Clinically Interpretable Extracellular Vesicle Gene Model for Non-Invasive Liver Cancer Diagnosis

Clinically Interpretable Extracellular Vesicle Gene Model for Non-Invasive Liver Cancer Diagnosis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Clinically Interpretable Extracellular Vesicle Gene Model for Non-Invasive Liver Cancer Diagnosis Yan Zhang, Zhengying Mo, Lei Zhang, Zhangming Zhou, Zhaohan Wei, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8217267/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 14 Feb, 2026 Read the published version in Scientific Reports → Version 1 posted 8 You are reading this latest preprint version Abstract Hepatocellular carcinoma (HCC) is a major cause of cancer death worldwide, underscoring the need for early non-invasive diagnostics. This study developed an interpretable model using extracellular vesicle (EV)-derived RNA signatures from exoRBase 2.0. Six machine learning algorithms were compared, with the Deep Neural Network (DNN) achieving superior performance (AUC = 0.8877). Ten diagnostic mRNAs (MTRNR2L8, HBB, PF4, FTL, MTRNR2L12, TMSB4X, PPBP, OST4, ACTB, and S100A9) were identified, among which MTRNR2L8 was the most significant predictor. SHAP and Kolmogorov-Arnold Networks (KAN) revealed nonlinear feature–outcome relationships and potential biomarkers. An online platform was created for real-time clinical use. This tool offers a robust, interpretable, and non-invasive method for early HCC detection, potentially improving diagnostic timeliness and decision-making. Further validation in prospective cohorts is warranted. Health sciences/Biomarkers Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Oncology Diagnostic biomarker Hepatocellular carcinoma Extracellular vesicle Interpretability Non-invasive Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Hepatocellular carcinoma (HCC) represents the predominant form of primary liver cancer and stands as the third leading cause of cancer-related mortality worldwide, with an increasing incidence in many regions 1 , 2 . The prognosis of HCC remains poor, with a 5-year survival rate below 20% in most countries, primarily due to late diagnosis when curative treatments are no longer effective 3 – 5 . Early detection of HCC is crucial for improving patient outcomes, as it enables timely intervention with potentially curative therapies such as surgical resection, liver transplantation, or ablation techniques. Current diagnostic approaches for HCC rely heavily on imaging techniques (ultrasonography, computed tomography, magnetic resonance imaging) and invasive liver biopsies, which have limitations including radiation exposure, high cost, operator dependency, and sampling errors 6 , 7 . Conventional serum biomarkers such as alpha-fetoprotein (AFP) lack sufficient sensitivity and specificity for early-stage HCC detection 8 . Therefore, there is an urgent need for non-invasive, accurate, and accessible diagnostic tools for early HCC detection. Extracellular vesicles (EVs) have emerged as promising sources of cancer biomarkers 9 , 10 . These membrane-bound vesicles, including exosomes and microvesicles, are released by cells into body fluids and contain various molecular components including proteins, lipids, and nucleic acids that reflect the status of their cells of origin 11 – 13 . Recent studies have demonstrated that cancer cells, including HCC cells, secrete EVs containing specific RNA signatures that differ from those released by normal cells 14 – 17 . These EV-derived RNAs can be detected in liquid biopsies, providing a potential avenue for non-invasive cancer diagnosis. With the advancement of high-throughput sequencing technologies and machine learning algorithms, it is now possible to analyze complex EV-derived RNA expression patterns to identify diagnostic signatures for various diseases 18 , 19 . Machine learning models can integrate multiple biomarkers to improve diagnostic accuracy beyond what single biomarkers can achieve 20 . However, the "black box" nature of many machine learning algorithms limits their clinical application, as understanding the biological relevance and decision-making process is crucial for medical implementation 21 – 23 . In this study, we aimed to develop and validate an interpretable diagnostic model based on EV-derived gene signatures for non-invasive detection of HCC. We utilized RNA-seq data from the exoRBase 2.0 database, which contains comprehensive information on messenger RNA (mRNA) and long non-coding RNA (lncRNA) in EVs from healthy individuals and HCC patients 24 , 25 . We compared multiple machine learning algorithms to identify the optimal approach, selected key features based on their importance, and employed interpretability frameworks to explain the model's predictions. Additionally, we developed an online platform to facilitate the clinical application of our diagnostic model. Our approach addresses the need for accurate, non-invasive, and interpretable tools for early HCC detection, potentially contributing to improved patient outcomes through earlier intervention. Methods Data source The information utilized in this study is sourced from the exoRBase 2.0 database 24 , 25 , a dedicated repository concentrating on long RNAs (exLRs) within EVs. This database encompasses details regarding messenger RNA (mRNA), long non-coding RNA (lncRNA), and circular RNA (circRNA), which consolidates RNA-seq data from 905 EV samples obtained from four distinct human bodily fluids (blood, urine, bile and cerebrospinal fluid), offering comprehensive analysis and visualization of RNA expression patterns, as well as modifications at the functional pathway level and insights into the variability of circulating EV origins. Specifically, we retrieved the following datasets from the exoRBase 2.0 database ( http://www.exorbase.org/exoRBaseV2/download/toIndex ), including exosomal data from HCC, exosomal data from healthy individuals, and annotation data for lncRNA and mRNA. The dataset comprises mRNA and lncRNA expression data from both healthy individuals and HCC patients, with the healthy cohort consisting of 112 samples with 15,966 mRNA samples and 7,115 lncRNA samples; the HCC cohort contains 118 sampels with 16,023 mRNA samples and 1,118 lncRNA samples. The study design was shown in Fig. 1 . Data preprocessing To ensure data quality, raw sequencing data were processed using Python scripts. mRNA and lncRNA expression profiles were extracted and labeled according to HCC and healthy control samples. To eliminate scale differences across features, all expression values were standardized using Z-score normalization (mean = 0, standard deviation = 1) 26 . To reduce noise and enhance signal quality, features with expression levels below 10% of the overall median or with low variance across all samples were excluded from downstream analyses. Data Division The dataset was randomly divided into a training set (70%) and a testing set (30%). The training set was used for model construction and hyperparameter optimization, while the testing set served to evaluate model generalizability. A five-fold cross-validation strategy was applied to the training set to ensure model stability and reduce overfitting. Model Development Six machine learning algorithms were explored, including Logistic Regression, Random Forest, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Decision Tree, and Deep Neural Network (DNN). These models were trained on either EV-derived mRNA or lncRNA expression data. The outcome variable was defined as binary: 0 for healthy individuals and 1 for HCC patients. Input variables included normalized expression values of mRNAs and lncRNAs. SHapley Additive exPlanations (SHAP) were used to assess feature importance, and models were compared across different feature subsets to identify the optimal number and combination of predictors. Model evaluation To Model performance was comprehensively evaluated using multiple metrics: Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, precision, recall, and F1 score. AUC was used to assess the model’s ability to distinguish between HCC and healthy samples, with higher values indicating better discriminative performance. Accuracy reflected the proportion of correctly classified samples among all predictions. Precision indicated the proportion of true positive cases among all positive predictions made by the model. Recall measured the proportion of true positive samples correctly identified among all actual positive cases. F1 score, the harmonic mean of precision and recall, provided a balanced measure of overall model performance. Formulas for these evaluation metrics are provided in the supplementary materials. $$\:\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\frac{\text{T}\text{N}+\text{T}\text{P}}{\text{T}\text{P}+\text{T}\text{N}+\text{F}\text{P}+\text{T}\text{N}}$$ $$\:\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\frac{\text{T}\text{P}}{\text{F}\text{P}+\text{T}\text{P}}$$ $$\:\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}=\frac{\text{T}\text{P}}{\text{F}\text{N}+\text{T}\text{P}}$$ $$\:\text{F}1-\text{S}\text{c}\text{o}\text{r}\text{e}=\frac{2\times\:\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\times\:\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}+\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}$$ Among them, true positive (TP), false positive (FP), true negative (TN), and false negative (FN) represent the four possible outcomes of classification results. Internal validation was conducted through five-fold cross-validation and the test set was used for external validation to comprehensively estimate the predictive performance of the model. Model Interpretation SHapley Additive exPlanations (SHAP) To address the “black-box” nature of machine learning models, the SHapley Additive exPlanations (SHAP) framework was employed to interpret the predictions of the optimal model. SHAP is based on cooperative game theory and quantifies the marginal contribution of each feature to a model’s prediction 27 . It provides both global and local interpretability: Global interpretation was achieved via summary plots, ranking features by their average absolute SHAP values to indicate their overall impact on model output. Local interpretation was performed using force plots, which visualize how individual features contribute to the prediction for specific samples. In this study, SHAP values were used to evaluate the contribution of EV-derived gene features to the diagnostic classification of HCC, enabling the identification of potential key