Extracellular Vesicle Surface Protein Profiling in Metabolic Dysfunction‐associated Steatotic Liver Disease in Obesity Authors: Nahuel Aquiles Garcia, Kirstine Kløve‐Mogensen, Katja Lund Cliff, Rikke Bæk, Malwina Ulanowska, Javier Alejandro Donoso Quezada, Malene Møller Jørgensen, Aase Handberg, Maiken Mellergaard Published: 2026 ABSTRACT ABSTRACTMetabolic dysfunction‐associated steatotic liver disease (MASLD) is highly prevalent in obesity, but scalable tools for biological risk stratification remain limited. In this cross‐sectional, single‐centre discovery study (MULTISITE), we profiled circulating extracellular vesicle (cEV) surface proteins by EV Array in platelet‐poor plasma from individuals with obesity and MASLD (n = 36), individuals with obesity without MASLD (n = 24), and lean controls (n = 27). Liver fat was quantified by magnetic resonance imaging proton density fat fraction (MRI‐PDFF). We assessed group differences, associations with continuous PDFF, and exploratory discriminatory performance using single markers, pairwise ratios, and sparse logistic models. We also performed an exploratory PDFF‐extremes enrichment analysis within obesity by comparing the top versus bottom 30% of the PDFF distribution (n = 18 vs. n = 18), excluding the middle 40%. Across contrasts, single markers showed modest separation, whereas pairwise ratios provided stronger exploratory signals. In the discovery panel, uncorrected best‐ratio AUCs were 0.80–0.84 and sparse‐model AUCs were 0.79–0.84 under internal, non‐nested cross‐validation across MASLD‐focused contrasts. In the focused panel, CD36 and TREM2 were the strongest single‐marker signals for MASLD versus lean controls, whereas LRP‐1‐anchored ratios captured obesity‐ and liver fat‐associated patterns. In the PDFF‐extremes analysis, internally cross‐validated AUCs reached 0.87–0.89 in the discovery panel and 0.86 for LRP‐1/FATP5 in the focused panel; these estimates should be interpreted as exploratory, uncorrected discovery‐stage observations rather than validated classifier performance. Exploratory benchmarking showed that FLI, but not FIB‐4, discriminated MASLD‐related contrasts strongly, however, these analyses provide contextual benchmarking only and do not establish incremental clinical value of cEV markers beyond routine clinical predictors. These findings nominate biologically coherent cEV surface signatures for future validation in MASLD, anchored by lipid handling/scavenger pathways and complemented by stress, immune, and coagulation biology. Trial Registration: ClinicalTrials.gov: NCT05699863Keywords: biomarkers, EV array, extracellular vesicles, MASLD, MRI‐PDFF, obesity Introduction Metabolic dysfunction‐associated steatotic liver disease (MASLD), formerly termed non‐alcoholic fatty liver disease (NAFLD), is the most common chronic liver disease worldwide and is strongly linked to obesity and insulin resistance (Huang et al. 2025). Although many individuals with MASLD have mild steatosis, a substantial proportion progress to steatohepatitis, fibrosis, cirrhosis and hepatocellular carcinoma. Clinical management is challenged by limited access to accurate non‐invasive assessment tools. Liver biopsy remains the reference standard for histology but is invasive and unsuitable for population‐level screening or longitudinal monitoring (Munk Lauridsen et al. 2025). Magnetic Resonance Imaging (MRI)‐based proton density fat fraction (MRI‐PDFF) quantification of liver fat is sensitive and non‐invasive, but costly and not uniformly available, motivating scalable blood‐based biomarkers that could support biological risk stratification and future validation workflows rather than replace established diagnostic pathways (Masoodi et al. 2021). Circulating extracellular vesicles (cEVs) are membrane‐enclosed particles released by many cell types, including hepatocytes, immune cells, platelets and endothelium (Garcia et al. 2023). Because cEVs carry surface proteins and cargo reflective of their cells of origin, they provide a window into tissue stress responses, immune