DADA-EV: domain-adaptive diffusion autoencoder for estimating tissue- and cell-type-specific origin in extracellular vesicle transcriptomes.

DADA-EV: domain-adaptive diffusion autoencoder for estimating tissue- and cell-type-specific origin in extracellular vesicle transcriptomes - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Brief Bioinform . 2026 Jun 4;27(3):bbag281. doi: 10.1093/bib/bbag281 DADA-EV: domain-adaptive diffusion autoencoder for estimating tissue- and cell-type-specific origin in extracellular vesicle transcriptomes Shuilin Liao 1, 2, 2 , Haoxiang Yang 3, 2 , Shuting Xiao 4 , Zhiwen Li 5 , Qi Zhu 6 , Le Li 7, 8 , Shanghui Lu 9, 10 , Yong Liang 11, ✉ , Ming Dong 12, 13, ✉ Author information Article notes Copyright and License information 1 Faculty of Innovation Engineering, Macau University of Science and Technology, Avenida Wai Long, Taipa, Macau 999078, China 2 Guangzhou National Laboratory, No. 9 XingDaoHuanBei Road, Guangzhou International Bio Island, Guangzhou, Guangdong Province 510005, China 3 Guangzhou National Laboratory, No. 9 XingDaoHuanBei Road, Guangzhou International Bio Island, Guangzhou, Guangdong Province 510005, China 4 State Key Laboratory of Respiratory Disease, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou Medical University, Guangzhou, Guangdong 510120, China 5 Guangzhou Exope Medical Biotechnology Co., Ltd, China 6 State Key Laboratory of Respiratory Disease, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou Medical University, Guangzhou, Guangdong 510120, China 7 Faculty of Innovation Engineering, Macau University of Science and Technology, Avenida Wai Long, Taipa, Macau 999078, China 8 Basic Teaching Department, Zhuhai Campus of Zunyi Medical University, No. 368, Jinwan Road, Jinwan District, Zhuhai, Guangdong Province 519000, China 9 Faculty of Innovation Engineering, Macau University of Science and Technology, Avenida Wai Long, Taipa, Macau 999078, China 10 School of Mathematics and Physics, Hechi University, No.42, Longjiang, Hechi, Guangxi 546300, China 11 Chinese Medicine Guangdong Laboratory, No. 2522, Huandao North Road, Hengqin New Area, Xiangzhou District, Zhuhai, Guangdong Province 519031, China 12 Guangzhou National Laboratory, No. 9 XingDaoHuanBei Road, Guangzhou International Bio Island, Guangzhou, Guangdong Province 510005, China 13 State Key Laboratory of Respiratory Disease, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou Medical University, Guangzhou, Guangdong 510120, China ✉ Corresponding authors. Ming Dong, Guangzhou National Laboratory, No. 9 Xing Dao Huan Bei Road, Guangzhou International Bio Island, Guangzhou, Guangdong Province 510005, China. E-mail: dong_ming@gzlab.ca.cn ; Yong Liang, Chinese Medicine Guangdong Laboratory, Hengqin, Guangdong 519031, China. E-mail: yongliangresearch@gmail.com 2 Shuilin Liao and Haoxiang Yang co-first authors. Received 2025 Oct 19; Revised 2026 Apr 10; Accepted 2026 May 11; Collection date 2026 May. © The Author(s) 2026. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License ( https://creativecommons.org/licenses/by-nc/4.0/ ), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact reprints@oup.com for reprints and translation rights for reprints. All other permissions can be obtained through our RightsLink service via the Permissions link on the article page on our site—for further information please contact journals.permissions@oup.com. PMC Copyright notice PMCID: PMC13235744  PMID: 42242685 Abstract Tracing the tissue and cell-type origins of extracellular vesicles (EVs) in blood is critical for liquid biopsy and precision medicine, yet existing deconvolution methods remain limited by the need for labor-intensive reference signatures and poor adaptability to distribution shifts between tissue/cell-type datasets and EV transcriptomes. We introduce DADA-EV (Domain-Adaptive Diffusion Autoencoder for EVs), a hybrid deep learning framework that combines an autoencoder backbone with a generative simulation module and adversarial domain adaptation. DADA-EV features three key innovations: (1) a reference-free design that eliminates reliance on predefined signatures; (2) cross-domain generalization by aligning feature distributions between source (tissue/cell-type) and target (EV) domain; and (3) reduced dependence on source data during target-domain training. Extensive evaluations on pseudo-EV data show that DADA-EV consistently outperforms existing approaches, yielding accurate fraction estimates across diverse tissues and gene sets. Validation using in vitro cell-line mixtures further confirms its reliability in resolving complex compositions, demonstrating high sensitivity in detecting low-abundance targets. Applied to real EV transcriptomes, it reveals tissue- and cell-type heterogeneity across patient groups. In summary, DADA-EV provides a robust, reference-free, and generalizable solution for EV origin tracing, with strong potential to advance diagnosis, prognosis, and treatment monitoring via liquid biopsy. Keywords: extracellular vesicles, origin tracing, transcriptome deconvolution, deep learning, domain adaptation Graphical Abstract Graphical Abstract. Open in a new tab Introduction Extracellular vesicles (EVs) are nanoscale, membrane-bound particles secreted by most eukaryotic cells, carrying diverse biomolecules such as proteins, lipids, metabolites, and RNAs [ 1–3 ]. They mediate intercellular communication and regulate processes ranging from immune responses to tumor progression [ 4 ]. Plasma-derived EVs, in particular, encapsulate signals from diverse tissues, thereby reflecting the body’s physiological or pathological state [ 5–7 ]. This makes EVs a promising, minimally invasive source for molecular profiling in precision medicine, where tracing their tissue or cell origins could aid diagnosis, prognosis, and treatment monitoring [ 8 , 9 ]. In the broader field of transcriptomics, computational deconvolution has evolved significantly to resolve cellular heterogeneity [ 10–14 ]. Mainstream approaches predominantly rely on machine learning (ML) algorithms, such as Support Vector Regression (e.g., CIBERSORT [ 10 ]) and non-negative least squares (e.g., MuSiC [ 15 ]). These methods operate on the core concept that bulk gene expression is a linear combination of cell-type-specific signatures, requiring curated marker gene lists or reference matrices to utilize regression models for estimating cell proportions. While effective for tissues, their accuracy is highly dependent on the quality and completeness of the reference signature matrix [ 16 ]. To overcome the reliance on static references, deep learning (DL) strategies have recently gained traction. Methods like Scaden [ 17 ] and TAPE [ 15 ] employ deep neural networks and autoencoder architectures to infer cell fractions. By training on large-scale simulated datasets, these reference-free approaches can capture complex, non-linear relationships within gene expression data, offering improved robustness over traditional linear models. Despite these advancements, few deconvolution methods are specifically tailored for EVs, and directly applying existing tissue-based algorithms presents significant challenges. Among the limited EV-specific efforts, EV-origin [ 5 ] represents a pioneering approach. However, it remains a reference-based method that assumes a linear relationship between cellular and EV transcriptomes, leaving it vulnerable to reference bias. Conversely, while general DL methods like TAPE [ 15 ] eliminate the need for references, they fail to account for the inherent biological disparity between cellular (source) and EV (target) domains. EV transcriptomes differ substantially from their cells of origin due to selective RNA packaging, distinct RNA biotype enrichment, and variations in RNA stability [ 2 , 18 ]. This results in a ‘domain shift’, a distributional difference between the training data (cell/tissue RNA-seq) and the inference data (EV transcriptomes). Current methods lack formal domain adaptation mechanisms to bridge this gap, leading to biased predictions when cell-trained models are applied to EV data. To address these challenges, we developed Domain-Adaptive Diffusion Autoencoder for EVs (DADA-EV), a novel deconvolution framework specifically designed to bridge the gap between tissue- or cell-type data and EV transcriptomes. DADA-EV integrates a deep autoencoder backbone with a diffusion-based generative module and adversarial domain adaptation strategies. By leveraging simulated data derived from tissue and cell transcriptomes, our method effectively transfers knowledge from the cellular domain to the EV domain. The key contributions of this study are summarized as follows: (1) Novel framework design: we propose DADA-EV, the first deep learning-based deconvolution model that explicitly incorporates domain adaptation to handle the biological discrepancy between tissue or cellular sources and EV targets. (2) Reference-free learning: Unlike traditional approaches that rely on pre-constructed signature matrices or curated marker genes, DADA-EV automatically learns complex feature representations directly from the data, eliminating the dependency on static references. (3) Source-data-free inference: By leveraging a diffusion-based generative module to capture the latent distribution of the source domain, our framework enables domain alignment and inference on target data without requiring concurrent access to the original source data. (4) Superior performance: DADA-EV demonstrates significantly higher precision and robustness in predicting tissue- or cell-type proportions from pseudo and real EV transcriptome data compared to both state-of-the-art general deconvolution methods and existing EV-specific approaches. Method Datasets and preprocess Data collection and domain definition We downloaded RNA-seq data of different tissues from two data sources: the GTEx Portal ( https://gtexportal.org ) and the HPA ( https://www.proteinatlas.org ). From both datasets, we downloaded the transcript per million data. Specifically, GTEx includes data from 31 tissue types and 48 231 tissue samples. HPA includes data from 39 tissue types and 187 human tissue samples. We selected only the 23 tissue types that are common to both datasets ( Table S1 ). In this study, we treated GTEx data as source-domain data and HPA data as target-domain data. Two PBMC single cell (scRNA-seq) datasets were obtained from 10× Genomics ( https://support.10xgenomics.com/single-cell-gene-expression/datasets ), where they were identified as ‘6k PBMCs from a Healthy Donor’ and ‘Donor C Frozen PBMCs’ , respectively. Here, we used the 6 k PBMCs dataset as the source-domain data and the Donor C dataset as the target-domain data. Real plasma EV RNA-seq data were obtained from exoRBase 2.0 ( http://www.exoRBase.org ) [ 3 ]. Five categories were included, each with more than 100 samples. In total, 664 samples were selected for deconvolution ( Table S2 ), specifically including 118 samples for healthy state (Healthy), 130 samples for benign disease (Benign), 140 samples for breast cancer (BRCA), 112 samples for hepatocellular carcinoma (HCC), and 