AI-driven biomarker learning for the early diagnosis of neurodegenerative diseases: ABLEDx

Abstract

Background

Tears are an easily accessible biofluid that reflects both emotional states and disease conditions. They are particularly enriched in extracellular vesicles (EVs), which carry proteins and nucleic acids relevant to neurological health. This makes tear EVs a promising source for biomarker discovery. However, limited sample volume and variability pose challenges for identifying reliable biomarkers for clinical diagnosis.

Results

We present AI-driven Biomarker Learning for the Early Diagnosis of Neurodegenerative Diseases (ABLEDx), which applies a conditional variational autoencoder (cVAE) to enhance proteomic analysis of tear EVs. This approach effectively addresses sample limitations and improves the identification of disease-associated biomarkers. Our results reveal that tear EVs capture molecular signals along the eye–brain axis, reflecting contributions from both ocular and central nervous system cells. ABLEDx identified clinically relevant protein modules, which were consistently elevated in patients with neurodegenerative diseases. Moreover, we recognize that KRAS is highly expressed in patients with Alzheimer’s disease, Parkinson’s disease, and ocular myasthenia gravis, and tear-EV-associated LRG1 and HSPG2 exhibit differentiation between Alzheimer’s disease and Parkinson’s disease.

Conclusions

ABLEDx demonstrates the utility of combining AI with tear-EV proteomics for non-invasive biomarker discovery. This strategy enables early and real-time detection of neurodegenerative and ocular diseases, offering new opportunities for clinical diagnostics and translational medicine.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12951-026-04072-3.

Background

Tears are closely associated with daily activities and emotions, providing a window to our thoughts, emotional states (e.g., happiness and sadness), and overall physical health [1–3]. Tears are a filtrate of blood plasma, primarily composed of water, electrolytes, and various biomolecules. In addition to being approximately 98% water, tears contain extracellular vesicles (EVs), low-molecular-weight proteins, lipids, and microRNA [4–6]. This unique composition makes tears a promising, non-invasive source for identifying novel biomarkers and exploring the intricate connections between emotions, neurological activity, and disease states [7–9]. Given the potential of tears as a source of biomarkers, they hold promise for addressing the growth burden of neurodegenerative diseases, which are becoming increasingly prevalent with aging populations. As life expectancy increases in many countries, the prevalence of neurodegenerative diseases is projected to rise significantly, posing a growing public health challenge [10, 11]. Early diagnosis and continuous monitoring of these conditions are essential for implementing effective interventions and optimizing patient management [12–14]. Among neurodegenerative conditions, Alzheimer’s disease (AD), the most common cause of dementia [15], affects about one in nine individuals (10.8%) aged 65 and older [16], highlighting the significant burden this disease places on the aging population. Similarly, Parkinson’s disease (PD) ranks as the second-most common neurodegenerative condition, with over 6 million individuals affected worldwide [17]. As the prevalence of these diseases continues to rise, addressing the challenges associated with early diagnosis becomes increasingly critical. There is an urgent need for improved diagnostic tools, including advanced omics techniques and biomarker discovery, to facilitate the identification of these conditions at their onset.

Tear EVs are a diverse group of membrane-enclosed nanoparticles widely found in various biofluids [18, 19]. Their ability to cross the blood-brain barrier and reflect molecular changes in the central nervous system makes them invaluable for identifying disease-specific biomarkers [20]. For example, Tau and TDP-43 proteins on plasma-derived EVs can be diagnostic biomarkers for frontotemporal dementia and amyotrophic lateral sclerosis [21]. Additionally, CatB detected in cerebrospinal fluid (CSF)-derived EVs may be considered a promising biomarker and therapeutic target in amyloid pathology associated with AD. [22] Detecting blood-related biomarkers often requires sensitive methods to counteract the influence of highly abundant proteins [23]. Recent studies on tear EVs demonstrate that these non-invasively collected nanoparticles carry critical molecular information from various tissues and cells, providing unique insights into precision diagnostics [24]. Tear EVs hold significant potential as a dynamic window into both physiological and pathological processes. A recent study indicates that the transport of β-amyloid from the brain to the eye may contribute to retinal degeneration in AD patients [25]. As a result, tear components may reflect changes not only in the eye but throughout the body [4]. This makes tear EVs a promising proxy for investigating molecular dynamics and understanding complex eye-to-brain signaling pathways. Additionally, we have previously developed an ultrafiltration-based method that enables the efficient recovery of EVs from tear samples [4, 26]. By harnessing the diagnostic potential of tear EVs, researchers can obtain valuable information on various conditions, paving the way for non-invasive and real-time approaches to the early diagnosis of neurodegenerative diseases.