activation and lipid handling pathways relevant to metabolic liver disease. Surface protein profiling is attractive for biomarker discovery because it is compatible with biofluids and multiplex immunoassays and may nominate candidate signatures for future validation. We developed a comprehensive in silico strategy for identifying EV surface proteins as biomarkers for NAFLD (Garcia et al. 2023). In this study, by using EV Array technology (Bæk and Jørgensen 2017; Jørgensen et al. 2013, 2015), we investigated whether cEV surface proteins could nominate candidate EV‐associated biomarker signatures for MASLD in obesity. Using blood samples from the prospective MULTISITE clinical protocol, we assessed cEV protein group‐level differences, correlations with continuous liver fat content, and exploratory discriminatory performance using single markers, ratios and multivariate models. The MULTISITE clinical trial was initiated before the MASLD nomenclature was introduced and therefore includes legacy NAFLD‐ and metabolic syndrome (MetS)‐based classifications. In the present work, we focus on MASLD as the primary clinical framework. We present the legacy NAFLD + MetS analyses as Supporting Information to preserve linkage to the original protocol and to assess robustness of key EV signatures to alternate labelling. Moreover, we recently showed that CD36+ cEVs correlate with PDFF values in this same cohort of participants (Mellergaard et al. 2026). This motivated us to explore the potential of cEV surface proteins as discovery‐stage candidate signatures for MASLD in obesity from an EV Array perspective. Our primary objective was to nominate exploratory cEV surface protein patterns associated with MASLD status and liver fat severity measured by MRI‐PDFF, rather than to validate a diagnostic classifier. We further aimed to identify parsimonious ratio‐based candidate signatures for future validation. We further aimed to develop parsimonious ratio‐based signatures suitable for future validation. Methods For an extended version of methods, see the Supporting Information Methods. Study Design and Participants This cross‐sectional case‐control study was conducted at Aalborg University Hospital, Aalborg, Denmark, within the MULTISITE protocol (ethics N‐20200013). Participants were 30–60‐years‐old. Obesity was defined as BMI 30.0–39.9 kg/m2 and lean status as Body Mass Index (BMI) 18.5–25.0 kg/m2. The MASLD‐focused analyses included obesity with MASLD (n = 36), obesity without MASLD (n = 24) and lean controls (n = 27). MASLD was defined by current consensus criteria (MRI‐PDFF ≥ 5.0% plus cardiometabolic risk). Full eligibility and classification details are provided in the Supporting Information Methods. Legacy NAFLD/MetS definitions from protocol initiation are reported only as supplementary robustness context (Figure S3; Tables S3–S6). For clarification, accordingly with the MASLD criteria (Rinella et al. 2023), we included one individual with overweight (BMI 29) in the obesity with MASLD group. MASLD was defined by Rinella et al. (2023): presence of liver steatosis (defined based on proton density fat fraction by MRI [PDFF] ≥ 5.0%) along with one of the following criteria: elevated waist circumference (≥94 cm in men or ≥80 cm in women) or BMI ≥ 25 kg/m2, elevated blood pressure (systolic blood pressure ≥130 mmHg or diastolic blood pressure ≥85 mmHg or antihypertensive drug), elevated fasting plasma glucose (≥5.6 mmol/L [100 mg/dL] or glucose‐lowering drugs), elevated triglycerides (≥1.7 mmol/L [150 mg/dL] or lipid‐lowering drug), low High‐Density Lipoprotein (HDL) cholesterol (<1.03 mmol/L [40 mg/dL] in men or <1.29 mmol/L [50 mg/dL] in women or lipid‐lowering drug). Clinical Characterisation and MRI‐PDFF Participants underwent standardised fasting assessments, including anthropometry and blood pressure. Venous blood was collected after overnight fasting. Insulin resistance was estimated using the Homeostasis Model Assessment 2 (HOMA2) calculator. Liver fat content was quantified by MRI proton density fat fraction (PDFF) on a 3T scanner using multi‐point Dixon sequences to generate PDFF maps. Liver PDFF was determined as the mean of four regions of interest on PDFF maps (Mellergaard et al. 2026; Gjela et al. 2024). EV Array Surface Protein Profiling Platelet‐poor plasma was prepared by sequential centrifugation and stored at –80°C. cEV surface profiling was performed using EV Array in a two‐stage design comprising a 90‐marker discovery panel and a 12‐marker focused same‐cohort follow‐up panel. The discovery and focused panels were processed independently as separate assay panels. In the 90‐marker discovery panel, markers were distributed into two 45‐marker runs/slides per sample, with CD9 and CD81 included on both slides and their summed signal used for batch correction/normalisation between runs; samples were analysed as mixed samples and distributed across 10 slides per setup. The 12‐marker focused panel was processed independently, with samples analysed as mixed samples and distributed across nine slides. The discovery panel was based on a priori in silico candidate prioritisation and the available EV Array antibody repository in combination with our previous research (Garcia et al. 2023); the focused panel was a targeted follow‐up enriched for consistent discovery trends and, motivated by our recent paper, included a functional CD36 antibody (Mellergaard et al. 2026), since we showed that CD36+ cEV correlates with PDFF in this cohort of participants. Complete antibody compositions are provided in Tables S1 and S2, extended performance summaries and model equations are provided in Tables S5–S10, and marker‐level imputed‐frequency summaries are provided in Tables S12 and S13. Signals were background‐corrected, handled across replicates as in prior EV Array publications (Bæk and Jørgensen 2017, Jørgensen et al. 2013, 2015), and log10 transformed for analysis. To provide orthogonal assay‐level corroboration of selected EV‐associated markers, high‐resolution flow cytometry (hFCM) was performed on three independent obesity EDTA platelet‐poor plasma pools, including pooled material with partial overlap to the EV Array clinical cohort; detailed acquisition settings, antibody/isotype information, calibration procedures, detergent controls and MIFlowCyt/MIFlowCyt‐EV reporting are provided in the Supporting Information Methods and hFCM Tables SA1–SA6 Statistical Analysis Clinical variables were summarized as median (range). For EV performance, ROC AUCs were calculated for single markers and all pairwise ratios within each panel. Because all candidate ratios were screened within the available cohort, best‐ratio AUCs are reported as uncorrected discovery maxima rather than validated performance estimates. Sparse logistic models used small EV feature sets selected from univariate ranking within each contrast; performance was reported as apparent training AUC and 10‐fold cross‐validated AUC. Feature selection was not nested within cross‐validation; therefore, these cross‐validated AUCs should be interpreted as internal, potentially optimistic estimates. No permutation correction was applied to the ratio‐search procedure. All AUC estimates are therefore reported as internal discovery‐stage estimates and should not be interpreted as externally validated performance or evidence of clinical readiness. As an exploratory secondary analysis within obesity, PDFF extremes were compared (bottom 30% vs. top 30%; n = 18 vs. n = 18), with the middle 40% excluded to create a biological enrichment contrast rather than a routine clinical screening scenario. Wilcoxon rank‐sum testing with Benjamini–Hochberg false discovery rate (FDR) control was used for group‐difference screening and is detailed in the Supporting Information and Tables S5–S10. As exploratory benchmarking against routine clinical indices, FIB‐4 and FLI were calculated from available clinical and biochemical variables using established formulae (Sterling et al. 2006; Bedogni et al. 2006), and their discriminatory performance was evaluated by ROC analysis in the same contrasts (Table S11). Results Cohort Characteristics and Liver Fat Burden Participants with obesity and MASLD and those with obesity without MASLD had comparable BMI distributions, while lean controls had lower BMI by design (Table 1). MASLD was accompanied by higher MRI‐PDFF and a less favourable cardiometabolic profile, including higher blood pressure, triglycerides, glucose and HOMA2‐IR together with lower HDL cholesterol. Within the obesity cohort, PDFF tracked a metabolic severity gradient; detailed PDFF‐extremes baseline characteristics are provided in Table S4. Legacy NAFLD/MetS cohort characterisation is provided for robustness context in Figure S3 and Table S3. TABLE 1Baseline characteristics of the MASLD‐focused analysis groups (median, range).GroupBMI, kg/m2 Sex (M/F), n/n Age, yearsPDFF, %SBP, mmHgDBP, mmHgHDL, mmol/LTG, mmol/LGlucose, mmol/LHOMA2‐IRIndividuals with obesity and MASLD35.8 (29.0–40.9)16/2047 (31–59)12.6 (5.1–33.1)138.6 (76.5–198.0)92.7 (78.0–130.5)1.1 (0.7–1.7)1.6 (0.7–3.6)5.8 (4.7–7.4)3.2 (0.9–6.2)Individuals with obesity without MASLD35.3 (30.4–39.5)5/1945 (33–58)2.6 (1.6–4.1)128.0 (109.5–154.5)86.1 (75.0–103.5)1.4 (1.0–2.2)1.0 (0.5–1.7)5.4 (4.9–6.5)1.6 (0.9–3.3)Lean controls22.7 (19.6–24.9)16/2046 (30–59)2.1 (1.2–4.8)126.4 (100.0–155.0)78.8 (63.5–101.5)1.5 (0.9–2.2)0.9 (0.4–1.9)5.2 (4.4–5.7)0.8 (0.2–2.1) Baseline characteristics of the MASLD‐focused analysis groups (median, range). Candidate cEV Signatures Show Exploratory Separation by MASLD Status and Association With MRI‐PDFF We profiled cEV surface proteins using two complementary EV Array panels: a 90‐marker discovery panel and a 12‐marker focused same‐cohort follow‐up panel. Across MASLD‐focused contrasts, single‐marker discrimination was modest, whereas pairwise cEV ratios provided stronger exploratory signals. In the 90‐marker discovery panel, top ratio features included tumour necrosis factor‐related apoptosis‐inducing ligand (TRAIL)/Heat Shock Protein 70 (Hsp70) for MASLD versus lean controls and CD8a/Fatty Acid Transport Protein 1 (FATP1) for MASLD versus obesity controls, and sparse two‐feature models reached non‐nested internally cross‐validated AUCs of 0.79–0.84 (Table 2; Table S7). In the 12‐marker focused panel, CD36 and Triggering Receptor Expressed on Myeloid cells 2 (TREM2) were the strongest single‐marker signals for MASLD versus lean controls, while low‐density lipoprotein receptor‐related protein 1 (LRP‐1)‐anchored ratios were most informative for obesity‐related contrasts (Table 3; Table S8). Representative ROC curves for logistic models are shown in Figure 1. Legacy NAFLD/MetS analyses are reported only as supplementary robustness context and showed broadly overlapping patterns (Tables S5 and S6). Orthogonal hFCM corroboration in three independent obesity EDTA platelet‐poor plasma pools, including pooled material with partial overlap to the EV Array clinical cohort, showed detectable CD36+, FATP1+ and FATP1+CD36+ events within a CD9+CD81+CD63+Lactadherin+ EV gate (Figures S1 and S2; hFCM Tables SA1–SA6). These data support marker detectability/co‐expression and gating plausibility at the assay level but do not represent participant‐level clinical validation of the EV Array‐derived classifiers. TABLE 2EV Array 90‐marker discovery panel performance for MASLD‐focused contrasts.Contrast (90‐marker panel)Best single AUCBest ratio AUCRepresentative logistic model (features)Train AUC10‐fold CV AUCMASLD vs. leanCD38 0.688TRAIL/Hsp70 0.804TRAIL/Hsp70; CD322/Hsp700.8120.786MASLD vs. obesity control groupFATP1 0.703CD8a/FATP1 0.839CD8a/FATP1; CD14/CD1710.8350.799Obesity (all) vs. leanCD11a 0.637Notch1/ADAMTS13 0.820Notch1/ADAMTS13; FATP5/CD1630.8760.841 Note: Best single and best ratio AUCs are univariate ROC AUCs used to rank features for modelling. Best‐ratio AUCs are uncorrected discovery maxima from the available feature space. Logistic models are sparse and evaluated by apparent training AUC and non‐nested 10‐fold cross‐validated AUC; these are internal estimates and should not be interpreted as validated classifier