biomarkers associated with disease onset. Kolmogorov-Arnold Networks (KAN) To further enhance model interpretability and predictive performance, Kolmogorov–Arnold Networks (KAN) were implemented as an alternative to traditional multilayer perceptrons (MLPs). Based on the Kolmogorov–Arnold representation theorem 23 . KAN introduces learnable activation functions along the edges of the network, forming a dual representation to the node-centric activation in MLPs. This architecture facilitates the derivation of explicit mathematical expressions describing the relationships between input variables and outcomes, offering improved model transparency through a sparse and interpretable network structure. Online Computing Platform To enable real-time clinical application, the final predictive model was deployed as a web-based platform using the Streamlit framework. This tool allows users to input EV-derived gene expression values and receive immediate diagnostic predictions, facilitating non-invasive early detection of HCC in practice. Statistical analysis All analyses and computations were conducted using R (version 4.4.1) and Python (version 3.9.12), with a two-sided P value < 0.05 considered statistically significant. For machine learning analyses, the XGBoost algorithm was implemented using the eXtreme Gradient Boosting package (version 2.0.3). Other algorithms, including Deep Neural Network (DNN), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), decision tree (DT), were executed using the scikit-learn package (version 1.3.2), alongside pandas (version 1.5.3) and numpy (version 1.23.5) for data manipulation and numerical computations. The DNN model was developed with the TensorFlow framework (version 1.13.1), while the Kolmogorov–Arnold Network (KAN) was constructed using the pykan package (version 0.2.8). Results Model development and comparison Firstly, we constructed diagnostic models using six algorithms based on extracellular vesicles (EV)-derived lncRNA and mRNA transcriptome data of HCC, respectively, and evaluated the performance of the models. As shown in Supplementary Table 1, DNN algorithm performed remarkable classification capability on the EV-derived mRNA data in the test set, achieving the highest accuracy of 0.8816, the precision of 0.9375, F1 score of 0.8696, and the AUC of 0.9324 compared to other five algorithms including logistic regression, random forest, SVM, XGBoost and decision tree. The diagnostic model constructed based on lncRNA data also demonstrated excellent classification performance with DNN, with F1-score of 0.7765, and the recall of 0.8919 (Supplementary Table 2). Despite Random Forest and XGBoost algorithms performed similarly to DNN on certain metrics, DNN algorithm demonstrated better stability and overall performance based on both mRNA and lncRNA features. Thus, the optimal model was selected based on EV-derived mRNA and lncRNA characteristics. Features selection To enhance the practicality of the diagnostic model, we evaluated the model performance within different number of EV-derived mRNAs and lncRNAs characteristics including top 20, top 15, top 10, and top 5 features for model training based on SHAP feature importance rankings. As shown in Fig. 2 , the AUCs of the mRNAs model on the test set were 0.9324 with all features, 0.9279 with top 20 features, 0.9089 with top 15 features, 0.8877 with top 10 features and 0.8628 with 5 features. The AUCs of lncRNAs model on the test set were 0.8656 with all features, 0.8663 with top 20 features, 0.8614 with top 15 features, 0.8150 with top 10 features, 0.7159 with top 5 features (Supplementary Fig. 1). In addition to the detailed evaluation metrics as shown in Supplementary Table 3, the diagnostic model with top10 genes was identified as the optimal model considering both the number of features and model performance. Model Interpretation Through the SHAP framework, we performed an interpretability analysis of the model's predictive outcomes and identified genes that were pivotal in the diagnosis of HCC. The key variables contributing to HCC early diagnosis were MTRNR2L8, HBB, PF4, FTL, MTRNR2L12, TMSB4X, PPBP, OST4, ACTB and S100A9, among which MTRNR2L8 was the most significant contributor (Fig. 3 and Supplementary Fig. 2). On the other hand, SHAP local interpretability revealed the predictive outcome for individuals. The diagnostic result was 1.00, indicating that the probability of being diagnosed with HCC was high (Fig. 4 ), The decision diagram reflected the decision-making process of key features on the output results (Fig. 4 a). The force plot visualized the direction and magnitude of each feature's contribution, of which the features of HBB (11,977.223) and MTRNR2L8 (10,055.035) made a significant positive contribution to the model output, pushing the diagnostic result to HCC (Fig. 4 b). Moreover, feature correlation analysis revealed that there was strong relationship between MTRNR2L8 and MTRNR2L12 (0.93), TMSB4X and OST4 (0.90), TMSB4X and PPBP (0.91), indicating these genes may work synergistically in HCC occurrence. PPBP and FTL showed a negative correlation (-0.55), which may reflect their antagonistic effects in the pathology of HCC (Supplementary Fig. 3). To verify the reliability of the model, the HCC early diagnostic model was interpreted and optimized by KAN. Figure 5 showed the performance and structural analysis of the KAN model in HCC diagnosis, highlighting its classification ability and interpretability. Figure 5 a presented the ROC curve of the KAN model with an AUC value of 0.86, indicating that the model displayed a good classification performance and could effectively distinguish between patients with HCC and healthy individuals. Figure 5 b illustrated the structure of the KAN network, where the input layer contained gene features including MTRNR2L8, PF4, TMSB4X, HBB, PPBP, MTRNR2L12, FTL, OST4, S100A9, and ACTB, connected to the hidden layer node (1,0) through activation functions on the edges, ultimately outputting the prediction results, showing the sparse structure of the KAN model. Figure 5 c showed the activation function fitting from MTRNR2L8 to node (1,0), with blue dots representing data points, and the orange line as the fitting curve ( R² =1.00), where red markers indicated sample points, demonstrating its perfect capture of the nonlinear relationship. Figure 5 d presented the activation function fitting from FTL to node (1,0), with the fitting curve (R²=0.92) also showing a good correspondence between the data and the fitting curve. These results indicate that the KAN model, through edge activation functions and a sparse structure, not only enhances prediction performance but also improves the model's interpretability, providing reliable theoretical support for HCC diagnosis. Additionally, the mathematical expression formula generated by KAN was displayed in Supplementary Fig. 4. Online prediction platform To enhance the clinical application of the HCC diagnostic model, an online prediction platform has been developed based on the Streamlit framework as shown in Fig. 6 . The accessible link was as follows, https://ai-model-j7uqkifaxp6etnfsf4r4t8.streamlit.app/ . Clinicians or patients could input EV-derived gene expression levels to obtain the predicted probability of HCC and an individual sample decision plot based on SHAP. This platform provides clinicians with a convenient diagnostic tool, accelerating the promotion of non-invasive early diagnosis of HCC. Discussions In this study, we developed and validated an interpretable machine learning model for non-invasive diagnosis of hepatocellular carcinoma based on EV-derived gene signatures. Our findings demonstrate that EV-derived mRNAs can serve as effective biomarkers for HCC detection, offering a promising alternative to conventional diagnostic methods. The superior performance of DNN compared to other machine learning algorithms in our study highlights the complex, non-linear relationships between EV-derived gene expression patterns and HCC status. This complexity necessitates sophisticated modeling approaches that can capture intricate interactions among multiple features. Our optimized model, incorporating ten key mRNAs (MTRNR2L8, HBB, PF4, FTL, MTRNR2L12, TMSB4X, PPBP, OST4, ACTB, and S100A9), achieved a robust performance with an AUC of 0.8877 in the test set, demonstrating its potential clinical utility for non-invasive HCC detection. The identified key genes have biological relevance to HCC pathogenesis. MTRNR2L8 (Humanin-like 8), identified as the most significant contributor to our model, is a peptide-encoding gene with reported anti-apoptotic and cytoprotective functions 28 , 29 . Previous studies have suggested that humanin and its analogs may play roles in cancer development through regulating cell survival pathways 30 – 32 . HBB (Hemoglobin Beta) has been associated with various cancers, potentially reflecting tumor-related angiogenesis, and oxygen transport alterations 33 , 34 . PF4 (Platelet Factor 4) is known for its anti-angiogenic properties and has been implicated in tumor microenvironment regulation 35 , 36 . S100A9, a calcium-binding protein, has been reported to promote inflammation and cancer progression, including in HCC 37 – 39 . The identification of these genes as key diagnostic markers not only enhances the model's predictive performance but also provides insights into potential biological mechanisms underlying HCC development. Our approach to model interpretability using SHAP and KAN frameworks addresses a critical limitation of many machine learning models in healthcare—the "black box" problem. By elucidating how individual features contribute to predictions at both global and local levels, our model provides transparency that is essential for clinician trust and adoption. The strong correlations observed between certain genes (e.g., MTRNR2L8 and MTRNR2L12, TMSB4X and OST4) suggest potential functional interactions or co-regulation mechanisms that warrant further investigation. The antagonistic relationship between PPBP and FTL may reflect competing biological processes in HCC pathogenesis. The development of an online prediction platform represents a significant step toward clinical implementation of our diagnostic model. This user-friendly interface allows clinicians to input patient-specific EV-derived gene expression data and obtain immediate diagnostic predictions with visual explanations, potentially facilitating integration into clinical workflows. Such accessibility could contribute to broader adoption of non-invasive HCC screening, particularly