164 samples for pancreatic adenocarcinoma (PAAD). In relation to the source-domain data used for training, the real EV data for inferring was considered as the target-domain data. Simulation of bio-enhanced pseudo-EV data DADA-EV requires a large number of training RNA-seq samples with known tissue or cell fractions (ground truth), referred to as pseudo-EV data. Our simulation framework builds upon established pseudo-bulk methodologies [ 15 ] but extends them to incorporate EV-specific biological discrepancies, including selective packaging, non-Gaussian distributional skewness, and signal-dependent zero-inflation. Let N denote the total number of pseudo-EV samples, denote the number of tissue/cell types and denote the number of genes. Cell Fraction Generation and Sparsity : First, we generated a proportion matrix using a Dirichlet distribution. To mimic the realistic sparsity of EV contents (where not all tissue or cell type release EVs into a specific sample), we applied a random sparsity mask. For each pseudo-EV sample, specific tissue or cell types were randomly assigned a fraction of zero with probability (set to 0.5), and the remaining proportions were renormalized such that for each pseudo-EV sample i ( i = 1, …, N ). Baseline Pseudo-bulk Aggregation : To construct the baseline expression matrix , we employed a stratified sampling approach. For the i -th pseudo-EV sample, the specific count of tissues or cells for type k , denoted as , was calculated based on a fixed total sample count : (1) We then randomly sampled single-tissue or cell expression profiles from the source data (e.g., GTEx datasets) for each type k . For example, assuming and two tissue types (Heart and Lung) with generated proportions , we would draw 350 Heart profiles (500 * 0.7) and 150 Lung profiles (500 * 0.3) from the source dataset. The baseline pseudo-bulk profile for sample i (denoted as ) is defined as the aggregation of these selected samples: (2) where is the set of indices for the selected single tissue/cell profiles of type k for sample i , and is the expression vector of tissue/ cell c . Bio-enhanced EV Transformation : To bridge the domain gap between cellular and EV data, we transformed into bio-enhanced pseudo-EV data ( ) through a multi-stage stochastic process: (a) Selective Packaging Simulation (Gene-specific Scaling) EV cargo loading is a selective and regulated process mediated by RNA-binding proteins that recognize specific sequence motifs, leading to preferential enrichment or exclusion of transcripts [ 19 ]. Chikara et al. [ 20 ] observed that chemical abundances across numerous cells follow a Log-Normal distribution. Therefore, based on these observations, we modeled gene-specific packaging coefficients using a Log-Normal distribution. We modeled the preferential packaging or exclusion of specific transcripts into EVs. For the j -th gene ( j = 1, …, M ), we assigned a packaging coefficient drawn from a Log-Normal distribution: (3) where controls the variance of selective packaging. Genes with represent enriched transcripts, while represent depleted transcripts. The packaged expression profile for gene j in sample i is calculated as (4) (b) Distributional skewness and overdispersion (Gamma noise) We observed that real EV data, derived from blood exosome transcriptomes in various disease conditions, is characterized by a long-tailed, non-Gaussian distribution ( Fig. S2A ). On the other hand, RNA-seq data are well known to exhibit overdispersion, commonly modeled using the Negative Binomial (NB) distribution [ 21 , 22 ]. The NB distribution can be interpreted as a Poisson–Gamma mixture, where the Gamma component captures biological variability across genes and samples. Motivated by this interpretation, we introduced multiplicative Gamma noise to simulate dispersion and skewness observed in EV transcriptomic data: (5) (6) Here, the shape parameter controls the skewness (e.g., creates a heavy-tailed distribution, while larger values approach a Gaussian). (c) Signal-dependent Dropout (Zero-Inflation) Due to the low abundance, limited copy numbers, and heterogeneous distribution of nucleic acid targets within EVs, coupled with the limited sensitivity of isolation and detection platforms, some genuine molecules may fall below the limit of detection [ 23 ]. This results in the ‘dropout’ phenomenon, where genes are present but fail to be detected. Similar zero-inflation phenomena have been extensively studied in single-cell RNA-seq data [ 24–26 ]. Rather than assuming random or uniform dropout, previous studies have shown that dropout events arise from a combination of low transcript abundance and limited capture efficiency, leading to a strong dependence between detection probability and underlying expression level. In particular, genes with low expression are more likely to be missed during reverse transcription and amplification, resulting in higher dropout rates. This motivates the use of a signal-dependent probabilistic model in our study. In this study, we adopted a logistic (sigmoid) function to model the retention probability, and the retention probability for gene j in sample i is defined as follows: First, a dynamic threshold is computed based on the global distribution of log-transformed expression values: (7) where defines the dropout strength, and denotes the 20th percentile of all elements in the matrix . The retention probability for each entry is then given by a sigmoid function: (8) where is the slope parameter controlling the sharpness of the transition (set to in this study). A Bernoulli mask is then sampled based on this retention probability: (9) To simulate sequencing depth fluctuations, a global scaling factor is applied to each sample. The final simulated expression value is obtained as: (10) In total, we generated 5000 samples for the source-domain pseudo-EV data and 500 samples for the target-domain pseudo-EV data, each containing known fractional compositions. Prior to model training, the source-domain dataset was partitioned into training and testing sets in an 8:2 ratio. Data preprocessing and normalization To ensure optimal performance and convergence of the deep learning model, we implemented a standardized preprocessing pipeline adapted from the TAPE framework [ 15 ]. First, we addressed the high dynamic range of transcriptomic data by performing a log-transformation. Specifically, we calculated for every entry in the expression matrix (for both simulated and real bulk data). This transformation stabilizes variance and prevents numerical issues associated with zero values. Subsequently, to align the feature distributions and facilitate gradient descent during training, we applied feature scaling. We normalized the expression data linearly into the range [0, 1]. This step prevents genes with inherently high absolute expression levels from dominating the loss function. It is important to note that identical preprocessing steps were applied to the source-domain (simulated) and target-domain (real) data to maintain consistency for the domain adaptation task. Problem definition We formally describe our data and tasks as follows. The source-domain pseudo-EV data is defined as a matrix with corresponding known fraction . We split the source data into a training set ( ) and a testing set ( ). The target-domain pseudo-EV data is denoted as with corresponding known fraction , while real EV data is represented as (where fractions are unknown). Our objectives using the DADA-EV model are twofold: (1) Reconstruction: to generate reconstructed expression profiles, denoted as and . These reconstructions are utilized for model optimization (via reconstruction loss). (2) Prediction: to estimate the cell-type fractions for the source-test, target, and real datasets, denoted as and , respectively. Architecture of DADA-EV Inspired by the explicit mitigation of distribution shifts in the scpDeconv model [ 27 ], we proposed DADA-EV, a domain-adaptive deep learning framework designed to explicitly mitigate distribution shifts in deconvolution without relying on original reference data during the adaptation phase. The overall model architecture is shown in Fig. 1 . It contains three stages: (1) Pre-training for source-domain; (2) Domain-adversarial training and deconvolution; (3) Inference for target-domain. Figure 1. Open in a new tab The overall architecture of DADA-EV. DADA-EV contains three stages: (1) Pre-training for source-domain, where the model learns latent representations from source-domain pseudo-EV data, jointly optimizing reconstruction and fraction prediction; a diffusion generator is trained to produce source-like embeddings from noise. (2) Domain-adversarial training and deconvolution, where source and target distributions are aligned via a discriminator with gradient reversal. (3) Inference for target-domain, where the trained encoder, decoder, and predictor are applied to target-domain pseudo-EV or real EV data to reconstruct expression profiles and estimate proportions. (1) Stage 1: pre-training for source-domain. In this stage, only source-domain data is used for training. High-dimensional gene expression data from the source-domain is input into an encoder to obtain a latent representation. Let denote the high-dimensional gene expression profile from the source domain. We first map to a lower-dimensional latent representation via a non-linear encoder: (11) This latent embedding serves as the input for three parallel branches. The first branch is a predictor module that estimates the cell-type fractions , while the second is a decoder module that reconstructs the gene expression profile . (12) (13) To guide the learning process, we optimize the network by minimizing a composite loss function consisting of and the fraction prediction loss (MSE between and the ground truth ) and the reconstruction loss (MSE between and ). Then, we freeze the encoder, decoder, and predictor. (14) (15) In the third branch, we train a lightweight diffusion generator to synthesize latent representations directly from noise, ensuring that source-domain characteristics can be reproduced without accessing the original data (more details are in Diffusion generator module section) (16) (2) Stage 2: domain-adversarial training and deconvolution. In this stage, we address the distribution shift between the source and target domains. We employ a domain-adversarial learning strategy [ 28 ] to fine-tune the model. Concretely, firstly, we get the synthetic source embeddings generated by the diffusion module, the reconstructed expression data of by the decoder, and the target-domain embeddings by the encoder. (17) (18) (19) Next, we use the true domain labels and , along with the opposite domain labels and , which have the same number of samples. These two types of labels are merged into and , respectively. We then train a domain discriminator to obtain the predicted labels and . The binary cross-entropy ( bce ) loss is then computed between the discriminator’s predictions and the true domain labels. (20) Additionally, the latent embedding of target-domain data and the reconstructed target-domain expression data are