performance. EV Array 90‐marker discovery panel performance for MASLD‐focused contrasts. Note: Best single and best ratio AUCs are univariate ROC AUCs used to rank features for modelling. Best‐ratio AUCs are uncorrected discovery maxima from the available feature space. Logistic models are sparse and evaluated by apparent training AUC and non‐nested 10‐fold cross‐validated AUC; these are internal estimates and should not be interpreted as validated classifier performance. TABLE 3EV Array 12‐marker focused panel performance for MASLD‐focused contrasts.Contrast (12‐marker panel)Best single AUCBest ratio AUCRepresentative logistic model (features)Train AUC10‐fold CV AUCMASLD vs. leanCD36 0.718 (TREM2 0.682)FATP5/TREM2 0.699Not fitted——MASLD vs. obesity control groupLRP‐1 0.706LRP‐1/FATP5 0.748 (LRP‐1/CD9 0.726)Not fitted——Obesity (all) vs. leanLRP‐1 0.690LRP‐1/CD36 0.790LRP‐1/CD36; LRP‐1/CD9; LRP‐1/TREM20.7910.742 Note: In MASLD versus lean, CD36 and TREM2 showed the strongest single‐marker signals in the focused same‐cohort panel. No multivariable models were fitted for the focused‐panel MASLD versus lean or MASLD versus obesity‐control contrasts to avoid additional overfitting in small groups when univariate discrimination was modest. Where multivariable models are shown, performance is reported as apparent training AUC and non‐nested 10‐fold cross‐validated AUC. EV Array 12‐marker focused panel performance for MASLD‐focused contrasts. Note: In MASLD versus lean, CD36 and TREM2 showed the strongest single‐marker signals in the focused same‐cohort panel. No multivariable models were fitted for the focused‐panel MASLD versus lean or MASLD versus obesity‐control contrasts to avoid additional overfitting in small groups when univariate discrimination was modest. Where multivariable models are shown, performance is reported as apparent training AUC and non‐nested 10‐fold cross‐validated AUC. Representative ROC curves for EV Array models in MASLD‐focused contrasts. (A) 90‐Marker discovery panel models: MASLD versus lean controls (n = 36 vs. n = 27; TRAIL/Hsp70 and CD322/Hsp70), MASLD versus obesity controls (n = 36 vs. n = 24; CD8a/FATP1 and CD14/CD171) and all obesity versus lean controls (n = 60 vs. n = 27; Notch1/ADAMTS13 and FATP5/CD163). (B) 12‐Marker focused panel model for all obesity versus lean controls (n = 60 vs. n = 27; LRP‐1/CD36, LRP‐1/CD9 and LRP‐1/TREM2). Numerical performance metrics are summarised in Tables 2 and 3. AUC, area under the curve; ROC, receiver operating characteristic. Exploratory ROC benchmarking against routine clinical indices showed that FLI consistently outperformed FIB‐4 across the evaluated contrasts (Table S11). For MASLD versus obesity controls, FLI achieved an AUC of 0.773 (95% CI: 0.638–0.881), whereas FIB‐4 showed limited discrimination (AUC 0.542, 95% CI: 0.389–0.688). For MASLD versus lean controls, FLI achieved an AUC of 1.000 (95% CI: 1.000–1.000), consistent with the strong contribution of BMI/waist‐based variables to this contrast, whereas FIB‐4 achieved an AUC of 0.630 (95% CI: 0.473–0.766). In the PDFF‐extremes contrast, FLI achieved an AUC of 0.849 (95% CI: 0.704–0.949) and FIB‐4 an AUC of 0.606 (95% CI: 0.395–0.772). These analyses provide contextual benchmarking only and do not test or establish incremental clinical value of EV markers beyond routine clinical predictors or established non‐invasive indices. Such incremental value will require independent cohorts and multivariable models that include routine anthropometric and biochemical predictors. PDFF‐Extremes Enrichment Analysis Within the Obesity Cohort Because MRI‐PDFF is resource‐intensive, we explored whether cEV surface features differ between individuals with obesity at the upper and lower extremes of liver fat burden. This PDFF‐extremes analysis should be interpreted as a biological enrichment analysis, not as a validated routine triage scenario. It deliberately compared the top versus bottom 30% of the PDFF distribution and excluded the middle 40%, thereby maximising contrast and likely inflating