in resource-limited settings where conventional imaging techniques may be less available. Several limitations of our study should be acknowledged. First, while the exoRBase 2.0 database provided a substantial dataset for model development, external validation with prospective, multi-center cohorts would strengthen the evidence for clinical applicability. Second, standardization of EV isolation and RNA quantification methods remains challenging and could affect the reproducibility of gene expression measurements across different laboratories. Third, our model focused exclusively on RNA signatures and did not incorporate other potentially valuable clinical data such as age, gender, etiology of liver disease, or conventional biomarkers like AFP. Integration of these factors might further enhance diagnostic performance. Future directions should include prospective validation in diverse patient populations, particularly those with various liver disease backgrounds, to assess the model's performance across different HCC etiologies. Investigation of the temporal dynamics of EV-derived signatures during HCC development could provide insights into their utility for early detection. Additionally, functional studies of the identified key genes would enhance our understanding of their roles in HCC pathogenesis and potentially reveal novel therapeutic targets. In conclusion, our study demonstrates the feasibility and potential clinical utility of an interpretable machine learning model based on EV-derived gene signatures for non-invasive HCC detection. By combining sophisticated algorithm selection, feature optimization, and model interpretability with accessible implementation, our approach addresses current diagnostic challenges and represents a step toward improved early detection and management of HCC. The integration of such liquid biopsy-based approaches into clinical practice could potentially transform HCC screening paradigms, ultimately contributing to better patient outcomes through earlier intervention. Abbreviations HCC, hepatocellular carcinoma; EVs, extracellular vesicles; SHAP, SHapley Additive exPlanations; KAN, Kolmogorov-Arnold Networks; DNN, Deep Neural Networks; AFP, alpha-fetoprotein. Declarations Funding provided by Natural Science Foundation of Chongqing, China (Grant No. CSTB2024NSCQ-MSX0490); by Joint project of Chongqing Health Commission and Science and Technology Bureau (Grant No. 2025MSXM075); by Scientific and Technological Research Program of Chongqing Municipal Education Commission (Grant No. KJZD-M202300102); by Development Program of Chongqing university, Jiangjin hospital (Grant No. 2023LJXM003, 2024LJXM001); by Research Launch Project of Chongqing university, Jiangjin hospital (Grant No. 2023qdjfxm006). Declaration of interests The authors declare no potential conflicts of interest. Author contributions All authors reviewed the manuscript. Acknowledgements We thank Guoyu Tan and members of the Lun Lun Research Institute of AI FOR SCIENCE for technical support and suggestions that improved the manuscript and experiments. Financial support Funding provided by Natural Science Foundation of Chongqing, China (Grant No. CSTB2024NSCQ-MSX0490); by Joint project of Chongqing Health Commission and Science and Technology Bureau (Grant No. 2025MSXM075); by Scientific and Technological Research Program of Chongqing Municipal Education Commission (Grant No. KJZD-M202300102); by Development Program of Chongqing university, Jiangjin hospital (Grant No. 2023LJXM003，2024LJXM001); by Research Launch Project of Chongqing university, Jiangjin hospital (Grant No. 2023qdjfxm006). Data availability statement The data analyzed in this study are available in the exoRBase 2.0 at http://www.exorbase.org/exoRBaseV2/download/toIndex. The protocol and statistical analysis methods used in the study can be requested directly from the corresponding author after approval. Supporting Information The online version contains supplementary material available at References Villanueva, A. Hepatocellular Carcinoma. N. Engl. J. Med. 380 , 1450–1462 (2019). Galle, P. R. et al. EASL Clinical Practice Guidelines: Management of hepatocellular carcinoma. J. Hepatol. ; 69 . (2018). Sangro, B. et al. Durvalumab with or without bevacizumab with transarterial chemoembolisation in hepatocellular carcinoma (EMERALD-1): a multiregional, randomised, double-blind, placebo-controlled, phase 3 study. Lancet 405 , 216–232 (2025). Ganne-Carrié, N. & Nahon, P. Differences between hepatocellular carcinoma caused by alcohol and other aetiologies. J. Hepatol. 82 , 909–917 (2025). Yang, Y. et al. Genomic and the tumor microenvironment heterogeneity in multifocal hepatocellular carcinoma. Hepatol. Dec. (2024). Marquardt, J. U., Galle, P. R. & Teufel, A. Molecular diagnosis and therapy of hepatocellular carcinoma (HCC): An emerging field for advanced technologies. J. Hepatol. ; 56 . (2012). El-Serag, H. B. et al. Diagnosis and Treatment of Hepatocellular Carcinoma. Gastroenterology 2008;134. Tzartzeva, K. et al. Surveillance Imaging and Alpha Fetoprotein for Early Detection of Hepatocellular Carcinoma in Patients With Cirrhosis: A Meta-analysis. Gastroenterology 2018;154. Xu, R. et al. Extracellular vesicles in cancer — implications for future improvements in cancer care. Nat. Rev. Clin. Oncol. ; 15 . (2018). Bebelman, M. P. et al. Biogenesis and function of extracellular vesicles in cancer. Pharmacol. Ther. ; 188 . (2018). ichiro Ohno, S., Ishikawa, A. & Kuroda, M. Roles of exosomes and microvesicles in disease pathogenesis. Adv. Drug Deliv Rev. ; 65 . (2013). Taylor, D. D. & Gercel-Taylor, C. Exosomes/microvesicles: mediators of cancer-associated immunosuppressive microenvironments. Semin Immunopathol. ; 33 . (2011). Park, K-S., Lässer, C. & Lötvall, J. Extracellular vesicles and the lung: from disease pathogenesis to biomarkers and treatments. Physiol. Rev. 105 , 1733–1821 (2025). Wang, Y. & Ding, S. Extracellular vesicles in cancer cachexia: deciphering pathogenic roles and exploring therapeutic horizons. J. Transl Med. 22 , 506 (2024). Dorado, E. et al. Extracellular vesicles as a promising source of lipid biomarkers for breast cancer detection in blood plasma. J. Extracell. Vesicles ; 13 . (2024). Russo, M. N. et al. Extracellular vesicles in the glioblastoma microenvironment: A diagnostic and therapeutic perspective. Mol. Aspects Med. 91 , 101167 (2023). Zhao, C. et al. Extracellular vesicle digital scoring assay for assessment of treatment responses in hepatocellular carcinoma patients. J. Experimental Clin. Cancer Res. 44 , 136 (2025). Yang, P. et al. Non-canonical small noncoding RNAs in the plasma extracellular vesicles as novel biomarkers in gastric cancer. J. Hematol. Oncol. 18 , 39 (2025). Shao, X. et al. Extracellular vesicle-derived miRNA-mediated cell–cell communication inference for single-cell transcriptomic data with miRTalk. Genome Biol. 26 , 95 (2025). Wang, G. et al. Integrating single-cell RNA sequencing, WGCNA, and machine learning to identify key biomarkers in hepatocellular carcinoma. Sci. Rep. 15 , 11157 (2025). Pinsky, M. R. et al. Use of artificial intelligence in critical care: opportunities and obstacles. Crit. Care . 28 , 113 (2024). Chen, H., Lundberg, S. M. & Lee, S-I. Explaining a series of models by propagating Shapley values. Nat. Commun. 13 , 4512 (2022). Liu, Z. et al. KAN: Kolmogorov-Arnold Networks. April (2024). Li, S. et al. ExoRBase: A database of circRNA, lncRNA and mRNA in human blood exosomes. Nucleic Acids Res. ; 46 . (2018). Lai, H. et al. exoRBase 2.0: An atlas of mRNA, lncRNA and circRNA in extracellular vesicles from human biofluids. Nucleic Acids Res. ; 50 . (2022). Pedregosa, F. et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. ; 12 . (2011). LUNDBERG S. A unified approach to interpreting model predictions. arXiv preprint (2017). Shen, Y. et al. Epigenome-Wide Association Study Indicates Hypomethylation of MTRNR2L8 in Large-Artery Atherosclerosis Stroke. Stroke. ;50:1330–1338. (2019). Huth, A. et al. Single cell transcriptomics of cerebrospinal fluid cells from patients with recent-onset narcolepsy. J. Autoimmun. 146 , 103234 (2024). Gordon-Lipkin, E. M. et al. Primary oxidative phosphorylation defects lead to perturbations in the human B cell repertoire. Front. Immunol. ; 14 . (2023). Bangert, C. et al. Comprehensive Analysis of Nasal Polyps Reveals a More Pronounced Type 2 Transcriptomic Profile of Epithelial Cells and Mast Cells in Aspirin-Exacerbated Respiratory Disease. Front. Immunol. ; 13 . (2022). Iparraguirre, L. et al. Whole-Transcriptome Analysis in Peripheral Blood Mononuclear Cells from Patients with Lipid-Specific Oligoclonal IgM Band Characterization Reveals Two Circular RNAs and Two Linear RNAs as Biomarkers of Highly Active Disease. Biomedicines 8 , 540 (2020). Maman, S. et al. The Beta Subunit of Hemoglobin (HBB2/HBB) Suppresses Neuroblastoma Growth and Metastasis. Cancer Res. 77 , 14–26 (2017). Zheng, Y. et al. Expression of β-globin by cancer cells promotes cell survival during blood-borne dissemination. Nat. Commun. 8 , 14344 (2017). Çakan, E. et al. TLR9 ligand sequestration by chemokine CXCL4 negatively affects central B cell tolerance. J. Exp. Med. ; 220 . (2023). Zhang, C. et al. ITGB6 modulates resistance to anti-CD276 therapy in head and neck cancer by promoting PF4 + macrophage infiltration. Nat. Commun. 