obtained. The reconstruction loss, defined as the MSE between target-domain inputs and was calculated. (21) (22) (23) (3) Stage 3: inference for target domain. After domain-adversarial training and deconvolution in stage 2, the trained encoder, decoder, and predictor are finally fixed and utilized to reconstruct the expression matrix and infer the tissue or cell-type proportions of target-domain samples. The equations for the architecture in stage 3 are: (24) (25) (26) Diffusion generator module To synthesize high-fidelity latent representations of the source domain, we designed a conditional diffusion probabilistic model parametrized by a neural network . We adopt a Residual Network (ResNet) backbone to effectively capture complex data distributions. The main framework of the diffusion generator is shown in Fig. S1 . The goal is to learn how to reverse the corruption process caused by adding noise over time, using a framework adapted from the diffusion probabilistic model [ 29 ]. To avoid confusion with the target-domain notation ( ), we denote the latent variable in the diffusion process as and the timestep as . The diffusion framework contains two processes: the forward diffusion process and the reverse diffusion process. The forward process is a fixed Markov chain that gradually adds Gaussian noise into the clean latent data (where ) over timesteps. We utilize a linear variance schedule , where increases linearly from to 0.02. To facilitate sampling at an arbitrary timestep k , we defined and the cumulative product . Consequently, the noisy sample can be expressed in closed form as: (27) The reverse process iteratively denoises using time-dependent conditioning. The goal of the training is to learn the reverse process to denoise samples. Instead of directly predicting the clean data or the distribution mean, our network is trained to predict the noise component added at each timestep k . The training objective minimizes the mean squared error between the actual Gaussian noise and the predicted noise: (28) where is the noisy input derived from Equation ( 27 ). To effectively capture the temporal dynamics, the network employs a stack of time-aware residual blocks. As implemented in our architecture, the timestep k is first mapped to a high-dimensional embedding vector and projected to match the feature dimensions. This time embedding is injected into each residual block to shift the intermediate features, allowing the model to dynamically adjust its denoising strategy based on the current noise level. Once the model is trained, we can synthesize new data by reversing the diffusion process. The generation starts by sampling a latent vector from the standard normal distribution, . We then iteratively refine this vector from k = K down to 1 using the learned noise predictor. The sampling update rule is derived as: (29) where is random noise introduced to maintain the stochasticity of the generative process (with =0 for k = 1), and is the posterior variance variance term. Through K iterations, the initial Gaussian noise is gradually transformed into a high-fidelity latent representation that follows the distribution of the source domain. Module network structure The detailed architecture and hyperparameters for all modules are summarized in Table S3 . The encoder maps the high-dimensional gene expression input into a low-dimensional latent embedding . It consists of three fully connected (FC) layers. The first layer maps input to 1024 dimensions followed by batch normalization and SiLU activation. The second layer compresses features to 512 dimensions, utilizing layer normalization, SiLU activation, and dropout for regularization. A final linear layer maps the features to the latent dimension . Designed to reconstruct the input from , the decoder consists of three FC layers. The first layer maps the latent features to 512 dimensions followed by CELU activation. The second layer expands features to 1024 dimensions, utilizing CELU activation. The final linear layer projects the features back to the original gene expression input dimension for reconstruction. The predictor operates on the latent vector to estimate cell type fractions. It is composed of three FC layers: the first maps the latent input to 128 dimensions, and the second compresses it to 64 dimensions. Both layers utilize CELU activation and dropout for regularization. The final linear layer projects the features to an output dimension corresponding to the number of tissue or cell types. The raw output is subsequently processed by ReLU activation and normalization to ensure the predicted fractions are non-negative and sum to one. To align the feature distributions across domains, the discriminator takes the latent vector as input. It comprises three FC layers. The first layer maps the latent input to 128 dimensions, followed by CELU activation and dropout. The second layer compresses the features to 64 dimensions, utilizing CELU activation and dropout. The final linear layer yields a scalar output, which is passed through a sigmoid activation function to predict the probability of the domain source. The noise prediction network within diffusion generator module estimates the noise component added to the latent data, conditioning on the noisy state and the timestep . The architecture incorporates a time embedding module and a residual backbone. First, the timestep is mapped to sinusoidal positional encodings (128 dimensions) and processed by a multi-layer perceptron consisting of a linear layer, SiLU activation, and a second linear layer to produce a 256-dimensional time embedding. In the main branch, the noisy input is projected to a 512-dimensional feature space via an initial linear layer. This is followed by three stacked time-aware residual blocks, which process the features while integrating the time embedding to modulate the signal. The final linear layer projects the 512-dimensional features back to the original latent dimension to yield the predicted noise. Ablation study of DADA-EV components To fully evaluate the contribution of all components in DADA-EV, we compared the full model against three ablation variants. The specific configurations of these variants are detailed below. Firstly, to establish a reliable baseline and estimate the performance domain adaptation brings, we utilized a standard autoencoder approach (referred to as ‘Stage1AE’). This model corresponds strictly to the pre-training phase (Stage 1) of DADA-EV. In this setting, pseudo-EV data is processed directly by the model trained solely in Stage 1, without any subsequent domain adaptation steps. The second variant (referred to as ‘SIG’) investigates the method of deriving tissue or cell type fractions. In this approach, the model is pre-trained on source-domain data (identical to Stage1AE). We then compute a signature matrix using the reconstructed expression profiles and predicted fractions from the source training data. Specifically, is derived via the pseudoinverse of the predicted fractions, as shown in Equation ( 30 ). During inference, test data is passed through the encoder to generate latent embeddings , which are then decoded to obtain reconstructed expression profiles . Finally, fractions for the test data are estimated by projecting onto the pseudoinverse of the pre-computed signature matrix . The process is defined by the following equations: (30) (31) (32) (33) Both Stage1AE and signature matrix (SIG) operate without domain adaptation. They are evaluated in two scenarios: (1) Source-domain evaluation: Trained on and tested on . (2) Target-domain evaluation: Trained on and tested on . The final ablation variant, denoted as ‘DADA_noDF,’ retains the DADA-EV architecture but bypasses the diffusion generation step. In this model, source-domain latent embeddings are derived directly from the input data through the Encoder rather than being synthesized from noise by the diffusion generator. Consequently, Equation ( 17 ) is replaced by Equation ( 31 ). Unlike the previous variants, DADA_noDF undergoes domain adaptation. It is evaluated exclusively on target-domain data (trained on , tested on ). (34) Evaluation metrics As used in other deconvolution methods [ 15 , 17 , 27 ], we employed two metrics to evaluate the performance of the model: Lin’s concordance correlation coefficient (CCC), mean absolute error (MAE). Lin’s CCC evaluates the agreement between two variables, with values ranging from −1 to 1, reflecting the consistency of trends between them. The formula is calculated as follows: (35) The MAE quantifies the magnitude of prediction errors and is defined as: (36) In these equations, represents the sample number. and are the ground-truth values and predicted values for the sample, respectively. denotes the ground-truth values and are the predicted value for all samples. and represent their mean values, while and denote their variances. represents the covariance between and . In general, higher CCC values combined with lower MAE values indicate better deconvolution performance. To further assess the monotonic relationship between the predicted and ground-truth proportions, we utilized Spearman’s rank correlation coefficient (Rho). (37) (38) where denotes the rank of the variable, and represents the difference between the ranks of the ground-truth and predicted values for the i -th sample. Finally, to evaluate the model’s detection accuracy and binary classification performance, we employed the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC). The ROC curve plots the True Positive Rate (TPR) against the False Positive Rate (FPR) at various threshold settings. The components of the ROC curve are defined as follows: TPR: Also known as sensitivity or recall, TPR measures the proportion of actual positives that are correctly identified. (39) FPR: This measures the proportion of actual negatives that are incorrectly identified as positives. (40) where , , , and represent True Positives, False Positives, True Negatives, and False Negatives, respectively. The AUC provides an aggregate measure of performance across all possible classification thresholds, where an AUC of 1.0 represents a perfect classifier and 0.5 represents a random classifier. (41) where represents the True Positive Rate as a function of the False Positive Rate, and denotes the differential element of the False Positive Rate along the x-axis. In vitro controlled EV mixture experiments To establish in vitro benchmarks with ground-truth compositions, we generated synthetic EV mixtures derived from a diverse panel of human cell lines. This panel included pancreatic ductal epithelial cells (hTERT-HPNE), normal liver cells (L02), human embryonic kidney cells (293 T), and bronchial epithelial cells (BEAS-2B). Detailed procedures regarding vitro cell culture, the isolation of EVs from both cell culture and sample mixing are provided in the Supplementary Methods. EVs were individually isolated from the conditioned media of each cell line and quantified using Nanoparticle Tracking Analysis. To create defined mixtures, purified EVs were combined based on strict particle count ratios. We designed three