CNS) A\u0026szlig;42 and other amyloid-beta species from the brain declines by 400% with aging from the 30s through the 80s, and further A\u0026szlig;42 clearance is specifically altered with amyloidosis, a process strongly correlated with disease progression (Patterson et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eMachine learning is a powerful tool for navigating the high dimensionality and complexity of large biological data sets, such as proteomics, to reveal insightful patterns often imperceptible through traditional statistical methods. This technology has been successfully leveraged to create numerous aging clocks, from large scale omics data, that have increased our understanding of how aging impacts disease (Horvath, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Melendez et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Oh et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Wang et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2026\u003c/span\u003e). The past decade has seen an explosion of large proteomic data sets generated by high-throughput platforms like SomaLogic, Olink, and Alamar (Bycroft et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Imam et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Large-scale studies utilizing the SomaScan platform have identified thousands of proteins associated with AD and other neurodegenerative diseases, distinguishing shared versus disease-specific molecular pathways (Ali, Erabadda, et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Work using Olink technology has successfully employed multiplex panels to identify high-performance, multi-protein signatures capable of classifying AD from controls with high accuracy (Jiang et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eOur primary goal was to gain biological insights into the age-associated molecular mechanisms underlying amyloidosis. To achieve this, we employed an adaptive elastic net machine learning model with a novel feature selection strategy to predict CSF A\u0026szlig;42/40, a key clinical biomarker of amyloidosis, solely from CSF-derived proteomics (SomaLogic 7k platform). By extracting the features from this model, we were able to better understand the biological basis of the predictions. Furthermore, we hypothesized that the \u003cem\u003eAPOE ε4\u003c/em\u003e carrier status, the most significant genetic risk factor for amyloidosis, may modify the progression of A\u0026szlig; accumulation. Therefore, we included \u003cem\u003eAPOE\u003c/em\u003e ε4 carrier status as a feature in the model. The final model, which integrates chronological age, \u003cem\u003eAPOE\u003c/em\u003e ε4 status, and proteomics data, accurately estimates CSF A\u0026szlig;42/40 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFinally, by extracting the protein features used by the algorithm, we gained biological insights into the age-associated factors underlying amyloidosis. The proteins identified by the model as most crucial for predicting amyloidosis represent key candidates for further study. We specifically examined the subset of these proteins that were correlated with aging and then performed pathway analysis on this age-associated set (and their highly correlated partners) to identify the age-related biological pathways that may be driving the decline in A\u0026szlig;42 clearance with age and the onset of amyloidosis.\u003c/p\u003e \u003cp\u003eOur study illustrates the potential of machine learning not only in biomarker approximation but also in extracting predictive features to reveal biological insights. The identified pathways and proteins in this study are potential targets for intervention in research studies to test causality of these pathways. This study also demonstrates approaches identifying potential predictive proteins and pathways across multiple diseases using standardized proteomic data sets.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eWe developed a machine learning model to analyze a large proteomic dataset for insights into age associated pathways and proteins predictive of amyloidosis. Our goal was to train models predicting a key AD biomarker and then interrogate the models to understand how their predictions were made. We hypothesized that age associated pathways enriched in model proteins would reveal key markers of the disease. As amyloid plaque formation is one of the earliest stages thought to lead to AD, our model was designed to predict the CSF A\u0026szlig;42/40 ratio, a key biomarker for amyloidosis in Alzheimer's disease pathology. Finally, we extracted protein features from this model and performed pathway analysis to analyze age associated proteomic signatures predictive of amyloidosis. The results from our model, along with our feature selection and age associated protein pathway analyses, are presented below.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eMachine learning to predict CSF A\u0026szlig;42/40\u003c/h2\u003e \u003cp\u003eTo investigate age-associated molecular factors linked to amyloidosis, we developed a machine learning approach to predict the CSF A\u0026szlig;42/40 ratio directly from proteomic data. To do this we trained an elastic net machine learning model on 7,008 protein features in CSF from the SomaLogic 7k assay from 856 participants from the Knight Alzheimer\u0026rsquo;s Disease Research Center (ADRC) to predict Lumipulse measured CSF A\u0026szlig;42/40 ratio, a widely used clinical biomarker used to identify amyloidosis in Alzheimer's disease patients. We then used an algorithm developed in our lab that we call \u0026ldquo;adaptive feature selection\u0026rdquo; to simultaneously improve model performance while minimizing protein features used (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). This created a highly predictive model using a minimal number of protein features (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eModel performance on the Knight ADRC cohort, used to train the models, was assessed by testing on 100 different train test splits made up of 80% training and 20% testing data held aside for each model. The median performing model had a 0.88 correlation (P\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) to Lumipulse measured CSF A\u0026szlig;42/40 (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA) and a mean absolute error (MAE) from the participants measured CSF A\u0026szlig;42/40 of 0.009 (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). The model used just 93 protein features along with the participants age and presence or absence of the \u003cem\u003eAPOE ε4\u003c/em\u003e allele to make this prediction (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC) (Model weights are available in Supplementary Materials Table\u0026nbsp;1). In addition to its predictive performance, we also evaluated the model's utility as a classification tool. This was accomplished by assessing its ability to distinguish between individuals with and without amyloidosis. To do this we used the established CSF A\u0026szlig;42/40 ratio cutoff from the Knight ADRC cohort, where a ratio below 0.0673 indicates amyloidosis positivity and a ratio above this threshold indicates a negative status (Barth\u0026eacute;lemy et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). When tested against this criterion, the median-performing model demonstrated robust performance, achieving a high accuracy of 94% in classifying amyloidosis positivity versus negativity.