Artificial intelligence virtual extracellular vesicles (AIVEVs)

Abstract

Recent progress in artificial intelligence (AI) has given rise to AI virtual cells (AIVCs), which are digital twins of predictable or dynamic biological cells. This model can simulate, predict and replicate the behavior of real cells in digital software. Extracellular vesicles (EVs) are nanoscale phospholipid bilayer structures released by cells and are important for intercellular communication. To fully leverage digital models in EVs research, we propose the interdisciplinary concept of AI virtual EVs (AIVEVs). This review systematically outlines the construction of AIVEVs through both knowledge-driven (white-box) and data-driven (black-box) modeling paradigms, integrating multi-omics data to simulate EVs biogenesis, cargo sorting, and intercellular communication. Moreover, we highlight how AIVCs drive models to predict the composition of AIVEVs, analyze cell communication behavior, construct diagnostic atlases of pathological virtual cells, and enhance the ability to trace vesicle origins. Furthermore, we also present a closed-loop workflow from in silico prediction to experimental validation and project the developmental trajectory of AIVEVs toward clinical translation. We firmly believe that AIVEVs can accelerate the development of EVs-based disease diagnosis and treatment, thereby opening a new era of intercellular communication research.

Graphical abstract

Highlights

• •Introduction of AIVEVs as a Novel Digital Twin Framework: This review pioneers the concept of AIVEVs, integrating AIVCs with EVs biology to create a predictive digital model for simulating EVs biogenesis, composition, and intercellular communication.
• •Comprehensive Simulation of EV-Mediated Communication: AIVEVs enable the dynamic modeling of donor–EVs–recipient interactions, allowing for the simulation of disease-specific communication networks, prediction of therapeutic EVs candidates, and identification of diagnostic biomarkers in silico.
• •Bridging Computational Predictions with Experimental Validation: The proposed framework combines multi-omics data integration, AI-driven prediction, and experimental feedback, offering a closed-loop research paradigm that enhances the accuracy, scalability, and clinical translatability of EVs-based studies.

1. Introduction

Cellular communication leads to molecular control and control of various biological activities, and cell signal control maintains normal organizational function, control of intercellular communication, and physiological and pathological signal basis [1,2]. In recent years, extracellular vesicles (EVs), phospholipid bilayer membrane structures secreted by cells and rich in various bioactive molecules such as proteins, nucleic acids, and lipids, have been identified as potential mediators for cell-to-cell signal transduction [[3], [4], [5], [6], [7]]. EVs possess stable drug delivery capacity, good biocompatibility, and modifiability, making them novel drug delivery vehicle with significant clinical application potential [8].

Meanwhile, the rise of artificial intelligence (AI) has also significantly advanced computational biology [9,10]. In particular, in 2024, the concept of AI Virtual Cell (AIVC) was proposed by Bunne in Cell [11]. AIVC is defined as a cell model based on a large-scale, multi-scale, multimodal neural network that can dynamically simulate various behaviors of cells, thereby accelerating drug discovery, disease mechanism research, and personalized medicine [[11], [12], [13], [14]]. Recently, our team innovatively proposed the interdisciplinary concept of AI virtual organoids (AIVOs), achieving a high-fidelity leap from physical entities to digital twins [15]. Therefore, based on this background, we innovatively propose the concept and method of artificial intelligence virtual extracellular vesicles (AIVEVs).

In this review, we first summarize in detail the construction, function, applications, and challenges of AIVCs. Then, we present the biogenesis, structure, composition, internalization, and isolation of EVs. Furthermore, we propose the knowledge-driven (white-box) and data-driven (black-box) modeling paradigms for simulating EV biogenesis, cargo sorting, and intercellular communication through the integration of multi-omics data (Fig. 1). The significance of AIVEVs is to build a digital world of intercellular communication that can make predictions and programming. This can greatly speed up the development of EVs-based diagnostics and therapies.

Fig. 1.

The generation and functions of AIVEVs. Build AIVEVs using artificial intelligence virtual modeling technology based on basic data (cell and/or EVs data). AIVEVs can be used to predict potential therapeutic EVs, analyze the intercellular communication produced by EVs, and simulate the characteristics of EVs related diagnostic markers in diseases.

2. Artificial intelligence virtual cells (AIVCs)

Cells are the basic units of life, and their structures have complex biological networks. Cellular systems form a multiscale network, from proteins to organelles, whose homeostasis relies on controlled intercellular communication [16]. To conduct a comprehensive analysis of nonlinear, dynamic, and heterogeneous cellular biological systems, it is necessary to integrate multimodal data [17,18]. Cell sensing and signal transduction can affect migration and differentiation progression. In physiological or pathological environments, cells adapt to changes through metabolic reprogramming and transcriptional regulation. Conventional approaches to model these adaptations, however, are often lengthy cycles, high costs, and low reproducibility [[19], [20], [21]].

Therefore, AIVCs have been introduced as a strategic alternative, which can predict, manipulate, and simulate cellular activity, thus providing insights into cellular biological processes [22]. AIVCs predict cellular processes gained from real biological large-scale data. Digital cell models can accurately predict real cell structure, functionality, and responses of real cells to environmental disturbances [11,23]. As an advanced computational framework integrating AI and systems biology, AIVCs aim to create “digital twins” of cells [11] (Fig. 2).

Fig. 2.

Comparison of real cells and AIVCs. Real cells rely on the central dogma to generate the raw materials for life activities, while AIVCs rely on basic data such as sequencing and genomics to form the molecular information foundation. Real cells have organelles and subcellular structures, while AIVCs use modeling models to simulate life activities within cells. Real cells interact and influence each other through intercellular connections and cross cell communication, while AIVCs simulate intercellular communication based on bioinformatics analysis of interactions.

In summary, AIVCs display three properties: multimodal integration, multiscale modeling, and predictive simulation and generation. Firstly, AIVCs demonstrate multimodal integration, where they fuse multimodal biological data from different sources such as genomics, transcriptomics, proteomics, metabolomics, imaging, and microscopic imaging. Thereby, AIVCs provide biological knowledge that goes further than superficial analyses [11,24]. Moreover, AIVCs demonstrate multiscale modeling, simulating cell biological activity across different levels, such as molecules, organelles, cellular, and tissue. Using multiscale models, AIVCs demonstrate cross-scale interactions where biological models operate on multiscale models, such as those depicting different biomolecular interaction events [11,25]. Finally, the capabilities of AIVCs include predictability, simulation, or generation. Virtual cell models can predict cellular activities in response to genetic perturbations, environmental disturbances, or pharmacological interventions, thus providing a virtual laboratory whereby biological hypotheses can be tested.

2.1. Construction of AIVCs

Generation of AIVCs represents more than the mere development of a single model, but a series of systems engineering tasks involving with data systems, model systems, simulation systems, as well as experimental systems. This process requires the system-oriented fusion of multimodal biological data [25,26]. A prevailing approach within the development of AIVCs depends on the use of large-scale AI models as engines that operate on the fused, representative, vast multimodal biological datasets.

AIVCs accomplish the simulation of strongly nonlinear, dynamic, and stochastic cellular events that often cannot be properly modeled or must be in traditional models. To achieve this high-fidelity digital representation, the underlying AI architectures must exhibit a set of key technical characteristics. First, must be large-scale and high-capacity are essential: the architecture itself needs to be a large-scale neural network to accommodate and integrate massive, multimodal data from the molecular to cellular and tissue levels, thereby enabling high-fidelity simulation. Second, multimodal fusion capability is crucial. The architecture needs specific modules to align and process different types of data, such as genomic sequences, spatial proteomics, and cell images [23]. This often requires combining or innovating neural components for different data modalities. In addition, cross-scale model mechanisms are also required. The model architecture should be able to integrate mechanisms across different biological scales (for example, transmitting molecular interaction information to predict cell behavior), which may involve hierarchical model design or the use of specific neural structures. Finally, interpretability and support for causal reasoning are crucial. When conducting interpretable simulation experiments using “virtual tools,” this architecture may need to incorporate modules for intervention and counterfactual reasoning, such as graph networks [24] or structural causal models.

To meet the requirements of the architecture, it is achieved by using and integrating specialized deep learning models, and each model operates for specific data structures and biological problems. Currently the construction of AIVCs is mainly based on different deep learning models [26]. Models based on Transformer (such as Geneformer, scGPT) take self-attention mechanism as the core, and can model the dependencies of complex sequence data, such as DNA/RNA. Basic biological models infer gene regulatory links from a large number of unlabeled single-cell data, which is helpful for AIVCs [22]. Graph neural networks (GNNs)-based models excel at processing molecular interaction graphs; they can approximate cell state distributions and thereby contribute to effective AIVC modeling, including the simulation of dynamic events like cell differentiation [24]. Meanwhile, convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are the main methods to analyze the spatial and temporal information of cells from images and sequences, respectively. These complementary parts together create complete and useful AIVCs.

2.2. Functions of AIVCs

The AIVCs' functional framework highlights its primary value as a digital model of biological life [27]. Its applications range from the unification of multiple modalities of data as well as the modeling of high-dimensional system states, through the accurate prediction of cellular dynamics [28,29], the realizability of virtual experiments, the production of new states through generative algorithms, and the versatility of personalized application in patient digital twin models [30]. By incorporating algorithms of deep learning, generative algorithms, dynamic simulation algorithms, as well as knowledge embedding algorithms, explainable, actionable, predictive, and scalable AIVCs may be constructed [29]. This would allow researchers to pursue intricate in silico simulations, witness unobservable events, deduce predictions of future states [27].

Multimodal integration, as well as the representation of states, form the most basic as well as essential capabilities of AIVCs [28]. Using this method, this system can encode inputs derived from different technological platforms, such as gene sequence, protein structure, microscopic images, spatial transcriptomics, single-cell RNA sequencing, epigenomic information, as well as three-dimensional organelle structures, into a shared space [28,31]. Prediction of cellular behaviors is considered the most central capability through which AIVCs endows simulated cells with biological dynamics, allowing the simulation of cell behaviors in response to different forms of cell disturbances, such as drug treatment, gene knockout, gene overexpression, environmental stress, or cell-cell interactions.

Virtual simulation is one of the most groundbreaking aspects of AIVCs, allowing researchers to conduct complex simulation studies in the virtual realm that would be difficult, if not impossible, to conduct in real space [27]. With its integrated state formulation and dynamic simulator, AIVCs facilitate scalable drug screening simulation, allowing researchers to simulate the response of different cell types to different drug combinations and concentrations [27,32]. Also, with its unique ability, AIVCs predict rare or hard-to-detect cell states [33], cell lineage inference [34], as well as modeling complex system dynamics such as tumor clonal evolution, immune cell infiltration, metabolic shift, and resistance evolution [35]. Also, with its simulation capabilities, AIVCs further generate biologically plausible cell states that remain unexplored or unobserved in experiments [36,37]. Indeed, digital twins remain one of the key applications of AIVCs technology in clinical medicine as well as precision therapy [38]. With its simulation of multiple patient cell types, patient-specific digital cell models can be derived that allow for personalized predictions as well as simulation modeling of therapeutic protocols. Indeed, digital twin models capture the unique gene expression, mutational, metabolic, as well as immune system properties of the patient's somatic cells, allowing for the simulation modeling of the patient's pathological processes within the computer [35,38].