distinct mixtures to simulate various biological scenarios. Mixed EV samples were then subjected to RNA extraction and transcriptome sequencing (see Supplementary Methods). Results Validation of the diffusion module performance using simulated cross-database scenarios We first analyzed the distributional characteristics of real EV expression data, observing that it follows a long-tailed distribution heavily populated with zero values (zero-inflated, Fig. S2A ). However, given the lack of ground truth regarding tissue/cell proportions in real EV data, we leveraged two external databases (GTEx and HPA) to simulate a domain adaptation task. Specifically, one dataset served as the source data and the other as the target data to mimic the process of a model trained on simulated data (source) predicting real EV data (target) ( Fig. S2B ). Firstly, to evaluate the robustness of our simulation framework, we conducted a sensitivity analysis by systematically varying key parameters, including the variance of selective packaging , the degree of overdispersion , and the dropout strength . The distributional similarity between simulated and real EV data was quantified using the Wasserstein distance (WD), where smaller WD values indicate higher similarity ( Fig. S2C ). The parameter ranges were set as follows: , , and . Our results revealed the following trends: (i) Larger values yielded smaller WD, indicating greater similarity to real data. This suggests that higher inter-gene variability more closely recapitulates the heterogeneity observed in real EV expression profiles. (ii) controls the global dispersion and tail behavior of the expression distribution. Larger values led to larger WD (poorer similarity), whereas smaller values generated heavier-tailed distributions that better matched real EV data. This indicates that simulating heavy-tailed expression distributions is critical for faithfully reproducing real data characteristics. (iii) determines the sparsity and zero-inflation level of the simulated data. Larger values resulted in smaller WD, indicating that stronger zero inflation improves similarity to real EV data. Notably, when increased from 1.0 to 1.5, WD exhibited a sharp decline, suggesting the existence of a critical threshold beyond which zero inflation becomes essential for capturing the sparsity pattern of real EV data. Using this simulation pipeline, we adopted two distinct strategies to generate datasets representing different distributional characteristics: (i) Zero-inflated / Long-tailed Distribution ( =1.2, =0.8, ): This configuration mimics high-noise data characterized by severe sparsity and heavy skewness. (ii) Gaussian Distribution ( =0.6, =2.5, ): This configuration represents a more centered, symmetric distribution with moderate noise. This process yields synthetic data that retains the ground-truth tissue- or cell-type proportions while introducing the distributional shifts, sparsity, and non-linear distortions characteristic of real EV transcriptomics. We designed three simulation scenarios ( Fig. S2D ): (1) Scene 1, where real EV data follows a zero-heavy long-tailed distribution, while both GTEx pseudo-EV and HPA pseudo-EV follow normal distributions; (2) Scene 2, where real EV and HPA pseudo-EV follow zero-heavy long-tailed distributions, while GTEx pseudo-EV follows a normal distribution; and (3) Scene 3, where real EV, GTEx pseudo-EV, and HPA pseudo-EV all exhibit zero-heavy long-tailed distributions. To evaluate the performance of the data generated by the diffusion generator module across these distinct distributions, we calculated the WD between the generated samples and the original data samples. We compared our approach against two simpler latent modeling techniques: Simple Gaussian and Gaussian Mixture Model. The results showed that, whether dealing with normal distributions or more complex long-tailed distributions, the diffusion method achieved smaller distances compared to the other two methods ( Fig. S2E ), demonstrating the superior capability of the diffusion model in accurately capturing and reconstructing complex data distributions. Furthermore, among the three constructed scenarios, we found that prediction performance was highest in Scene 1 and lowest in Scene 3 ( Fig. S2F ). Notably, across all three scenarios, the performance of the diffusion model remained consistent with baselines using original data as input. In contrast, the performance of the alternative methods degraded significantly in the complex Scene 3. These findings further demonstrate the superiority and robustness of using diffusion to implement source-free domain adaptation. Superior performance on pseudo-EV data across different domains We further tested two ablations within the source domain: a stage1-only autoencoder (Stage1AE) and a method using a domain-specific SIG On the GTEx test set. Both methods performed well ( Fig. 2A ), with Stage1AE slightly outperforming SIG. Figure 2. Open in a new tab Ablation study on pseudo-EV data across different domains. (A) Performance on GTEx-simulated test data, which represents source-domain data of Stage1AE and SIG methods. (B) Performance on HPA-simulated data, which represents target-domain data of Stage1AE, SIG, DADA_noDF, and DADA-EV methods. (C) Deconvolution performance across 23 tissues by the four ablations on HPA pseudo-EV. (D) t-SNE visualization of latent embeddings from source- and target- domain data before and after domain adaptation. (E) Comparison between DADA-EV and other three representative domain adaptation techniques. On the HPA set, we further compared the full DADA-EV model with a variant lacking the diffusion generator (DADA_noDF), in which stage 2 embeddings were obtained directly from source-domain expression data via the encoder (see Methods). Cross-domain evaluation showed that Stage1AE and SIG suffered substantial performance loss, whereas DADA-EV and DADA_noDF maintained performance at the same level as in the source domain, achieving higher CCC and lower MAE ( Fig. 2B ). Remarkably, DADA-EV achieved an average MAE of 0.02 and CCC of 0.83 for fraction prediction, comparable to DADA_noDF (MAE 0.02, CCC 0.81), demonstrating that diffusion-generated embeddings are essentially indistinguishable in effectiveness from encoder-derived source embeddings. To assess the impact of domain adaptation at specific tissue level, we compared the deconvolution performance of the four methods across 23 tissues. The largest performance differences were observed in the Small Intestine, Skin, Prostate, and Pancreas ( Fig. 2C ). To quantitatively assess the alignment between source and target latent distributions, we calculated the WD before and after domain adaptation. As shown in Fig. 2D , the WD decreased (e.g., from 0.838 to 0.541), indicating that the domain shift was effectively mitigated. This result is further corroborated by t-SNE visualization, which shows that the initially disjoint source and target embeddings became well-aligned and overlapped following our GRL-based adaptation. Together, these metrics validate the effectiveness of our proposed strategy. To rigorously validate the superiority of our approach against established baselines, we compared DADA-EV with three representative domain adaptation techniques: DeepCORAL [ 30 ], CoGAN [ 31 ], and ADDA [ 32 ]. As illustrated in Fig. 2E , DADA-EV consistently outperformed all baselines, followed by DeepCORAL and ADDA, while CoGAN exhibited the lowest performance. To investigate the underlying mechanism, we analyzed the alignment intensity by calculating the reduction in WD ( ) between source and target latent distributions before and after adaptation. Interestingly, CoGAN achieved the largest , implying the most aggressive distribution alignment. However, its poor deconvolution performance suggests that such excessive alignment forced distinct tissue-type features to collapse, leading to negative transfer. Conversely, ADDA showed limited alignment capabilities. DADA-EV achieved a balanced reduction in while maintaining the highest prediction accuracy. This indicates that by integrating adversarial learning with reconstruction-based regularization, our strategy successfully mitigates domain shifts while preserving the intrinsic biological manifold structure required for accurate deconvolution. DADA-EV outperforms existing deconvolution methods To assess the efficacy of our approach, we performed a comprehensive comparison between DADA-EV and other established deconvolution methods: TAPE [ 15 ], Scaden [ 17 ], EV-origin [ 5 ], CIBERSORT [ 10 ], Bisque [ 12 ], DeconRNASeq [ 13 ], MuSiC [ 14 ], and DWLS [ 33 ], and DestVI [ 34 ] (Supplementary Methods). To ensure a fair and consistent comparison, we unified the input feature space by constructing a pre-selected gene expression profile (GEP) using the TSS-DEP-MCN approach (Supplementary Methods) [ 5 , 35 ]. These selected genes served as input features for DADA-EV and as the signature matrix for all competing methods. Before evaluating deconvolution performance, we verified whether DADA-EV captures biologically meaningful tissue signatures. For the source domain data prior to domain adaptation, we utilized the latent space obtained from the autoencoder. Specifically, we derived a signature matrix directly from the Stage1AE and compared it with the manual constructed tissue-specific gene expression profile via TSS-DEP-MCN method (referred to as the manual method). This manual benchmark was rigorously created using tissue specificity scoring and condition number optimization (detailed in Supplementary Methods). Visually, the autoencoder-derived matrix exhibited distinct tissue-specific patterns highly similar to the manual baseline ( Fig. S3A ). Quantitatively , correlation analysis confirmed a strong correspondence between the learned signatures and the manual profiles ( Fig. S3B ). This suggests that baseline autoencoder effectively captures intrinsic tissue identity with higher specificity. Building on these robust tissue representations, our model excelled in the source domain validation: Stage1AE achieved the lowest MAE (average 0.011) and the highest CCC (average 0.916) on the domain-free GTEx test data (source domain data), surpassing the second-best method, Scaden ( Fig. 3A , Table S4 ). Figure 3. Open in a new tab Comparison of DADA-EV with existing deconvolution methods on pseudo-EV data. (A) Performance of Stage1AE (DADA-EV) versus other methods on GTEx-simulated test data. CCC represents the Lin’s concordance correlation coefficient, measuring the concordance between the predicted fraction and the ground truth. MAE represents mean absolute error, measuring the accuracy of prediction. Boxplots show the median (center lines), interquartile range (hinges) and 1.5 times the interquartile range (whiskers). (B) Performance of DADA-EV versus other methods on HPA-simulated test data. (C) CCC comparison of DADA-EV against other methods for each tissue on GTEx- and HPA-simulated test data. The CCC for each tissue is represented by the average CCC value of that tissue across all samples. When introducing domain shifts (HPA data), the performance