apparent discriminatory performance relative to an unselected obesity population. In this enriched contrast, ratio‐based models achieved the highest internal discrimination estimates. In the 90‐marker discovery panel, single‐ratio models based on CD280/Catenin Beta‐1 (CTNNB1), TRAIL/CD280 and CD163/CD280 achieved non‐nested cross‐validated AUCs of approximately 0.87–0.89 (Figure 2 and Table 4; Table S9). In the 12‐marker focused panel, LRP‐1/FATP5 was the top ratio feature and achieved a non‐nested cross‐validated AUC of approximately 0.86; adding FATP5/Interleukin‐27 receptor subunit alpha (IL27RA) yielded minimal incremental gain (Table 5; Table S10). Importantly, the high‐ and low‐PDFF strata differed in sex distribution (high fat: 11/7 M/F; low fat: 3/15 M/F) and broader cardiometabolic severity; therefore, the PDFF‐extremes EV signatures should not be interpreted as independent of sex or independent of metabolic severity. Representative ROC curves for EV Array models in the exploratory PDFF‐extremes enrichment analysis within obesity (top 30% vs. bottom 30% PDFF; n = 18 vs. n = 18). (A) 90‐Marker discovery panel models: CD280/CTNNB1, TRAIL/CD280, CD163/CD280 and CD171/ADAMTS13 + CXCL16/CTNNB1 + TRAIL/CD280. (B) 12‐Marker focused panel models: LRP‐1/FATP5 and LRP‐1/FATP5 + FATP5/IL27RA. Numerical performance metrics are summarised in Tables 4 and 5. AUC, area under the curve; ROC, receiver operating characteristic. TABLE 4PDFF‐extremes enrichment performance within the obesity cohort (90‐marker discovery panel).ModelEV features (predictors) Training AUC (95% CI) 10‐Fold CV AUC (95% CI) 1CD280/CTNNB10.909 (0.874–0.944)0.873 (0.737–1.000)2TRAIL/CD2800.897 (0.863–0.931)0.889 (0.778–1.000)3CD163/CD2800.895 (0.858–0.932)0.883 (0.757–1.000)4CD171/ADAMTS13; CXCL16/CTNNB1; TRAIL/CD2800.948 (0.926–0.970)0.886 (0.758–1.000) Note: Outcome is high fat (top 30% PDFF) versus low fat (bottom 30% PDFF), n = 18 versus n = 18. This is an enriched exploratory contrast excluding the middle 40% of the obesity cohort. Training and 10‐fold cross‐validated AUCs are internal, non‐nested estimates and should not be interpreted as expected performance in an unselected obesity population. PDFF‐extremes enrichment performance within the obesity cohort (90‐marker discovery panel). Training AUC (95% CI) 10‐Fold CV AUC (95% CI) Note: Outcome is high fat (top 30% PDFF) versus low fat (bottom 30% PDFF), n = 18 versus n = 18. This is an enriched exploratory contrast excluding the middle 40% of the obesity cohort. Training and 10‐fold cross‐validated AUCs are internal, non‐nested estimates and should not be interpreted as expected performance in an unselected obesity population. TABLE 5PDFF‐extremes enrichment performance within the obesity cohort (12‐marker focused panel).ModelEV features (predictors) Training AUC (95% CI) 10‐fold CV AUC (95% CI) 1LRP‐1/FATP50.883 (0.842–0.924)0.855 (0.712–0.998)2LRP‐1/FATP5; FATP5/IL27RA0.881 (0.841–0.920)0.852 (0.714–0.989) Note: Outcome is high fat (top 30% PDFF) versus low fat (bottom 30% PDFF), n = 18 versus n = 18. This is an enriched exploratory contrast excluding the middle 40% of the obesity cohort. Training and 10‐fold cross‐validated AUCs are internal, non‐nested estimates and should not be interpreted as expected performance in an unselected obesity population. PDFF‐extremes enrichment performance within the obesity cohort (12‐marker focused panel). Training AUC (95% CI) 10‐fold CV AUC (95% CI) Note: Outcome is high fat (top 30% PDFF) versus low fat (bottom 30% PDFF), n = 18 versus n = 18. This is an enriched exploratory contrast excluding the middle 40% of the obesity cohort. Training and 10‐fold cross‐validated AUCs are internal, non‐nested estimates and should not be interpreted as expected performance in an unselected obesity population. Discussion In this MASLD‐focused analysis of a deeply phenotyped obesity cohort with MRI‐PDFF, cEV surface protein profiling identified candidate signatures that distinguish MASLD from controls and capture variation in liver fat burden. Across both