2.3. Applications of AIVCs

Being a new digital model of biological life, AIVCs have gained much attention within the scientific community. It represents a key tool in future research in the field of biological studies [11,39]. Applications of this tool range from basic biological research, studies of cancer, drug discovery, synthetic biology, and automated laboratory platforms.

In basic biological studies, AIVCs make it possible to infer gene regulatory networks, decode developmental paths, and recreate cell lineage maps in a unified computational system [39,40]. Although traditional biological studies only focus on cell states within a narrowed scope, AIVCs provide continuous mappings of cell fate paths in space, thus determining crucial cell types as transition phases in cell development or regeneration [27,41]. Malignant tumors represent systemic pathologies at the cellular level, so AIVCs technology is suitable for modeling the tumor cell ecosystem [36,42], and can simulate tumor cell clonal composition, metabolic changes, or immune cell interactions [43]. In drug discovery or early development, with the help of cell microspaces and dynamic unified models, AIVCs can find drug targets, screen drugs, understand mechanisms of action, or predict toxic effects in simulation models. This reduces the reliance on animal and cellular testing of candidate drugs. In synthetic cell biology, when designing cells, AIVCs transform the design from trial and error to a informed design-test-optimization cycle and combine with predictive simulation [44]. AIVCs reflect a paradigm shift toward automated biological experimentation. With integration with automated robotic platforms, high-throughput screening equipment, and real-time acquisition infrastructure, AIVCs technology currently corresponds to the role of “an intelligent controller" in basic scientific studies.

3. Extracellular vesicles (EVs)

In the field of cell biology, intercellular communication is a fundamental process that helps maintain tissue homeostasis, regulate development, and respond appropriately to the microenvironment. EVs, membrane-bound vesicles that are actively released by cells, contain biological cargo, such as proteins, nucleic acids, and lipids, and play a central role as messenger in long-distance intercellular communication [45]. EVs can be divided into three subgroups based on their release mechanism, size, or other properties: microvesicles, apoptotic bodies, and exosomes [46]. EVs have shown great therapeutic potential as drug nanocarriers [[47], [48], [49], [50]]. EV-based therapies have achieved extensive breakthroughs in human diseases [51,52]. Multiple studies have shown that EV-mediated cell signaling exhibit high specificity and efficiency under various pathological conditions [52,53].

3.1. Biogenesis, structures, and compositions of EVs

The biogenesis of EVs is generally considered a three-step process (Fig. 3). First, the cell internalizes external substances and membrane-bound proteins via endocytosis, accompanied by plasma membrane invagination, thus forming early endosomes. These early endosomes may interact with other organelles to produce intraluminal vesicles (ILVs). Finally, ILVs mature into late endosomes, also known as multivesicular bodies (MVBs). MVBs may fuse with autophagosomes or lysosomes for degradation, or they may fuse with the plasma membrane to release their contents extracellularly [54,55]. Based on their biogenesis, EVs can currently be divided into three main categories: exosomes (30–150 nm), microvesicles (100–1000 nm), and apoptotic bodies (500–5000 nm) [56].

Fig. 3.

The biogenesis, release, composition, internalization, and functional delivery of EVs. (1) Biogenesis and release (donor cells): Early endosomes develop into late endosomes/multivesicular bodies (MVBs), which are ultimately released to form exosomes—as a classic pathway for EVs production. MVB has several fates: some fuse with lysosomes and are degraded, while others fuse with the cell membrane, releasing their contents into the extracellular space. Some common EVs of different sizes include apoptotic bodies (500–2000 nm), microvesicles (100–1000 nm), and exosomes (30–150 nm). (2) EV composition: The main components of extracellular vesicles include proteins (such as CD63, CD81, CD9, TSG101, Alix, etc.), nucleic acids (such as miRNA), and lipids. (3) Internalization and functional delivery (target cells): Extracellular vesicles enter target cells through various mechanisms, including membrane fusion, clathrin mediated endocytosis, phagocytosis, macropinocytosis, and ligand receptor interactions. After entering the target cell, cargo such as miRNA may regulate target mRNA, and proteins can modulate signaling pathways to achieve functional delivery.

EVs are rich in a variety of bioactive cargo, including nucleic acids (DNA, mRNA, and microRNA), proteins (membrane-associated proteins, cytoplasmic proteins, and signaling pathway proteins), and lipids (cholesterol, sphingomyelin, and phosphatidylserine) [57]. Among them, the tetraspanic membrane protein superfamily (such as CD63, CD9, and CD81) is involved in cargo sorting and EV formation. Therefore, these proteins are classic biomarkers of EVs [58]. In addition, proteins related to the endosome sorting complex (such as Alix and TSG101) are involved in the formation of multivesicular endosomes and the biogenesis of EVs [58].

3.2. Internalization of EVs

EVs are generally believed to deliver genetic material to target cells through three distinct pathways: fusion, endocytosis, and ligand-receptor interactions (Fig. 3) [[59], [60], [61], [62]]. Endocytosis remains the primary mechanism by which EVs enter host cells [[63], [64], [65]]. Endocytic pathways include micropinocytosis, phagocytosis, clathrin-mediated endocytosis, cavitin-dependent endocytosis, lipid raft-dependent pathways, and clathrin- or cavitin-independent pathways [[66], [67], [68], [69]]. Among these, phagocytosis remains the dominant endocytic mechanism [70,71].

Furthermore, EVs can interact with a variety of cell types and activate a variety of receptors, including Toll-like receptors (TLRs), such as TLR1, TLR2, TLR4, and TLR6, thereby generating signals for intercellular communication [72]. Although endocytosis is considered the primary uptake mechanism, these three pathways coexist. Once EVs enter the host cell, they can induce a variety of cellular responses, depending on the material they carry, and the amount of material released in the early endosome compartment [73].

3.3. Heterogeneous characteristics of EVs

3.3.1. Heterogeneity in source

EVs are nanoscale membrane vesicles released by almost all cell types, and the cell type from which they originate has a significant impact on their properties and functions [1]. Different types of cells produce EVs with different biophysical characteristics and biomolecular compositions [8]. For example, EVs released by tumor cells are often rich in specific tumor markers, which can be used for early diagnosis and prognostic assessment of cancer [63]. In addition, the activation state of cells, microenvironment, and biogenetic pathways also affect the properties of EVs, resulting in significant differences in size, morphology, and function of EVs from different cell sources [74]. The origin of cells is relatively complex, which makes EVs have complex situations in intercellular communication and biomarker discovery, but this also brings new opportunities for precision medicine [1].

3.3.2. Heterogeneity in cargo composition

The carrier component of vesicles is central to their function [60]. There are significant differences in the composition of biomolecules (such as proteins, lipids, RNA, etc.) among vesicles from different sources [61] This difference not only influence vesicle function, but also determines their specific roles in disease progression. For example, certain vesicles may be rich in specific miRNAs or proteins that can regulate gene function in other cells [50], thereby driving processes such as cancer metastasis and immune response [7]. In addition, the physiological state of cells themselves and external environmental stimuli can also affect vesicles cargo composition, resulting in significant differences in the vesicle components released by the same type of cells under different conditions [4]. The differences in the composition of this carrier offer novel insights into the functions of vesicles in physiological and pathological processes.

3.3.3. Heterogeneity in size and morphology

The size and shape of EVs are important biological characteristics. Different types of EVs differ in these two aspects. The size of EVs is usually between 50 and 1000 nm; EVs of different sizes may have different biological functions and cellular uptake mechanisms. For example, small exosomes are internalized by target cells via endocytosis, whereas larger microvesicles might enter cells through direct membrane fusion [47]. Also, the shape of EVs (spherical, elliptical, etc.) can impact their biodistribution and function; this shape is often related to their cellular origin and the physiological state of the parent cells [70]. Therefore, it is necessary to understand the difference in size and morphology of EVs so as to know how they work in the body.

3.4. The relationship between EVs and cell homeostasis

3.4.1. The role of EVs in cell metabolism

The role of EVs in cell metabolism is relatively important. They are carriers of intercellular signals thereby regulating cellular metabolism. Studies have shown that EVs can carry a variety of bioactive molecules, such as proteins, lipids, and nucleic acids, which play key roles in cellular metabolic pathways. For example, EVs from tumor cells can transport specific metabolic enzymes and signaling molecules, reprogram the metabolic state of surrounding cells, and promote tumor growth and metastasis [7]. In addition, studies indicate that EVs can regulate the energy metabolism of host cells and enhance cellular antioxidant capacity, so they play an important role in maintaining cellular homeostasis. A deep understanding of the role of EVs in cell metabolism will not only help reveal the mechanism of cell-cell communication but also identify novel therapeutic strategies for treating metabolism-related diseases.

3.4.2. The role of EVs in immune regulation

EVs are relatively important for immune regulation. They carry immune-related molecules (cytokines, antigens, etc.) to regulate the immune reaction. It has been found that EVs secreted by tumor cells can suppress the activity of T cells, cause immune tolerance, and help tumors evade the immune system [33]. Also, immune cell-derived EVs can enhance the immune response and encourage anti-tumor immune activation. EVs also play a regulatory role in autoimmune diseases, they deliver immunosuppressive signals to keep the immune system stable and avoid overactive immune reactions. EVs are not just messengers between immune cells, they are also important regulators of immune responses and may have clinical applications.

3.4.3. The relationship between EVs and cellular stress response

EVs play a key role in regulating cell stress. When a cell is under stress, it can release EVs that carry stress signals, these signals change how nearby cells work. Studies have found that cells produce more EVs when they experience oxidation or other forms of harm. EVs contain antioxidants and repair proteins to manage stress. And EVs help cells protect themselves and heal by controlling how messages travel between cells that got the EVs, so cells can fit into their surroundings. Stress response mechanisms are vital for cellular survival, and they might be significant for disease development. Studying EVs in cell stress response will deepen our understand how a cell reacts to changes in its environment and find new ways to treat related illnesses.

3.5. Effects of cell state on EVs

3.5.1. The relationship between cell cycle and EVs release

The stages of the cell cycle can influence the release of EVs. Studies have shown that cells regulate the number of vesicle types at different cycle stages. For example, during mitosis: cells divide and release more vesicles containing cell cycle - related signal molecules. This may be related to the change in the metabolic capacity of cells during mitosis. Cells need to rely on the release of vesicles to regulate intercellular signals and material exchange, so as to maintain normal division and proliferation. In addition, key regulatory factors of the cell cycle, such as cyclins/kinases, may be involved in the regulation of this situation by affecting the biosynthesis and release of vesicles. This discovery provides a new understanding of the interaction between cells and the environment through EVs, and it may also potentially find new targets for treating diseases such as cancer [2].