of most methods declined. Notably, DADA-EV achieved the most substantial gains in this cross-domain scenario, significantly outperforming all competitors. Its CCC on the HPA dataset (0.751) exceeded that of Scaden (0.649) by a margin of 0.1 ( Fig. 3B , Table S4 ), demonstrating that DADA-EV delivers state-of-the-art (SOTA) accuracy in handling distributional shifts. To further dissect these results, we conducted a stratified analysis of estimation accuracy across individual tissue. DADA-EV consistently achieved higher CCC scores across the majority of tissue types in both datasets ( Fig. 3C ). Crucially, breakdown analysis reveals that DADA-EV excels in identifying tissues where other methods struggle, such as Adipose Tissue and Cervix Uteri in both GTEx and HPA, and Ovary in HPA. This highlights the model’s versatility: it maintains high specificity for major, complex contributors (e.g., Adipose Tissue) while simultaneously exhibiting the sensitivity required to detect minority contributors whose signals are typically masked in bulk mixtures. Accurate deconvolution across diverse gene variance levels and tissue contexts To evaluate the effectiveness and reliability of DADA-EV under varying conditions, we systematically assessed its performance across different gene variance levels, gene set sizes, numbers of tissue types, and ratio configurations. This comprehensive analysis enabled us to probe the robustness and adaptability of the model under diverse simulated scenarios. Under various conditions of gene variance levels and gene set sizes ( Fig. S4 , see Supplementary Methods), DADA-EV consistently demonstrated strong deconvolution performance in domain adaptation ( Fig. 4A ). Across different variance levels and gene set sizes, the model achieved an overall mean MAE of 0.03 and CCC of 0.96 for reconstructed expression profiles, and a mean MAE of 0.02 and CCC of 0.76 for fraction prediction. Notably, genes with low variance were predicted with the highest accuracy and reliability, as indicated by the lowest MAE and highest CCC values. Within each variance group, performance improved as the number of genes increased, with both low- and high-variance groups showing reduced MAE and increased CCC. This pattern is likely to reflect the design of DADA-EV: low-variance genes provide stable, domain-invariant signals that are easier to align in the latent space, while larger gene sets enhance the encoder’s capacity to capture discriminative features, thereby improving reconstruction and fraction estimation. Figure 4. Open in a new tab Performance evaluation on target-domain pseudo-EV data across diverse gene variance levels and tissue contexts. (A) Deconvolution performance of DADA-EV across different variance levels (high, medium, and low variance level) and gene set sizes (1000, 2000, and 3000 genes). (B) Deconvolution performance with three tissue types under different ratio settings. Three configurations were tested: T3C1 (50% adipose tissue, 35% muscle, 15% lung), T3C2 (35% adipose tissue, 15% muscle, 50% lung), and T3C3 (15% adipose tissue, 50% muscle, 35% lung). (C) Deconvolution performance with ten tissue types (adipose tissue, muscle, lung, liver, skin, colon, bladder, kidney, small intestine, and esophagus) under different ratio settings. Three configurations were applied: T10C1 (equal distribution of approximately 10% for each tissue), T10C2 (gradually increasing distribution from 1% to 19% across tissues), and T10C3 (gradually decreasing distribution from 19% to 1% across tissues). Meanwhile, to assess how DADA-EV performs under varying tissue complexities and fraction distributions, we evaluated scenarios with both three and ten tissue types, each tested under three ratio configurations (see Supplementary Methods). Across all settings, DADA-EV consistently achieved high performance, with correlation coefficients exceeding 0.9 ( Fig. 4B and C ), and the predicted fractions closely matched the predefined ratios. These results demonstrate that DADA-EV maintains robust deconvolution accuracy regardless of whether tissue compositions involve few or many components, and whether fractions are evenly or unevenly distributed. The consistent performance across different tissue configurations highlights DADA-EV’s ability to capture tissue-specific EV release patterns, which is crucial for understanding EV biology. The model can also detect subtle contributions from rare or low-abundance cell types, which is biologically significant as even small cell populations can influence intercellular communication and disease progression. DADA-EV deciphers disease-specific heterogeneity through biologically interpretable latent representations We next examined the heterogeneity in tissue origin contributions to plasma EV across various pathological states. We analyzed a public EV RNA-seq dataset comprising samples from healthy individuals, patients with benign diseases, and patients with three types of cancer: BRCA, HCC, and PAAD [ 3 ]. When applied to real-world EV RNA-seq data, DADA-EV demonstrated accurate reconstruction of expression profiles ( Fig. 5A ). Compared with other deconvolution algorithms, DADA-EV produced a more balanced distribution across the 23 tissue types ( Fig. 5B ). In contrast, several existing methods showed highly skewed predictions ( Fig. S5 ). For example, TAPE assigned 58% of the total contribution to only spleen and pancreas, EVOrigin and CIBERSORT attributed 58% to spleen and skin, DeconRNASeq predicted 47% from spleen and liver, and DWLS assigned more than 80% to spleen alone. Figure 5. Open in a new tab Tissue-level deconvolution of plasma-derived EVs across disease groups through DADA-EV. (A) Reconstruction performance on real EV datasets. (B) Stacked bar plots of 23 inferred tissue fractions across healthy, benign, and three cancer cohorts (BRCA, HCC, PAAD). (C) Group-wise comparisons of predicted breast fractions (BRCA versus others), liver fractions (HCC versus others), and pancreas fractions (PAAD versus others). Two-sided Student’s t-tests are used to analyze the differences between groups. Asterisks indicate significant differences compared to the control group (asterisks indicate significant differences: * P < .05, ** P < .01, *** P < .001). (D) ROC curves using predicted breast, liver, and pancreas fractions to distinguish BRCA, HCC, and PAAD from healthy, benign, non-cancerous, and other cancer groups. (E) Heatmap illustrating the count of shared genes among the top 20 driver genes identified across the 23 tissue types and the 256-dimensional latent space. Only latent dimensions sharing more than 10 genes (>50%) are displayed. (F) Bar chart showcasing the number of overlapping genes between three specific tissues (breast, liver, pancreas) and their corresponding latent dimensions (identified in E). The specific names of these intersected driver genes are detailed in the heatmap to the right. (G) GO functional enrichment analysis of the driver genes identified in Figure E, highlighting key biological processes associated with each tissue. When comparing samples within the same tissue type, DADA-EV also revealed clear and biologically meaningful differences across disease conditions ( Fig. 5C ). BRCA samples showed significantly higher contributions from breast tissue, HCC samples exhibited increased liver tissue contributions, and PAAD samples showed elevated pancreatic tissue contributions relative to the other cohorts. Importantly, none of the other methods were able to simultaneously distinguish all three tissue-specific differences ( Fig. S6 ). To further assess disease-discriminative potential, we performed ROC analysis based on breast, liver, and pancreas fractions ( Fig. 5D ). Among these, the liver tissue fraction effectively distinguished HCC patients from Healthy, Benign, Non-cancerous (Healthy + Benign), and Other cancerous (BRCA + PAAD) groups, with AUCs of 0.759, 0.692, 0.724, and 0.622, respectively. To validate that DADA-EV captures meaningful biological signals rather than technical artifacts on real-world EV data, we employed Integrated Gradients, a formal attribution method that assigns importance scores to input features by integrating gradients along a path from a baseline to the input [ 36 ]. This analysis was systematically applied to both the deconvolution predictor and the encoder to validate biological relevance through two distinct lenses: Predictor Attribution (Gene-to-Proportion): By calculating the mean absolute attribution scores across the 23 tissue types, we identified the top 20 driver genes for each lineage ( Table S5 ). Encoder Attribution (Gene-to-Latent Space): We interrogated the 256-dimensional latent space by attributing individual dimensions back to input genes to understand how the model structures its internal representations ( Table S6 ). Our analysis revealed several key insights into the model’s internal logic: (1) Within the 256-dimensional latent space, 35 dimensions were found to share more than 10 attributed genes (>50%) across the 23 tissue types ( Fig. 5E ). These dimensions exhibited a clear correspondence with specific tissues, suggesting the model learns a structured biological map rather than random features. (2) We observed significant gene overlaps between specific tissues and their corresponding latent dimensions, specifically 18 genes for Breast, 10 for Liver, and 16 for Pancreas ( Fig. 5F ). For instance, Liver fraction predictions were attributed by genes including SHMT2, ZC3H13, and ST6GAL1, Breast fraction predictions by RBP7, OXTR, and MYB, and Pancreas fraction predictions by TPST2, SEL1L, and PDCD4. (3) The attributed genes were enriched in distinct functional modules that are closely associated with the pathophysiology of their respective organs ( Fig. 5G ). Specifically, Breast-associated genes were enriched in lipid transport and response to steroid hormone, Liver-associated genes were linked to serine family amino acid biosynthetic process and enzyme inhibitor activity, and Pancreas-associated genes were primarily involved in endoplasmic reticulum protein-containing complex and cellular response to hypoxia. Together, these findings demonstrate that DADA-EV not only achieves accurate deconvolution but also captures disease-associated alterations in tissue origins. By maintaining a robust focus on intrinsic biological identities rather than technical noise or simple mathematical projections, DADA-EV learns an interpretable, lower-dimensional representation of the EV-transcriptome that is deeply grounded in authentic biology. Cross-domain PBMC cell-type deconvolution in EVs powered by DADA-EV It has been reported that multiple PBMC cell types of release EVs, and their cargo and surface markers retain source-specific signatures. Using deconvolution approaches, the composition of immune cell types in EVs, such as monocytes, B cells, NK cells, NKT cells, and CD4/CD8 T cells, can be inferred, providing insights into immune constitution. To this end, we downloaded two PBMC single-cell RNA-seq datasets from 10× Genomics (‘6k PBMCs from a Healthy Donor’ as source-domain and ‘Donor C Frozen PBMCs’ as target-domain), normalized