the 90‐marker discovery panel and the 12‐marker focused same‐cohort follow‐up panel, selected pairwise ratios showed higher internal discrimination estimates than single markers. These findings should be interpreted as discovery‐stage candidate prioritisation rather than validated classifier performance. Methodologically, this EV Array strategy differs from untargeted EV cargo, miRNA or bulk proteomic approaches because it interrogates surface‐exposed epitopes on captured cEV populations, enabling multiplexed immunophenotyping and ratio construction while not resolving intravesicular cargo or definitive cellular origin. The biology implicated by the top‐performing features is coherent with known pathways in metabolic liver injury. Lipid uptake and trafficking markers (CD36, FATP5, FATP1) (Garcia et al. 2026; Perumal et al. 2024; Povero et al. 2020) and the endocytic scavenger receptor LRP‐1 suggest altered handling of fatty acids and lipoprotein remnants (Hamlin et al. 2018). Stress and apoptosis markers (TRAIL‐ and Hsp70‐related ratios) are compatible with hepatocellular injury and cEV release under lipotoxic conditions (Hirsova et al. 2017; Di Naso et al. 2015; Angelini et al. 2022). Immune activation markers (TREM2, CD163) (Hou et al. 2021; Kawanaka et al. 2023; Skytthe et al. 2026) implicate macrophage‐related responses, and coagulation/adhesion pathways (CD280, ADAMTS13, CD171) point towards vascular and haemostatic perturbations that commonly accompany advanced metabolic dysfunction in obesity (Kuo et al. 2025; Geys et al. 2017; Potze et al. 2016; Boel et al. 2025). From a translational perspective, the PDFF‐extremes analysis within the obesity cohort provides a hypothesis‐generating enrichment model for identifying EV surface features associated with high liver fat burden. However, because the analysis excludes the middle 40% of the obesity cohort, the reported AUCs cannot be interpreted as performance estimates for routine clinical screening or for patients with intermediate liver fat. Exploratory benchmarking also showed strong performance of FLI, particularly in contrasts involving lean controls, highlighting that future studies must test whether EV signatures add mechanistic or clinical information beyond routine anthropometric and biochemical indices. Interpretation is further constrained by the sex imbalance and concomitant cardiometabolic severity gradient between the high‐ and low‐PDFF strata. Therefore, these EV patterns may reflect liver fat burden together with sex‐linked and broader metabolic‐severity effects, rather than PDFF‐specific or sex‐independent signatures. Robustness across nomenclatures is important for studies initiated under earlier diagnostic frameworks. Cohort‐level alignment across MASLD and legacy NAFLD+MetS groupings is summarised in Figure S3 and Tables S3–S6. The substantial overlap of top‐performing ratio features across frameworks supports the interpretation that the dominant cEV signatures are driven primarily by liver fat burden and cardiometabolic injury rather than by the diagnostic label itself. Conclusions cEV surface protein phenotyping nominates mechanistically coherent candidate signatures of MASLD in obesity. Across panels and contrasts, selected ratio features showed higher internal discrimination estimates than single markers, particularly in an exploratory PDFF‐extremes enrichment design within the obesity cohort. Because feature selection was not nested within cross‐validation and the ratio search was not permutation‐corrected, these results should be interpreted as uncorrected discovery‐stage observations. The findings support further validation of compact cEV ratio signatures, anchored by lipid uptake/scavenger pathways (CD36, FATP5, LRP‐1) and complemented by stress, immune and coagulation biology, in larger independent cohorts and longitudinal metabolic liver disease studies. Limitations This is a single‐centre discovery‐stage case‐control study with modest sample size, and performance metrics are based on internal, non‐nested