\u003c/p\u003e \u003cp\u003eNext, we evaluated the pre-trained median Knight ADRC model on an external proteomics data set from ADNI which was locked and not visible to the model during training. The ADNI validation cohort used the Elecsys platform to measure CSF A\u0026szlig;42/40 in contrast to the Knight ADRC training data which used Lumipulse, however measurements form both platforms have been reported to be highly concordant with each other (Dakterzada et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) and interchangeable. We found no harmonization was needed for our Lumipulse trained model to predict Elecsys measured A\u0026szlig;42/40 ratios. Here the model had a Pearson correlation of 0.86 between predicted CSF A\u0026szlig;42/40 ratio and Elecsys measured CSF A\u0026szlig;42/40 ratio in the ADNI validation cohort (48 participants) and mean estimated error of 0.013 from measured values. Additionally, the model was 88% accurate in classifying amyloidosis positive vs negative participants in this validation data (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eAge association of proteins used in the model\u003c/h3\u003e\n\u003cp\u003eAging is a primary risk factor for Alzheimer's disease (AD) and amyloidosis, with the incidence of AD doubling every five years after the age of 65 (Qiu et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). The body's ability to clear amyloid-beta from the brain declines significantly with age, a process that is strongly correlated with the progression of AD. This age-related decline in clearance is a critical factor in the accumulation of amyloid plaques, a hallmark of AD pathology. As such we sought to examine what proteins in our machine learning model predicting CSF A\u0026szlig;42/40 had strong associations with age in the Knight ADRC cohort. We hypothesize that these proteins and associated pathways may be drivers of declining amyloid beta clearance capacity or compensatory mechanisms to combat the negative effects of this decline.\u003c/p\u003e \u003cp\u003eOur model made use of 93 proteins, in addition to chronological age and APOE E4 presence or absence, to predict CSF A\u0026szlig;42/40. As the model was designed to aggressively drop predictively redundant proteins, we looked for proteins that were highly correlated with these model proteins prior to performing pathway analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). Of the 93 model selected proteins, 728 proteins strongly correlated with them (Pearson correlation\u0026thinsp;\u0026gt;\u0026thinsp;=\u0026thinsp;0.7) (Supplementary Materials Table\u0026nbsp;2). Of these 821 proteins 37.5% (308) were correlated with age at a p-value of 0.0001 or lower in the Knight ADRC cohort giving us an estimated false discovery rate (FDR) of 0.027%. We considered a protein age associated if it was significantly (P\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) correlated with chronological age in either pre-amyloidosis (CSF A\u0026szlig;42/40\u0026thinsp;\u0026gt;\u0026thinsp;0.0673) or post-amyloidosis participants (CSF A\u0026szlig;42/40\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0.0673).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003ePathway analysis of age associated amyloidosis predictive proteins:\u003c/h3\u003e\n\u003cp\u003eWe next sought to determine which pathways enriched for age associated proteins in our CSF A\u0026szlig;42/40 predictive model. As the set of pre-amyloidosis aging proteins captured nearly all post-amyloidosis aging proteins, and our primary interest was in finding potential age-related drivers of amyloidosis, we focused on the 306 pre-amyloidosis age-related proteins for pathway analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB, C). Pathway analysis was performed using the bioinformatic tool Metascape (metascape.org) on the Reactome pathway knowledgebase using proteins measured in the Somalogic 7k assay as a background (reactome.org) (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). Our analysis showed signals of dysregulated protein quality control, increased DNA damage repair, increases in cellular stress responses, and dysregulated cell cycle proteins, increased with age and amyloidosis. Top pathways that were enriched included Membrane Trafficking (R-HSA-199997), Autophagy (R-HSA-9612973) and Signaling by Rho GTPases (R-HSA-194315), Formation of Incision Complex in GG-NER (R-HSA-5696395), Attenuation phase (R-HSA-3371568) and Deregulated CDK5 triggers multiple neurodegenerative pathways in Alzheimer's disease models (R-HSA-8862803). Fold change enrichment for each pathway, Reactome pathway identifiers, log10 q-values (p-value adjusted for false discovery rate for each enriched pathway), proteins enriched in each pathway, protein model weights and protein correlations with age in non-amyloidosis participants can be found in Supplementary Materials Table\u0026nbsp;3.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe onset and progression of Alzheimer\u0026rsquo;s disease is inextricably linked to aging, with amyloidosis representing one of the earliest detectable pathological stages. Understanding the age-related biological factors that drive the accumulation of A\u0026szlig; is crucial for identifying therapeutic targets. In this study, we leveraged the power of machine learning to process large-scale, high-dimensional proteomic data to approximate a clinical biomarker of amyloidosis, CSF A\u0026szlig;42/40 ratio, and extract novel biological insights related to aging and amyloidosis.\u003c/p\u003e\n\u003ch3\u003eMachine Learning Robustly Approximates Clinical Biomarkers from Proteomics\u003c/h3\u003e\n\u003cp\u003eA key challenge in utilizing newly available, large proteomic datasets (such as those generated by SomaLogic, Olink, and Alamar) is the frequent lack of accompanying gold-standard biomarkers (e.g. CSF A\u0026szlig;42/40) necessary for comprehensive disease analysis. To address this issue, we demonstrated that machine learning models could accurately predict the value of a critical biomarker of Alzheimer's Disease (AD) research, CSF A\u0026szlig;42/40, using CSF proteomics data. Our primary goal was to gain insights into aging and amyloidosis progression. Therefore, we needed a model capable of predicting a wide range of amyloidosis states, from non-amyloidosis to amyloidosis positivity.