Tear fluid from patients has shown significant potential as a diagnostic tool for detecting AD. [27] However, several technological challenges impede the effective utilization of EVs derived from this fluid. Despite their accessibility, a significant challenge in tear-EV research lies not in the volume of individual samples but in the limited availability of patient cohorts and the small number of EVs obtainable per sample. Additionally, the complex omics data generated from these samples pose significant challenges in data analysis and interpretation. Addressing these issues is crucial for fully exploiting the diagnostic potential of tear-derived EVs in neurodegenerative disease research and clinical applications.

Advances in proteomic technologies have facilitated precision diagnostics by focusing on the proteins contained within EVs [28, 29]. Nevertheless, EV preparation and omics analysis processes can introduce substantial noise and variability, particularly in rare disease research where sample sizes are often limited. This presents complications in applying bioinformatics to EV-derived omics data, making it difficult to determine the molecular relevance of EVs to human diseases [30]. High-dimensional omics datasets often require dimensionality reduction to align with specific disease phenotypes. Techniques such as Principal component analysis (PCA) and independent component analysis (ICA) are commonly used. Still, they may need to capture the nonlinear effects in complex biological systems adequately. In contrast, Variational Autoencoders (VAE) excel at uncovering hidden information by learning low-dimensional representations and the underlying data distributions [31, 32]. This technique can effectively compress and abstract dynamic features from EV proteomic profiles, facilitating efficient and comprehensive analysis of cohorts.

In this study, we present an AI-driven Biomarker Learning platform for the Early Diagnosis of neurodegenerative diseases (ABLEDx). This intelligent platform utilizes an advanced data-processing framework, employing conditional variational autoencoders (cVAE), to reduce high-dimensional EV proteomic data to a lower-dimensional latent space. Additionally, it generates numerous simulated samples through deep learning. In this proof-of-concept study, we explore the proteomic dynamics of tear EVs from patients with Alzheimer’s Disease (AD), Ocular Myasthenia Gravis (MG), Parkinson’s Disease (PD), Neuromyelitis Optica Spectrum Disorders (NMOSD), Optic Neuritis (ON), and Healthy Controls (HC). ABLEDx offers a versatile platform for extracting critical features from diverse omics datasets, dramatically enhancing precision diagnostics through molecular analysis of tear-derived EVs.

Methods

Clinical sample collection and tear EV isolation

Tear fluid was collected using Schirmer test strips (Jingming, China) and placed in the lower eyelid for 5 to 10 min at room temperature. The saturated strips were then transferred into 2 mL of cold phosphate-buffered saline (PBS, Thermo Fisher Scientific). The PBS-diluted tear fluid was centrifuged at 500 g for 10 min, followed by a second centrifugation at 2,000 g for another 10 min to remove cells and debris. The resulting tear suspension was filtered through a 0.25 μm membrane filter (Millipore). EVs from the tear fluid were isolated using the EXODUS method, with vacuum actuation set at −20 kPa and a conversion time of 10 s. The purified EV sample was stored at −80 °C until further use.

Nanoparticle tracking analysis (NTA)

Tear EV nanoparticles were characterized using a NanoSight NS300 (Malvern) with a 488 nm laser. The samples were diluted with phosphate-buffered saline (PBS) to achieve an optimal nanoparticle tracking analysis (NTA) concentration range between 10⁷ and 10⁹ particles per mL. A syringe pump introduced the samples into the microfluidic chamber. Each sample was analyzed in triplicate under optimal parameters, ensuring that there were 20 to 50 particles per frame.