3.5.2. Regulation of vesicle properties by the cellular microenvironment

The microenvironment of cells has changed. This change will affect the characteristics and functions of EVs. Physicochemical factors in the microenvironment (such as oxygen concentration, pH value, cell-cell interaction) will affect the biosynthesis, release and content selection of EVs. For example, under hypoxic conditions, EVs released by tumor cells are enriched with pro-tumor miRNAs, proteins, etc., which are helpful for tumor growth and metastasis. In addition, inflammation in the microenvironment can change the composition of EVs, making its immunoregulatory ability stronger. This regulation is of great significance in the occurrence and metastasis of tumors. This shows that EVs are not only information transmitters, but also reflectors and regulators of microenvironment changes [50].

3.5.3. Regulation of cargo selection

Cell signaling pathways are essential regulators of EV cargo selection. Specific signaling cascades can modulate the molecular composition of EVs, thereby directing the packaging of functional biomolecules. For instance, transcription factors such as NF-κB, p53, and HIF-1α, as well as signal transduction proteins like Ras, Akt, and ERK1/2, are known to regulate genes involved in EV biogenesis and cargo sorting. These regulators can alter the loading of microRNAs, proteins, and lipids into EVs. Moreover, in response to external stimuli—including growth factors (e.g., EGF, TGF-β) and cytokines (e.g., TNF-α, IL-6)—cells can actively direct specific biomolecules into EVs through defined pathways, thereby influencing intercellular communication. This selective packaging mechanism not only governs the biological roles of EVs but also informs the development of EV-based therapeutic strategies [69].

4. Artificial intelligence virtual extracellular vesicles (AIVEVs)

With the development of AI, its ability to simulate cellular dynamics and predict cellular behavior has improved, providing a solid foundation for studying AIVEVs [39]. AIVEVs are an extension of AIVCs. Based on the state of AIVCs, researchers can deduce the possible molecular makeup of AIVEVs that are emitted by AIVCs and then recreate the entire “donor cell-EVs-recipient cell” communication circuit within a computer setting [1]. Here, we aim to systematically review the research progress and application prospects of AIVEVs, focusing on exploring the potential of AIVCs in predicting AIVEVs components, analyzing communication mechanisms, constructing disease diagnostic profiles, and tracing vesicle origins (Fig. 4).

Fig. 4.

The workflow for generating and applying AIVEVs. The biogenesis pathway guides the generation of EVs containing cellular molecular sets and provides foundational biological data through multi-omics data. AIVEVs can be built after establishing AIVCs, or AIVEVs can be constructed directly through multi-omics data. Following computational modeling, AIVEV are predicted and validated, and its potential applications were demonstrated through functional simulations (pharmacological targeting, intercellular communication). This workflow outlines the overall framework of AIVEVs' “generation—prediction—application".

The fundamental value of AIVEVs is “in silico pre-experimentation", which integrates multi-omics data and simulates biological mechanisms to predict uncharacterized EV molecular markers, functional mechanisms, and disease associations [27]. It can reduce the blindness of traditional experiments, but it cannot replace the experimental verification—the biological function of EVs and the complex microenvironment interaction need to be verified through in vitro/in vivo experiments. In particular, AIVEVs have their own unique strengths such as being able to quickly produce testable hypotheses, prioritize valuable experimental directions, and simulate difficult-to-continuously-monitor dynamic biological processes; however, actual experiments still cannot be replaced because they are necessary to verify the biological authenticity of AIVEV predictions, clarify the underlying molecular mechanisms, and support the clinical application of related results.

Together they form a synergistic closed loop: AIVEVs and real EVs experiments create a loop of “digital guidance → experimental validation → model optimization” 35. AIVEVs solve the problems of traditional experiments such as long cycles, high costs, and lack of vision, and real experiments offer biological restrictions and data supplements to AIVEVs, promoting the advancement of EV research together. This position matches the current use of AI in life sciences, and it truthfully shows how AIVEVs work and what they can do.

To achieve such value and operational closed loop, the choice of proper AI architectures for AIVEVs modeling is not random; rather, it is primarily determined by the particular biological mechanisms that need to be simulated. Therefore, a strategy for integrating specialized models is required. Foundational transformer-based models such as ScGPT and Geneformer are needed for setting up the starting point for making AIVEVs. These models excel at interpreting single-cell transcriptomic data, enabling a comprehensive, context-aware understanding of the donor cell's state. However, while adept at capturing a cell's overall RNA profile, they were not specifically designed to replicate the mechanisms that govern the packaging of biomolecules into EVs. The second challenge—determining which biomolecules are selected for inclusion in EVs—involves deciphering complex, coordinated molecular interactions, such as those mediated by the ESCRT machinery. Consequently, GNNs are particularly well-suited to this task. By operating directly on graph structures, GNNs can represent and learn from the intricate interactions among proteins, nucleic acids, and sorting complexes thereby providing AIVEVs with a more functionally grounded understanding of cargo selection. Beyond GNNs, CNNs and RNNs or LSTMs [33] also offer important capabilities, such as integrating spatial information from imaging data and modeling the temporal dynamics of EV release and signaling, respectively. Therefore, the current paradigm for creating predictive AIVEVs relies on a blend of these fundamental parts, each handling a distinct aspect of the intricate biology.

Beyond the current foundation models, there are many advanced architectural paradigms that provide avenues to greatly improve the biological realism and function of AIVEVs. Generative models such as diffusion models and variational autoencoders (VAEs) [40] go beyond just making predictions to allow for the creation of new, biologically possible in-silico EV populations. This capability is essential for simulating diseases or abnormal physiological processes that are difficult to observe experimentally, and it enables the evaluation of prediction accuracy. Similarly, Graph Transformers aim to address the limitations of conventional GNNs [24] by incorporating self-attention mechanisms, which can dynamically assign importance to different nodes and edges in molecular interaction networks. The integration of relational and structural information holds significant potential for improving cargo packaging strategies, possibly revealing generalizable and interpretable master rules. Another frontier is Geometric Deep Learning, exemplified by equivariant GNNs [24], which incorporate the three-dimensional structure of biomolecules. This is crucial because molecular interactions adhere to principles of spatial complementarity, and overlooking atomic-level geometry diminishes the accuracy of predicting phenomena such as ligand-receptor binding—a process fundamental to AIVEV–target cell communication. Additionally, frameworks such as Neural Ordinary Differential Equations (Neural ODEs) are especially appropriate for modeling cellular processes that occur continuously over time. Different from discrete-time models, Neural ODEs can simulate the smooth and continuous dynamics of intercellular communication, giving a more natural setting for studying diseases such as chronic disease progression.

To summarize, the biological complexity of extracellular vesicle biogenesis, cargo sorting, and intercellular communication requires a hybrid method that cannot rely solely on one, monolithic AI structure [1]. Biological complexity of extracellular vesicle biogenesis, cargo sorting, and intercellular communication needs a hybrid solution. Future of this area is to have smart integration of all these complementary computational paradigms. An integrated system would leverage the contextual awareness of transformers, the mechanistic reasoning of graph networks, the spatial-temporal modeling of CNNs/RNNs, as well as the improved capabilities of generative, geometric, and continuous models. The final aim is to develop a complete digital twin which can duplicate the existing biology as well as find out the unknown rules of extracellular vesicle functions.

4.1. Multimodal integration for predicting AIVEVs composition and function

EVs as nanoscale carriers secreted by cells, inherit their components from the mother cell and possess unique sorting characteristics. EVs are rich in transmembrane proteins, specific lipids, and nucleic acids, but significantly lacking in membrane-bound organelles, nuclear proteins, and other aspects [4]. EVs have a complex molecular composition, so creating precise AIVEVs requires combining various omics data to mimic the active sorting process of cell contents.

AIVCs' multi-omics map integrates diverse molecular data from genome, transcriptome, and proteome into a unified model. This model leverages single-cell multi-omics technology to resolve cellular heterogeneity. This model uses bioinformatics algorithms to standardize and integrate cross-omics data, thus reconstructing the molecular level cellular function network, offering multi-dimensional data support for studying the mechanism of cell activity [12]. Compared to the traditional single-omics method, it can simulate the molecular changes occurring under normal and abnormal conditions dynamically, thus enabling the discovery of the connections between different omics levels that were previously obscured to uncover. Translating the cellular state into a predicted AIVEV cargo profile requires a structured multi-stage pipeline. This begins with preprocessing and batch-effect correction of multi-omics data to ensure quality and comparability. The most important thing is to use multimodal deep learning models project these heterogeneous data into a shared latent space, creating a complete image of the donor AIVC. Then there's the cargo prediction step utilizes GNNs operating on a bio-inspired graph. This graph formally encodes not just the main ESCRT-dependent route but also includes other sorting rules, such as the random inclusion of cytoplasmic contents, the exclusion of certain subcellular-localized proteins (such as nuclear proteins), and hitchhiking along with strongly interacting molecules on actively sorted cargo (for example, exosomes from myocardial cells contain a lot of HSP60 protein; proteins that bind strongly to HSP60 might also end up in exosomes). This integrated approach enables the model to emulate the complex, multifaceted sorting logic that confers a distinct molecular signature to each EV population, moving beyond simplistic, uniform sorting mechanisms.

The process begins with the essential steps of data pre-processing and harmonization. During this stage, raw multi-omics data undergo rigorous quality control and batch-effect correction to establish a reliable foundation for analysis. The harmonized data are then integrated in a multi-modal fusion step, facilitated by deep learning architectures such as multi-view variational autoencoders, which map distinct data types into a common latent space. This fusion results in a unified digital representation of the donor cell. Central to the predictive framework is the subsequent phase of knowledge-guided inference. Here we have a GNN being applied to a biologically structured graph which contains both known and proposed sorting rules. By propagating information through this graph, the GNN estimates the likelihood of cargo inclusion, thereby simulating the sorting mechanism. Lastly, the pipeline includes an iterative validation and refinement loop that compares the predictions against experimental EV datasets and uses the differences to refine the model to continuously improve its predictive accuracy, forming a learning cycle. Because of the great difference between different EVs, we need to do multi-omics integration analysis including transcriptomics, proteomics and metabolomics to understand what they do. Therefore, the production of AIVEVs mainly depends on computational models that deduce EV secretion patterns based on the complete molecular condition of the donor AIVCs.

In terms of technical implementation, the technical implementation of AIVEVs relies on the in-depth integration of established biological knowledge and mature AI architectures. Specifically, advanced AI frameworks have exhibited strong applicability to EV research scenarios: Transformer-based models (scGPT, Geneformer) for sequence data processing, GNNs for molecular interaction simulation, and generative models (diffusion models) for heterogeneous feature inference, all demonstrate considerable potential to address EV-related research challenges.