15 , 7077 (2024). Zhang, Z. et al. E-twenty-six-specific sequence variant 5 (ETV5) facilitates hepatocellular carcinoma progression and metastasis through enhancing polymorphonuclear myeloid-derived suppressor cell (PMN-MDSC)-mediated immunosuppression. Gut February 2025:gutjnl-2024-333944. Valdés-Mas, R. et al. Metagenome-informed metaproteomics of the human gut microbiome, host, and dietary exposome uncovers signatures of health and inflammatory bowel disease. Cell 188 , 1062–1083e36 (2025). Zhang, Q. et al. PGAM5 interacts with and maintains BNIP3 to license cancer-associated muscle wasting. Autophagy 20 , 2205–2220 (2024). Additional Declarations No competing interests reported. Supplementary Files SupportingInformation.docx Cite Share Download PDF Status: Published Journal Publication published 14 Feb, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 07 Jan, 2026 Reviews received at journal 06 Jan, 2026 Reviewers agreed at journal 17 Dec, 2025 Reviewers invited by journal 16 Dec, 2025 Editor assigned by journal 16 Dec, 2025 Editor invited by journal 02 Dec, 2025 Submission checks completed at journal 01 Dec, 2025 First submitted to journal 30 Nov, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8217267","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":561667260,"identity":"e9207dce-690c-46d6-a399-0a499cadc4a7","order_by":0,"name":"Yan Zhang","email":"","orcid":"","institution":"Chongqing University, Jiangjin Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Zhang","suffix":""},{"id":561667261,"identity":"08576c06-2707-40ff-bafe-6e1ecf8f2421","order_by":1,"name":"Zhengying Mo","email":"","orcid":"","institution":"Tai-He Hospital, Hubei University of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Zhengying","middleName":"","lastName":"Mo","suffix":""},{"id":561667262,"identity":"11c3f6fa-1608-45a2-893d-53ce03d02b82","order_by":2,"name":"Lei Zhang","email":"","orcid":"","institution":"Chongqing Health Center for Women and Children","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Zhang","suffix":""},{"id":561667263,"identity":"a19a86c6-4464-467e-9751-ec05aad7036a","order_by":3,"name":"Zhangming Zhou","email":"","orcid":"","institution":"Dujiangyan Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Zhangming","middleName":"","lastName":"Zhou","suffix":""},{"id":561667264,"identity":"d59e4a8c-168c-47c1-8ed9-8a56913a0056","order_by":4,"name":"Zhaohan Wei","email":"","orcid":"","institution":"School of Biomedical Engineering, Shenzhen University Medical School, Shenzhen University","correspondingAuthor":false,"prefix":"","firstName":"Zhaohan","middleName":"","lastName":"Wei","suffix":""},{"id":561667265,"identity":"b87dc1dd-ceac-4026-a953-9ffe85886e0a","order_by":5,"name":"Yuan Zhu","email":"","orcid":"","institution":"Beijing Diji Technology Co., Ltd.","correspondingAuthor":false,"prefix":"","firstName":"Yuan","middleName":"","lastName":"Zhu","suffix":""},{"id":561667266,"identity":"ecaecd94-0605-4096-bdb6-7e26a8d5d354","order_by":6,"name":"Yadong Wang","email":"","orcid":"","institution":"Beijing Diji Technology Co., Ltd.","correspondingAuthor":false,"prefix":"","firstName":"Yadong","middleName":"","lastName":"Wang","suffix":""},{"id":561667267,"identity":"ab8eaedd-8922-4fbe-9e4f-66d922699fb2","order_by":7,"name":"Hu Wang","email":"","orcid":"","institution":"Jingzhou Third People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hu","middleName":"","lastName":"Wang","suffix":""},{"id":561667268,"identity":"2847813d-b565-423a-a5af-287d760d5073","order_by":8,"name":"Debing Xiang","email":"","orcid":"","institution":"Chongqing University Jiangjin Hospital","correspondingAuthor":false,"prefix":"","firstName":"Debing","middleName":"","lastName":"Xiang","suffix":""},{"id":561667269,"identity":"f88ea7ca-4c26-4668-9866-09b1720ff0b0","order_by":9,"name":"Jing Huang","email":"","orcid":"","institution":"Chongqing University Jiangjin Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Huang","suffix":""},{"id":561667270,"identity":"2f399909-19c3-4093-a287-58c1aaa36186","order_by":10,"name":"Qingle Liang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4UlEQVRIiWNgGAWjYBACPmYwJcHAwM7A+CChwoawFja4FmYGZoMHZ9KI0AJnMTOwST5sO0SEFnbmZw+/lFnkyTszH6tIYDvAwN/enUDAYWzmxjLnJIoND7Ol3UjgucMgcebsBgJaGMykJdskEjc285jdSJB4xmAgkUtIC/s3qBb+bwUJBoeJ0cJjJvkRqGU+Mw8bQ0ICcVrKpBnOSSRuYGYzlkg4kMZD0C/8/Me3Sf4oq0uc39788OPPfzZy/O29+LWAAMhJDAYHIBwegspBgPEHUIt8A1FqR8EoGAWjYCQCAGdqQCDeD81hAAAAAElFTkSuQmCC","orcid":"","institution":"Chongqing University, Jiangjin Hospital","correspondingAuthor":true,"prefix":"","firstName":"Qingle","middleName":"","lastName":"Liang","suffix":""}],"badges":[],"createdAt":"2025-11-27 02:53:26","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8217267/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8217267/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-40020-9","type":"published","date":"2026-02-14T15:59:01+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":98638226,"identity":"6c93a705-0b7c-4b29-b41b-a65cdfc5fb5b","added_by":"auto","created_at":"2025-12-19 17:45:35","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1282642,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/b8437f763a31f49167d5cbc4.docx"},{"id":98638208,"identity":"55d8523f-446d-4ba4-bbad-626b86858e7a","added_by":"auto","created_at":"2025-12-19 17:45:35","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10450,"visible":true,"origin":"","legend":"","description":"","filename":"fd3425c6dbd148bd83547d3b9ced6413.json","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/1d06aaee76e0a80ba9d9d186.json"},{"id":98638227,"identity":"148bf676-3bf4-452f-a40e-cd414197e9d4","added_by":"auto","created_at":"2025-12-19 17:45:37","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1533422,"visible":true,"origin":"","legend":"","description":"","filename":"SupportingInformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/d85c81f798b77a1d9289504b.docx"},{"id":98775163,"identity":"3173a94b-82da-4784-bbd8-aade1a6c7c6e","added_by":"auto","created_at":"2025-12-22 12:18:42","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":84754,"visible":true,"origin":"","legend":"","description":"","filename":"fd3425c6dbd148bd83547d3b9ced64131enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/5161e03519204d67023050bb.xml"},{"id":98775723,"identity":"6a3d817d-f884-4d7b-9e52-0042ca328988","added_by":"auto","created_at":"2025-12-22 12:20:55","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":274966,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/e5eb7e0e443404addff99e0a.jpeg"},{"id":98775760,"identity":"dd5ac45e-a628-4aaa-8863-23df1e924568","added_by":"auto","created_at":"2025-12-22 12:21:00","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":165358,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/7b502d604c17562858bcf702.png"},{"id":98775160,"identity":"bbc1c769-aac8-42b4-9f71-0e04e04c8a58","added_by":"auto","created_at":"2025-12-22 12:18:41","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":118031,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/407bb952098a0580a80eea47.png"},{"id":98774865,"identity":"4100effd-8c03-444d-b348-644f9cb43d5b","added_by":"auto","created_at":"2025-12-22 12:15:52","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":175090,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/f5cfaa1b5df5f549982d33ec.png"},{"id":98638214,"identity":"51cfd77f-0d32-49e1-99ed-343135c05f43","added_by":"auto","created_at":"2025-12-19 17:45:35","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":326271,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/d69501f59f89ed9298669624.png"},{"id":98638220,"identity":"7aad2271-ac88-4a91-bdf5-230e9c80eb7b","added_by":"auto","created_at":"2025-12-19 17:45:35","extension":"jpeg","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":198432,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/9dd859a8237fdd3b34b14e55.jpeg"},{"id":98775890,"identity":"5ac4dcc9-b498-4413-b46a-a378306ee79d","added_by":"auto","created_at":"2025-12-22 12:21:31","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":34616,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/c5b268e9e3f7dae0b58163ed.png"},{"id":98774825,"identity":"5d8acf10-4005-4175-8959-cff4b30548a5","added_by":"auto","created_at":"2025-12-22 12:15:08","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":30548,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/b8117b867c2c2991ae3d5b26.png"},{"id":98775424,"identity":"a8282ca7-5ea8-4f40-9d33-13643ef486f6","added_by":"auto","created_at":"2025-12-22 12:19:53","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24833,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/fba766e51f62c0d0883f0af6.png"},{"id":98775303,"identity":"53c65506-e8a8-4da3-b563-57438ba315c5","added_by":"auto","created_at":"2025-12-22 12:19:17","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29131,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/c3b772f6094884a7dfa049a0.png"},{"id":98638225,"identity":"cb2d96fe-c95a-47a4-8c8f-fbea074ebecb","added_by":"auto","created_at":"2025-12-19 17:45:35","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":56646,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/92a1b2b0e99f06463eb6a5e5.png"},{"id":98774848,"identity":"68b33c65-686a-4b7f-a7b9-b505ce46da97","added_by":"auto","created_at":"2025-12-22 12:15:44","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28346,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/16f8098872f9cf74e026d690.png"},{"id":98775345,"identity":"46165f3a-81bd-49c4-87d6-0900ab6535ec","added_by":"auto","created_at":"2025-12-22 12:19:29","extension":"xml","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":83319,"visible":true,"origin":"","legend":"","description":"","filename":"fd3425c6dbd148bd83547d3b9ced64131structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/54bf0d3e9a6d7438b0c8bcd9.xml"},{"id":98775031,"identity":"7a31cb04-8ce3-4de7-8194-ece1bce10b14","added_by":"auto","created_at":"2025-12-22 12:17:59","extension":"html","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":96836,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/8a4ff1deaf29b573005e21c7.html"},{"id":98775330,"identity":"3b31d460-6871-455f-9a80-787e4f5a534d","added_by":"auto","created_at":"2025-12-22 12:19:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":127057,"visible":true,"origin":"","legend":"\u003cp\u003eThe workflow of the study.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/832d1d3d2b4411888c6cd692.png"},{"id":98638203,"identity":"85ed3cb2-65c0-4e83-9fc0-dbbd36cf2094","added_by":"auto","created_at":"2025-12-19 17:45:35","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":103951,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves of models with different numbers of mRNA features (a) train set (b) test set.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/d6495097153d599d83eaf9aa.png"},{"id":98638205,"identity":"15255b92-d004-4f89-b84a-8d6361ff6700","added_by":"auto","created_at":"2025-12-19 17:45:35","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":86217,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP global interpretation (a) Feature importance rank by SHAP swarm diagram (b) Feature importance distribution by donut diagram.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/1390d220296cb4642475ec33.png"},{"id":98775863,"identity":"8cac8416-66fd-46d5-8fc6-c15680d3a376","added_by":"auto","created_at":"2025-12-22 12:21:23","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":122631,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP local interpretation (a) SHAP decision plot (b) SHAP force plot.