Transmission electron microscopy (TEM)

The morphology of tear extracellular vesicles (EVs) was assessed using transmission electron microscopy (TEM). A 20 µL sample of EVs was mixed with 4% paraformaldehyde (PFA) and spotted onto parafilm. Formvar/carbon-coated copper grids were placed on the sample droplet for 30 min to allow adsorption. Afterward, the sample on the grid was washed with phosphate-buffered saline (PBS) and fixed with 1% glutaraldehyde for 5 min. It was then negatively stained with 2% uranyl acetate for 45 s. Excess stain was carefully removed using filter paper, and the grid was air-dried. Imaging was conducted using a transmission electron microscope operating at 200 kV (Talos F200).

Western blot (WB) analysis

Western blot (WB) analysis was performed using 4–12% precast polyacrylamide mini-gels (Tri-glycine, pH 8.3) in a Mini Trans-Blot module (Bio-Rad). Samples with equal protein mass (3 µg, measured using Qubit Protein Assay Kits) were loaded onto the gels. Proteins were transferred to a polyvinylidene fluoride (PVDF) membrane using the Trans-Blot Turbo Transfer System (Bio-Rad). After blocking with 5% nonfat dry milk in PBS containing 0.1% Tween 20 for 1 h, the membrane was incubated overnight at 4 °C with the primary antibody. The membrane was washed and incubated with HRP-conjugated anti-mouse IgG (7076, Cell Signaling Technology) or HRP-conjugated anti-rabbit IgG (7074, Cell Signaling Technology) as the secondary antibody. Enhanced chemiluminescence (Peiqing Science & Technology) was utilized for immunodetection. The following primary antibodies were used: anti-Alix (sc-53540, Santa Cruz Biotechnology), anti-CD63 (Abcam), and anti-CD81 (Santa Cruz Biotechnology). All primary antibodies were diluted 1:1000, while the secondary antibodies were diluted 1:3000.

LC-MS/MS-based proteomic analysis

Label-free liquid chromatography-tandem mass spectrometry (LC-MS/MS) was used to analyze tear extracellular vesicle (EV) samples isolated from each patient. The proteins extracted from the EVs were digested in 50 mM ammonium bicarbonate with 10 ng/mL trypsin at 37 °C for 10 h. The resulting peptides were desalted, lyophilized, and then reconstituted in 0.1% (v/v) formic acid in ddH2O. These peptides were subsequently subjected to LC-MS/MS analysis, utilizing a nanoElute liquid chromatography system (Bruker Daltonics) coupled with a timsTOF Pro 2 mass spectrometer (Thermo Fisher Scientific). The separation was performed using a 25 cm x 75 μm column (1.6 μm, C18, ion optics). The mobile phases consisted of Phase A, which was water with 0.1% (v/v) formic acid, and Phase B, composed of acetonitrile (ACN), water, and formic acid (80%:19.9%:0.1% v/v/v). The linear gradient was set as follows: from 0 to 45 min, 5% to 27% B; from 45 to 50 min, 27% to 46% B; from 50 to 55 min, 46% to 100% B; and from 55 to 60 min, 100% B. Data-dependent acquisition (DDA) mode was employed for information collection.

Principal component analysis (PCA)

PCA was applied to the newly generated samples to assess similarities between groups. The process involved calculating the covariance matrix, determining the eigenvalues and eigenvectors, selecting the leading eigenvectors, and projecting the data onto these components. This allowed for the evaluation of group similarities and the identification of clusters with high similarity.

Construction and training of the cVAE model

The cVAE model consists of an encoder, a latent space, and a decoder. It undergoes conditional training using quality-controlled protein abundance data imputed for missing values, standardized, and conditioned on one-hot encoded group labels.

The data were derived from quantitative mass spectrometry-based proteomics of tear fluid exosomes from individuals with five neurological diseases and healthy controls. Protein abundance data were processed using the proteoDA software. Any proteins with abundance values marked as NA in more than two-thirds of the samples were excluded. The remaining data were normalized using the Cyclic Loess (Cycloess) method to eliminate systematic biases. Missing values were imputed using the RandomForestRegressor module from Python’s sklearn library. The preprocessed protein abundance data were combined with one-hot encoded group labels for each sample.