But there are still some technical challenges because of two major problems: unanswered biological questions (such as how ESCRT-independent cargo sorting works at the molecular level, and what the real-time regulatory network looks like for single-vesicle release) and inherent limitations in AI algorithms (such as integration bias in multimodal data analysis and poor capture of low-abundance molecular features). To address these issues proactively, AIVEVs adopt a tiered strategy: first, by validating models of established biological processes, then, leveraging existing knowledge to formulate hypotheses about unresolved aspects, and finally prioritizing long-term research trajectories based on the current prominence of key technological concepts (see Table 1).

Table 1.

Boundaries of AIVEV modeling capability.

Modeling Categories	Content currently amenable to precise modeling	Content that can be inferred based on existing knowledge	Long-Term Vision
Cargo Sorting	1.ESCRT-dependent pathway-mediated sorting of nucleic acids (miRNA, mRNA) and proteins (e.g., CD63, TSG101-associated cargo); 2.Sorting of membrane proteins mediated by tetramembrane proteins (CD63, CD81, CD9); 3.Sorting of lipid cargo driven by known lipid-protein interactions (e.g., cholesterol binding to CD81).	1.Sorting trends for cargoes associated with non-ESCRT pathways (e.g., neutral lipid synthesis pathways); 2.Probability distribution of cargo “hitchhiking effects” (e.g., sorting of proteins bound to HSP60); 3.Preferential changes in cargo sorting under pathological conditions (e.g., tumors).	1.Real-time dynamic regulation of cargo sorting at the single vesicle level; 2.Precise simulation of sorting mechanisms driven by cellular microenvironment signals (e.g., hypoxia, inflammatory factors); 3.Predictive sorting specificity for rare cargo (e.g., low-abundance regulatory RNAs).
Heterogeneity characteristics	1.Particle size-based EV subclassification (exosomes 30–150 nm, microvesicles 100–1000 nm); 2.Molecular composition of subclasses defined by classical markers (CD63+, CD81+, CD9+); 3.Molecular heterogeneity supported by bulk EV multi-omics data (proteome, transcriptome).	1.Core molecular characteristics of functional subpopulations (e.g., “metastasis-promoting EVs,” “immunomodulatory EVs”); 2.Heterogeneity trends among EVs from different cellular sources (stem cells, tumor cells, immune cells); 3.Dynamic changes in the proportion of EV subpopulations during disease progression.	1.Precise replication of single EV horizontal multi-omics heterogeneity (including low-abundance molecules); 2.Simulation of dynamic transition mechanisms among functional subpopulations; 3.Personalized prediction of EV heterogeneity driven by individual variations.
Dynamic Release Rate	1.Baseline EV release rate under cellular homeostasis (based on time-series data from batch cultures); 2.Correlation between expression levels of core regulatory factors (e.g., VPS4, ESCRT-III subunits) and release rates; 3.Trends in release rate responses to drug perturbations (e.g., ESCRT inhibitors).	1.The magnitude of variation in EV release rates under pathological stimuli (e.g., viral infection, DNA damage); 2.The fluctuation range of EV release rates across different cell cycle phases; 3.Spatial variation trends in EV release within the tissue microenvironment.	1.Dynamic monitoring and simulation of real-time EV release rates in single cells; 2.Molecular network simulation of multi-signal pathway co-regulation of release rates; 3.Spatiotemporal specificity prediction of EV release in complex in vivo microenvironments (e.g., tumor stroma).
Functional Effects	1.Activation of core signaling pathways mediated by EVs (e.g., PI3K-AKT, MAPK pathways); 2.Phenotypic changes in target cells corresponding to known cargo (e.g., proliferation, apoptosis); 3.Targeted delivery efficiency of engineered EVs (e.g., surface-modified targeting peptides).	1.Speculation on synergistic functional effects of EV cargo combinations (e.g., protein + miRNA); 2.Comparative analysis of functional differences among EVs from distinct subpopulations targeting recipient cells; 3.Prediction of diagnostic efficacy (AUC, sensitivity, specificity) for EVs derived from clinical samples.	1.Molecular-level simulation of EV interactions with target cells (e.g., receptor binding kinetics); 2.Precise prediction of in vivo EV metabolism and clearance pathways; 3.Dynamic assessment of efficacy and safety for personalized EV therapies.

The white box model takes the EV biological mechanism as a deterministic constraint, with the main purpose of converting the known EV production and sorting rules into the basic framework of an AI model to achieve accurate prediction. On one hand, it defines the roles of key molecules such as the ESCRT complex (VPS32, VPS4, etc.) and transmembrane proteins (CD63, CD81) to elucidate EV cargo sorting, for example by excluding nuclear proteins and forming specific interactions with signaling molecules. At the same time, it includes a feature database of EV isolation techniques, using weighted algorithms to correct for enrichment biases of EV subpopulations induced by techniques such as ultracentrifugation and affinity capture. Additionally, it uses computational labeling strategies to deal with the fact that there are lack intrinsic markers for different subgroups of EVs. In the end, all these constraints guarantee that the predictions match the EV creation rules (for instance, ESCRT gene activity is positively related to EV secretion amounts), thereby effectively reducing nonspecific noise. A typical example is the ESCRT machinery is universally conserved across diverse cell types and organisms where it facilitates intercellular communication, pathogen transmission, and nutrient recycling by sorting specific cargo into MVBs for EVs release [75]. For instance, depletion of ESCRT-associated proteins, such as Vps4, significantly reduces EVs release and cargo loading [76]. Similarly, it was also observed in archaeon Sulfolobus islandicus that when both ESCRT-III and Vps4 were absent, EVs generation and gene transfer capability similarly ceased, which demonstrates the evolutionary conservation of ESCRT-mediated regulation of EVs biogenesis [77]. Therefore, virtual cells can predict EVs secretion levels by monitoring the expression of ESCRT core components such as VPS4 and ESCRT-III subunits.

In the context of ESCRT-regulated cargo sorting, the ESCRT pathway selectively packages specific molecules into EVs. Mechanistically, ESCRT machinery directly regulates nucleic acid cargo sorting through protein-protein interactions [78]. Therefore, AIVCs can use expression data of ESCRT-related genes (like ESCRT-II components) and miRNA regulatory networks to predict possible miRNAs or other nucleic acid cargoes in EVs. Furthermore, ESCRT-dependent EVs biogenesis is also regulated by intracellular and extracellular signals [79]. For example, during EV71 virus infection, the ESCRT-dependent pathway promotes the packaging of viral particles into EVs, thereby enhancing viral replication [80]. By mimicking these signaling pathways, AIVCs can predict the presence of pathogen-associated proteins or immunomodulatory factors in EVs.

Building on the above mechanisms, the virtual cell models can construct a knowledge-driven computational framework that predicts EVs secretion levels by watching the expression or activity of core ESCRT components. Additionally, this framework also uses expression data of ESCRT-associated genes and miRNA regulatory networks to predict possible miRNA [[81], [82], [83]] or other nucleic acid cargoes in EVs. When AIVCs produce higher levels of EVs biogenesis-related proteins, such as CD63, than under normal conditions, the model increases the chance that these molecules will be sorted into EVs. By comparing the AIVEVs components generated by the model with the real EVs components [84,85], we can continuously optimize the prediction algorithm parameters, thereby improving the accuracy of the model component predictions.

In addition to predicting components, the AIVEVs prediction system provides a reasonable framework for EVs treatment plans, fundamentally altering the traditional low-throughput, time-consuming trial-and-error R&D approach. The platform can systematically compare the predicted compositions of EVs derived from different cell types such as mesenchymal stem cells, tumor cells, immune cells, and engineered cell types. For instance, tumor cell-derived EVs can be used for targeted delivery of anticancer drugs, thereby improving therapeutic efficacy overcoming drug resistance [47]. Moreover, immune cell-derived EVs are being studied for cancer immunotherapy by changing the tumor microenvironment to boost antitumor immune responses. Additionally, genetic modification of cells enables further optimization of EVs cargo content and targeting specificity, thereby broadening their therapeutic applicability. Engineered EVs have entered clinical trials for conditions including cancer, neurodegenerative disorders, and metabolic syndrome, demonstrating encouraging prospects for clinical use [28].

To further enhance prediction accuracy, integrating multi-omics data and examining the spatial association between ESCRT components and cargo molecules can refine EVs cargo forecasts. For instance, upregulating VPS32—a component of the ESCRT-III complex—not only boosts EV secretion but also modifies their protein profile [73]. This computational screening approach allows prioritization of AIVEVs predicted to be enriched with specific anti-inflammatory cytokines, growth factors, or therapeutic RNAs while lacking immunogenic proteins, substantially reducing the uncertainty and resource demands associated with conventional experimental screening. Overall, this strategy offers an efficient pathway for developing novel EVs-based therapies in areas, such as tissue regeneration, immune modulation, and drug delivery.

Unlike this mechanism-based “white box” method, another one uses generative AI models, such as diffusion models, that are trained straightaway on huge collections of EVs actual compositions. These “black-box" models learn the complex distribution of EVs cargo, enabling them to generate novel AIVEVs profiles without explicit knowledge of sorting pathways. While powerful in data-driven discovery, this approach offers less interpretability than that of knowledge-guided counterpart. The future likely belong to hybrid model that can fully leverage the strengths of both paradigms (Fig. 5). The rapid advancement of single-cell multi-omics technologies has let researchers capture transcriptomic, proteomic, and epigenomic data from individual cells at the same time, generating high-resolution molecular profiles This provides core data support for building AIVCs that can accurately show cellular states. High-throughput data, such as protein, RNAs, lipids, and metabolite profiles can be obtained from single-EV multi-omics databases [86]. By integrating transcriptomic, proteomic, and metabolomic data the molecular composition of AIVCs can be determined, thereby mimicking the physiological state of real cells [87]. The digital AIVCs incorporates specific RNA expression levels and surface protein profiles, which form the foundation of predictive models used to infer which molecules will be sorted into EVs.

Fig. 5.

Knowledge-driven and data-driven paradigms for AIVEV cargo prediction and the future direction of hybrid modeling. This figure illustrates two distinct paradigms for predicting the cargo of AIVEVs and outlines the future direction of hybrid modeling. (1) Knowledge-driven “white-box" paradigm: Starting with donor cell multi-omics data, this approach processes and integrates multimodal data. It constructs a knowledge graph based on established biological rules (e.g., ESCRT machinery, sorting rules, cargo interactions) and uses GNNs for cargo prediction. The predicted AIVEV cargo profile is then validated and refined iteratively using experimental EV cargo data. (2) Data-driven “black-box" paradigm: This paradigm utilizes two primary data sources: donor cell multi-omics data (for the AIVCs-to-AIVEVs path) and bulk/single EV cargo datasets (for the direct EV-to-AIVEVs path). A generative model is trained on this data to sample and generate a synthetic AIVEV cargo profile. The output undergoes statistical validation for iterative improvement. (3) Future direction: Hybrid AIVEV model: The integrated hybrid model combines the knowledge- and data-driven paradigms, leveraging the strengths of both to form a more robust framework for AIVEV cargo prediction. This figure highlights the workflow of each paradigm and the trend towards integrative, hybrid models for enhanced predictive accuracy.