cross‐validation. Accordingly, confidence intervals around AUC estimates are wide, and model rankings may change in independent cohorts. The candidate‐ratio search was high‐dimensional, feature selection was performed within the same cohort, and no permutation correction was applied to the ratio‐selection process; therefore, best‐ratio AUCs should be interpreted as uncorrected discovery maxima rather than validated classifier performance. EV Array provides relative rather than absolute quantification, and we did not evaluate multi‐centre analytical robustness, longitudinal stability, or lot‐to‐lot variability. Although the EV Array run/slide structure and CD9/CD81‐based normalisation between the two discovery runs are now reported in the Supporting Information Methods, no additional statistical batch‐correction term was included in the downstream classifier models; residual slide/run‐related technical variation therefore cannot be fully excluded. The focused 12‐marker panel was evaluated in the same participant cohort and therefore provides same‐cohort follow‐up, not independent validation. The PDFF‐extremes analysis deliberately compared the top and bottom 30% of liver fat within obesity and excluded the middle 40%; consequently, these AUCs do not estimate expected performance in an unselected obesity population. Sex distribution differed between groups, most prominently in the PDFF‐extremes analysis, where the low‐fat stratum included 3/15 M/F and the high‐fat stratum included 11/7 M/F. The high‐PDFF stratum also showed greater cardiometabolic severity. Therefore, the PDFF‐extremes EV signatures cannot be interpreted as independent of sex or broader metabolic severity, and larger cohorts will be required for sex‐stratified and covariate‐adjusted modelling. An important EV‐specific limitation relative to MISEV2023 (Welsh et al. 2024) is that orthogonal EV characterisation was not performed systematically across the full individual‐level study cohort because of limited sample material. Selected EV‐associated markers (including CD36 and FATP1) were examined by hFCM in three independent obesity EDTA platelet‐poor plasma pools that included pooled material with partial overlap to the EV Array clinical cohort, whereas the EV Array experiments used citrate plasma. The hFCM workflow included antibody/isotype controls, antibody‐mix filtration to reduce aggregates, detergent lysis controls, scatter and fluorescence calibration, compensation controls and MIFlowCyt/MIFlowCyt‐EV reporting. These orthogonal data support marker detectability/co‐expression and gating plausibility, but they do not constitute full participant‐level same‐sample validation of the EV Array‐derived classifiers or definitive EV subtype attribution. Finally, exploratory comparison with FIB‐4 and FLI showed strong FLI performance, underscoring that future work must evaluate incremental clinical value of cEV profiling beyond routine risk factors and established non‐invasive indices. Marker‐level imputed‐frequency summaries are now provided in Tables S12 and S13, but missingness/imputation was not formally modelled in the classifier pipeline; potential influence of non‐random failed triplicates on variance estimates and ROC ranking is therefore acknowledged as a technical limitation. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Supporting Information: jex270183‐sup‐0001‐SuppMat.docx Garcia, N. A. , Kløve‐Mogensen K., Cliff K. L., et al. 2026. “Extracellular Vesicle Surface Protein Profiling in Metabolic Dysfunction‐associated Steatotic Liver Disease in Obesity.” Journal of Extracellular Biology 5, no. 9: e70183. 10.1002/jex2.70183 N. A. Garcia and K. Kløve‐Mogensen shared authorship. Data Availability Statement The data that support the findings of this study are available in the supplementary material of this article References Associated Data Supplementary Materials Supporting Information: jex270183‐sup‐0001‐SuppMat.docx Data Availability Statement The data that support the findings of this study are available in the supplementary material of this article