For training, we employed the loss function of a variational autoencoder (VAE), which comprises the reconstruction error and the Kullback-Leibler (KL) divergence, to compute the complete Evidence Lower Bound (ELBO) loss. The Adam optimizer was employed to minimize the reconstruction error and KL divergence, optimizing the model parameters. This process allows the model to effectively generate samples from the latent space that resemble the input data distribution. The reconstruction error, measured by Mean Squared Error (MSE), quantifies the difference between the data generated by the decoder and the actual input data. The KL divergence measures the difference between the distribution of the latent variables and a standard distribution. To manage gradient explosion issues during training, log variance values falling outside the range of −10 to 10 were clipped to the nearest boundary value. We used Principal Component Analysis (PCA) to assess the training process, which was halted when the samples were distinctly separated in a PCA plot.

After training, the decoder generated 2,000 new samples for each group, thereby enhancing the diversity and robustness of the dataset. To visualize the protein abundances of the simulated samples, we employed the Uniform Manifold Approximation and Projection (UMAP) method for dimensionality reduction. The steps involved were as follows:

• (i)Calculating Mean Protein Abundances: We computed the mean abundance for each protein across the simulated samples within each of the six groups: Alzheimer’s Disease (AD), Parkinson’s Disease (PD), Optic Neuritis (ON), Neuromyelitis Optica Spectrum Disorder (NMOSD), and Healthy Controls (HC).
• (ii)Dimensionality Reduction with UMAP: Using the calculated mean abundances, we applied the UMAP method to reduce the high-dimensional protein abundance data to a two-dimensional space, resulting in a UMAP plot where each point represents a specific protein.
• (iii)Visualization and Coloring: The UMAP plot was colored based on the Z-score normalized mean abundances of the proteins in each group, providing a visual representation of the characteristic features of the samples from each group.

A detailed description of the model-building and training methods can be found in the supplementary information file.

WGCNA analysis

We utilized the WGCNA R package to identify modules within the simulated six-group samples and assess their correlation with various phenotypes. Next, we conducted a functional enrichment analysis on the genes within these modules using the g: GOSt API from g: Profiler, employing the default parameters. Finally, we summarized the main functions of each identified module. We acquired human protein-protein interaction (PPI) data from the STRING database for visualization. The WGCNA analysis was conducted with a power of 12, utilizing a signed network and the bicolor method for correlation. We focused on the modules that showed a strong correlation with Alzheimer’s Disease (AD), Parkinson’s Disease (PD), and Myasthenia Gravis (MG) and selected the genes from these modules. We then reran WGCNA to identify new modules and calculate their correlations with AD, PD, and MG phenotypes. The significantly enriched functions for each module were summarized. For plotting the protein-protein interactions, we applied the following criteria: we only included interactions with a score greater than 400 and those in which at least one of the interacting proteins was part of the module that showed a strong correlation with the healthy control (HC) group.

Protein signature identification

Simulated samples were analyzed using the Wilcoxon Rank-Sum Test to identify proteins that were significantly differentially expressed. The abundance values for each sample were log2-transformed and normalized across samples using cyclic loess normalization. For each protein, the log-fold change (LogFC) between sample groups was calculated using the following formula:

P-values were computed using the Wilcoxon rank-sum test, and adjusted p-values were generated using the Benjamini-Hochberg procedure. The differentially expressed proteins (DEPs) were ranked based on LogFC with a p-value threshold of 0.05. Groups with high similarity were defined as clusters, and biomarkers were identified within each group to ensure accurate and reliable cluster identification.

ELISA validation

ELISA was conducted to assess the expression levels of selected protein signatures, specifically LRG1 and HSPG2, in tear extracellular vesicles (EVs). The protein concentrations in each sample (100 µL) were measured following the manufacturer’s instructions: HSPG2 (CUSABIO, China, CSB-EL010868HU) and LRG1 (Abcam, ab260066-1 × 96 t). The signals were normalized to the total amount of loaded protein.

Statistical analysis

Statistical comparisons were conducted using a one-tailed t-test to assess whether the values in the case group were significantly higher than those in the control group. The statistical analysis was performed using GraphPad Prism (version 8.0.1, GraphPad Software). A P-value of less than 0.05 was deemed statistically significant.