Another modeling approach based on EVs cargo data, also following a “black-box" paradigm, is called the “EV-direct" modeling approach. This method follows a direct data-driven logic research approach, completely abandoning explicit dependence on the multi-omics status of donor cells or known biological sorting rules, and instead treating the EVs population as a complex system. Moreover, this method directly applies deep learning to large-scale EVs cargo-omics profiling data obtained through high-throughput experimental techniques [88], which encompass information at multiple molecular levels, such as proteomics, RNA, and lipidomics. From a theoretical perspective, the establishment of this model is based on an important assumption: Although the sorting process of individual EVs exhibits obvious randomness, as a macroscopic population, their molecular composition distribution is actually collectively shaped by a large number of unresolved intracellular biophysical and chemical constraints.

The implementation of this modeling paradigm mainly relies on the following advanced generative AI architectures: Firstly, VAE, which compresses high-dimensional EVs cargo data into a low dimensional, continuous, and distributed latent space (such as Gaussian distribution) through an encoder network, and then reconstructs the data from this latent space through a decoder network [[89], [90], [91]]. This type of machine learns the most representative features from data, rather than simply memorizing them. When generating new AIVEVs, by sampling within the latent space or traversing along a specific trajectory, and the decoder can output a brand new EVs molecular spectrum that is similar in distribution to the training data. Next is generative adversarial networks (GAN), which includes a generator and a discriminator [92,93]. The task of the generator is to generate AIVEVs data that is “indistinguishable from real data", while the discriminator strives to distinguish between real data and generated data. Both evolve together in the adversarial game, ultimately enabling the generator to produce samples that are almost indistinguishable statistically from the distribution of real EVs goods. Another type is the diffusion models, which is currently the most outstanding architecture in the field of image and biological data generation. The diffusion model gradually adds noise to real EVs data through a forward process until it becomes completely random noise, and then trains a denoising network to learn the reverse process, that is, how to gradually recover structured EVs data from random noise. Once trained, the model can generate highly realistic and diverse AIVEVs by starting the denoising process from pure noise [90,92].

Direct modeling of EVs may make outstanding contributions in revealing the cargo composition of EV subpopulations, such as using single vesicle proteomic data, the model can distinguish subsets of vesicle markers and infer their associated cargo profiles. After extensive data learning, AIVEV has the potential to infer the cargo content in the vesicles represented by a certain biomarker, so that we can have a deeper understanding of a subgroup hidden in the total EVs. Of course, this is currently an idea, as single vesicle proteomics is based on advanced isolation and uses an array of 550 proteins for detection. Therefore, the existing data types are not sufficient for robust EV-direct modeling to facilitate the study of EV subpopulations. Another source of modeling data is analysis of clinical EV samples. Due to individual differences in patients and unknown pathogenic mechanisms, directly modeling the cargo profiles of pathological EVs to generate disease-related AIVEVs is expected to provide EVs biomarkers for disease diagnosis. To achieve this, EVs data can also be used without relying on multi-omics data of cells.

However, this paradigm also faces its own challenges, one of which is the limited interpretability of the model. Due to its decision-making process resembling a “black box" operation, it is difficult to provide clear biological mechanism explanations for the discovered feature combinations, which to some extent limits the translation of research results into practical applications. Another limitation is its heavy reliance on data, where the performance of this method is highly dependent on the size and quality of the training data. The representativeness and completeness of the dataset directly determine the model's generalization ability and reliability.

The “black-box” approach bypasses the need for a complete full intermediate biological process analysis. Instead, it creates a direct link between “source cell features" and “EV output features" by learning from large amounts of multi-omics information [37]. A key advantage is that it circumvents uncertainties in the middle steps. The specific implementation is as follows:

Black-box model addresses key challenges in EV study: Single-vesicle heterogeneity, generative AI models such as diffusion models and VAE learn the distribution pattern of molecules of EVs among various states of cells using single-cell multi-omics data and single-EV database (SVAtlas) [26]. It outputs directly the predicted molecular profiles of individual vesicles and automatically detects functional subpopulations via unsupervised clustering without interference from meaningless heterogeneity noise. To deal with the separation method bias, transfer learning creates a general model that works for all different separation methods. Pre-training on gold-standard data (e.g., density gradient centrifugation) captures the main biological features of EVs [62], then is fine-tuned on limited target method data to learn method-specific biases, which can be used to precisely correct any separation method. To solve the problem of the lack of ground-truth labels for EV subpopulations, semi-supervised and contrastive learning are used together. Black-box model extracts features from unlabeled multi-omics EV data, contrastive learning amplifies discriminative features. Using a few functionally labeled EV samples as anchors allows unsupervised labeling of EV subpopulations, which is closer to their biological nature.

To address model uncertainty and cross-multi-omics error aggregation, the black-box model uses a basic principle of “quantification-tracking-correction-validation” to resolve them. This method is compatible with the data-driven aspect and ensures reliable outcomes: To quantify and control model uncertainty, generative AI such as diffusion models and VAE use probabilistic modeling [44] to produce confidence intervals for prediction results, quantifying uncertainties due to data noise, model architecture, and biological randomness. This is combined with multi-model ensembles [90] and domain-adaptive metrics to evaluate the consistency of predictions among different models and the discrepancies between target and reference data. High-uncertainty results, such as those related to low-abundance molecules or rare EV subpopulations, are highlighted for prioritized experimental verification, thus facilitating efficient optimization of model parameters and improving the credibility of predictions. Cross-omics error aggregation tracing and correction [29] is done with black-box models to find out the errors across different omics areas (for example, proteomics has batch effect and transcriptomics has varying sequencing depths). Variance decomposition calculates how much each type of single-omics error contributes, plus how much inter-omics correlations contribute. Multi-view variational autoencoders (MV-VAEs) and contrastive learning are employed to automatically mitigate omics-specific errors during integration while amplifying functionally relevant features to reduce error aggregation interference. At the same time, taking advantage of low-error omics data such as genomic sequence as references, transfer learning passes on the stable features to correct the bias in high error omics such as lipidomics. In the closed-loop validation phase, a “computational prediction - experimental validation - model fine-tuning” cycle feeds clinical sample measurements back into the model to optimize error correction parameters. Model generalization is validated using EV data across laboratories and separation methods, ensuring robust uncertainty quantification and error correction stability. This provides a solid foundation for the clinical application of AIVEVs [94].

4.2. Simulating AIVEVs-mediated intercellular communication and therapeutic effects

EVs participate in disease processes by delivering signaling molecules, and their molecular characteristics can serve as a basis for predicting disease communication networks [95]. For example, in cancer, tumor-derived EVs can carry specific microRNAs and proteins to remodel the tumor microenvironment and promote immune evasion [96]. In autoimmune diseases, complexes formed by EVs and autoantibodies not only participate in antigen presentation and immune regulation but also serve as predictive biomarkers to help assess disease activity and treatment response [97]. By characterizing EVs through high-throughput technologies, disease-specific communication networks can be inferred from their molecular signatures.

AIVCs-based AIVEVs models can systematically simulate abnormal communication mechanisms under pathological conditions by integrating disease-specific multi-omics data. For example, introducing an EVs communication modules into established virtual tumor microenvironment models can simulate the process by which cancer cell-derived EVs modulate endothelial or immune cells via the delivery of miR-664a-3p [98]. This help to find identify aberrant events and provide optimal choices for testing in labs. Furthermore, in neurodegenerative diseases, EVs from neurons or glial cells transmit abnormal proteins or miRNAs to other healthy cells, causing problems to spread [97]. This method of modeling can discover important disease-related communication events. By leveraging EVs that are adept at communicating with neighboring cells, it provides a robust opportunity to identify disease indicators and therapeutic strategies, thereby enhancing prediction accuracy.

Beyond disease modeling, this framework also applicable to analyzing growth and development processes. These processes have closely coordinated spatiotemporal communication with EVs as important signal carriers. This framework also helps predict cell develop after EVs release, thus identifying which EVs are responsible for guiding tissue growth into right shapes. It also enables investigation into how niche cells modulate stem cell decisions via EVs-borne transcription factors or non-coding RNAs. These applications collectively provide testable molecular hypotheses for regenerative medicine and developmental biology. By establishing a simulated “donor cell-EVs-recipient cell” system, AIVCs can advance cellular communication research from fragmented experimental observations to comprehensive mechanistic modeling and quantitative prediction across the entire communication pathway. Despite ongoing challenges in model integration complexity [11,20], EVs communication models provide a powerful new research paradigm for deepening understanding of dynamic principles of intercellular communication under both physiological and pathological conditions.

EVs exert therapeutic effect is the core behavior mediated by EVs through intercellular communication, which is their core function. Current research has widely confirmed that functionalizing EVs through genetic engineering or chemical modification can significantly enhance their active targeting ability towards diseased tissues. For example, by displaying the GE11 peptide targeting the epidermal growth factor receptor on the membrane of EVs, its recognition and drug delivery efficiency towards EGFR positive tumor cells can be effectively improved. Compared with unmodified EVs and free drugs, it exhibits stronger anti-tumor activity and higher safety [99]. In the field of cardiovascular diseases, by binding single chain antibody fragments against fibroblast activation protein to EVs, engineered EVs can be constructed that can specifically target activated myofibroblasts in the fibrotic region of the heart, providing a platform for precise delivery of anti fibrotic drugs (such as miR-29b mimetics), and their efficacy in reducing cardiac fibrosis has been validated in animal models [100]. In addition, drawing on the homing characteristic of cancer cell metastasis, a biomimetic delivery system constructed by encapsulating nanoparticles in EVs membrane derived from melanoma cells also demonstrated specific targeting ability towards primary tumors and lung metastases [101]. The essence of these strategies relies on the specific interactions between ligands and receptors. In this context, the value of AIVEVs is highlighted. When the model can accurately predicts the component expression profile of EVs based on the state of AIVCs, its surface protein composition can be systematically analyzed to evaluate its natural active targeting potential. Furthermore, researchers can selectively perturb AIVCs (such as overexpression of specific targeting peptides) to predict whether the EVs generated after this design can achieve the desired targeted modification results. This provides a new computational driving path for rational design of high-performance EVs delivery carriers.