them, and obtained expression profiles of seven immune cell types ( Fig. S7 , see Supplementary Methods). Based on these profiles, we simulated pseudo-EV data with varying cell proportions ( Fig. 6A ) and applied DADA-EV along with other benchmark methods for deconvolution. Figure 6. Open in a new tab Cell-level deconvolution of pseudo-EV data and plasma-derived EVs across disease groups through DADA-EV. (A) Simulation workflow for PBMC cell-type-level pseudo-EV data. (B) Performance on simulated data from donor C PBMC scRNA-seq (target domain) for Stage1AE, SIG, DADA_noDF, and DADA-EV. (C) Performance comparison of DADA-EV with existing methods on target-domain PBMC pseudo-EV data. (D) Stacked bar plots of five inferred PBMC cell-type fractions across healthy, benign, and three cancer cohorts (BRCA, HCC, PAAD). (E) Group-wise comparisons of predicted dendritic cells fractions (BRCA versus others), B cells fractions (HCC versus others), and monocytes fractions (PAAD versus others). Two-sided Student’s t-tests are used to analyze the differences between groups. Asterisks indicate significant differences compared to the control group ( * P < .05, ** P < .01, *** P < .001). Consistent with our tissue-level analyses, DADA-EV demonstrated superior domain adaptation capabilities when applied to cross-source training and testing data. As shown in Fig. 6B , it achieved the lowest error (mean MAE = 0.02) and highest concordance (mean CCC, 0.98), comparable to the DADA_noDF baseline. Furthermore, under identical gene sets, DADA-EV significantly outperformed other classical approaches, achieving SOTA performance ( Fig. 6C ). In the stratified analysis of individual cell types using simulated domain data, DADA-EV consistently ranked first, followed by Scaden and TAPE ( Fig. S8 ). It showed particularly strong performance in identifying Monocytes, Plasma cells, and Platelets. Notably, although plasma cells and platelets constituted only minor fractions (~5% and ~ 4%, respectively) of the single-cell reference ( Fig. S7G ), DADA-EV maintained high deconvolution accuracy. This suggests that DADA-EV is robust to class imbalance and effectively captures distinct feature signatures even from minority cell populations. Finally, we applied DADA-EV to estimate PBMC-derived cells in real EV data. Intriguingly, while plasma cells and platelets were rare in the cellular reference, their predicted EV fractions increased significantly to averages of ~20% and ~ 30%. With the exception of TAPE and Scaden, the other comparison methods predicted substantially lower fractions of plasma cells and platelets, resulting in generally weaker discriminative power between groups compared to DADA-EV ( Fig. 6D and Fig. S9 ). Furthermore, DADA-EV revealed distinct immune profile variations across disease cohorts: Dendritic Cells (DCs) were enriched in breast cancer, B cells were depleted in HCC, and Monocytes were elevated in pancreatic cancer. In contrast, competing methods failed to consistently resolve these significant differences across the disease groups ( Fig. 6E and Fig. S10 ). Effective deconvolution performance of DADA-EV in synthetic cell-line EV mixtures To validate the deconvolution accuracy of DADA-EV, we conducted in vitro mixture experiments using EVs derived from four distinct cell lines representing major organs: hTERT-HPNE (pancreas), L02 (liver), 293 T (kidney), and BEAS-2B (lung). We designed three specific mixing scenarios to simulate varying biological complexities ( Fig. 7A ): a Balanced set (Mix1) to assess baseline identification when sources contribute equally; a Heterogeneity set (Mix2) containing low-abundance targets (~5%) to evaluate sensitivity; and a Gradient set (Mix3) with varying proportions to test the model’s dynamic quantification range. Figure 7. Open in a new tab Validation of DADA-EV deconvolution accuracy using synthetic cell-line EV mixtures. (A) Configuration of the in vitro EV mixture experiment design. (B) DADA-EV estimation of each cell-line EV fraction in the mixture sample. (C) Stacked bar charts comparing the predicted (pred) versus real (real) EV source proportions for each sample. (D) Performance assessment across three distinct mixing scenarios. (E) Boxplots showing the distribution of prediction residuals (predicted - real) for each cell type. (F) Scatter plot illustrating the global concordance between DADA-EV predicted fractions (y-axis) and ground truth values (x-axis) across all samples. The predicted fractions of cell-type-derived EVs are shown in Fig. 7B . Visual reconstruction of the sample compositions (as shown in the stacked bar plots, Fig. 7C ) confirms that DADA-EV effectively recovers the major proportion profiles of complex samples. We further dissected the model’s performance across specific biological scenarios (Mix1–Mix3, Fig. 7D ). The model achieved its most robust performance in the Heterogeneity scenario (Mix2), which simulates rare cell population detection, with a Pearson correlation coefficient (PCC) of 0.91. Notably, DADA-EV accurately resolved the low-abundance hTERT-HPNE targets (~5%) from the dominant background, confirming its high sensitivity for trace signal detection. While the Balanced (Mix1) and Gradient (Mix3) scenarios showed moderate deviation from the ground truth, this behavior is largely attributable to a systematic bias revealed in the sample reconstruction ( Fig. 7E ). Specifically, the model tends to overestimate 293 T-derived signals while underestimating contributions from L02 or hTERT-HPNE. This inherent bias skews the relative proportions in identical-proportional (Mix1) or gradient (Mix3) settings, yet the model remains effective at identifying the presence of distinct cellular components. Overall, DADA-EV demonstrated strong concordance with the ground truth, achieving a PCC of 0.66 across all samples ( Fig. 7F ). This correlation indicates that the model reliably captures cell type-specific EV signatures within mixed populations, suggesting DADA-EV is a robust tool for dissecting cellular origins in complex EV mixtures. Discussion and conclusion In the development of liquid biopsy traceability models, only a few prior studies [ 5 , 37–39 ] have been reported. These works mainly relied on existing algorithmic frameworks, adapting them by modifying signature gene-based matrices for EV or tissue/cell deconvolution. However, none of them made specific algorithmic adjustments to account for the differences between liquid biopsy data and tissue/cell-source data. In contrast, we developed DADA-EV, which represents a significant advancement by integrating diffusion-based generation and domain adaptation. Unlike previous methods, DADA-EV does not depend on predefined signature matrices, which is a major limitation of conventional approaches, and bridges the distributional gap between tissue/cell-type and EV transcriptomes. This enables accurate estimation of tissue- and cell-type-specific contributions and provides a generalizable strategy applicable to other cross-domain biomedical integration tasks, such as multi-omics fusion or cross-platform single-cell analysis. Diffusion module is capable of learning the full data distribution and progressively reconstructing signals from noisy observations, making them particularly suitable for handling complex and highly heterogeneous EV transcriptomic data. Unlike conventional latent-variable models that impose simplified distributional assumptions, diffusion models can flexibly capture non-Gaussian, long-tailed, and multimodal distributions, which are commonly observed in EV-derived RNA profiles. Combined with the results of DADA-EV in Fig. S2F and Fig. 2B and the consistency with baselines using original data as input, these findings collectively demonstrate that the diffusion module successfully captures the intrinsic latent distribution of the source domain, allowing the model to discard original reference data without compromising performance. While diffusion model effectively captures complex data distributions, it cannot inherently guarantee cross-domain consistency. In our task, the discrepancy between tissue/cell-derived profiles and EV-derived profiles is not merely a simple distribution shift, but involves nonlinear distortions, signal-dependent dropout, and compositional effects. To address this, we incorporate adversarial domain adaptation to explicitly enforce alignment between tissue/cell and EV representations. The adversarial objective encourages the learned latent space to be domain-invariant, thereby enabling accurate transfer of tissue/cell-type-specific signatures across domains. Hence, the key advantage of our design lies in the complementary roles of the two components: (i) The diffusion module models the data generation process, enabling accurate reconstruction of complex, noisy, and incomplete transcriptomic signals. (ii) The adversarial module enforces cross-domain alignment, ensuring that the learned representations are transferable between tissue and EV domains. Such a design is also supported by recent studies in other domains. For example, in computer vision, Zhang et al. [ 40 ] used diffusion to generate target-domain samples while leveraging domain classifiers for alignment. In the biomedical domain, Yang et al. [ 41 ] integrated diffusion-based generation with adversarial learning to bridge the gap between simulated and real medical images. These studies collectively suggest that combining generative modeling with adversarial alignment is particularly effective in scenarios with large domain gaps. When using deep learning-based deconvolution models like Scaden [ 17 ] and TAPE [ 15 ], an important preprocessing step is to filter out genes with low variance, as the model may not predict fractions effectively otherwise. Models tend to perform better when the number of genes is limited to around 10,000. One likely reason for this is that less variable genes often contain a significant amount of noise. However, in our DADA-EV, we achieved good deconvolution results across varying levels of variance, whether with a smaller subset of genes or the full set. Importantly, using all genes (MAE 0.018, CCC 0.828) achieved better than gene-selected settings (MAE 0.022, CCC 0.751) while eliminating the need for labor-intensive gene selection. This flexibility enhances usability and broad applicability. Additionally, the diffusion generator by synthesizing source-like embeddings from noise, reduces dependence on reference data and enabling effective training even with scarce or low-quality references, which might be a notable advantage in contexts such as rare diseases or underrepresented tissues. Our plasma EV analysis supports existing evidence that hematopoietic cells, particularly platelets and leukocytes, are dominant contributors to the circulating EV pool [ 42 , 43 ]. Although these populations are relatively minor in single-cell datasets, DADA-EV correctly identifies them as major EV sources. Furthermore, disease-specific alterations inferred by DADA-EV align with previous reports. For example, reduced dendritic cells in breast cancer are consistent with findings that