Results

Tears are a valuable biofluid for noninvasive disease detection and contain a high concentration of EV particles. Using the EXODUS method, we isolated high-quality tear EVs from the tear fluids of individual patients. The resulting EV product was further evaluated according to the MISEV 2023 guideline [33], which indicated a typical cup-shaped morphology, characteristic protein markers, and sizes ranging from 30 to 200 nm (Fig. S1). We then performed label-free MS-based proteomics on these samples, detecting 3015 proteins from tear EVs (Data S1). Using the Human Protein Atlas dataset, we studied and mapped the tissue and cell origins of EV proteins within the ocular system. Our findings revealed that the genesis of tear EV proteins is closely related to physically adjacent tissues, including the eye, brain, choroid, and retina, as well as various types of neural cells, such as glial cells, neurons, astrocytes, oligodendrocytes, and oligodendrocyte precursor cells (OPC) (Fig. 1a).

Fig. 1.

Proteomic investigation of disease information via the origins of tear-derived extracellular vesicles. (A) Demonstration of intricate proteomic connections observed between tear EVs and tissues and neuronal cells. The proteomic profiles of tissues and cells were aligned with the Human Protein Atlas dataset. (B) Differential co-expression analysis of neuronal cell-specific proteins detected in tear EVs between different disease groups. (C) The heat map shows a differential analysis of tissue-specific proteins detected in tear EVs from the disease and control groups. (D) AI-driven cVAE training and simulation to capture the essential characteristics of the complex proteomic system related to diseases

Next, we conducted a differential protein analysis of neural cell-specific proteins across 5 groups of patients with different diseases and healthy controls. This analysis showed a significant overlap in the differential proteins among the groups (Fig. 1b). This suggests that tear EVs may have a multifaceted role in various conditions, affecting both central nervous system disorders and ocular diseases. By expanding the tissue sources analyzed for differential proteins, we discovered a more intricate network of tear EVs originating from both nearby and distant organs (Fig. 1c). The distinct patterns observed in the heatmap for each group indicate that tear EVs are involved in a broad range of physiological and pathological processes. We propose utilizing an AI-based model to accurately identify biomarkers within this complex system. This model, conditional VAE (cVAE), consists of a learning encoder and decoder that leverages deep learning techniques to simulate and analyze the proteomic matrix (Fig. 1d), providing a powerful tool for biomarker discovery.

Before modeling, the proteomic data underwent pre-processing, which included calibration and Cyclic Loess normalization. This process resulted in 2017 valid proteins available for analysis and model training (Fig. 2a). The cVAE was trained using group-encoded data, where the encoder mapped the inputs to a latent space, and the decoder reconstructed them from this latent representation. After training, the decoder generated new samples by selecting random points from the learned latent space and mapping them back to the high-dimensional space of protein abundances. To manage data complexity and reduce noise, we processed the protein abundance data using one-hot encoded group labels through the input layer (Fig. 2a). This supervised approach significantly enhanced the model’s learning capability. In the PCA analysis of actual samples, the disease groups were indistinguishable (Fig. 2b). Similarly, when applying a VAE model without one-hot encoded group labels, no distinct clusters were observed, irrespective of the depth of the deep learning process (Fig. S2). This highlights the critical importance of a supervised approach. Even when group labels were used without sample-generating functions, the disease groups remained indistinguishable (Fig. S3). Using our trained cVAE decoder, we generated 2,000 new samples for each group. The deep learning model, trained over 180 epochs, progressively improved sample classification, resulting in distinct clusters (Fig. 2c). The cluster distances reflected disease similarities among the diseases, with the MG cluster being closer to AD and PD than to ON and NMOSD, despite all three being autoimmune diseases. Additionally, the Upset map indicated that these three groups shared more differentially expressed proteins than the others (Fig. S4).

Fig. 2.

