Circulating extracellular vesicle-based multianalyte biomarker signatures accurately distinguish glioblastoma from brain metastasis patients before surgery

Circulating extracellular vesicle-based multianalyte biomarker signatures accurately distinguish glioblastoma from brain metastasis patients before surgery | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 14 April 2025 V1 Latest version Share on Circulating extracellular vesicle-based multianalyte biomarker signatures accurately distinguish glioblastoma from brain metastasis patients before surgery Authors : Susannah M. Hallal , Ágota Tűzesi , Abhishek Vijayan , Vineet Gorolay , Brindha Shivalingam , Hao-Wen Sim , Michael E. Buckland , Laveniya Satgunaseelan , Fatemeh Vafaee , and Kimberley Alexander 0000-0002-7239-039X [email protected] Authors Info & Affiliations https://doi.org/10.22541/au.174465021.16864881/v1 361 views 212 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Distinguishing glioblastoma (GBM) from secondary brain metastases (BMET) remains a common diagnostic challenge due to overlapping radiographic features and diverse clinical presentations. Non-invasive liquid biopsies offer a promising alternative to image-based diagnostics to guide neoadjuvant treatments and improve patient outcomes. Extracellular vesicles (EVs), membranous nanoparticles, serve as valuable biomarker reservoirs with unique properties, making them ideal for liquid biopsy development. Here, we describe a standardised method to isolate circulating-EVs from 1 mL plasma for tandem high-dimensional proteomic and transcriptomic profiling by data-independent mass spectrometry and next-generation-sequencing, respectively. In this pilot study, a machine-learning based biomarker discovery approach identified GBM- and BMET-specific biomarker signatures, comprised of a panel of 23 proteins and 4 microRNAs that could accurately distinguish GBM and BMET, achieving a high discriminatory power with area under the curves (AUCs) of 0.99 and 0.912. Notably, further analysis of the BMET specimens demonstrated that plasma-EV protein signatures can stratify BMET patients based on the primary tumour malignancy. The performance of the plasma-EV biomarkers illustrates the promising role of EV-based liquid biopsies for minimally invasive and highly accurate brain tumour diagnostics and stratification prior to surgery. However, further validation in larger independent cohorts is warranted to establish their clinical utility. Circulating extracellular vesicle-based biomarker signatures accurately distinguish glioblastoma from brain metastasis before surgery Susannah M. Hallal 1,2,3 , Ágota Tűzesi 2,3 , Abhishek Vijayan 4 , Vineet Gorolay 1,5,6 , Hao-Wen Sim 7,8,9,10 , Brindha Shivalingam 1 , Michael E. Buckland 2,3 , Laveniya Satgunaseelan 1,3,7 , Fatemeh Vafaee 4 and Kimberley L. Alexander 1,2,3 1. Brain Cancer Research, Neurosurgery Department, Chris O’Brien Lifehouse, Camperdown, New South Wales, Australia 2. School of Medical Sciences, Faculty of Medicine and Health Sciences, The University of Sydney, New South Wales, Australia 3. Department of Neuropathology, Royal Prince Alfred Hospital, Camperdown, New South Wales, Australia 4. School of Biotechnology and Biomolecular Sciences, The University of New South Wales, Kensington, New South Wales, Australia 5. Department of Radiology, Royal Prince Alfred Hospital, Camperdown, New South Wales, Australia 6. Department of Radiology, Royal North Shore Hospital, St Leonards, New South Wales, Australia 7. Sydney Medical School, Faculty of Medicine and Health Sciences, The University of Sydney, New South Wales, Australia 8. Department of Medical Oncology, Chris O’Brien Lifehouse, Camperdown, New South Wales, Australia 9. School of Clinical Medicine, Faculty of Medicine and Health, The University of New South Wales, Kensington, New South Wales, Australia 10. NHMRC Clinical Trials Centre, The University of Sydney, Camperdown, New South Wales, Australia Corresponding author: Kimberley L. Alexander PhD Brain Cancer Research, Chris O’Brien Lifehouse 119-143 Missenden Road Camperdown NSW 2050 PO BOX M33 Missenden Road NSW 2050 Email: [email protected] Abstract (max 200 words) Distinguishing glioblastoma (GBM) from secondary brain metastases (BMET) remains a common diagnostic challenge due to overlapping radiographic features and diverse clinical presentations. Non-invasive liquid biopsies offer a promising alternative to image-based diagnostics to guide neoadjuvant treatments and improve patient outcomes. Extracellular vesicles (EVs), membranous nanoparticles, serve as valuable biomarker reservoirs with unique properties, making them ideal