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/652000831dd2ab7dc0af86ed.png"},{"id":98774842,"identity":"c0a2c9e9-1225-4c29-80de-125523e741c1","added_by":"auto","created_at":"2025-12-22 12:15:37","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":186902,"visible":true,"origin":"","legend":"\u003cp\u003eKAN interpretation (a) ROC curve (b) The network analysis (c) The activation function fitting from MTRNR2L8 to node (1,0), with blue points representing the data and the orange line representing the fitting curve. (d) The activation function fitting from FTL to node (1,0), with blue points representing the data and the orange line representing the fitting curve.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/e8d83fc5a65dd36c69723529.png"},{"id":98774998,"identity":"e60b36c6-ae99-435d-a947-bc2970e1b68b","added_by":"auto","created_at":"2025-12-22 12:17:50","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":130099,"visible":true,"origin":"","legend":"\u003cp\u003eThe screenshot of the online diagnostic model for HCC.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/204791ec1d7adc889cb5d539.png"},{"id":102786725,"identity":"0db06c16-595a-41ac-a4a9-a671d563bfeb","added_by":"auto","created_at":"2026-02-16 16:14:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1303599,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/93eed2de-401d-4d7b-8ff6-a5a520bac569.pdf"},{"id":98775366,"identity":"438835c7-81e4-4ca0-8cb5-89c48deb7e17","added_by":"auto","created_at":"2025-12-22 12:19:40","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1533422,"visible":true,"origin":"","legend":"","description":"","filename":"SupportingInformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-8217267/v1/ee7ce8bf73cdca6156bc5b4c.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Clinically Interpretable Extracellular Vesicle Gene Model for Non-Invasive Liver Cancer Diagnosis","fulltext":[{"header":"Introduction","content":"\u003cp\u003eHepatocellular carcinoma (HCC) represents the predominant form of primary liver cancer and stands as the third leading cause of cancer-related mortality worldwide, with an increasing incidence in many regions\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. The prognosis of HCC remains poor, with a 5-year survival rate below 20% in most countries, primarily due to late diagnosis when curative treatments are no longer effective\u003csup\u003e\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Early detection of HCC is crucial for improving patient outcomes, as it enables timely intervention with potentially curative therapies such as surgical resection, liver transplantation, or ablation techniques.\u003c/p\u003e \u003cp\u003eCurrent diagnostic approaches for HCC rely heavily on imaging techniques (ultrasonography, computed tomography, magnetic resonance imaging) and invasive liver biopsies, which have limitations including radiation exposure, high cost, operator dependency, and sampling errors\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Conventional serum biomarkers such as alpha-fetoprotein (AFP) lack sufficient sensitivity and specificity for early-stage HCC detection\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Therefore, there is an urgent need for non-invasive, accurate, and accessible diagnostic tools for early HCC detection.\u003c/p\u003e \u003cp\u003eExtracellular vesicles (EVs) have emerged as promising sources of cancer biomarkers\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. These membrane-bound vesicles, including exosomes and microvesicles, are released by cells into body fluids and contain various molecular components including proteins, lipids, and nucleic acids that reflect the status of their cells of origin\u003csup\u003e\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Recent studies have demonstrated that cancer cells, including HCC cells, secrete EVs containing specific RNA signatures that differ from those released by normal cells\u003csup\u003e\u003cspan additionalcitationids=\"CR15 CR16\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. These EV-derived RNAs can be detected in liquid biopsies, providing a potential avenue for non-invasive cancer diagnosis.\u003c/p\u003e \u003cp\u003eWith the advancement of high-throughput sequencing technologies and machine learning algorithms, it is now possible to analyze complex EV-derived RNA expression patterns to identify diagnostic signatures for various diseases\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Machine learning models can integrate multiple biomarkers to improve diagnostic accuracy beyond what single biomarkers can achieve\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. However, the \"black box\" nature of many machine learning algorithms limits their clinical application, as understanding the biological relevance and decision-making process is crucial for medical implementation\u003csup\u003e\u003cspan additionalcitationids=\"CR22\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn this study, we aimed to develop and validate an interpretable diagnostic model based on EV-derived gene signatures for non-invasive detection of HCC. We utilized RNA-seq data from the exoRBase 2.0 database, which contains comprehensive information on messenger RNA (mRNA) and long non-coding RNA (lncRNA) in EVs from healthy individuals and HCC patients\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. We compared multiple machine learning algorithms to identify the optimal approach, selected key features based on their importance, and employed interpretability frameworks to explain the model's predictions. Additionally, we developed an online platform to facilitate the clinical application of our diagnostic model. Our approach addresses the need for accurate, non-invasive, and interpretable tools for early HCC detection, potentially contributing to improved patient outcomes through earlier intervention.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData source\u003c/h2\u003e \u003cp\u003eThe information utilized in this study is sourced from the exoRBase 2.0 database\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, a dedicated repository concentrating on long RNAs (exLRs) within EVs. This database encompasses details regarding messenger RNA (mRNA), long non-coding RNA (lncRNA), and circular RNA (circRNA), which consolidates RNA-seq data from 905 EV samples obtained from four distinct human bodily fluids (blood, urine, bile and cerebrospinal fluid), offering comprehensive analysis and visualization of RNA expression patterns, as well as modifications at the functional pathway level and insights into the variability of circulating EV origins. Specifically, we retrieved the following datasets from the exoRBase 2.0 database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.exorbase.org/exoRBaseV2/download/toIndex\u003c/span\u003e\u003cspan address=\"http://www.exorbase.org/exoRBaseV2/download/toIndex\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), including exosomal data from HCC, exosomal data from healthy individuals, and annotation data for lncRNA and mRNA. The dataset comprises mRNA and lncRNA expression data from both healthy individuals and HCC patients, with the healthy cohort consisting of 112 samples with 15,966 mRNA samples and 7,115 lncRNA samples; the HCC cohort contains 118 sampels with 16,023 mRNA samples and 1,118 lncRNA samples. The study design was shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eData preprocessing\u003c/h3\u003e\n\u003cp\u003eTo ensure data quality, raw sequencing data were processed using Python scripts. mRNA and lncRNA expression profiles were extracted and labeled according to HCC and healthy control samples. To eliminate scale differences across features, all expression values were standardized using Z-score normalization (mean\u0026thinsp;=\u0026thinsp;0, standard deviation\u0026thinsp;=\u0026thinsp;1)\u003csup\u003e26\u003c/sup\u003e. To reduce noise and enhance signal quality, features with expression levels below 10% of the overall median or with low variance across all samples were excluded from downstream analyses.\u003c/p\u003e\n\u003ch3\u003eData Division\u003c/h3\u003e\n\u003cp\u003eThe dataset was randomly divided into a training set (70%) and a testing set (30%). The training set was used for model construction and hyperparameter optimization, while the testing set served to evaluate model generalizability. A five-fold cross-validation strategy was applied to the training set to ensure model stability and reduce overfitting.\u003c/p\u003e\n\u003ch3\u003eModel Development\u003c/h3\u003e\n\u003cp\u003eSix machine learning algorithms were explored, including Logistic Regression, Random Forest, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Decision Tree, and Deep Neural Network (DNN). These models were trained on either EV-derived mRNA or lncRNA expression data. The outcome variable was defined as binary: 0 for healthy individuals and 1 for HCC patients. Input variables included normalized expression values of mRNAs and lncRNAs. SHapley Additive exPlanations (SHAP) were used to assess feature importance, and models were compared across different feature subsets to identify the optimal number and combination of predictors.\u003c/p\u003e\n\u003ch3\u003eModel evaluation\u003c/h3\u003e\n\u003cp\u003eTo Model performance was comprehensively evaluated using multiple metrics: Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, precision, recall, and F1 score. AUC was used to assess the model\u0026rsquo;s ability to distinguish between HCC and healthy samples, with higher values indicating better discriminative performance. Accuracy reflected the proportion of correctly classified samples among all predictions. Precision indicated the proportion of true positive cases among all positive predictions made by the model. Recall measured the proportion of true positive samples correctly identified among all actual positive cases. F1 score, the harmonic mean of precision and recall, provided a balanced measure of overall model performance. Formulas for these evaluation metrics are provided in the supplementary materials.