In addition to transforming EVs themselves into therapeutic carriers, combining them with biomaterials to construct composite delivery systems is another highly anticipated strategy that can significantly improve the local retention of EVs and achieve controlled release. Studies has shown that the surface physicochemical properties of EVs, such as surface charge, are key factors mediating their interaction with biomaterials. Under physiological conditions, small EVs usually carry negative charges, allowing them to effectively bind to positively charged materials (such as polylysine modified scaffolds) through electrostatic adsorption [102]. For example, negatively charged EVs can be simply mixed with positively charged PEGylated gold nanostars through electrostatic interactions to construct a composite for photothermal therapy [103]. A more forward-looking strategy is to utilize the membrane structure of EVs to endow synthetic materials with biomimetic functions. For example, by encapsulating tumor derived EVs membranes on the surface of mesoporous silica nanoparticles, a hybrid system can inherit the targeting properties of EVs and significantly enhance the targeted delivery efficiency of chemotherapy drugs. In the field of tissue engineering, EVs are loaded into implantable scaffolds such as hydrogels, which can be used to build a composite system for promoting tissue regeneration, and realize the slow release and long-term effect of EVs on the injured site [104]. In the optimized design of these materials, AIVEVs can play a key predictive role. By analyzing the known interacting molecules contained in the expression profile of the EVs generated, the potential binding affinity of the EVs with specific materials can be predicted. This allows for the computational of EVs subtypes with optimal compatibility and functional synergy with the target biomaterial, thus guiding the rational construction of high-performance composite materials.

It should be noted that although the physical properties such as surface charge of EVs exceed the focus of AIVEVs on biomolecules, recent research has found that the composition of their surface proteins directly affects their electrical properties. More importantly, the surface properties of EVs in real biological fluids are dynamically changing, and one of the core influencing factors is the formation of the protein corona. Protein corona refers to a complex structure composed of a layer of proteins spontaneously adsorbed on the surface of EVs after entering the environment such as plasma and tissue fluid [105]. Composition relies on both the source environment and the membrane characteristics of EVs themselves. For example, the protein corona of EVs in plasma is rich in lipoproteins, while the corona of EVs from platelets is rich in proangiogenic factors [105]. This dynamic protein corona will cover the original surface molecules of EVs, greatly changing how they interact with target cells and materials [105]. In such a complex scenario, the ability of AIVEVs is to integrate multiple modes. On one hand, AIVEVs that are created according to AIVCs can help us precise forecasts about what their membranes are made of. On other hand, there is a more straightforward modeling approach known as “EV-direct modeling” that has great potential. This paradigm utilizes large-scale EVs data obtained from clinical samples for training, and the generated AIVEVs can more directly simulate the real EVs state in the complex microenvironment of the body, including modifications such as protein coronas. This allows for a more accurate description of its final physical and chemical properties and biological behavior, providing a more reliable model basis for predicting their actual fate and efficacy in vivo.

4.3. Pathological AIVEVs for inspiring diagnosis

During disease progression, intercellular communication mediated by EVs is critically important. This interaction frequently amplifies pathological processes in real-world cellular environments and EV-mediated communication exacerbating disease through transferring biomolecules to modulate immune responses or facilitate viral transmission. It has also emerging as a key area for the development of new therapeutic targets [106]. This is reflected in the association between specific immune cell subsets and patient survival [107], their involvement in inflammation, tissue remodeling, and functional decline in chronic diseases [108], and their induction of catabolism and metabolic disorders in distant tissues in cachectic conditions [109].

In cancer, the tumor can cause a loss of appetite and muscle atrophy, while EVs can induce catabolism in distal skeletal muscle and fat. Through the transmission of proinflammatory and cachexia factors such as TNF-α and IL-6, these EVs can promote muscle breakdown wasting and metabolic disorders, thus accelerating disease progression [108]. This EVs-driven communication not only makes the disease worse, but also become a major area for developing new therapeutic targets [109]. These examples from real cellular environment suggest that EVs-mediated communication can exacerbate disease by altering immune responses or promoting viral transmission through the transfer of biomolecules. Therefore, we hypothesize that AIVEVs derived from pathological AIVCs can precisely characterize their disease specificity properties.

Two main directions in EVs research are elucidating the disease regulatory mechanisms mediated by EVs, and whether EVs can be used as disease diagnostic biomarkers. Profiling EVs obtained from pathological AIVCs allows for a description of their disease-specific features such as composition, structural phenotype, and functional activity by simulating the regulatory networks [110,111]. This method takes multi-omics data and major pathological stimuli in the pathological microenvironment as input and combines dynamic simulation of disease-related signaling pathway to reconstruct the relationship between EVs characteristics and disease stages [112].

We can speed up the discovery of AIVEVs biomarkers with a method focused on pathological AIVCs. These computer models are built by combining multi-omics data and capable of simulating disease-specific cellular characteristics. The related computer process has a several key stages: building and setting up disease models, simulating EV formation and molecular contents in response to disease signals, finding changed molecules through comparative analysis, and finally ranking candidate biomarkers with bioinformatics tools. The final process ends with experimental validation which is essential to turn these computer predictions into clinically applicable tools.

In addition to the “top-down" prediction path based on donor cell digital models mentioned above, a more direct “from effect to effect" modeling strategy - EV-direct modeling-provides a powerful supplement for constructing disease-specific AIVEVs. The core of this method lies in completely avoiding complex inferences about the multi-omics status of donor cells, and directly utilizing EVs cargo data obtained from high-throughput sequencing of clinical patient samples for modeling [[113], [114], [115]]. By analyzing EVs data from a large number of patient cohorts, generative AI models can learn and capture the overall statistical distribution patterns of EVs molecular features under disease states [116]. The advantage of disease-specific AIVEVs constructed is that they can effectively average the noise and bias caused by individual differences and technical errors in a single patient or a small sample, thereby extracting more stable and representative disease core EVs features. The emergence of AIVEVs directly from clinical EVs big data is not a simple replication of a specific patient sample, but a condensed and summarized molecular portrait of EVs for a certain type of disease (such as lung cancer, Alzheimer's disease). Therefore, it can more directly and realistically reflect the essential characteristics of the EVs population in the context of the disease, providing a more reliable digital basis for discovering highly robust diagnostic biomarkers.

Simulations of intercellular interaction networks can infer EVs-mediated communication patterns in disease context, while mass spectrometry enables high-throughput profiling of the lipidome and proteome of EVs to identify disease-specific biomarkers such as oncogenic mutant proteins or glycosylation patterns [117,118]. These analyses support predicting how EVs modulate intercellular communication. By simulating the interactions between recipient cells and EVs, potential activation signaling pathways can be inferred and the contribution of EVs to disease progression can be pre-assessed [119]. Integrating these virtual projections allowing more accurate inference of EVs-mediated communication behaviors, thus providing a basis for targeted therapy, offering potential evidence for studying the pathology of disease progression and identifying therapeutic targets. Furthermore, it addresses the challenges of low abundance, difficult extraction, and frequent neglect of EVs in tissue samples, thereby advancing mechanistic understanding of how EVs mediate intercellular communication.

Intercellular communication mediated by AIVEVs is critically important, with. Such interactions frequently amplify pathological processes in real-world cellular environments and EVs-mediated communication exacerbating disease through transferring biomolecules (proteins or nucleic acids) to modulate immune responses or facilitate viral spread. It has also emerging as a key area for the development of new therapeutic targets [106], as evidenced by associations between specific immune cell subsets and patient survival [107], involvement in inflammation, tissue remodeling, and functional decline in chronic diseases [108], and induction of catabolism and metabolic disorders in distant tissues in cachectic conditions [109] (Fig. 6).

Fig. 6.

Workflow and applications of disease-specific AIVEVs modeling. The process begins with pathological conditions (e.g., hypoxia, inflammation), which drive the generation of pathological AIVEVs. Computational simulation then generates a disease-specific AIVEV profile, containing unique biomarkersand functional EV subgroups. The core applications of this profile are: (1) Diagnostic biomarker discovery: Machine learning models (e.g., Random Forest, Support Vector Machine) screen highly specific biomarkers from liquid biopsy panels for early disease detection, subtyping, and prognosis. (2) It decodes how pathological AIVEVs cause miscommunication with target cells, leading to phenomena such as immune suppression, tumor metastasis, or neurodegeneration. Traditional EV clinical research has many problems, such as untrackable origin, invasive sampling, and invisible rare subgroups. But this AIVEV method can solve all these problems. It is trackable, non-invasive, scalable for simulation, and able to address rare subgroups.

Furthermore, EVs are highly diverse. They originate from different biological sources and have distinct subpopulations with varying sizes and formation patterns. Researchers can use pathological AIVCs to simulate cell changes during disease. This helps determine if these changes cause cells to release EVs with unique molecular contents through specific pathways. When we apply disease-related stimuli to AIVCs, the model can predict whether the cells activate a non-classical EVs secretion pathways [120]. If activated, the model uses computer-based analysis to define a potential new EVs subpopulation. This generates computer-based evidence for the existence of the new subset. This prediction can then help design a appropriate isolation strategy. This method can search for and confirm the presence of the predicted subset in patient serum [121]. These attempts give us some clues as to how to diagnose certain clinically relevant results by using more specific EVs subsets. When we combine subset-related multi-omics data, the AIVCs model can also trace the cellular origins and formation paths of previously undiscovered EVs subsets, which help us find out the etiology of diseases.

4.4. AIVEVs supplement and promote traditional EV research

The traditional research of EVs has some basic problems in component analysis because there are too many differences among EVs themselves, leading to biased results. First, EVs from different biological pathways and cell types have notable differences in size, composition, and function [122,123]. Second, the molecular composition of EVs is highly complicated. For instance, the glycosylation modification on their surface is constantly changing and varied, greatly impacting the creation, loading, and connection with target cells of EVs. Traditional methods are difficult to provide a comprehensive analysis [122]. More importantly, the physiological or pathological state in which cells are located can significantly alter the EVs components they release. Studies has shown that EVs produced by different modes of cell death exhibit significant differences in protein content, enriched pathways, and ribosomal RNA methylation patterns [123]. These factors collectively result in highly heterogeneous mixtures of EVs isolated from biological fluids, making it difficult to accurately trace their cellular origin and analyze their true functional composition using the traditional experimental based separation before analysis model. As a result, the accuracy and reproducibility of the analysis results are limited. In addition, although proteomics and transcriptomics detection technologies at the level of individual EVs have been developed, they are mainly based on limited detection targets and cannot achieve high-throughput single EV analysis. Currently, it remains difficult to meet the precise analysis of EVs components.

To overcome the above bottlenecks, the AIVEVs model achieves accurate analysis of EVs components through “inside-out" prediction approach. AIVEV first use AIVCs multi-omics maps to simulate the biological processes of EVs under specific pathological conditions, thereby predicting their theoretical components and construct disease specific AIVEVs reference maps. During this process, deep learning models were applied to correct batch effects and technical biases in multi-omics data, significantly improving the accuracy of feature matching [124,125].

Subsequently, by comparing and evaluating the EVs omics data measured in actual samples with virtual prediction maps, the probability of their association with specific cell sources or biological occurrence states can be quantified. However, before the comparative analysis, AIVEVs-based component analysis may have already been completed, requiring only low-throughput target validation of real EVs. This method can effectively integrate and interpret complex information that is difficult to handle with traditional techniques, such as specific molecular features of EVs with different glycosylation patterns or cell death related EVs [122,123], thereby elevating EVs composition analysis from a rough description of heterogeneous mixtures to accurate prediction and traceability of specific functional subgroups. This lays a solid foundation for understanding the specific mechanisms of EVs in diseases and promoting their clinical translation.