for liquid biopsy development. Here, we describe a standardised method to isolate circulating-EVs from 1 mL plasma for tandem high-dimensional proteomic and transcriptomic profiling by data-independent mass spectrometry and next-generation-sequencing, respectively. In this pilot study, a machine-learning based biomarker discovery approach identified GBM- and BMET-specific biomarker signatures, comprised of a panel of 23 proteins and 4 microRNAs that could accurately distinguish GBM and BMET, achieving a high discriminatory power with area under the curves (AUCs) of 0.99 and 0.912. Notably, further analysis of the BMET specimens demonstrated that plasma-EV protein signatures can stratify BMET patients based on the primary tumour malignancy. The performance of the plasma-EV biomarkers illustrates the promising role of EV-based liquid biopsies for minimally invasive and highly accurate brain tumour diagnostics and stratification prior to surgery. However, further validation in larger independent cohorts is warranted to establish their clinical utility. 7.3 Statement of significance of the study Brain tissue biopsy is still the only reliable method to differentiate glioblastoma (GBM) from brain metastasis (BMET). This study introduces a strategised, blood-based approach to stratify GBM and BMET patients, potentially eliminating the need for invasive neurosurgery. By isolating extracellular vesicles (EVs), 30–1000 nm membranous nanoparticles, from pre-operative plasma, we leveraged advanced multi-omic profiling and machine learning-based feature selection to analyse their molecular cargo. Our approach identified highly specific protein and small RNA biomarker panels capable of accurately distinguishing GBM from BMET and, notably, stratifying BMET patients by their primary malignancy. These findings have significant clinical implications. A precise, non-invasive diagnostic method could allow some BMET patients to avoid neurosurgery and instead receive targeted, brain-penetrant therapies. Similarly, early identification of GBM could enable the safe trialling of neoadjuvant treatments, offering a potential breakthrough in a disease that has had stagnant survival outcomes for over 30 years. As EV-based liquid biopsies gain traction in oncology, this study highlights their immense potential for routine clinical integration, to reduce diagnostic delays, minimise unnecessary interventions, and improve patient outcomes. Introduction Brain metastasis (BMET), where tumours have spread to the brain from another organ, is a devastating complication that affects 15-30% of cancer patients [1, 2]. The majority of BMET tumours (75-90%) originate from cancers of the lung, breast, and cutaneous melanoma [2, 3]. Among these tumours, melanoma has the greatest propensity to metastasise to the brain; nearly 75% of melanoma patients die with evidence of intracranial disease [4]. In contrast, glioblastoma IDH-wildtype (GBM) is an aggressive primary brain tumour that rarely spreads beyond the central nervous system. Both BMET and GBM are often first identified using imaging techniques such as magnetic resonance imaging (MRI), however differentiating the two entities remains a frequent diagnostic challenge. Gadolinium-enhanced MRI is the current standard of care for detection of BMET, especially in patients with multiple brain lesions and a known history of cancer [5, 6]. However, one-third of BMET patients present with a solitary lesion in the brain and in many instances have no apparent systemic disease [7]. Similarly, GBM typically manifests as a solitary brain lesion and can be difficult to distinguish from BMET at presentation. Further exacerbating these challenges, up to 20% of GBM cases exhibit multifocal disease, which is a pattern that is more commonly associated with BMET [8]. GBMs can also develop in patients with a history of systemic malignancies [9], possibly as a consequence of prior cancer therapy. Misdiagnoses of these brain tumours have been documented to lead to the administration of inappropriate treatment [10], which further compromises outcomes for patients. Problematically, GBM and BMET share similar morphologic features on imaging, such as extensive oedema, contrast ring-enhancement and necrotic areas [11]. Multiple advancements in imaging modalities have been made to improve MRI diagnostic capability, enabling more accurate evaluations of the tumour microenvironment and improved differentiation of GBMs from BMETs [5]. Improvements to MRI include diffusion-based techniques [12], MR perfusion [13], MR spectroscopy [14], neurite orientation dispersion and density imaging [15] and automated machine learning [5, 16]. In particular, studies have demonstrated that perfusion MRI is a promising technique to discriminate GBM from BMET [17], yet predictive multi-parametric imaging models have reported conflicting features and have not permitted reliable differentiation between the two tumour entities [18-21]. A recent radiomics study that combined multiple classifiers from 321 MRI features in peri-enhancing oedema regions was also unable to reliably distinguish GBM from BMET patients [22]. In most instances, open neurosurgery is required for a definitive tissue-based diagnosis of brain tumours. However, many metastatic malignancies now have treatment options that can effectively target intracranial metastases without surgery. If an accurate, minimally invasive diagnostic method were available, a proportion of patients with BMET could avoid surgery altogether and be managed with upfront stereotactic radiosurgery, immunotherapy and/or targeted systemic therapy. Likewise, if GBM could be reliably diagnosed prior to surgery, neoadjuvant therapies could be trialled in a patient population that has seen no improvement in overall survival for more than 30 years. Liquid biopsies that measure tumour-derived factors in body fluids offer new and promising avenues for cancer detection and diagnosis, as well as for guiding treatments and implementing precision care models. It is, therefore, a research priority to develop novel liquid biopsy strategies that measure sensitive and specific tumour biomarkers. Extracellular vesicles (EVs) are sub-micron (30-1000 nm) nanoparticles that carry an array of selectively packaged molecules (i.e., proteins, nucleic acids, lipids, glycans and metabolites) that reflect the identity and molecular state of their cell-of-origin [23]. EVs readily cross physical membranes, e.g., the blood-brain-barrier, and are easily recovered from the blood circulation in large numbers, making them easily accessible for analysis [24, 25]. We and others, have explored EVs as reservoirs of GBM biomarkers in multiple bio-compartments and detected robust EV-associated molecular signatures from neurosurgical fluids [26, 27], cerebrospinal fluid (CSF) [28, 29], peripheral blood [30, 31] and recently, the urine [32, 33]. We have also shown that EV proteomes can distinguish different primary brain tumour types [31] and detect evolutionary changes in brain tumours associated with tumour progression and treatment resistance [26, 30, 31, 34-36]. As EVs are 10 times more abundant than circulating tumour cells [37] and comprise a lipid bilayer membrane that protects cargoed molecules from enzymatic degradation in the circulation, EVs hold great promise as superior biomarker reservoirs for the diagnosis of metastatic brain cancers. In this study, we aim to define diagnostic molecular profiles for brain tumours cargoed by circulating EVs. To achieve this, we standardised and strategised an automated, scalable size exclusion chromatography method for EV isolation that allows sufficient EV yields from just 1 mL of blood-plasma for multi-omics analysis. The plasma-EVs were isolated into six fractions 7-12, and for each sample, the first portion (fractions 7-9) was combined for quantitative proteomic analyses using data-independent acquisition mass spectrometry (DIA-MS), in conjunction with a custom glioma spectral library comprised of 8602 proteins and a targeted data-extraction strategy [31]. The second portion (fractions 10-12) was pooled for transcriptomic analyses by next-generation-sequencing of small RNA species. Our findings demonstrate that plasma-EV biomarkers hold remarkable diagnostic power for sensitively discerning patients with GBM and BMET prior to surgery. Additionally, plasma-EV protein profiles have a unique capacity to distinguish BMET patients by their primary malignancy, offering a high level of precision for liquid-biopsy-based brain tumour diagnostics. Nevertheless, as our sample cohorts are limited in size, the biomarkers identified in this study require further validation with large external cohorts of plasma from GBM and BMET patients to ensure their generalisability, reproducibility and clinical applicability. Materials and Methods: 2.1 Patient cohorts and plasma specimens Platelet-depleted (3000 x g ) plasma (1 mL) was accessed from the Sydney Brain Tumour Bank and Mark Hughes Foundation Biobank and analysed under The University of Sydney Human Research Ethics Committee protocol 2019/705. Plasma collected from a total of 26 GBM and 21 BMET patients prior to the surgical removal of their tumours was compared with plasma from 21 healthy controls (HC), who reported no significant history of