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\text{A}\\text{c}\\text{c}\\text{u}\\text{r}\\text{a}\\text{c}\\text{y}=\\frac{\\text{T}\\text{N}+\\text{T}\\text{P}}{\\text{T}\\text{P}+\\text{T}\\text{N}+\\text{F}\\text{P}+\\text{T}\\text{N}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}=\\frac{\\text{T}\\text{P}}{\\text{F}\\text{P}+\\text{T}\\text{P}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}=\\frac{\\text{T}\\text{P}}{\\text{F}\\text{N}+\\text{T}\\text{P}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:\\text{F}1-\\text{S}\\text{c}\\text{o}\\text{r}\\text{e}=\\frac{2\\times\\:\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}\\times\\:\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}}{\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}+\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eAmong them, true positive (TP), false positive (FP), true negative (TN), and false negative (FN) represent the four possible outcomes of classification results. Internal validation was conducted through five-fold cross-validation and the test set was used for external validation to comprehensively estimate the predictive performance of the model.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eModel Interpretation\u003c/h2\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003eSHapley Additive exPlanations (SHAP)\u003c/h2\u003e \u003cp\u003eTo address the \u0026ldquo;black-box\u0026rdquo; nature of machine learning models, the SHapley Additive exPlanations (SHAP) framework was employed to interpret the predictions of the optimal model. SHAP is based on cooperative game theory and quantifies the marginal contribution of each feature to a model\u0026rsquo;s prediction\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. It provides both global and local interpretability: Global interpretation was achieved via summary plots, ranking features by their average absolute SHAP values to indicate their overall impact on model output. Local interpretation was performed using force plots, which visualize how individual features contribute to the prediction for specific samples. In this study, SHAP values were used to evaluate the contribution of EV-derived gene features to the diagnostic classification of HCC, enabling the identification of potential key biomarkers associated with disease onset.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e\n\u003ch3\u003eKolmogorov-Arnold Networks (KAN)\u003c/h3\u003e\n\u003cp\u003eTo further enhance model interpretability and predictive performance, Kolmogorov\u0026ndash;Arnold Networks (KAN) were implemented as an alternative to traditional multilayer perceptrons (MLPs). Based on the Kolmogorov\u0026ndash;Arnold representation theorem\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. KAN introduces learnable activation functions along the edges of the network, forming a dual representation to the node-centric activation in MLPs. This architecture facilitates the derivation of explicit mathematical expressions describing the relationships between input variables and outcomes, offering improved model transparency through a sparse and interpretable network structure.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eOnline Computing Platform\u003c/h2\u003e \u003cp\u003eTo enable real-time clinical application, the final predictive model was deployed as a web-based platform using the Streamlit framework. This tool allows users to input EV-derived gene expression values and receive immediate diagnostic predictions, facilitating non-invasive early detection of HCC in practice.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eAll analyses and computations were conducted using R (version 4.4.1) and Python (version 3.9.12), with a two-sided P value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 considered statistically significant. For machine learning analyses, the XGBoost algorithm was implemented using the eXtreme Gradient Boosting package (version 2.0.3). Other algorithms, including Deep Neural Network (DNN), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), decision tree (DT), were executed using the scikit-learn package (version 1.3.2), alongside pandas (version 1.5.3) and numpy (version 1.23.5) for data manipulation and numerical computations. The DNN model was developed with the TensorFlow framework (version 1.13.1), while the Kolmogorov\u0026ndash;Arnold Network (KAN) was constructed using the pykan package (version 0.2.8).\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eModel development and comparison\u003c/h2\u003e \u003cp\u003eFirstly, we constructed diagnostic models using six algorithms based on extracellular vesicles (EV)-derived lncRNA and mRNA transcriptome data of HCC, respectively, and evaluated the performance of the models. As shown in Supplementary Table\u0026nbsp;1, DNN algorithm performed remarkable classification capability on the EV-derived mRNA data in the test set, achieving the highest accuracy of 0.8816, the precision of 0.9375, F1 score of 0.8696, and the AUC of 0.9324 compared to other five algorithms including logistic regression, random forest, SVM, XGBoost and decision tree. The diagnostic model constructed based on lncRNA data also demonstrated excellent classification performance with DNN, with F1-score of 0.7765, and the recall of 0.8919 (Supplementary Table\u0026nbsp;2). Despite Random Forest and XGBoost algorithms performed similarly to DNN on certain metrics, DNN algorithm demonstrated better stability and overall performance based on both mRNA and lncRNA features. Thus, the optimal model was selected based on EV-derived mRNA and lncRNA characteristics.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eFeatures selection\u003c/h2\u003e \u003cp\u003eTo enhance the practicality of the diagnostic model, we evaluated the model performance within different number of EV-derived mRNAs and lncRNAs characteristics including top 20, top 15, top 10, and top 5 features for model training based on SHAP feature importance rankings. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the AUCs of the mRNAs model on the test set were 0.9324 with all features, 0.9279 with top 20 features, 0.9089 with top 15 features, 0.8877 with top 10 features and 0.8628 with 5 features. The AUCs of lncRNAs model on the test set were 0.8656 with all features, 0.8663 with top 20 features, 0.8614 with top 15 features, 0.8150 with top 10 features, 0.7159 with top 5 features (Supplementary Fig.\u0026nbsp;1). In addition to the detailed evaluation metrics as shown in Supplementary Table\u0026nbsp;3, the diagnostic model with top10 genes was identified as the optimal model considering both the number of features and model performance.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eModel Interpretation\u003c/h2\u003e \u003cp\u003eThrough the SHAP framework, we performed an interpretability analysis of the model's predictive outcomes and identified genes that were pivotal in the diagnosis of HCC. The key variables contributing to HCC early diagnosis were MTRNR2L8, HBB, PF4, FTL, MTRNR2L12, TMSB4X, PPBP, OST4, ACTB and S100A9, among which MTRNR2L8 was the most significant contributor (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Supplementary Fig.\u0026nbsp;2). On the other hand, SHAP local interpretability revealed the predictive outcome for individuals. The diagnostic result was 1.00, indicating that the probability of being diagnosed with HCC was high (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), The decision diagram reflected the decision-making process of key features on the output results (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). The force plot visualized the direction and magnitude of each feature's contribution, of which the features of HBB (11,977.223) and MTRNR2L8 (10,055.035) made a significant positive contribution to the model output, pushing the diagnostic result to HCC (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). Moreover, feature correlation analysis revealed that there was strong relationship between MTRNR2L8 and MTRNR2L12 (0.93), TMSB4X and OST4 (0.90), TMSB4X and PPBP (0.91), indicating these genes may work synergistically in HCC occurrence. PPBP and FTL showed a negative correlation (-0.55), which may reflect their antagonistic effects in the pathology of HCC (Supplementary Fig.\u0026nbsp;3).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo verify the reliability of the model, the HCC early diagnostic model was interpreted and optimized by KAN. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e showed the performance and structural analysis of the KAN model in HCC diagnosis, highlighting its classification ability and interpretability. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea presented the ROC curve of the KAN model with an AUC value of 0.86, indicating that the model displayed a good classification performance and could effectively distinguish between patients with HCC and healthy individuals. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb illustrated the structure of the KAN network, where the input layer contained gene features including MTRNR2L8, PF4, TMSB4X, HBB, PPBP, MTRNR2L12, FTL, OST4, S100A9, and ACTB, connected to the hidden layer node (1,0) through activation functions on the edges, ultimately outputting the prediction results, showing the sparse structure of the KAN model. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec showed the activation function fitting from MTRNR2L8 to node (1,0), with blue dots representing data points, and the orange line as the fitting curve (\u003cem\u003eR\u0026sup2;\u003c/em\u003e=1.00), where red markers indicated sample points, demonstrating its perfect capture of the nonlinear relationship. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ed presented the activation function fitting from FTL to node (1,0), with the fitting curve (R\u0026sup2;=0.92) also showing a good correspondence between the data and the fitting curve. These results indicate that the KAN model, through edge activation functions and a sparse structure, not only enhances prediction performance but also improves the model's interpretability, providing reliable theoretical support for HCC diagnosis. Additionally, the mathematical expression formula generated by KAN was displayed in Supplementary Fig.\u0026nbsp;4.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eOnline prediction platform\u003c/h2\u003e \u003cp\u003eTo enhance the clinical application of the HCC diagnostic model, an online prediction platform has been developed based on the Streamlit framework as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. The accessible link was as follows, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ai-model-j7uqkifaxp6etnfsf4r4t8.streamlit.app/\u003c/span\u003e\u003cspan address=\"https://ai-model-j7uqkifaxp6etnfsf4r4t8.streamlit.app/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Clinicians or patients could input EV-derived gene expression levels to obtain the predicted probability of HCC and an individual sample decision plot based on SHAP. This platform provides clinicians with a convenient diagnostic tool, accelerating the promotion of non-invasive early diagnosis of HCC.