For research on EVs, whether exploring their intercellular communication methods or diagnosing diseases, figuring out original source is crucial for understanding their function. However, traditional characterization techniques have significant limitations that severely restrict the precise traceability of EVs. Moreover, it is often difficult to obtain accurate EVs information in complex biological samples [126,127]. Secondly, EVs exhibit high heterogeneity, and different subtypes differ in their formation mechanisms and sizes. However, traditional techniques are difficult to directly distinguish these functionally diverse subgroups [128]. Finally, different separation techniques such as ultracentrifugation and immune capture can significantly affect the purity and abundance of EVs subtypes, leading to biased detection results of molecular markers relied upon for traceability, and even revealing individual differences in the expression of classical markers, challenging their applicability as standard EVs markers [129].

To address the challenges of traditional EVs research, AIVEV is expected to provide digital and predictive research information. Firstly, AIVEV is used to predict potential EVs source specific markers, the omics characteristics of disease cells, and the EVs subpopulations and component features that the cells will produce can be predicted. Subsequently, targeted low throughput detection can be performed during the experimental validation phase. Reverse engineering can also be used to match the source cells that may produce pathogenic EVs through AIVEVs and AIVCs after detecting their compositional characteristics in experiments. The key advantage of this new technology is its ability to directly predict and validate EVs in unpurified samples with high sensitivity and multiple biomarkers. With only a small number of samples, dozens of surface antigens can be analyzed, greatly reducing traceability errors caused by sample preprocessing and effectively addressing the heterogeneity problem of EVs [130,131].

Base on this, AIVEVs provide specific solutions to overcome the technical bottleneck of EVs separation and characterization [132]. Firstly, by simulating the surface proteomic characteristics of EVs secreted by cells under specific conditions, AIVEVs can predict dominant surface markers, providing a theoretical basis for optimizing immune capture or affinity separation methods. For example, when the model predicts that a certain treatment-related EV subgroup is rich in specific membrane proteins, it can directly guide the selection of antibody immunoaffinity chromatography schemes, achieving a transition from empirical screening to rational design. Zhang et al. [53]used the mitochondrial outer membrane protein TOMM20 for immunoaffinity chromatography to isolate mitochondrial derived vesicles (MDVs). However, in this study, the possibility of ignoring mitochondrial inner membrane MDVs and CD63⁺TOMM20⁺MDVs (MDVs encapsulated by CD63⁺membrane) was emphasized. We think that the AIVEV model could give more exact subpopulation separation markers for comparable tests. Moreover, AIVEVs can act as a virtual validation platform for new analytical technologies. Simulating how virtual EVs with specific physical and chemical characteristics would interact with different microfluidic chip designs or sensor surfaces helps assess crucial aspects such as capture efficiency and flow rate before making the actual devices, which saves a lot of time and reduces the costs associated with trial and error [32]. The most revolutionary one is showing off complex EVs mixtures virtually. Researchers could create a digital reference map library by means of AIVEVs, each representing the presumed molecular features of purified EVs subgroups originating from particular cell sources or biological pathways. High-throughput omics tests are carried out on the roughly separated samples, after which deconvolution occurs using the spectral library to quantitatively examine the proportionate parts of different functional subgroups within the mixture. This method avoids the limitations of physical separation and provides unmatched analytical throughput and resolution accuracy for heterogeneity studies [132]. In this framework, the experimental validation results and the computational predictions give each other feedback. Consistency results can improve the model, while inconsistencies suggest new biological problems or technical issues that need to be iteratively improved throughout the entire traceability analysis system. This method can both overcomes the shortcomings of traditional methods in terms of sensitivity, flux, and specificity, and promotes the transition of EVs traceability from relying on a single, unstable biomarker to recognizing them based on multidimensional, large-scale data. This lays a solid technical foundation for exploring the precise functions of EVs in complex pathological conditions such as immune regulation and tumors [80,99,105].

Traditional EVs research is undergoing a methodological change. This change stems from the deep integration of predictive computational models, AI technologies, and rigorous experimental verification [102]. This closed loop, repetitive approach has significantly enhance EVs tracking tech, especially regarding identifying their origin and modes of generation (biosynthetic pathways) at the cellular level, thus driving the field forward. In addition, AI technology can rapidly process many kinds of information, including proteins, RNA, chemical objects [133], and therefore can discover unique combinations of signals that are difficult to identify using traditional methods.

AI technology can process multi-omics data to explore cooperative biomarkers, which are difficult to detect by traditional methods and need to have high specificity and high sensitivity [133,134]. For example, in the early diagnosis of colorectal cancer, AI models have achieved almost perfect diagnostic performance (AUC up to 0.986) by integrating EVs derived miRNAs (such as miR-23a-3p) with traditional tumor markers CEA [135].This marks a significant upgrade of predictive models from single biomarker screening to multi-dimensional feature recognition.

Based on the predicted results, researchers can design more targeted experimental plans for verification. For example, AI-assisted multiparameter flow cytometry can be used for fine classification of EV subpopulations [136], or nano pattern technology based on freeze thaw induction combined with random forest algorithm can be used to identify EVs from breast cancer with high accuracy [137]. These AI-enabled experimental techniques not only improve the efficiency and specificity of verification, but more importantly, they can provide quantifiable performance indicators (such as accuracy, AUC value, sensitivity, etc.), providing accurate basis for subsequent model feedback and optimization.

In the future, the iterative cycle of “prediction verification feedback" using AIVEVs will continue to promote continuous exploration in the EVs field. When AIVEVs-based prediction highly matches the experimental results, the reliability of the model is strengthened, and the biological laws revealed (such as the correlation between specific surface antigens and heart transplant rejection [138]) are also confirmed, thus laying the foundation for clinical translation. Although still facing challenges such as data standardization, model generalization, and algorithm black boxes [139], AI-driven prediction validation integration is undoubtedly accelerating the transformation of EVs basic discovery into clinical application scenarios such as early disease diagnosis, treatment monitoring, and personalized treatment [135,138,140].

4.5. Challenges and future perspectives

The development of AIVEVs presents a significant methodological challenge for computational biology that tries to create forecastable digital versions of EVs’ creation and operation. But this effort encounters substantial challenges on multiple fronts: technical, empirical, and evaluation and ethical. Subsequent discussions will address these main issues by adding important points to form a path for developing strong and reliable AIVEV models.

AIVEV model building is inherently restricted by substantial biological unknowns and methodological constraints, thus affecting the scalability of multiscale simulations, treatment of randomness, and interpretability of the model. A fundamental challenge arises from different EVs are made in various, usually overlapping ways [98,111]. Without a full grasp of how each of those separate pathways work and what kind of things they sort out for their cargo, white box models that depend on structured knowledge graphs have built-in uncertainty. They cannot achieve perfect fidelity because they're only as good as the incomplete biological knowledge they contain. On the other hand, data-driven black-box models such as GANs or diffusion models are limited by the resolution and bias of current experimental omics data. Current bulk EV profiling techniques tend to overlook the differences between different EV subgroups, and the technological hurdle of performing high-throughput, single-vesicle multi-omics evaluation along with exact cell-of-origin tracking poses a significant barrier to developing highly accurate models. Also, both modeling approaches have to deal with the inherent randomness of cargo loading, which involves the unintentional, passive inclusion of nearby molecules into the developing vesicle [133]. This stochastic aspect is hard to separate from the actively sorted cargo, complicates the task of making certain, fully interpretable predictions about EV contents.

Besides the above modeling-related restrictions, another important but frequently neglected bottleneck is that the current public EV datasets do not provide enough support for developing good AIVEV models, so it is necessary to create a new standard database specifically for AIVEV training and validation. The mainstream repositories (ExoCarta etc.) have significant technical heterogeneity in isolation and characterization methods which leads to a 30–50 % CV in the protein abundance profile of EVs from the same cell line, which directly impacts the stability of model feature learning. Top-tier AI models need multi-dimensional data along the “molecular composition-spatial distribution-functional effects" axis, but existing databases are inadequate: SVAtlas has single-EV multi-omics profiles but no functional validation or clinical data [126], and Vesiclepedia focuses on bulk single-omics data, missing the high-res single-EV-level info needed to predict EV heterogeneity [136]. Also, clinical samples constitute less than 20 % of the present datasets, there is a lack of big-scale disease group data [96], and outdated entries because of infrequent updates (for instance, ExoCarta) restricts clinical application possibilities. In order to solve all the above problems, we need to establish a standardized database with three major parts: unified technical standards, including isolation protocol (ultracentrifugation combined with density gradient centrifugation), characterization threshold (mass spectrometry resolution greater than or equal to 70,000, sequencing depth greater than or equal to 100 times), pre-processing pipeline (median normalization), etc., to reduce the inter-study CV to less than 10 % [116]. It needs to include the entire data chain: single-EV multi-omics data [100], functional validation data (such as EV internalization efficiency, in vivo delivery efficiency), and clinical metadata (such as disease type, treatment effect); and better usability features such as prioritizing large-scale clinical groups to connect EV molecular characteristics with phenotypes [96], using standardized data formats (FASTQ/CSV) and open sharing methods, setting up a double review (machine + expert) quality control system with monthly updates to guarantee data accuracy and timeliness, fitting AIVEV's “molecular-functional-clinical" prediction idea, ensuring data is accurate and timely.

Data support has a basic framework in place, so moving AIVEVs from a predictive framework to a validated discovery tool needs a definite, repetitive loop among in silico prediction and in vitro/vivo experimental validation. AIVEVs can be seen as a strong generator of hypotheses [134], for example, finding possible disease-specific EV biomarkers or looking for EV subgroups that might contain more of certain helpful things, such as particular cytokines or regulatory RNAs. But the accuracy of those predictions has to be proven by doing experiments. The validation cycle that I propose should have experimental confirmation as a necessary endpoint. A good plan would be to use genetic engineering to change how important parts inside cells work, such as pieces of something called the ESCRT machine or things that might be carried by EVs, and then see what happens to the EVs and how they work because of this change. The outcomes that the AIVEV model predicts, having been simulated under the same perturbation conditions, are then directly compared with these experimental results. Tech that advances validation to the level of single vesicle [131], such as nano flow cytometry and advanced microfluidic platform for EV manipulation and analysis, will become necessary to do more rigorous, higher-res tests. They permit investigation into the differences among EVs and follow-up on the results in a controlled setting, supplying the high-resolution data needed to improve and test the AIVEV model's forecasts [141], thus completing the circle between simulation and living beings.