\u003c/p\u003e \u003cp\u003eInstead of creating a classifier that determined whether a participant was amyloidosis positive or negative, we chose to predict the actual CSF A\u0026szlig;42/40 biomarker ratio. We hypothesized that a continuous model that could estimate the ratio of these A\u0026szlig; peptides would better capture the progression from non-amyloidosis to amyloidosis. Currently the only way to measure this biomarker ratio accurately is via mass spectrometry or specialized antibody assays such as Lumipulse or Elecsys. This continuous approach would provide more nuanced insights into the biology, compared to classifiers that rely on binary differentiation of amyloidosis positivity versus negativity. We think that the largest differentiators identified by binary classifiers were more likely to represent disease endpoints and miss proteins and pathways important in transitional states. Thus, our aim was to build a machine learning model that was as accurate as possible at estimating an amyloidosis biomarker, the idea being that the better performing models would be more likely to identify proteins and pathways relevant to the underlying biology during the feature extraction stage.\u003c/p\u003e \u003cp\u003eTo accomplish this, we employed our modified elastic net model, utilizing our novel \"adaptive feature selection\" approach to reduce features and improve model accuracy. This method successfully estimated CSF A\u0026szlig;42/40, a clinical measure reflecting the onset of amyloidosis. Since CSF A\u0026szlig;42/40 is clinically used to classify amyloidosis positivity vs negativity, though not our primary goal, we also reported how well our model estimated CSF A\u0026szlig;42/40 compared to measured CSF A\u0026szlig;42/40. The model achieved a high correlation of R\u0026thinsp;=\u0026thinsp;0.88 and 94% classification accuracy on the internal Knight ADRC testing cohort. Crucially, this predictive capability generalized well to the unseen external ADNI validation cohort, maintaining a strong correlation of R\u0026thinsp;=\u0026thinsp;0.86 and 88% classification accuracy, despite the ADNI cohort using a different but comparable assay platform (Elecsys vs. Lumipulse). This robust cross-cohort validation provides compelling evidence that the protein signatures identified by the model captures fundamental, cohort-independent proteomics changes predictive of A\u0026szlig; accumulation, rather than cohort-specific overfitting.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eAge-Associated Pathways Driving Amyloidosis: Failure to Complete Autophagy May Lead to Increased Amyloidosis and Neurodegeneration\u003c/h2\u003e \u003cp\u003eThe Autophagy pathway (R-HSA-9612973) showed the highest level of significance for enriched pathways with a log10 q-value of \u0026minus;\u0026thinsp;3.4 and enrichment fold change over background of 4.8x. Autophagy contributes to cellular homeostasis by degrading damaged organelles and misfolded proteins. A wide range of proteins involved in autophagosome formation and function, including RNASE1, UBB, UBE2N, UBE2V1, GABARAP, GABARAPL2, GABARAPL1, MAP1LC3B, MAP1LC3A, MVB12B, and DYNLL2, were enriched in this pathway (Frake et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Gong et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Xie \u0026amp; Klionsky, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2007\u003c/span\u003e). All model proteins enriched in the pathway had negative model weights indicating that increased amounts of the proteins was associated with amyloidosis. Additionally, these proteins all had positive correlations with chronological age. Overall, we observe that the presence of proteins associated with the initiation of autophagy generally increases with age and predicts a worsening disease state. Initial assessment might suggest this appears contradictory, as an increase in autophagic proteins should result in improved clearance of amyloid beta. However, the literature shows that there is an increase in the number of immature autophagic vacuoles in the neurons of people with AD due to defective autophagosome-lysosome fusion (Boland et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Hussain et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Nixon et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). This aligns with our interpretation that the age-related increases in autophagic proteins detected in the CSF may reflect a pathological compensatory response. We hypothesize that age-related increase in these autophagic proteins in the CSF may be a proxy for a pathological compensatory response within cells attempting to initiate and ramp up the autophagic process in response to the mounting burden of misfolded proteins and A\u0026szlig;. However, failure of these autophagic vacuoles to fuse with lysosomes creates a bottleneck for the cellular machinery leading to a buildup of autophagic proteins. It is possible this lack of lysosomal fusion is a driver of amyloid beta clearance deficits leading to increased amyloid plaques with age, however more work will need to be done to see if this is a result or cause of amyloidosis and if what we observe in the CSF truly reflects intracellular and tissue changes. If autophagic impairment is a causal mechanism, an exciting avenue of research may be to correct this fusion failure phenotype and determine if amyloid beta clearance improves, or to target rejuvenation of the lysosomes directly. Interestingly, mutations in Presenilin 1 (PS1), the protein that forms the catalytic core of the gamma-secretase enzyme and the most commonly mutated protein in autosomal dominant Alzheimer's disease (ADAD), have also been implicated in a mechanism independent of A\u0026szlig; production, causing the failure of autophagosomes to fuse with and be degraded by the lysosomes due to impaired lysosomal acidification (Lee et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2010\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTied with the Autophagy pathway, Membrane Trafficking (R-HSA-199997) was also highly enriched with a score log10 q-value of -3.4 and enrichment fold change over background of 2.6x. Like the Autophagy pathway, most membrane trafficking proteins had negative model weights indicating increased abundance was predictive of worsening pathology, and all model proteins enriched in the pathway were positively correlated with age. Proteins from the 14-3-3 family (YWHAZ, YWHAB, YWHAE, YWHAG, YWHAH) were highly enriched in both the Membrane Trafficking pathway and Signaling by Rho GTPases pathway (R-HSA-194315), the third most highly enriched pathway with a log10 q-value of -3.1 and enrichment fold change over background of 2.5x. These proteins are well-established CSF biomarkers in the Alzheimer's disease continuum, have been identified as being significantly dysregulated in other multi-cohort proteomics studies (Ali, Timsina, et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Qiang et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) and are recognized as central players at the intersection of aging and neurodegenerative disease (Fan et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). 