Deep learning-driven conditional variational autoencoders (cVAE) enhance multi-task performance in EV-proteomics through data augmentation. (A) Construction and training of the cVAE model for proteomic data augmentation. The cVAE learns the essential characteristics of data, maps them to a latent space, and generates simulated samples for downstream multi-tasking. (B) PCA analysis of actual samples. (C) cVAE was conducted with 2000 newly generated samples before and after the training. The different depths of the deep learning process were investigated, and distinct clusters were observed in epoch 180. The dots represent samples

We conducted a weighted correlation network analysis (WGCNA) on the generated proteomic data, identifying 11 functional modules closely associated with clinical relevance (Fig. S5). These modules participate in distinct KEGG pathways (Fig. 3a) and biological processes (Fig. S6). Notably, the unsigned module is functionally enriched in pathways related to neurodegeneration and infectious diseases, reflecting the shared characteristics of the diseases under study. The specific correlations between protein modules and disease phenotypes, illustrated in Fig. 3b, suggest that correlation scores may help differentiate diseases from a global functional perspective. This conclusion is further supported by the Uniform Manifold Approximation and Projection (UMAP) analysis shown in Fig. 3c. We utilized UMAP for dimensionality reduction to visualize the protein abundances of the simulated samples. The Z-score normalized mean protein abundances are represented by color in this plot, clearly highlighting the characteristic features of each group.

Fig. 3.

Investigation of clinical relevance and functions of protein modules. (A) WGCNA analysis identified 11 functional protein modules, along with corresponding KEGG enrichment analysis. Each node represents one protein, color-coded by the different modules. (B) Pearson correlation analysis between protein modules and disease phenotypes. (C) UMAP analysis for dimensionality reduction and visualization of protein expression patterns from EV proteomics across different diseases. The dots represent proteins

We further identified protein signatures for the tested diseases using the Wilcoxon Rank-Sum test. Our analysis revealed detectable protein markers between clusters within each disease group, as listed in Table S1. The top markers for each group included KRAS for AD, PD, and MG, LRG1 for AD, and HSPG2 for PD, all of which demonstrated strong discriminative abilities (Fig. 4a and b). Enrichment analysis of these markers highlighted disease-related pathways and GO terms (Fig. S7-S8). For example, metabolic deficits related to AD are associated with metabolic pathways and pyruvate metabolism [34]. Similarly, neurodegenerative diseases such as PD involve protein misfolding and accumulation, which leads to cellular dysfunction and brain damage [35, 36]. Additionally, neutrophils play a central role in the pathogenesis of NMOSD [37]. To further validate the cVAE model, we confirmed the effectiveness of the top markers LRG1 and HSPG2 in classifying AD and PD. ELISA tests indicated that LRG1 significantly differentiates AD patients (14 samples) from those with MG (15 samples) and HC individuals (12 samples) (Fig. 4c). Furthermore, HSPG2 effectively distinguishes AD patients (15 samples) from PD patients (6 samples) (Fig. 4c).

Fig. 4.

Investigation of protein signatures of diseases based on cVAE modeling. (A) Heatmap showing the top 10 protein markers for disease groups, including AD_PD_MG (AD_PD_MG vs. HC_ON_NMOSD), sub_AD (AD vs. PD_MG_HC), and sub_PD (PD vs. AD). (B) Relative abundance of the top protein marker in each group. (C) Overexpression of HSPG2 confirmed in PD patients compared to AD by ELISA (left, AD, n = 15; PD, n = 6); Overexpression of LRG1 confirmed in AD patients compared to HC and MG by ELISA (right, AD, n = 14; HC, n = 12; MG, n = 15). (D) Overexpression of LRG1 was characterized in AD patients compared to HC, PD, MG, ON, and NMOSD by LC-MS analysis (n = 3, pooled samples)

The KRAS mutation in human brain endothelial cells is a crucial factor in the development of sporadic cerebral arteriovenous malformation [38], which may contribute to the pathogenesis of neurodegenerative diseases by altering signaling pathways and affecting cellular function and behavior. Research has shown that heat shock proteins (HSPs) play a significant role in PD, helping to maintain protein homeostasis and preventing protein aggregation by properly folding and activating intracellular proteins associated with the disease [39]. Additionally, we applied a targeted proteomic approach using three pooled samples per group and observed that LRG1 was markedly elevated in tear EVs from AD patients compared with those from the PD, MG, ON, NMOSD, and HC groups, as determined by LC-MS–based PRM (Parallel Reaction Monitoring) analysis (Fig. 4d). Although the small sample size limits statistical power and the p-values do not reach strong significance, the results show biologically meaningful trends across groups. This finding is consistent with previous research indicating LRG1’s important role in various conditions, including cancer, ocular, cardiovascular, and neurodegenerative disease [40]. Our data suggest that LRG1 carried by tear EVs may serve as a non-invasive marker for AD diagnostics; however, more clinical samples are needed for further validation.