dysfunctional DCs impair immune surveillance and promote metastasis [ 44 ], and that cDC1 infiltration is preferentially enriched in triple-negative breast cancer [ 45 ]. Similarly, reduced B cells in HCC agree with evidence that B cell subsets are significantly decreased in tumor tissues compared to adjacent live [ 46 ], and that memory B cell percentages decline during HCC progression [ 47 ]. Finally, increased monocytes in pancreatic cancer reflect clinical observations linking elevated monocyte levels to tumor aggressiveness and poor prognosis [ 48 ], as well as enrichment of complement-high monocytes and tumor-associated macrophages [ 49 ]. Together, these consistencies confirm that DADA-EV not only achieves technical improvements but also yields biologically credible insights into EV origins. In the field of liquid biopsy, several studies have explored tissue-of-origin inference using alternative omics signals. For example, transcription start site coverage patterns and methylation atlases have been successfully applied to cell-free DNA for tissue tracing [ 38 , 50 ]. However, for EVs, which are membrane-bound particles carrying diverse molecular cargo including DNA, RNA, methylation signatures, and proteins, current tissue deconvolution approaches are almost exclusively based on transcriptomic data [ 5 ]. The DADA-EV framework provides a potentially generalizable strategy for cross-domain biomedical inference. Its diffusion-based generative module and domain adaptation design are, in principle, compatible with other omics layers in EVs, provided that appropriate tissue- or cell-type-specific reference atlases are available. Looking forward, we anticipate that extending DADA-EV to multi-omics settings will be particularly valuable. Integrating deconvolution results across different molecular layers may enable cross-validation and yield a more comprehensive and robust understanding of EV tissue/cell-of-origin. In addition, a promising direction is dual layer deconvolution, where DADA-EV could first resolve tissue-level origins and then further dissect the cellular composition within each tissue. Several limitations should be acknowledged. First, a fundamental challenge in liquid biopsy is the lack of absolute ground truth for the tissue-of-origin composition of circulating EVs in humans. Unlike synthetic mixtures with predefined proportions, it remains technically infeasible to isolate and quantify EVs from every organ within a patient, making direct validation of model predictions impossible. Second, the current framework does not explicitly account for inter-individual variability. Differences in EV secretion patterns, physiological states, and disease contexts across individuals may introduce biases that are not fully captured by reference atlases. Third, our evaluation is based on bulk EV transcriptomic profiles, which represent averaged signals across heterogeneous EV populations. This limits the resolution of the model, as it cannot capture variability at the level of individual EVs or distinct EV subpopulations. Looking ahead, we envision DADA-EV as a versatile tool with broad clinical applications. First, it may provide valuable insights into the diagnosis and assessment of systemic diseases. Many conditions, such as metabolic syndrome and diabetes, involve coordinated dysfunction across multiple organs rather than isolated organ pathology. By quantifying shifts in the tissue-of-origin composition of EVs, DADA-EV could reveal the relative contributions of different organs during disease progression. For example, an increase in adipose-derived EVs indicates inter-organ communication and metabolic imbalance. These EVs carry bioactive molecules that can reprogram distal tissue metabolism and promote insulin resistance and macrophage activation in obesity [ 51 , 52 ]. Second, DADA-EV holds promise for evaluating metastatic potential and predicting metastatic sites in cancer. Tumor-derived EVs often prepare pre-metastatic niches in specific target organs such as lung, liver and brain [ 53 ]. Thus, an EV tissue-origin profile enriched not only for lung-derived EVs but also for liver-derived EVs in a lung cancer patient may signal heightened metastatic risk. Additionally, quantitative EV tracing could help infer the primary tumor site in cases of cancer of unknown primary. Third, DADA-EV could facilitate early monitoring and precise management of treatment. For example, Shi et al. [ 54 ] analyzed the cell types present in plasma EVs from individuals with metastatic melanoma and used this information to create a tool for tracking tumor/non-tumor ratios in response to treatment receiving immune checkpoint inhibitors. Therefore, performing transcriptome deconvolution on EVs to pinpoint the originating cell types opens a new avenue for monitoring tumor persistence and immune activation. We anticipate that DADA-EV will provide a versatile framework for liquid biopsy, enabling systemic disease evaluation, metastasis risk prediction, and treatment monitoring. More broadly, its ability to bridge EV biology with clinical translation may help advance precision medicine. Key Points Domain-Adaptive Diffusion Autoencoder for EVs (DADA-EV) integrates autoencoder learning, generative simulation, and adversarial adaptation to achieve reference-free deconvolution of EV transcriptomes for estimating tissue- or cell-type-specific origin. Adversarial adaptation enables robust cross-domain generalization between tissue/cell-type and EV datasets. Generative simulation reduced dependence on source data during target-domain training. DADA-EV outperforms existing methods and accurately estimates tissue- and cell-type fractions in pseudo-EV. It reveals tissue and cellular heterogeneity in blood, offering a generalizable tool for EV-based liquid biopsy and precision medicine. Supplementary Material supplementary-material_bbag281 supplementary-material_bbag281.zip (2.9MB, zip) Contributor Information Shuilin Liao, Faculty of Innovation Engineering, Macau University of Science and Technology, Avenida Wai Long, Taipa, Macau 999078, China; Guangzhou National Laboratory, No. 9 XingDaoHuanBei Road, Guangzhou International Bio Island, Guangzhou, Guangdong Province 510005, China. Haoxiang Yang, Guangzhou National Laboratory, No. 9 XingDaoHuanBei Road, Guangzhou International Bio Island, Guangzhou, Guangdong Province 510005, China. Shuting Xiao, State Key Laboratory of Respiratory Disease, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou Medical University, Guangzhou, Guangdong 510120, China. Zhiwen Li, Guangzhou Exope Medical Biotechnology Co., Ltd, China. Qi Zhu, State Key Laboratory of Respiratory Disease, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou Medical University, Guangzhou, Guangdong 510120, China. Le Li, Faculty of Innovation Engineering, Macau University of Science and Technology, Avenida Wai Long, Taipa, Macau 999078, China; Basic Teaching Department, Zhuhai Campus of Zunyi Medical University, No. 368, Jinwan Road, Jinwan District, Zhuhai, Guangdong Province 519000, China. Shanghui Lu, Faculty of Innovation Engineering, Macau University of Science and Technology, Avenida Wai Long, Taipa, Macau 999078, China; School of Mathematics and Physics, Hechi University, No.42, Longjiang, Hechi, Guangxi 546300, China. Yong Liang, Chinese Medicine Guangdong Laboratory, No. 2522, Huandao North Road, Hengqin New Area, Xiangzhou District, Zhuhai, Guangdong Province 519031, China. Ming Dong, Guangzhou National Laboratory, No. 9 XingDaoHuanBei Road, Guangzhou International Bio Island, Guangzhou, Guangdong Province 510005, China; State Key Laboratory of Respiratory Disease, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou Medical University, Guangzhou, Guangdong 510120, China. Author contributions S.L.: formal analysis, software, writing—original draft; H.X.Y.: software, writing—review & editing; S.T.X.: investigation; Z.W.L.: investigation; Q.Z.: writing—review & editing; L.L.: software, methodology; S.H.L.: software, methodology; Y.L.: conceptualization, supervision, writing—review & editing; M.D.: conceptualization, supervision, writing—review & editing. Conflicts of interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Funding This work was supported by Major Project of Guangzhou National Laboratory (Grant No. GZNL2024A03006), and the TianYuan funds for Mathematics of the National Science Foundation of China (Grant No. 12326604). Data availability Original datasets are obtained from the datasets of GTEx, HPA, exoRBase 2.0, and 10× Genomics website. The code, version dependencies, trained models, preprocessed data, and tutorials are available at https://github.com/ShuilinLiao/DADAEV/tree/liao . References 1. Kalluri  R, LeBleu  VS. The biology, function, and biomedical applications of exosomes. science n367 2020;367:eaau6977. 10.1126/science.aau6977 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Van Niel  G, d'Angelo  G, Raposo  G. Shedding light on the cell biology of extracellular vesicles. Nat Rev Mol Cell Biol 2018;19:213–28. 10.1038/nrm.2017.125 [ DOI ] [ PubMed ] [ Google Scholar ] 3. Lai  H, Li  Y, Zhang  H  et al. exoRBase 2.0: an atlas of mRNA, lncRNA and circRNA in extracellular vesicles from human biofluids. Nucleic Acids Res 2022;50:D118–28. 10.1093/nar/gkab1085 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Berumen Sánchez  G, Bunn  KE, Pua  HH  et al. Extracellular vesicles: mediators of intercellular communication in tissue injury and disease. Cell Communication and Signaling 2021;19:104. 10.1186/s12964-021-00787-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Li  Y, He  X, Li  Q  et al. EV-origin: enumerating the tissue-cellular origin of circulating extracellular vesicles using exLR profile. Comput Struct Biotechnol J 2020;18:2851–9. 10.1016/j.csbj.2020.10.002 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Barceló  M, Mata  A, Bassas  L  et al. Exosomal microRNAs in seminal plasma are markers of the origin of azoospermia and can predict the presence of sperm in testicular tissue. Hum Reprod 2018;33:1087–98. 10.1093/humrep/dey072 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Dou  Y-Q, Kong  P, Li  C-L  et al. Smooth muscle SIRT1 reprograms endothelial cells to suppress angiogenesis after ischemia. Theranostics 2020;10:1197–212. 10.7150/thno.39320 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Li  Y, Zhao  J, Yu  S  et al. Extracellular vesicles long RNA sequencing reveals abundant mRNA, circRNA, and lncRNA in human blood as potential biomarkers for cancer diagnosis. Clin Chem 2019;65:798–808. 10.1373/clinchem.2018.301291 [ DOI ] [ PubMed ] [ Google Scholar ] 9. Yu  S, Li  Y, Liao  Z  et al. Plasma extracellular vesicle long RNA profiling identifies a diagnostic signature for the detection of pancreatic ductal adenocarcinoma. Gut 2020;69:540–50. 10.1136/gutjnl-2019-318860 [ DOI ] [ PubMed ] [ Google Scholar ] 10. Newman  AM, Liu  CL, Green  MR  et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods 2015;12:453–7. 10.1038/nmeth.3337 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Hunt  GJ, Freytag  S, Bahlo  M  et al. Dtangle: accurate and robust cell type deconvolution. Bioinformatics 2019;35:2093–9. 