14-3-3 proteins are among the most abundant in the brain, functioning as central hub proteins that regulate signal transduction, protein transport, and cytoskeletal dynamics by binding to phosphorylated target proteins (Abdi et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Pennington et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) and are known to have a protective, chaperone-like role, suppressing the aggregation of tau and other proteins and regulating A\u0026szlig; clearance via the lysosomal pathway (Abdi et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). This fits with a picture of the brain mounting a compensatory response to failed amyloid beta clearance in the brain.\u003c/p\u003e \u003cp\u003ePaired with the observation of a failing autophagy system, we hypothesis that this increase in membrane trafficking proteins coupled with worsening amyloidosis could be a symptom of the same lysosomal bottleneck. Normally, APP is recycled from the membrane by internalization into early endosomes along with BACE1 and γ-secretase. This change in pH in the early endosome is conducive of amyloidogenic processing of APP by γ-secretase after which it is either transported to the lysosome for degradation (by fusion of the endosome to a lysosome) or released from the cell. Alternatively, APP is recycled to the membrane via retromers and the trans-Golgi network (Jiang et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Siegenthaler \u0026amp; Rajendran, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). We hypothesize that the longer residence time of APP in early endosomes, and the pro-amylogenic environment of those vesicles, leads to APP being more likely to be cleaved into A\u0026szlig; instead of being transported to the Golgi and recycled to the membrane. This could explain how failure of early endosomes to complete autophagy would lead to reduced clearance of A\u0026szlig; leading to upregulation of both membrane trafficking and autophagy machinery in an attempt to overcome this backup of endosome processing.\u003c/p\u003e \u003cp\u003eAnother key protein in the membrane trafficking pathway is IGF2R (insulin-like growth factor 2 receptor), which has a negative model weight and a positive age correlation. This protein is crucial for transporting lysosomal enzymes from the Golgi to the lysosome (Wang et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Since A\u0026szlig; is predominantly produced in endosomal and lysosomal compartments, proper trafficking of both APP and the enzymes that cleave it is essential for regulating A\u0026szlig; production and degradation. The dysregulation of a key trafficking protein like IGF2R could compromise the delivery of critical degradative enzymes to the lysosome. This would, in turn, impair the efficiency of the autophagy-lysosome pathway, leading to the accumulation of immature autophagosomes leading to the bottleneck discussed earlier, and reduced cellular clearance of A\u0026szlig;.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eOur findings support the use of machine learning as a powerful tool for biomarker approximation and biological discovery in the era of high throughput 'omics data. Here we have demonstrated that CSF A\u0026szlig;42/40 can be accurately inferred from standard SomaLogic 7k proteomics. This capability offers a path toward a single, integrated biofluid test where routine proteomic screens could simultaneously diagnose, monitor, or test for multiple diseases by simply changing the machine learning model applied to the output.\u003c/p\u003e \u003cp\u003eMoving forward, the specific proteins and pathways identified by the model, particularly those highly correlated with age, represent high-priority candidates for targeted functional studies. Future work should focus on validating the causal relationship between the identified age-related proteins and the mechanisms of A\u0026szlig; clearance and production. Specifically, the enrichment of autophagy and membrane trafficking proteins suggests a key pathological bottleneck in the endosomal-lysosomal pathway, characterized by a hypothesized age-related failure of autophagosome-lysosome fusion. This fusion failure may lead to both impaired A\u0026szlig; degradation and a prolonged residence time for APP in pro-amyloidogenic early endosomes, thereby increasing A\u0026szlig; production and driving pathology. Much work needs to be done to vet and validate this hypothesis, however Interventions targeting this fusion failure phenotype as a possible age-related link is a particularly exciting avenue for research. Such work could illuminate the molecular mechanisms by which aging increases AD risk, potentially leading to the development of novel therapeutic interventions aimed at restoring cellular quality control and mitigating age-related stress.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eCohort description\u003c/h2\u003e \u003cp\u003eCSF proteomics data was obtained for this study via a data request to the Knight Alzheimer Disease Research Center (ADRC) and the Alzheimer's Disease Neuroimaging Initiative (ADNI). The Knight ADRC cohort consisted of 856 participants while the ADNI cohort consisted of 48 participants. Participants in the Knight ADRC were aged 43\u0026ndash;99 years old and ranged from cognitively unimpaired (CN, CDR\u0026thinsp;=\u0026thinsp;0) to those with dementia (CDR\u0026thinsp;\u0026gt;\u0026thinsp;=\u0026thinsp;1) and covered a wide range of amyloidosis states including those of negative, positive and intermediate status as measured by the CSF A\u0026szlig;42/40 biomarker. Participants in ADNI ranged from ages 60 to 86 and likewise included both CN and CDR\u0026thinsp;\u0026gt;\u0026thinsp;=\u0026thinsp;1 participants over a similar range of amyloidosis states to the Knight ADRC participants. Participants from the ADRC and ADNI cohorts were selected based on availability of proteomic data and amyloid-β (Aβ) 42/40 ratio measurements obtained within 1 year of CSF collection. CSF 42/40 measurements were quantified using the Lumipulse platform for the Knight ADRC cohort, while they were quantified using the Elecsys immunoassay platform in the ADNI cohort. Proteomic profiling was performed using the SomaLogic SOMAscan 7k assay in both cases. \u003cem\u003eAPOE ε4\u003c/em\u003e status information was available for all participants through genetic testing.