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussions","content":"\u003cp\u003eIn this study, we developed and validated an interpretable machine learning model for non-invasive diagnosis of hepatocellular carcinoma based on EV-derived gene signatures. Our findings demonstrate that EV-derived mRNAs can serve as effective biomarkers for HCC detection, offering a promising alternative to conventional diagnostic methods.\u003c/p\u003e \u003cp\u003eThe superior performance of DNN compared to other machine learning algorithms in our study highlights the complex, non-linear relationships between EV-derived gene expression patterns and HCC status. This complexity necessitates sophisticated modeling approaches that can capture intricate interactions among multiple features. Our optimized model, incorporating ten key mRNAs (MTRNR2L8, HBB, PF4, FTL, MTRNR2L12, TMSB4X, PPBP, OST4, ACTB, and S100A9), achieved a robust performance with an AUC of 0.8877 in the test set, demonstrating its potential clinical utility for non-invasive HCC detection.\u003c/p\u003e \u003cp\u003eThe identified key genes have biological relevance to HCC pathogenesis. MTRNR2L8 (Humanin-like 8), identified as the most significant contributor to our model, is a peptide-encoding gene with reported anti-apoptotic and cytoprotective functions\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e,\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. Previous studies have suggested that humanin and its analogs may play roles in cancer development through regulating cell survival pathways\u003csup\u003e\u003cspan additionalcitationids=\"CR31\" citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. HBB (Hemoglobin Beta) has been associated with various cancers, potentially reflecting tumor-related angiogenesis, and oxygen transport alterations\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. PF4 (Platelet Factor 4) is known for its anti-angiogenic properties and has been implicated in tumor microenvironment regulation\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e,\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e. S100A9, a calcium-binding protein, has been reported to promote inflammation and cancer progression, including in HCC\u003csup\u003e\u003cspan additionalcitationids=\"CR38\" citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e. The identification of these genes as key diagnostic markers not only enhances the model's predictive performance but also provides insights into potential biological mechanisms underlying HCC development.\u003c/p\u003e \u003cp\u003eOur approach to model interpretability using SHAP and KAN frameworks addresses a critical limitation of many machine learning models in healthcare\u0026mdash;the \"black box\" problem. By elucidating how individual features contribute to predictions at both global and local levels, our model provides transparency that is essential for clinician trust and adoption. The strong correlations observed between certain genes (e.g., MTRNR2L8 and MTRNR2L12, TMSB4X and OST4) suggest potential functional interactions or co-regulation mechanisms that warrant further investigation. The antagonistic relationship between PPBP and FTL may reflect competing biological processes in HCC pathogenesis.\u003c/p\u003e \u003cp\u003eThe development of an online prediction platform represents a significant step toward clinical implementation of our diagnostic model. This user-friendly interface allows clinicians to input patient-specific EV-derived gene expression data and obtain immediate diagnostic predictions with visual explanations, potentially facilitating integration into clinical workflows. Such accessibility could contribute to broader adoption of non-invasive HCC screening, particularly in resource-limited settings where conventional imaging techniques may be less available.\u003c/p\u003e \u003cp\u003eSeveral limitations of our study should be acknowledged. First, while the exoRBase 2.0 database provided a substantial dataset for model development, external validation with prospective, multi-center cohorts would strengthen the evidence for clinical applicability. Second, standardization of EV isolation and RNA quantification methods remains challenging and could affect the reproducibility of gene expression measurements across different laboratories. Third, our model focused exclusively on RNA signatures and did not incorporate other potentially valuable clinical data such as age, gender, etiology of liver disease, or conventional biomarkers like AFP. Integration of these factors might further enhance diagnostic performance.\u003c/p\u003e \u003cp\u003eFuture directions should include prospective validation in diverse patient populations, particularly those with various liver disease backgrounds, to assess the model's performance across different HCC etiologies. Investigation of the temporal dynamics of EV-derived signatures during HCC development could provide insights into their utility for early detection. Additionally, functional studies of the identified key genes would enhance our understanding of their roles in HCC pathogenesis and potentially reveal novel therapeutic targets.\u003c/p\u003e \u003cp\u003eIn conclusion, our study demonstrates the feasibility and potential clinical utility of an interpretable machine learning model based on EV-derived gene signatures for non-invasive HCC detection. By combining sophisticated algorithm selection, feature optimization, and model interpretability with accessible implementation, our approach addresses current diagnostic challenges and represents a step toward improved early detection and management of HCC. The integration of such liquid biopsy-based approaches into clinical practice could potentially transform HCC screening paradigms, ultimately contributing to better patient outcomes through earlier intervention.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eHCC, hepatocellular carcinoma; EVs, extracellular vesicles; SHAP, SHapley Additive exPlanations; KAN, Kolmogorov-Arnold Networks; DNN, Deep Neural Networks; AFP, alpha-fetoprotein.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eprovided by Natural Science Foundation of Chongqing, China (Grant No. CSTB2024NSCQ-MSX0490); by Joint project of Chongqing Health Commission and Science and Technology Bureau (Grant No. 2025MSXM075); by Scientific and Technological Research Program of Chongqing Municipal Education Commission (Grant No. KJZD-M202300102); by Development Program of Chongqing university, Jiangjin hospital (Grant No. 2023LJXM003, 2024LJXM001); by Research Launch Project of Chongqing university, Jiangjin hospital (Grant No. 2023qdjfxm006).\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eDeclaration of interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no potential conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors reviewed the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank Guoyu Tan and members of the Lun Lun Research Institute of AI FOR SCIENCE for technical support and suggestions that improved the manuscript and experiments.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFinancial support\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFunding provided by Natural Science Foundation of Chongqing, China (Grant No. CSTB2024NSCQ-MSX0490); by Joint project of Chongqing Health Commission and Science and Technology Bureau (Grant No. 2025MSXM075); by Scientific and Technological Research Program of Chongqing Municipal Education Commission (Grant No. KJZD-M202300102); by Development Program of Chongqing university, Jiangjin hospital (Grant No. 2023LJXM003，2024LJXM001); by Research Launch Project of Chongqing university, Jiangjin hospital (Grant No. 2023qdjfxm006).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data analyzed in this study are available in the exoRBase 2.0 at http://www.exorbase.org/exoRBaseV2/download/toIndex. The protocol and statistical analysis methods used in the study can be requested directly from the corresponding author after approval.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupporting Information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe online version contains supplementary material available at\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eVillanueva, A. Hepatocellular Carcinoma. \u003cem\u003eN. Engl. J. Med.\u003c/em\u003e \u003cb\u003e380\u003c/b\u003e, 1450\u0026ndash;1462 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGalle, P. R. et al. EASL Clinical Practice Guidelines: Management of hepatocellular carcinoma. \u003cem\u003eJ. Hepatol.\u003c/em\u003e ;\u003cb\u003e69\u003c/b\u003e. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSangro, B. et al. Durvalumab with or without bevacizumab with transarterial chemoembolisation in hepatocellular carcinoma (EMERALD-1): a multiregional, randomised, double-blind, placebo-controlled, phase 3 study. \u003cem\u003eLancet\u003c/em\u003e \u003cb\u003e405\u003c/b\u003e, 216\u0026ndash;232 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGanne-Carri\u0026eacute;, N. \u0026amp; Nahon, P. Differences between hepatocellular carcinoma caused by alcohol and other aetiologies. \u003cem\u003eJ. Hepatol.\u003c/em\u003e \u003cb\u003e82\u003c/b\u003e, 909\u0026ndash;917 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, Y. et al. Genomic and the tumor microenvironment heterogeneity in multifocal hepatocellular carcinoma. \u003cem\u003eHepatol. Dec.\u003c/em\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarquardt, J. U., Galle, P. R. \u0026amp; Teufel, A. Molecular diagnosis and therapy of hepatocellular carcinoma (HCC): An emerging field for advanced technologies. \u003cem\u003eJ. Hepatol.\u003c/em\u003e ;\u003cb\u003e56\u003c/b\u003e. (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEl-Serag, H. B. et al. Diagnosis and Treatment of Hepatocellular Carcinoma. Gastroenterology 2008;134.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTzartzeva, K. et al. Surveillance Imaging and Alpha Fetoprotein for Early Detection of Hepatocellular Carcinoma in Patients With Cirrhosis: A Meta-analysis. Gastroenterology 2018;154.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu, R. et al. Extracellular vesicles in cancer \u0026mdash; implications for future improvements in cancer care. \u003cem\u003eNat. Rev. Clin. Oncol.\u003c/em\u003e ;\u003cb\u003e15\u003c/b\u003e. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBebelman, M. P. et al. Biogenesis and function of extracellular vesicles in cancer. \u003cem\u003ePharmacol. Ther.