In order to make sure that the validation results are valid, we need to clearly specify what kind of data sets can be used for validation, because the main advantage of AIVEV is that it can accurately simulate the molecular composition and functions of EVs. And thus, the validation dataset has to meet these three requirements: high resolution, multi-dimensional and traceable, and match exactly with the AI model's prediction of molecular profile, spatial distribution pattern and function outcome. By relying on the most recent advancements in EV research and the special advantages of authoritative databases, the following experimental datasets have been confirmed to have good compatibility for AIVEV validation:

Among them, nFCM-MS-generated data are the most convincing. This integrated platform allows for qualitative and quantitative profiling of proteins within each individual vesicle, which matches perfectly with the protein abundance heterogeneity anticipated by AIVEV. And Hou et al. have created a single-EV multi-omics database called SVAtlas that includes 8120 protein entries from 137 million individual EVs. It also confirms the possibility of applying it directly in the characterization of EV subpopulations and biomarker screening, and gives evidence for the validation of AIVEV. For example, when we target the CD63⁺CD81⁺ tumor EV subpopulation that was predicted by AIVEV, this technology can be used to experimentally measure the co-expression rate of both markers on a single-vesicle basis, thereby validating the accuracy of the subpopulation-specific prediction.

Also, copy number data of nucleic acid molecules such as mRNAs and miRNAs obtained by single-vesicle RNA sequencing is a crucial validation criterion for AIVEV's prediction of nucleic acid components. Besides its protein library, the SVAtlas database has documented 106 RNA entries, which include mRNAs, miRNAs, and lncRNAs. Data architecture indicates that single-EV-level nucleic acid heterogeneity datasets can distinguish molecular signatures of EVs from various disease conditions, offering direct experimental evidence for AIVEV-driven transcriptome forecasts [86].

Data produced through spatial transcriptomic systems (such as Visium tech) and spatial proteomic methods (such as MIBI-TOF technology) can strongly support the “EV–cell” spatial interaction designs forecasted by AIVEV. In particular, these datasets accurately define the spatial distribution of EVs within tissue microenvironments and how close they are to donor/recipient cells, thus resolving a major shortcoming of traditional bulk analysis which overlooks EV spatial specificity. For example within the tumor microenvironment, spatial profiling data could directly confirm the directionality of AIVEV-predicted migration of tumor cell derived EVs towards macrophages.

Additionally, there are three kinds of functional datasets that can directly validate AIVEV's prediction about the functional outcome: EV internalization efficiency (measured by flow cytometry to assess the rate of cellular uptake), target cell signaling pathway activation (for example, the degree of PI3K/AKT phosphorylation detected through Western blotting), and in vivo targeted delivery capability (assessed based on the tissue distribution curve obtained from in vivo fluorescence imaging). Previous studies have shown that integrating these functional readouts with molecular composition data greatly enhances the reproducibility of EV-related research results. This kind of approach aligns well with AIVEV's main idea of integrating “molecular–function" predictions.

To ensure that validation outcomes are reliable and comparable despite different datasets, a robust framework for quantitative comparison must be established. This framework should comprise multiple, hierarchically organized indicators. The system has to include small differences among molecules as well as large changes in how well things work all at once, following the rules of computational biology. Implementation methods are introduced below specifically:

Pearson correlation coefficient was applied to check the overall trend similarity between AIVEV-predicted molecule amounts and actual measurements, whereas Spearman correlation coefficient was utilized to examine the level of agreement between rank distributions. For the core molecules (such as biomarkers and functional molecules), the correlation coefficient had to be at least 0.7. Furthermore, absolute deviation was also measured by two measures, root mean square error (RMSE) and mean absolute error (MAE). According to the quality control standards of the ExoCarta database, the MAE value should be ≤ 0.15. For example the predicted abundance of miR-21 based on AIVEV, when the Pearson correlation coefficient between the measured and predicted values is 0.78 and the corresponding RMSE is 0.12, such results meet the set evaluation standards and indicate that AIVEV's forecasts are highly consistent with the actual experimental outcomes. Relative error (RE) was used to determine the difference between the AIVEV-predicted and experimentally determined functional effect values, which were calculated by the following equation:

$$\text{RE}=\frac{\left|\text{Predicted}\text{Value}‐\text{Measured}\text{Value}\right|}{\text{Measured}\text{Value}}\times 100\%$$

Based on established criteria for EV function assessment, the key functional metrics such as the efficiency of EV internalization and the suppression rate of target genes were required to have a RE ≤ 15 %. For example, if AIVEV predicts that the macrophage uptake efficiency of tumor-derived EVs will be 45 %, and the experimental measurement shows it to be 40 %, then the RE will be 12.5 %, which satisfies the set quantitative accuracy requirements.

For time-series data such as EV secretion kinetics and temporal dynamics of target cell response, we used the dynamic time warping (DTW) algorithm to evaluate the similarity between the predicted and measured curves. After the data was normalized, a DTW distance threshold of less than or equal to 0.2 was set for validation. Furthermore, this algorithm can be successfully applied to validate the dynamic model of EV secretion [100], which can robustly capture the temporal synchrony between the predicted and experimental datasets (Fig. 7).

Fig. 7.

Workflow example of the process from AIVEV prediction to experimental verification. Step 1 (In Silico): The AIVC model forecasts that AIVEV-X is enriched with protein A and miRNA-B under specific pathological conditions. Step 2 (In Vitro): (a) Employ size exclusion chromatography (SEC) to isolate authentic EVs. (b) Conduct physical characterization using NTA and TEM. (c) Detect the abundance of protein A and miRNA-B utilizing high-sensitivity mass spectrometry and RNA sequencing. Step 3 (Feedback): Evaluate the methods for quantifying the comparison between predictive models and experimental outcomes and provide a critical analysis of how factors such as detection limits and the purity of current experimental techniques may influence the accuracy of the verification process.

Beyond technical validation and quantification, setting up appropriate standards for AIVEVs and considering their ethical implications are two sides of the same coin; both strive to ensure the results are biologically feasible and prevent false conclusions. Benchmarking should be conducted along two distinc yet complementary directions. First, according to composition and biophysics, the model should be able to make EVs within the known biological constraints [126]. This includes verifying the absence of molecules not typically found in EVs (e.g., nuclear proteins lacking export signals), and implementing scoring systems that prioritize high-confidence, biologically relevant cargo over the stochastic incorporation of low-abundance background molecules. Second, from a functional and contextual standpoint, AIVEVs need to be evaluated in simulated interactive environments [102]. Futhermore, the predicted function, such as targeting a particular cell type or activating a downstream signaling pathway, should be supported by existing biological evidence or presented as a specific, testable hypothesis. Failure to meet these benchmarks raises ethical concerns, such as the generation of speculative or exaggerated claims [117]. If such erroneous in silico predictions guides experimental design, it might result in a lot of wasted scientific resources, wasting time, funds, and experiments effect on pursuing artifacts generated computationally. Therefore, rigorous benchmarking is not merely a technical task, but also an ethical responsibility to maintain the integrity of the research based on these virtual models.

With robust benchmarking criteria and ethical measures to guarantee biological accuracy, AIVEVs could successfully bridge basic research and clinical translation via an organized translational course. The main pathway that AIVEV uses to move EV research into the clinic can be described as a closed loop of “Digital Guidance - Experimental Validation - Clinical Implementation” [141]. For diagnostics, AIVEV first takes all kinds of information about EVs from different parts of the body and integrates them with established clinical knowledge to find and choose which things show sickness better than others. Following pilot clinical validation to optimize assay performance, it facilitates the development of user-friendly diagnostic kits. Subsequently, multi-center trials are conducted, culminating in regulatory approval and eventual hospital implementation. In terms of applying the therapy, AIVEV simulates natural or artificial therapeutic EV protocols [99], improves the scalability of production procedures such as cell culture, isolation, and purification, and establishes multidimensional quality control systems involving physical properties, biological activity, and safety [111]. Subsequently, nonclinical efficacy and safety assessments are conducted, followed by phased clinical trials (Phases I-III). This process, supported by regulatory engagement [111], aims to expedite market approval. Post-marketing (Phase IV) studies then monitor long-term effectiveness and safety, informing optimal clinical use. Additionally, the entire translation procedure relies upon a standardized EV database for dependable data assistance [113]. Furthermore, linking “AI Algorithm Teams-Basic Research Teams-Clinicians-Corporate R&D Teams” connects tech feasibility to clinical need. It addresses bottlenecks such as lack of technical standardization [102], insufficient regulatory framework, and difficulty in large-scale production, thereby realizing efficient transition from basic research to clinical application (Fig. 8) [133].

Fig. 8.

Developmental stages of AIVEV within 10 years: from proof-of-concept to clinical tool.

But it needs to be embedded all along in the development of AIVEVs with a complete set of ethical, privacy and security frameworks for it to work successfully. Proactively addressing these issues is not merely an adjunct task but a fundamental requirement for successful ethical review and approval. It calls for a multi-faceted method focused on strong governance, forward-looking risk mitigation, and open compliance. One important part of this effort is using robust protocols for data privacy protection and research conduct. Ethical data sourcing starts with required Institutional Review Board (IRB) approval and informed consent. Technically speaking, protecting patients' privacy means applying strong de-identification methods to get rid of personal details, and sticking to the idea of using just enough data for what you need, so you only take the information needed for your model's job. In parallel with privacy protection, a proactive stance on identifying and managing security risks is essential. Malicious usage possibilities exist, such as intentionally designing dangerous biological agents or altering the results of models. Concrete safeguards are needed for these. Effectively mitigate it by setting up a double-check system for risky model forecasts, making sure that any output related to pathogenicity or toxicity has been independently verified before any experiment takes place. And also keep strict control over the access and audit trails of the model generation tool to avoid misuse. Finally, adhering to the changing norms and supervisors' regulations leads to the acceptance by the clinic. It includes the production of clear documents, such as thorough model cards that explain the intended uses, limitations, and possible biases. Researchers would be able to greatly ease their way through the ethical review process and into clinical trials if they were to actively align AIVEV development with the emerging regulatory standards for AI-based medical devices. To summarize, incorporating this entire framework makes ethics and security go from being sideshows to being mainstays, as the core supports for doing good and useful translations of AIVEV technologies.

5. Conclusion

In summary, integrating AIVCs with multi-omics data and biological pathway analysis enables the construction of AIVEVs. AIVEVs can predict EVs component expression profiles, identify therapeutic EVs candidates, and model EVs-mediated intercellular communication, yielding potential EVs biomarkers and insights into disease-associated communication behaviors. When coupled with experimental verification, AIVEVs can offer a digital solution that improves the precision and efficiency of EVs analysis, accelerate the progress of EVs research in clinical translation of EV-based diagnostics and therapies, and reduce the time and cost of traditional experimental methods.

CRediT authorship contribution statement

Han Liu: Writing – review & editing, Writing – original draft, Project administration, Conceptualization. Shiyu Li: Writing – review & editing, Writing – original draft, Project administration, Conceptualization. Jian Wang: Validation, Investigation. Jiacan Su: Writing – review & editing, Conceptualization.

Ethics approval and consent to participate

Confirm that no ethical issues are involved.

Declaration of competing interest

All authors declare no conflict of interest.