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eAdaptive feature selection algorithm\u003c/h2\u003e \u003cp\u003eAdaptive feature selection is an algorithm we developed that reduces the number of features used in a machine learning model while simultaneously improving predictive performance. We developed this algorithm to solve the frequent problem in biological data sets of having large numbers of features with relatively small sample sizes. It can be applied to any machine learning model where feature impact can be measured. Here we will describe how we applied the algorithm to our elastic net machine learning model of CSF A\u0026szlig;42/40.\u003c/p\u003e \u003cp\u003eAn elastic net machine learning model was trained on 100 balanced testing and training splits of the CSF proteomics SomaLogic 7k data from the Knight ADRC. This generated 100 differently weighted elastic net models, each predicting the biomarker of interest. In each model the trained elastic net assigned a feature weight to each of the 7,008 proteins. Protein features with a weight of zero were not useful to the model, whereas features with positive weights were considered informative. For each of the 100 models we created a binary array of whether the feature was used (1 for informative, 0 for not used). For each feature we then asked what percentage of the time it was used across all models (Sum of times informative/100 models\u0026thinsp;=\u0026thinsp;percent used). From here we created feature sets based on thresholds of the percentage of features that were used. For instance, if a feature was used in 10% or more of the models it was included in the 10% feature set for that round. If a feature was used in 90% or more models, it was included in the 90% feature set. For simplicity and computational performance feature usage thresholds were set to 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70% 80%, 90%, and 100% or more usage across models. For each feature set, all other features were removed, and the elastic net was run on those features for 100 new train-test splits of the data. From here the median performing model of each feature set was measured via Person correlation of the model\u0026rsquo;s prediction vs real world measurements of the biomarker. The feature set with the best performing median model was chosen to move on to the next round. If the median model\u0026rsquo;s performance was better than the prior best median model, the entire adaptive feature selection process was repeated using that model\u0026rsquo;s features as the starting point. This progressively produced models with fewer features each round and improved performance until the best performing feature set selected no longer outperformed the previous round, at which point the features and weights of the median model would be set and the final, trained model produced.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eCSF A\u0026szlig;42/40 machine learning model\u003c/h2\u003e \u003cp\u003eCustom python code was written to train our elastic net-based machine learning algorithm to predict CSF A\u0026szlig;42/40 based on 7008 protein features measured in the SomaLogic 7k data set in human CSF along with chronological age and \u003cem\u003eAPOE ε4\u003c/em\u003e presence or absence. The elastic net algorithm was implemented via the scikit-learn machine learning library. Missing data points were imputed using a k nearest neighbor-based approach with \u003cem\u003eK\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2. Prior to training the data was scaled from 0 to 1 using the min-max feature scaler of scikit-learn, which gave the best performance of the data transformation approaches tried. For a single round of elastic-net training the data were randomly split into two class balanced sets: a training set consisting of 80% of the data and a test set consisting of 20% of the data. The training set was then again split into two groups consisting of 80% of the training set, the \u0026ldquo;sub-training set,\u0026rdquo; and 20% of the training set, the \u0026ldquo;sub-test set,\u0026rdquo; to find hyperparameters. The sub-training set and sub-test sets were used to train the elastic net and find the optimal tuning parameters (L1 and Alpha), while the original test set was used to gauge performance e for this stage of training. Each model was trained on the Knight ADRC SomaLogic 7k CSF proteomics cohort (N\u0026thinsp;=\u0026thinsp;856) and validated on the ADNI SomaLogic 7k CSF proteomics cohort (N\u0026thinsp;=\u0026thinsp;48 for CSF A\u0026szlig;42/40 biomarker measurements). Model performance was trained and assessed on Pearson correlation of the model\u0026rsquo;s output with clinical biomarker measurements (Lumipulse or Elecsys).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eMetascape Analysis\u003c/h2\u003e \u003cp\u003eA pathway enrichment analysis was conducted using the bioinformatic tool Metascape (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://metascape.org\u003c/span\u003e\u003cspan address=\"https://metascape.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) to identify biological pathways that were associated with proteins predictive of the CSF A\u0026szlig;42/40 ratio. The original analysis began with 93 protein features identified as predictive of A\u0026szlig;42/40 in the model. Additional proteins that were highly correlated (r\u0026thinsp;\u0026gt;\u0026thinsp;0.7) with these 93 protein features were included, resulting in 821 total proteins. Of these proteins, 308 were significantly correlated with age in non-amyloidosis participants and were selected for enrichment analysis.\u003c/p\u003e \u003cp\u003eMetascape analysis was performed using the Reactome knowledge base with all 7008 possible protein features used as the background protein list. Enrichment significance was evaluated using default Metascape parameters. Among the significantly enriched pathways, SARS-CoV-2 (Reactome: R-HSA-9755779) appeared highly enriched on the list but was chosen to be omitted as it was not deemed relevant to the analysis.