\u003c/em\u003e ;\u003cb\u003e188\u003c/b\u003e. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eichiro Ohno, S., Ishikawa, A. \u0026amp; Kuroda, M. Roles of exosomes and microvesicles in disease pathogenesis. \u003cem\u003eAdv. Drug Deliv Rev.\u003c/em\u003e ;\u003cb\u003e65\u003c/b\u003e. (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTaylor, D. D. \u0026amp; Gercel-Taylor, C. Exosomes/microvesicles: mediators of cancer-associated immunosuppressive microenvironments. \u003cem\u003eSemin Immunopathol.\u003c/em\u003e ;\u003cb\u003e33\u003c/b\u003e. (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark, K-S., L\u0026auml;sser, C. \u0026amp; L\u0026ouml;tvall, J. Extracellular vesicles and the lung: from disease pathogenesis to biomarkers and treatments. \u003cem\u003ePhysiol. Rev.\u003c/em\u003e \u003cb\u003e105\u003c/b\u003e, 1733\u0026ndash;1821 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, Y. \u0026amp; Ding, S. Extracellular vesicles in cancer cachexia: deciphering pathogenic roles and exploring therapeutic horizons. \u003cem\u003eJ. Transl Med.\u003c/em\u003e \u003cb\u003e22\u003c/b\u003e, 506 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDorado, E. et al. Extracellular vesicles as a promising source of lipid biomarkers for breast cancer detection in blood plasma. \u003cem\u003eJ. Extracell. Vesicles\u003c/em\u003e ;\u003cb\u003e13\u003c/b\u003e. (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRusso, M. N. et al. Extracellular vesicles in the glioblastoma microenvironment: A diagnostic and therapeutic perspective. \u003cem\u003eMol. Aspects Med.\u003c/em\u003e \u003cb\u003e91\u003c/b\u003e, 101167 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao, C. et al. Extracellular vesicle digital scoring assay for assessment of treatment responses in hepatocellular carcinoma patients. \u003cem\u003eJ. Experimental Clin. Cancer Res.\u003c/em\u003e \u003cb\u003e44\u003c/b\u003e, 136 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, P. et al. Non-canonical small noncoding RNAs in the plasma extracellular vesicles as novel biomarkers in gastric cancer. \u003cem\u003eJ. Hematol. Oncol.\u003c/em\u003e \u003cb\u003e18\u003c/b\u003e, 39 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShao, X. et al. Extracellular vesicle-derived miRNA-mediated cell\u0026ndash;cell communication inference for single-cell transcriptomic data with miRTalk. \u003cem\u003eGenome Biol.\u003c/em\u003e \u003cb\u003e26\u003c/b\u003e, 95 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, G. et al. Integrating single-cell RNA sequencing, WGCNA, and machine learning to identify key biomarkers in hepatocellular carcinoma. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 11157 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePinsky, M. R. et al. Use of artificial intelligence in critical care: opportunities and obstacles. \u003cem\u003eCrit. Care\u003c/em\u003e. \u003cb\u003e28\u003c/b\u003e, 113 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, H., Lundberg, S. M. \u0026amp; Lee, S-I. Explaining a series of models by propagating Shapley values. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 4512 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, Z. et al. KAN: Kolmogorov-Arnold Networks. April (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, S. et al. ExoRBase: A database of circRNA, lncRNA and mRNA in human blood exosomes. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e ;\u003cb\u003e46\u003c/b\u003e. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLai, H. et al. exoRBase 2.0: An atlas of mRNA, lncRNA and circRNA in extracellular vesicles from human biofluids. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e ;\u003cb\u003e50\u003c/b\u003e. (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePedregosa, F. et al. Scikit-learn: Machine learning in Python. \u003cem\u003eJ. Mach. Learn. Res.\u003c/em\u003e ;\u003cb\u003e12\u003c/b\u003e. (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLUNDBERG S. A unified approach to interpreting model predictions. \u003cem\u003earXiv preprint\u003c/em\u003e (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShen, Y. et al. Epigenome-Wide Association Study Indicates Hypomethylation of \u003cem\u003eMTRNR2L8\u003c/em\u003e in Large-Artery Atherosclerosis Stroke. Stroke. ;50:1330\u0026ndash;1338. (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuth, A. et al. Single cell transcriptomics of cerebrospinal fluid cells from patients with recent-onset narcolepsy. \u003cem\u003eJ. Autoimmun.\u003c/em\u003e \u003cb\u003e146\u003c/b\u003e, 103234 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGordon-Lipkin, E. M. et al. Primary oxidative phosphorylation defects lead to perturbations in the human B cell repertoire. \u003cem\u003eFront. Immunol.\u003c/em\u003e ;\u003cb\u003e14\u003c/b\u003e. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBangert, C. et al. Comprehensive Analysis of Nasal Polyps Reveals a More Pronounced Type 2 Transcriptomic Profile of Epithelial Cells and Mast Cells in Aspirin-Exacerbated Respiratory Disease. \u003cem\u003eFront. Immunol.\u003c/em\u003e ;\u003cb\u003e13\u003c/b\u003e. (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIparraguirre, L. et al. Whole-Transcriptome Analysis in Peripheral Blood Mononuclear Cells from Patients with Lipid-Specific Oligoclonal IgM Band Characterization Reveals Two Circular RNAs and Two Linear RNAs as Biomarkers of Highly Active Disease. \u003cem\u003eBiomedicines\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, 540 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaman, S. et al. The Beta Subunit of Hemoglobin (HBB2/HBB) Suppresses Neuroblastoma Growth and Metastasis. \u003cem\u003eCancer Res.\u003c/em\u003e \u003cb\u003e77\u003c/b\u003e, 14\u0026ndash;26 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng, Y. et al. Expression of β-globin by cancer cells promotes cell survival during blood-borne dissemination. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, 14344 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u0026Ccedil;akan, E. et al. TLR9 ligand sequestration by chemokine CXCL4 negatively affects central B cell tolerance. \u003cem\u003eJ. Exp. Med.\u003c/em\u003e ;\u003cb\u003e220\u003c/b\u003e. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, C. et al. ITGB6 modulates resistance to anti-CD276 therapy in head and neck cancer by promoting PF4\u0026thinsp;+\u0026thinsp;macrophage infiltration. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 7077 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, Z. et al. E-twenty-six-specific sequence variant 5 (ETV5) facilitates hepatocellular carcinoma progression and metastasis through enhancing polymorphonuclear myeloid-derived suppressor cell (PMN-MDSC)-mediated immunosuppression. Gut February 2025:gutjnl-2024-333944.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVald\u0026eacute;s-Mas, R. et al. Metagenome-informed metaproteomics of the human gut microbiome, host, and dietary exposome uncovers signatures of health and inflammatory bowel disease. \u003cem\u003eCell\u003c/em\u003e \u003cb\u003e188\u003c/b\u003e, 1062\u0026ndash;1083e36 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, Q. et al. PGAM5 interacts with and maintains BNIP3 to license cancer-associated muscle wasting. \u003cem\u003eAutophagy\u003c/em\u003e \u003cb\u003e20\u003c/b\u003e, 2205\u0026ndash;2220 (2024).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Diagnostic biomarker, Hepatocellular carcinoma, Extracellular vesicle, Interpretability, Non-invasive","lastPublishedDoi":"10.21203/rs.3.rs-8217267/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8217267/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHepatocellular carcinoma (HCC) is a major cause of cancer death worldwide, underscoring the need for early non-invasive diagnostics. This study developed an interpretable model using extracellular vesicle (EV)-derived RNA signatures from exoRBase 2.0. Six machine learning algorithms were compared, with the Deep Neural Network (DNN) achieving superior performance (AUC\u0026thinsp;=\u0026thinsp;0.8877). Ten diagnostic mRNAs (MTRNR2L8, HBB, PF4, FTL, MTRNR2L12, TMSB4X, PPBP, OST4, ACTB, and S100A9) were identified, among which MTRNR2L8 was the most significant predictor. SHAP and Kolmogorov-Arnold Networks (KAN) revealed nonlinear feature\u0026ndash;outcome relationships and potential biomarkers. An online platform was created for real-time clinical use. This tool offers a robust, interpretable, and non-invasive method for early HCC detection, potentially improving diagnostic timeliness and decision-making. Further validation in prospective cohorts is warranted.\u003c/p\u003e","manuscriptTitle":"Clinically Interpretable Extracellular Vesicle Gene Model for Non-Invasive Liver Cancer Diagnosis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-19 17:45:30","doi":"10.21203/rs.3.rs-8217267/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-07T17:26:17+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-06T08:19:50+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"228690065874117667326006708066070083010","date":"2025-12-17T06:42:21+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-16T21:53:53+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-12-16T09:57:10+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-12-02T18:26:35+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-12-01T09:56:41+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-12-01T01:35:26+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"6de62fb8-1d90-457c-b178-1d8a96ab7d75","owner":[],"postedDate":"December 19th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":59803960,"name":"Health sciences/Biomarkers"},{"id":59803961,"name":"Biological sciences/Cancer"},{"id":59803962,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":59803963,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-02-16T16:14:23+00:00","versionOfRecord":{"articleIdentity":"rs-8217267","link":"https://doi.org/10.1038/s41598-026-40020-9","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-02-14 15:59:01","publishedOnDateReadable":"February 14th, 2026"},"versionCreatedAt":"2025-12-19 17:45:30","video":"","vorDoi":"10.1038/s41598-026-40020-9","vorDoiUrl":"https://doi.org/10.1038/s41598-026-40020-9","workflowStages":[]},"version":"v1","identity":"rs-8217267","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8217267","identity":"rs-8217267","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}