cancer. Participant demographics and the integrated diagnosis and neuropathological assessments of excised GBM and BMET tumours are summarised in Table 1; refer to Supplementary Table 1 for further detail. Table 1. Overview of Experimental Cohorts Glioblastoma, IDH-wildtype (GBM) 26 64.9 (41.9 – 76.5) 18M, 8F Brain Metastasis (BMET) Metastatic melanoma (MMel) OTHER metastatic cancer (OTHER) 21 10 11 61.8 (33.1 – 82.4) 61.1 (33.1 - 82.4) 62.5 (36.1 – 81.4) 12M, 9F 7M, 3F 5M, 6F Healthy participants, non-cancer controls (HC) 21 57.8 (35.3-72) 12M, 9F 2.2 EV isolation, fractionation and characterisation Plasma was slowly thawed on ice, and platelets were depleted by centrifugation (3000 x g, 20 min, 4 °C). EVs were purified from plasma using iZON qEV original size exclusion chromatography columns as per manufacturer’s instructions with an Automated Fraction Collector V1 (AFC, iZON Science). Briefly, the AFC was set to collect 6 x 500 µL EV fractions from 500 µL plasma using qEV original columns and a default void volume of 2.85 mL. The qEV original columns were flushed with 15 mL PBS before loading 500 µL of platelet-free plasma onto the column. The void volume was discarded (F1-6; 2.85 mL) and 6 x 500 µL EV fractions were eluted in PBS (F7-12). For each sample, two sequential EV isolations were performed on the same column (total 1 mL plasma) with 30 mL PBS washes in between to circumvent signal carryover. The plasma-EV fractions were stored at −80 °C. We previously reported that Fractions 7-9 were the least contaminated with high abundance serum proteins and fractions 10-12 contained the highest numbers of EV particles [31], which is important for enhanced mass spectrometry analysis and isolating sufficient small RNA (sRNA) yields, respectively. EV fractions were therefore combined into two distinct pools and strategised for biomarker discovery by complementary ‘omic platforms, i.e., Fractions 7-9 for comprehensive proteomics (isolated from 500 μL plasma) and Fractions 10-12 for small RNA sequencing (isolated from 1 mL plasma). The workflow of the complementary ‘omics strategy is summarised in Figure 1. The size distributions and concentrations of the plasma-EV fractions were measured by nanoparticle tracking analysis (NTA) software (version 3.0) in triplicate using the NanoSight LM10-HS (NanoSight Ltd., Amesbury, UK), configured with a tuned 488 nm laser and a digital camera system (sCMOS trigger camera). EVs were diluted with sterile-filtered PBS (viscosity 1.09 cP) to ensure 20–100 particles were detectable within the field of view of the standard CCD camera of the microscope at a camera level of 11. The NTA software captured triplicate 60 s video recordings of the EVs at 25 frames per second with default minimal expected particle size, minimum track length, blur setting, with the temperature of the laser unit controlled to 25 °C. The videos were analysed by NTA3.3, which translates the Brownian motion and light scatter properties of the individual laser-illuminated particles into a size distribution (ranging from 10 to 1000 nm) and concentration (particles per mL) while simultaneously calculating their diameter using statistical methods [34]. The EV size distributions and concentrations were analysed in Microsoft Excel ® . EV samples were imaged by Cryo-electron Transmission Microscopy (Cyro-TEM) as described before [35]. Briefly, the EV samples were applied to copper 300-mesh lacey carbon grids and plunge frozen into ethane using a Vitrobot IV System (ThermoFisher Scientific). The grids were imaged at 45,000 x magnification using SerialEM 4.0 software [38] (Mastronarde, The Boulder Laboratory) on a Thermo Scientific TM Glacios Cryogenic Transmission Electron Microscope. The microscope was operated at 200 kV and equipped with a Falcon III direct electron detector (Thermo Scientific TM ). The cryo-TEM images were edited, and scale bars were generated in ImageJ 1.53K software. 2.3 EV proteome preparation and MS analysis Plasma-EV proteomes were extracted from Fractions 7-9, prepared for LC-MS/MS analysis and spiked with heavy-labelled PepCalMix peptides using previously established methods [26]. Briefly, desalted peptide (2 μg) mixtures prepared from GBM Cohort A, BMET and HC patient plasma-EVs (Figure 2A, Supplementary Table 1) were analysed by a TripleTOF®6600 Quadrupole Time-Of-Flight mass analyser (AB Sciex) coupled to an Eskpert TM NanoLC 425 autosampler system operating in SWATH/data-independent acquisition mode (Sydney Mass Spectrometry Facility) [31]. Desalted peptides (300 ng) prepared from GBM Cohort B patient plasma-EVs (Figure 2A, Supplementary Table 1) were analysed by an Orbitrap Eclipse Tribid