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eWe thank the entire Bateman lab for their valuable feedback and support. We also thank all the participants and their families, as well as the many institutions involved and their staff. This research was supported by Tracy Family Stable Isotope Labeling Quantitation Center established by the Tracy Family (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://silqcenter.wustl.edu/\u003c/span\u003e\u003cspan address=\"https://silqcenter.wustl.edu/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e ), Richard Frimel \u0026amp; Gary Werths, GHR Foundation, Pat and Jane Tracy, Anonymous, Anne \u0026amp; Ray Capestrain, Community Foundation Serving West Central Illinois and Northeast Missouri, JTL Family Fund, Payne Family, Mary \u0026amp; Jay Sullivan, Tracy Family Foundation, Catherine \u0026amp; Tom Tracy, Community Foundation for the Land of Lincoln, Jim \u0026amp; Jil Tracy, Joe \u0026amp; Jill Tracy, Sonja \u0026amp; Robert M. Willman, Boniface Foundation, Jean \u0026amp; Michael Buckley, Ann Liberman, Clemence S. Lieber Foundation, Mary Schoolman \u0026amp; Dr. James Hinrichs, and Susan \u0026amp; Scott Stamerjohn brought together by The Foundation for Barnes-Jewish Hospital. Research was also supported by the Healthy Aging and Senile Dementia P01 AG03991, Alzheimer\u0026rsquo;s Disease Research Center P30 AG066444, Adult Children Study P01 AG026276. Validation cohort data collection and sharing for this project was funded by the Alzheimer's Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U19 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). This work was also supported by access to equipment made possible by the Hope Center for Neurological Disorders, the Neurogenomics and Informatics Center (NGI: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://neurogenomics.wustl.edu/\u003c/span\u003e\u003cspan address=\"https://neurogenomics.wustl.edu/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) and the Departments of Neurology and Psychiatry at Washington University School of Medicine.\u003c/p\u003e\u003ch2\u003eData Availability Statement\u003c/h2\u003e \u003cp\u003eThe data that support the findings of this study are available on request from the corresponding author upon approval from the Knight Alzheimer Disease Research Center (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://knightadrc.wustl.edu/\u003c/span\u003e\u003cspan address=\"https://knightadrc.wustl.edu/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) and the Alzheimer's Disease Neuroimaging Initiative (ADNI). The data are not publicly available due to privacy or ethical restrictions. All trained models and software to run the CSF A\u0026szlig;42/40 model presented in this paper on custom data sets are available upon request from the corresponding author. Code to run the final model is also available through the Bateman lab GitHub at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/wusm-neurology-batmanlab/CSF_ABeta4240_Proteomics_Clock\u003c/span\u003e\u003cspan address=\"https://github.com/wusm-neurology-batmanlab/CSF_ABeta4240_Proteomics_Clock\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAlzheimer\u0026rsquo;s disease facts and figures (2025) Alzheimer\u0026rsquo;s Dement 21(4):e70235. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/alz.70235\u003c/span\u003e\u003cspan address=\"10.1002/alz.70235\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbdi G, Jain M, Patil N, Upadhyay B, Vyas N, Dwivedi M, Kaushal RS (2024) 14-3-3 proteins\u0026mdash;a moonlight protein complex with therapeutic potential in neurological disorder: In-depth review with Alzheimer\u0026rsquo;s disease. Front Mol Biosci 11:1286536. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fmolb.2024.1286536\u003c/span\u003e\u003cspan address=\"10.3389/fmolb.2024.1286536\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAli M, Erabadda B, Chen Y, Xu Y, Gong K, Liu M, Pichet Binette A, Timsina J, Western D, Yang C, Heo G, Vogel JW, Tijms BM, Krish V, Imam F, Hansson O, Winchester L, Cruchaga C (2025) Shared and disease-specific pathways in frontotemporal dementia and Alzheimer\u0026rsquo;s and Parkinson\u0026rsquo;s diseases. Nat Med 31(8):2567\u0026ndash;2577. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41591-025-03833-1\u003c/span\u003e\u003cspan address=\"10.1038/s41591-025-03833-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAli M, Timsina J, Western D, Liu M, Beric A, Budde J, Do A, Heo G, Wang L, Gentsch J, Schindler SE, Morris JC, Holtzman DM, Ruiz A, Alvarez I, Aguilar M, Pastor P, Rutledge J, Oh H, Cruchaga C (2025) Multi-cohort cerebrospinal fluid proteomics identifies robust molecular signatures across the Alzheimer disease continuum. Neuron 113(9):1363\u0026ndash;1379e9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.neuron.2025.02.014\u003c/span\u003e\u003cspan address=\"10.1016/j.neuron.2025.02.014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmft M, Ortner M, Eichenlaub U, Goldhardt O, Diehl-Schmid J, Hedderich DM, Yakushev I, Grimmer T (2022) The cerebrospinal fluid biomarker ratio Aβ42/40 identifies amyloid positron emission tomography positivity better than Aβ42 alone in a heterogeneous memory clinic cohort. Alzheimer\u0026rsquo;s Res Therapy 14(1):60. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s13195-022-01003-w\u003c/span\u003e\u003cspan address=\"10.1186/s13195-022-01003-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBarth\u0026eacute;lemy NR, Saef B, Li Y, Gordon BA, He Y, Horie K, Stomrud E, Salvad\u0026oacute; G, Janelidze S, Sato C, Ovod V, Henson RL, Fagan AM, Benzinger TLS, Xiong C, Morris JC, Hansson O, Bateman RJ, Schindler SE (2023) CSF tau phosphorylation occupancies at T217 and T205 represent improved biomarkers of amyloid and tau pathology in Alzheimer\u0026rsquo;s disease. Nat Aging 3(4):391\u0026ndash;401. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s43587-023-00380-7\u003c/span\u003e\u003cspan address=\"10.1038/s43587-023-00380-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoland B, Kumar A, Lee S, Platt FM, Wegiel J, Yu WH, Nixon RA (2008) Autophagy Induction and Autophagosome Clearance in Neurons: Relationship to Autophagic Pathology in Alzheimer\u0026rsquo;s Disease. J Neurosci 28(27):6926\u0026ndash;6937. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1523/JNEUROSCI.0800-08.2008\u003c/span\u003e\u003cspan address=\"10.1523/JNEUROSCI.0800-08.2008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBycroft C, Freeman C, Petkova D, Band G, Elliott LT, Sharp K, Motyer A, Vukcevic D, Delaneau O, O\u0026rsquo;Connell J, Cortes A, Welsh S, Young A, Effingham M, McVean G, Leslie S, Allen N, Donnelly P, Marchini J (2018) The UK Biobank resource with deep phenotyping and genomic data. Nature 562(7726):203\u0026ndash;209. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41586-018-0579-z\u003c/span\u003e\u003cspan address=\"10.1038/s41586-018-0579-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCampion D, Pottier C, Nicolas G, Le Guennec K, Rovelet-Lecrux A (2016) Alzheimer disease: Modeling an Aβ-centered biological network. Mol Psychiatry 21(7):861\u0026ndash;871. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/mp.2016.38\u003c/span\u003e\u003cspan address=\"10.1038/mp.2016.38\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDakterzada F, L\u0026oacute;pez-Ortega R, Arias A, Riba-Llena I, Ruiz-Juli\u0026aacute;n M, Huerto R, Tahan N, Pi\u0026ntilde;ol-Ripoll G (2021) Assessment of the Concordance and Diagnostic Accuracy Between Elecsys and Lumipulse Fully Automated Platforms and Innotest. Front Aging Neurosci 13:604119. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fnagi.2021.604119\u003c/span\u003e\u003cspan address=\"10.3389/fnagi.2021.604119\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFan X, Cui L, Zeng Y, Song W, Gaur U, Yang M (2019) 14-3-3 Proteins Are on the Crossroads of Cancer, Aging, and Age-Related Neurodegenerative Disease. Int J Mol Sci 20(14):3518. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/ijms20143518\u003c/span\u003e\u003cspan address=\"10.3390/ijms20143518\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFindeis MA (2007) The role of amyloid β peptide 42 in Alzheimer\u0026rsquo;s disease. Pharmacol Ther 116(2):266\u0026ndash;286. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.pharmthera.2007.06.006\u003c/span\u003e\u003cspan address=\"10.1016/j.pharmthera.2007.06.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrake RA, Ricketts T, Menzies FM, Rubinsztein DC (2015) Autophagy and neurodegeneration. J Clin Investig 125(1):65\u0026ndash;74. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1172/JCI73944\u003c/span\u003e\u003cspan address=\"10.1172/JCI73944\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGong B, Radulovic M, Figueiredo-Pereira ME, Cardozo C (2016) The Ubiquitin-Proteasome System: Potential Therapeutic Targets for Alzheimer\u0026rsquo;s Disease and Spinal Cord Injury. Front Mol Neurosci 9:4. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fnmol.2016.00004\u003c/span\u003e\u003cspan address=\"10.3389/fnmol.2016.00004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHaass C, Selkoe DJ (2007) Soluble protein oligomers in neurodegeneration: Lessons from the Alzheimer\u0026rsquo;s amyloid β-peptide. Nat Rev Mol Cell Biol 8(2):101\u0026ndash;112. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/nrm2101\u003c/span\u003e\u003cspan address=\"10.1038/nrm2101\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHorvath S (2013) DNA methylation age of human tissues and cell types. Genome Biol 14(10):R115. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/gb-2013-14-10-r115\u003c/span\u003e\u003cspan address=\"10.1186/gb-2013-14-10-r115\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHussain MS, Agrawal N, Ilma B, Nayak MMR, Kaur PP, Khachi M, Goyal A, Rekha K, Gupta A, Gupta S, G., Dua K (2025) Autophagy and Cellular Senescence in Alzheimer\u0026rsquo;s Disease: Key Drivers of Neurodegeneration. CNS Neurosci Ther 31(7):e70503. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/cns.70503\u003c/span\u003e\u003cspan address=\"10.1111/cns.70503\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eImam F, Saloner R, Vogel JW, Krish V, Abdel-Azim G, Ali M, An L, Anastasi F, Bennett D, Pichet Binette A, Boxer AL, Bringmann M, Burns JM, Cruchaga C, Dage JL, Farinas A, Ferrucci L, Finney CA, Frasier M, Lovestone S (2025) The Global Neurodegeneration Proteomics Consortium: Biomarker and drug target discovery for common neurodegenerative diseases and aging. Nat Med 1\u0026ndash;11. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41591-025-03834-0\u003c/span\u003e\u003cspan address=\"10.1038/s41591-025-03834-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang S, Li Y, Zhang X, Bu G, Xu H, Zhang Y (2014) Trafficking regulation of proteins in Alzheimer\u0026rsquo;s disease. Mol Neurodegeneration 9:6. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/1750-1326-9-6\u003c/span\u003e\u003cspan address=\"10.1186/1750-1326-9-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang Y, Zhou X, Ip FC, Chan P, Chen Y, Lai NCH, Cheung K, Lo RMN, Tong EPS, Wong BWY, Chan ALT, Mok VCT, Kwok TCY, Mok KY, Hardy J, Zetterberg H, Fu AKY, Ip NY (2022) Large-scale plasma proteomic profiling identifies a high-performance biomarker panel for Alzheimer\u0026rsquo;s disease screening and staging. Alzheimer\u0026rsquo;s Dement 18(1):88\u0026ndash;102. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/alz.12369\u003c/span\u003e\u003cspan address=\"10.1002/alz.12369\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee J-H, Yu WH, Kumar A, Lee S, Mohan PS, Peterhoff CM, Wolfe DM, Martinez-Vicente M, Massey AC, Sovak G, Uchiyama Y, Westaway D, Sisodia SS, Cuervo AM, Nixon RA (2010) Lysosomal Proteolysis and Autophagy Require Presenilin 1 and Are Disrupted by Alzheimer-Related PS1 Mutations. Cell 141(7):1146\u0026ndash;1158. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cell.2010.05.008\u003c/span\u003e\u003cspan address=\"10.1016/j.cell.2010.05.008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLewczuk P, Łukaszewicz-Zając M, Mroczko P, Kornhuber J (2020) Clinical significance of fluid biomarkers in Alzheimer\u0026rsquo;s Disease. Pharmacol Rep 72(3):528\u0026ndash;542. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43440-020-00107-0\u003c/span\u003e\u003cspan address=\"10.1007/s43440-020-00107-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLian P, Guo Y, Yu J (2025) Biomarkers in Alzheimer\u0026rsquo;s disease: Emerging trends and clinical implications. Chin Med J 138(9):1009\u0026ndash;1012. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1097/CM9.0000000000003579\u003c/span\u003e\u003cspan address=\"10.1097/CM9.0000000000003579\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa C, Hong F, Yang S (2022) Amyloidosis in Alzheimer\u0026rsquo;s Disease: Pathogeny, Etiology, and Related Therapeutic Directions. Molecules 27(4):1210. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/molecules27041210\u003c/span\u003e\u003cspan address=\"10.3390/molecules27041210\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMelendez J, Sung YJ, Orr M, Yoo A, Schindler S, Cruchaga C, Bateman R (2024) An interpretable machine learning-based cerebrospinal fluid proteomics clock for predicting age reveals novel insights into brain aging. Aging Cell 23(9):e14230. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/acel.14230\u003c/span\u003e\u003cspan address=\"10.1111/acel.14230\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNixon RA, Wegiel J, Kumar A, Yu WH, Peterhoff C, Cataldo A, Cuervo AM (2005) Extensive Involvement of Autophagy in Alzheimer Disease: An Immuno-Electron Microscopy Study. J Neuropathology Experimental Neurol 64(2):113\u0026ndash;122. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/jnen/64.2.113\u003c/span\u003e\u003cspan address=\"10.1093/jnen/64.2.113\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eO\u0026rsquo;Brien RJ, Wong PC (2011) Amyloid Precursor Protein Processing and Alzheimer\u0026rsquo;s Disease. Annu Rev Neurosci 34:185\u0026ndash;204. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1146/annurev-neuro-061010-113613\u003c/span\u003e\u003cspan address=\"10.1146/annurev-neuro-061010-113613\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOh HS-H, Rutledge J, Nachun D, P\u0026aacute;lovics R, Abiose O, Moran-Losada P, Channappa D, Urey DY, Kim K, Sung YJ, Wang L, Timsina J, Western D, Liu M, Kohlfeld P, Budde J, Wilson EN, Guen Y, Maurer TM, Wyss-Coray T (2023) Organ aging signatures in the plasma proteome track health and disease. Nature 624(7990):164\u0026ndash;172. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41586-023-06802-1\u003c/span\u003e\u003cspan address=\"10.1038/s41586-023-06802-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePatterson BW, Elbert DL, Mawuenyega KG, Kasten T, Ovod V, Ma S, Xiong C, Chott R, Yarasheski K, Sigurdson W, Zhang L, Goate A, Benzinger T, Morris JC, Holtzman D, Bateman RJ (2015) Age and amyloid effects on human central nervous system amyloid-beta kinetics. Ann Neurol 78(3):439\u0026ndash;453. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/ana.24454\u003c/span\u003e\u003cspan address=\"10.1002/ana.24454\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePennington KL, Chan TY, Torres MP, Andersen JL (2018) The dynamic and stress-adaptive signaling hub of 14-3-3: Emerging mechanisms of regulation and context-dependent protein\u0026ndash;protein interactions. Oncogene 37(42):5587\u0026ndash;5604. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41388-018-0348-3\u003c/span\u003e\u003cspan address=\"10.1038/s41388-018-0348-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQiang Q, Skudder-Hill L, Toyota T, Huang Z, Wei W, Adachi H (2023) CSF 14-3-3β is associated with progressive cognitive decline in Alzheimer\u0026rsquo;s disease. Brain Commun 5(6):fcad312. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/braincomms/fcad312\u003c/span\u003e\u003cspan address=\"10.1093/braincomms/fcad312\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQiu C, Kivipelto M, von Strauss E (2009) Epidemiology of Alzheimer\u0026rsquo;s disease: Occurrence, determinants, and strategies toward intervention. Dialog Clin Neurosci 11(2):111\u0026ndash;128. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.31887/DCNS.2009.11.2/cqiu\u003c/span\u003e\u003cspan address=\"10.31887/DCNS.2009.11.2/cqiu\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchindler SE, Bollinger JG, Ovod V, Mawuenyega KG, Li Y, Gordon BA, Holtzman DM, Morris JC, Benzinger TLS, Xiong C, Fagan AM, Bateman RJ (2019) High-precision plasma β-amyloid 42/40 predicts current and future brain amyloidosis. Neurology 93(17):e1647\u0026ndash;e1659. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1212/WNL.0000000000008081\u003c/span\u003e\u003cspan address=\"10.1212/WNL.0000000000008081\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiegenthaler B, Rajendran L (2012) Retromers in Alzheimer\u0026rsquo;s disease. Neurodegenerative Dis 10(1\u0026ndash;4) Article 1\u0026ndash;4. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1159/000335910\u003c/span\u003e\u003cspan address=\"10.1159/000335910\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang R, Qi X, Schmiege P, Coutavas E, Li X (2020) Marked structural rearrangement of mannose 6-phosphate/IGF2 receptor at different pH environments. Sci Adv 6(7):eaaz1466. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1126/sciadv.aaz1466\u003c/span\u003e\u003cspan address=\"10.1126/sciadv.aaz1466\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y, Xiao S, Liu B, Jiang R, Liu Y, Hang Y, Chen L, Chen R, Vitiello MV, Bennett D, Wang B, Lv J, Yu C, Haslam DE, Zheng Q, Gerszten RE, Bao Y, Shi J, Xie J, Chan AT (2026) Organ-specific proteomic aging clocks predict disease and longevity across diverse populations. Nat Aging 6(1):162\u0026ndash;180. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s43587-025-01016-8\u003c/span\u003e\u003cspan address=\"10.1038/s43587-025-01016-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXie Z, Klionsky DJ (2007) Autophagosome formation: Core machinery and adaptations. Nat Cell Biol 9(10):1102\u0026ndash;1109. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/ncb1007-1102\u003c/span\u003e\u003cspan address=\"10.1038/ncb1007-1102\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8791371/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8791371/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAlzheimer\u0026rsquo;s disease (AD) is clinically characterized by progressive memory loss and cognitive decline, with aging as the primary risk factor. Early AD pathology includes accumulation of amyloid-beta (A\u0026szlig;) in plaques. This study aims to investigate the molecular mechanisms by which aging contributes to brain amyloidosis using a machine learning-based approach on large proteomic datasets from cerebrospinal fluid (CSF). To accomplish this, we trained a machine learning model to predict CSF A\u0026szlig;42/40, a key biomarker for amyloidosis. Our modified elastic net model using adaptive feature selection achieved robust accuracy (Pearson Correlation of 0.86) predicting CSF A\u0026szlig;42/40 in our validation cohort. Pathway analysis of the model-utilized proteins (and proteins highly correlated to them) revealed age-associated alterations potentially linked to amyloidosis, particularly highlighting dysregulated autophagy and membrane trafficking pathways. These findings suggest that impaired autophagosome-lysosome fusion and endosomal processing may drive the decline in A\u0026szlig; clearance with aging. Our study highlights the power of machine learning in biomarker approximation and biological prediction, enabling insights into multiple diseases.\u003c/p\u003e","manuscriptTitle":"Identifying Age-Related Protein Mechanisms of Alzheimer's Disease Amyloidosis from Cerebrospinal Fluid Proteomics Using a Novel Machine Learning Approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-24 04:31:38","doi":"10.21203/rs.3.rs-8791371/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"65e7afad-9276-4165-a8eb-71b26111e7ba","owner":[],"postedDate":"February 24th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":63271775,"name":"Biological sciences/Neuroscience/Cognitive ageing"},{"id":63271776,"name":"Biological sciences/Computational biology and bioinformatics/Machine learning"},{"id":63271777,"name":"Biological sciences/Neuroscience/Diseases of the nervous system/Alzheimer's disease"},{"id":63271778,"name":"Biological sciences/Developmental biology/Ageing"}],"tags":[],"updatedAt":"2026-03-09T16:25:40+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-24 04:31:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8791371","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8791371","identity":"rs-8791371","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}