Discussion

Limited clinical samples present significant challenges in various research fields, particularly in omics analysis within biomedical and clinical studies. These fields rely on robust datasets to minimize noise and uncover meaningful biological insights. Artificial intelligence is transforming the way research is conducted in the biomedical sciences [41, 42]. In this work, we developed ABLEDx for diagnosing neurodegenerative diseases, which enhances data analysis of tear-EV proteomics. The samples generated using the cVAE model closely resemble the original data distribution while introducing variability. As a result, we can produce new samples that retain the latent features of the original data while introducing diversity, thereby increasing the robustness of the dataset. Unlike conventional approaches that directly utilize the latent space [43, 44], our method leverages the decoder for data augmentation. This enables us to create a higher-dimensional latent space that retains richer information and more complex features from the original data. By providing additional training data, cVAEs can help mitigate overfitting in machine-learning models built on limited clinical samples.

Our decoder-centric strategy offers two main advantages. First, it significantly reduces the dependency on large patient cohorts, a critical challenge in rare disease research and early diagnostic studies where samples are often scarce. Our approach generates statistically plausible “synthetic” samples that maintain the original dataset’s underlying distribution and modular structure, consequently enhancing the statistical power of downstream analyses. Second, the decoder-driven generation method is compatible with various established bioinformatics tools. Unlike latent embeddings, which often require additional interpretation or reconstruction steps, our synthetic samples can be directly integrated into standard analytical pipelines. This plug-and-play functionality simplifies workflows and encourages the adoption of our method across diverse omics studies, including genomics, proteomics, metabolomics, and transcriptomics, where sample scarcity and data noise are persistent challenges.

Despite these advantages, one limitation of our cVAEs is their reliance on expert-driven qualitative assessments to evaluate the model’s generative performance. These assessments are valuable but inherently subjective when determining whether synthetic data relationships align with established biological knowledge and clinical intuition. A more robust statistical framework is needed to assess the fidelity and uncertainty of generated samples quantitatively. Potential approaches could include formal hypothesis testing on latent representations, Bayesian calibration techniques, or novel metrics designed to measure how effectively synthetic distributions capture the original data’s complex correlation structures, network topologies, and pathway enrichments. Additionally, future work should focus on integrating robust uncertainty quantification into the generation pipeline. This capability would enable researchers to evaluate the reliability of synthetic samples and apply confidence thresholds before utilizing them in critical applications, such as clinical decision-making. Such advancements will enhance the utility and credibility of our approach, facilitating its broader adoption in high-stakes biomedical research.

Current results suggest that tear-EVs are derived from several organ sources, including the glands, retina, eye, and brain, and may play essential roles in maintaining ocular and brain health [24]. Our findings indicate that tear-EVs from patients with neurodegenerative diseases are closely associated with neuronal and glial cells (Fig. 1). This connection highlights their potential as biomarkers, reflecting the cellular and molecular changes in the nervous system. The strong relationship between tear-EVs and these cell types underscores their relevance in studying disease mechanisms and opens new avenues for non-invasive diagnostics and monitoring of neurodegenerative disorders. Furthermore, a recent study shows that β-amyloid from the brain can be transported to the eyes via cerebrospinal fluid (CSF) along the optic nerve, strongly supporting the idea that the eyes provide a unique window for detecting pathological changes in the brain [25]. In this study, β-amyloid was not detected in tear EVs, unlike in plasma EVs [23]. Interestingly, the amyloid precursor protein (APP) was significantly downregulated in AD patients compared to the healthy control group (Data S1). This finding suggests that alterations in APP levels in tear-EVs may be closely related to AD pathology, and the observed APP dysregulation highlights its role in the disease process and its potential for further exploration in the context of non-invasive diagnostics.