10.1093/bioinformatics/bty926 [ DOI ] [ PubMed ] [ Google Scholar ] 12. Jew  B, Alvarez  M, Rahmani  E  et al. Accurate estimation of cell composition in bulk expression through robust integration of single-cell information. Nat Commun 2020;11:1971. 10.1038/s41467-020-15816-6 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Gong  T, Szustakowski  JD. DeconRNASeq: a statistical framework for deconvolution of heterogeneous tissue samples based on mRNA-Seq data. Bioinformatics 2013;29:1083–5. 10.1093/bioinformatics/btt090 [ DOI ] [ PubMed ] [ Google Scholar ] 14. Wang  X, Park  J, Susztak  K  et al. Bulk tissue cell type deconvolution with multi-subject single-cell expression reference. Nat Commun 2019;10:380. 10.1038/s41467-018-08023-x [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Chen  Y, Wang  Y, Chen  Y  et al. Deep autoencoder for interpretable tissue-adaptive deconvolution and cell-type-specific gene analysis. Nat Commun 2022;13:6735. 10.1038/s41467-022-34550-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Vallania  F, Tam  A, Lofgren  S  et al. Leveraging heterogeneity across multiple datasets increases cell-mixture deconvolution accuracy and reduces biological and technical biases. Nat Commun 2018;9:4735. 10.1038/s41467-018-07242-6 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Menden  K, Marouf  M, Oller  S  et al. Deep learning–based cell composition analysis from tissue expression profiles. Sci Adv 2020;6:eaba2619. 10.1126/sciadv.aba2619 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Willms  E, Cabañas  C, Mäger  I  et al. Extracellular vesicle heterogeneity: subpopulations, isolation techniques, and diverse functions in cancer progression. Front Immunol 2018;9:738. 10.3389/fimmu.2018.00738 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Fabbiano  F, Corsi  J, Gurrieri  E  et al. RNA packaging into extracellular vesicles: an orchestra of RNA-binding proteins? J Extracell Vesicles 2020;10:e12043. 10.1002/jev2.12043 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Furusawa  C, Suzuki  T, Kashiwagi  A  et al. Ubiquity of log-normal distributions in intra-cellular reaction dynamics. Biophysics 2005;1:25–31. 10.2142/biophysics.1.25 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Robinson  MD, McCarthy  DJ, Smyth  GK. edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 2010;26:139–40. 10.1093/bioinformatics/btp616 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Love  MI, Huber  W, Anders  S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014;15:550. 10.1186/s13059-014-0550-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Narrandes  S, Xu  W. Gene expression detection assay for cancer clinical use. J Cancer 2018;9:2249–65. 10.7150/jca.24744 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Qiu  P. Embracing the dropouts in single-cell RNA-seq analysis. Nat Commun 2020;11:1169. 10.1038/s41467-020-14976-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Pierson  E, Yau  C. ZIFA: dimensionality reduction for zero-inflated single-cell gene expression analysis. Genome Biol 2015;16:241. 10.1186/s13059-015-0805-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Kharchenko  PV, Silberstein  L, Scadden  DT. Bayesian approach to single-cell differential expression analysis. Nat Methods 2014;11:740–2. 10.1038/nmeth.2967 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Wang  F, Yang  F, Huang  L  et al. Deep domain adversarial neural network for the deconvolution of cell type mixtures in tissue proteome profiling. Nature Machine Intelligence 2023;5:1236–49. 10.1038/s42256-023-00737-y [ DOI ] [ Google Scholar ] 28. Ganin  Y, Lempitsky  V. Unsupervised domain adaptation by backpropagation. In: Bach F, Blei D (eds.), International Conference on Machine Learning , pp. 1180–9. Lille, France: PMLR, 2015. [ Google Scholar ] 29. Ho  J, Jain  A, Abbeel  P. Denoising diffusion probabilistic models. Adv Neural Inf Proces Syst 2020;33:6840–51. [ Google Scholar ] 30. Sun  B, Saenko  K. Deep coral: Correlation alignment for deep domain adaptation. In: Leibe B, Matas J, Sebe N, Welling M (eds.), European Conference on Computer Vision. 2016, p. 443–50. Cham, Switzerland: Springer, 10.1007/978-3-319-49409-8_35 [ DOI ] [ Google Scholar ] 31. Liu  M-Y, Tuzel  O. Coupled generative adversarial networks. Adv Neural Inf Proces Syst 2016;29. [ Google Scholar ] 32. Tzeng  E, Hoffman  J, Saenko  K  et al. Adversarial discriminative domain adaptation. In: Liu Y, Rehg JM, Taylor CJ, Wu Y (eds.), Proceedings of the IEEE conference on computer vision and pattern recognition. Los Alamitos, CA, USA: IEEE Computer Society, pp. 7167–76, 2017. 10.1109/CVPR.2017.316 [ DOI ] 33. Tsoucas  D, Dong  R, Chen  H  et al. Accurate estimation of cell-type composition from gene expression data. Nat Commun 2019;10:2975. 10.1038/s41467-019-10802-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Lopez  R, Li  B, Keren-Shaul  H  et al. DestVI identifies continuums of cell types in spatial transcriptomics data. Nat Biotechnol 2022;40:1360–9. 10.1038/s41587-022-01272-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Akat  KM, Lee  YA, Hurley  A  et al. Detection of circulating extracellular mRNAs by modified small-RNA-sequencing analysis. JCI Insight 2019;4:e127317. 10.1172/jci.insight.127317 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Sundararajan  M, Taly  A, Yan  Q. Axiomatic attribution for deep networks. In: Doina  P, Yee  WT (eds.), Proceedings of the 34th International Conference on Machine Learning , pp. 3319–28. Sydney, Australia: Proceedings of Machine Learning Research: PMLR, 2017. [ Google Scholar ] 37. Larsen  JH, Jensen  IS, Svenningsen  P. Benchmarking transcriptome deconvolution methods for estimating tissue-and cell-type-specific extracellular vesicle abundances. J Extracell Vesicles 2024;13:e12511. 10.1002/jev2.12511 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Li  S, Zeng  W, Ni  X  et al. Comprehensive tissue deconvolution of cell-free DNA by deep learning for disease diagnosis and monitoring. Proc Natl Acad Sci 2023;120:e2305236120. 10.1073/pnas.2305236120 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Zhu  Q, Cheng  L, Deng  C  et al. The genetic source tracking of human urinary exosomes. Proc Natl Acad Sci 2021;118:e2108876118. 10.1073/pnas.2108876118 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Zhang  Y, Chen  S, Jiang  W  et al. Domain-guided conditional diffusion model for unsupervised domain adaptation. Neural Netw 2025;184:107031. 10.1016/j.neunet.2024.107031 [ DOI ] [ PubMed ] [ Google Scholar ] 41. Yang  Y, Ning  G, Zhong  C, Liao  H. Adversarial diffusion model for domain-adaptive depth estimation in bronchoscopic navigation. In: Linguraru MG, Dou Q, et al. (eds.), International Conference on Medical Image Computing and Computer-Assisted Intervention. 2024, p. 46–56. Cham, Switzerland: Springer, 10.1007/978-3-031-72089-5_5. [ DOI ] [ Google Scholar ] 42. Auber  M, Svenningsen  P. An estimate of extracellular vesicle secretion rates of human blood cells. Journal of Extracellular Biology 2022;1:e46. 10.1002/jex2.46 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Newman  L, Rowland  A. Detection and isolation of tissue-specific extracellular vesicles from the blood. Journal of Extracellular Biology 2025;4:e70059. 10.1002/jex2.70059 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. da  Cunha  A, Michelin  MA, Murta  EF. Pattern response of dendritic cells in the tumor microenvironment and breast cancer. World J Clin Oncol 2014;5:495–502. 10.5306/wjco.v5.i3.495 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Hubert  M, Gobbini  E, Couillault  C  et al. IFN-III is selectively produced by cDC1 and predicts good clinical outcome in breast cancer. Sci Immunol 2020;5:eaav3942. 10.1126/sciimmunol.aav3942 [ DOI ] [ PubMed ] [ Google Scholar ] 46. Zhang  Z, Ma  L, Goswami  S  et al. Landscape of infiltrating B cells and their clinical significance in human hepatocellular carcinoma. Oncoimmunology 2019;8:e1571388. 10.1080/2162402X.2019.1571388 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Wang  X, Wang  L, Ji  F  et al. Decreased CD27 on B lymphocytes in patients with primary hepatocellular carcinoma. J Int Med Res 2012;40:307–16. 10.1177/147323001204000131 [ DOI ] [ PubMed ] [ Google Scholar ] 48. Hansen  FJ, David  P, Akram  M  et al. Circulating monocytes serve as novel prognostic biomarker in pancreatic ductal adenocarcinoma patients. Cancers 2023;15:363. 10.3390/cancers15020363 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Kemp  SB, Steele  NG, Carpenter  ES  et al. Pancreatic cancer is marked by complement-high blood monocytes and tumor-associated macrophages. Life Sci Alliance 2021;4:4. 10.26508/lsa.202000935 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Li  L, Li  X, Li  Y  et al. Tracing the tissue origin of cell-free DNA through open chromatin footprint. Commun Biol 2025;8:1845. 10.1038/s42003-025-09232-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Clement  E, Lazar  I, Attané  C  et al. Adipocyte extracellular vesicles carry enzymes and fatty acids that stimulate mitochondrial metabolism and remodeling in tumor cells. EMBO J 2020;39:e102525. 10.15252/embj.2019102525 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Deng  Z-b, Poliakov  A, Hardy  RW  et al. Adipose tissue exosome-like vesicles mediate activation of macrophage-induced insulin resistance. Diabetes 2009;58:2498–505. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Hoshino  A, Costa-Silva  B, Shen  T-L  et al. Tumour exosome integrins determine organotropic metastasis. Nature 2015;527:329–35. 10.1038/nature15756 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Shi  A, Kasumova  GG, Michaud  WA  et al. Plasma-derived extracellular vesicle analysis and deconvolution enable prediction and tracking of melanoma checkpoint blockade outcome. Sci Adv 2020;6:eabb3461. 10.1126/sciadv.abb3461 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials supplementary-material_bbag281 supplementary-material_bbag281.zip (2.9MB, zip) Data Availability Statement Original datasets are obtained from the datasets of GTEx, HPA, exoRBase 2.0, and 10× Genomics website. The code, version dependencies, trained models, preprocessed data, and tutorials are available at https://github.com/ShuilinLiao/DADAEV/tree/liao . Articles from Briefings in Bioinformatics are provided here courtesy of Oxford University Press ACTIONS View on publisher site PDF (3.2 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• DivideFold+: an AI-based tool for RNA secondary structure prediction with subdomains identification and visualization and data augmentation#268471
• PCIPG2.0: multi-omics fusion and structure-aware graph autoencoding for protein complex identification.#437234
• Dissecting Alzheimer's proteomic landscape through NULISA profiling of brain cell-specific extracellular vesicles.#480744
• Minimal information for studies of extracellular vesicles (MISEV2023): From basic to advanced approaches#1079274
• MolMAE: A Surface-Centric Multimodal Masked Autoencoder for Molecular Representation Learning#448052