Despite this minimal dataset, the cVAE model was able to learn class-specific latent distributions and generate augmented samples that effectively enhanced downstream analyses. Notably, even with only four real samples per class, the augmented data improved the reproducibility of differential protein signatures, allowing the model to recapitulate key clinical characteristics of these disease types. The validation assay demonstrated strong concordance with the ABLEDx predictions (Fig. 4). Although KRAS was identified as a top-ranked candidate by our AI-based analysis, we did not pursue ELISA-based validation due to its markedly lower abundance in the proteomic dataset compared to LRG1 and HSPG2. Moreover, KRAS is a small GTPase primarily localized to intracellular membranes, and its presence on the surface or within tear-derived EVs is expected to be limited, making reliable ELISA detection challenging. Therefore, experimental validation focused on LRG1 and HSPG2, which exhibit higher EV-associated expression and are supported by robust, commercially available ELISA assays. In summary, our work represents an early-stage effort to identify tear–EV–derived proteins with differential expression in AD and PD. The diagnostic potency of LRG1 and HSPG2 will be further evaluated in larger cohorts with dedicated training and validation sets.

While the potential of tear EVs as diagnostic markers for neurodegenerative disorders is promising, further research is essential to fully understand the mechanisms involved in their biogenesis and their association with disease states. Exploring how these vesicles are formed and their specific cellular sources is necessary. In conclusion, we have developed the ABLEDx platform to facilitate the early diagnosis of neurodegenerative diseases using tear-derived EVs. The cVAE model has demonstrated its ability to capture essential characteristics while efficiently minimizing noise, effectively addressing the challenges of poor sampling and noise in EV proteomic data for disease diagnostics. Our model is versatile and capable of handling various omics data, including proteomics, transcriptomics, and metabolomics. This approach meets a critical need to uncover the molecular characteristics of human diseases through comprehensive omics research. Translating these experimental advances into clinical practice remains a significant challenge. It will require a multidisciplinary effort that integrates cell biology, bioinformatics, and clinical research to realize their diagnostic potential fully.

Conclusions

Tears serve as a valuable biofluid containing a high concentration of EVs, which hold great potential for the early detection of neurological diseases. However, isolating EVs and analyzing omics data present significant challenges, including high variability and noise—issues that are particularly pronounced in rare disease research, where clinical samples are often limited. In this study, we developed the ABLEDx platform for the early diagnosis of neurodegenerative diseases using tear-derived EVs. Our findings demonstrate that the conditional variational autoencoder framework effectively captures essential disease-related features while mitigating noise, thereby addressing key limitations associated with sample scarcity and variability in EV proteomic data. Thus, the proposed cVAE-based workflow has the potential to moderately shorten the time-to-result and reduce the frequency of repeat runs compared with current practice. Furthermore, our model provides a versatile and scalable solution applicable to various omics datasets, including proteomics, transcriptomics, and metabolomics. By enabling the comprehensive discovery of molecular signatures, ABLEDx represents a powerful tool for advancing precision diagnostics and broadening the understanding of disease mechanisms through integrative omics analysis.

Abbreviations

AD

Alzheimer’s Disease

cVAE

Conditional variational autoencoders

DDA

Data-dependent acquisition

EVs

Extracellular vesicles

ICA

Independent component analysis

MG

Myasthenia Gravis

NMOSD

Neuromyelitis Optica Spectrum Disorders

NTA

Nanoparticle tracking analysis

ON

Optic neuritis

PBS

Phosphate-buffered saline

PCA

Principal component analysis

PD

Parkinson’s Disease

UMAP

Uniform manifold approximation and projection

WGCNA

Weighted correlation network analysis

Author contributions

The manuscript was written through the contributions of all authors. All authors have approved the final version of the manuscript. †These authors contributed equally. S.W. and Z.L. collected and analyzed the clinical samples. Q.Z., P.H., Z.L., Q.S. and X.Z. processed the whole experiment. Q.Z., S.W., P.H. and F.L. prepared the manuscript and Figures. L.P.L. and F.L. edited the manuscript. All experiments were conducted under the supervision of L.P.L. and F.L.

Funding

The work was primarily supported by a research fund provided by the National Natural Science Foundation of China 22474092 (to QZ), and the Shenzhen Science and Technology Innovation Commission 202323257 (to PH).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Tear samples were collected from patients according to protocols approved by the Research Ethics Committee at Xi’an No. 1 Hospital. Informed consent was obtained from all participants.

Consent for publication

All authors have read and approved the final manuscript.

Competing interests

Dr. Fei Liu is a co-founder of Huixin Lifetech.