mass spectrometer (Thermo Scientific, MA, USA) coupled to an Ultimate 3000 UHPLC and autosampler system (Dionex, Amsterdam, Netherlands) in DIA-mode. For both mass spectrometers, the peptide specimens were injected onto an in-house 15 cm C 18 reversed-phase column (75 µm diameter and 5 µm particle size). The HPLC solvent system was comprised of buffer A: 0.1% ( v/v ) FA (Thermo Scientific, Cat.No. 85178) and buffer B: 80% ( v/v ) ACN (Thermo OPTIMA LC/MS grade, Cat.No. 34851-4), 0.1% ( v/v ) FA. As the two sample batches were analysed on different mass spectrometry platforms, separate DIA-methods were required to maximise peptide coverage using time-of-flight (TOF; TripleTOF 6600) and orbitrap (Orbitrap Eclipse Tribid) instruments. The SWATH-method on the TripleTOF6600 covered a total of 159 custom, variable-sized windows (with a 1.0 Da window overlap) over a precursor mass range of 350–1750 m/z (Supplementary Table 2). Peptides were eluted over a 120 min gradient (2-35% B for 80 min, 35-95% B for 19 min, 95% B for 5 min, and 2%B for 16 min). The DIA-method on the Orbitrap Eclipse Tribid covered a total of 23 variably-sized windows (including 1.0 Da window overlap) over a precursor mass range of 350–1650 m/z (Supplementary Table 2), and peptides were eluted over a 140 min linear gradient of 5% B to 42% B with a constant flow rate of 300 nL min -1 . To ensure the repeatability and reproducibility of our MS quantitation approach, four repeat injections of a plasma-EV sample was performed across the mass spectrometry acquisition in SWATH-mode on the TripleTOF®6600 (Int Ctrl, Supplementary Figure 1) and in DIA-mode on the Orbitrap Eclipse Tribid (SB, Supplementary Figure 2). The protein normalised abundances were plotted as scatter plots for all replicate pairs and the Pearson’s correlation coefficient (r2) was calculated in GraphPad Prism 10.0.0. Stable, reproducible, and repeatable MS/MS quantitation was observed between injections on the TripleTOF®6600 (Supplementary Figure 1, r2 > 0.9697) and the Orbitrap Eclipse Tribid (Supplementary Figure 2, r2>0.9218). 2.4 Targeted data extraction of plasma-EV proteins Chromatographic peaks for the plasma-EV samples were extracted by aligning the SWATH-MS and DIA-MS acquisitions to a comprehensive glioma spectral library comprised of 8602 proteins as described before [31] in Skyline 22.1 (Maccoss Labs); the SWATH-MS and DIA-MS data acquisitions were extracted in separate Skyline documents. Firstly, the spectral library was imported to Skyline 22.1 (Maccoss Lab, University of Washington, Seattle, WA, USA) and only peptides with a q-value < 0.05 were included in the target list from the 8602 proteins for extraction of SWATH/DIA-MS data. Before importing the SWATH-MS and DIA-MS data to Skyline, an isolation scheme was created according to the variable isolation windows that were used to acquire SWATH-MS data on the TripleTOF ® 6600 and DIA-MS data on the Orbitrap Eclipse Tribid Mass Spectrometer (Supplementary Table 2). The skyline peptide and transition settings were prepared for data analysis as reported before [31]. The raw SWATH-MS (.wiff) and DIA-MS (.raw) acquisitions of plasma-EV samples were imported to Skyline and aligned to the spectral library using an indexed retention time (iRT) calculator that was calibrated to the retention times of the heavy-labelled PepCalMix peptides that were spiked-in to both the spectral library and peptide samples. The MS peaks were used if their retention time was within a 5 min window of their predicted retention time. The mProphet algorithm was applied to the MS data to determine the confidence that a peak corresponds to its targeted peptide. Peaks with a q-value < 0.05 were extracted and integrated for analysis if at least half of the transitions contributed to a precursor peak). The plasma-EV proteomic data was exported from Skyline as MSstats input files and processed with MSstats R package [39]. The process involved conversion to a quantitative data format, log2 transformation, and filtering out proteins with less than 50% missing values across all samples, followed by imputation of missing values using the missForest R package [40]. Data captured by separate mass spectrometers were combined using the 4117 common proteins, followed by quantile normalisation and batch correction using the ComBat function from sva R package as described in section 2.6 [41]. 2.5 EV small RNA transcriptome preparation and analysis Plasma-EV Fractions 10-12, isolated from GBM Cohort A, GBM Cohort B, BMET and HC specimens, were pooled for next-generation sRNA sequencing. Plasma-EV fractions 10-12 were treated with RNase A (37 °C for 10 min; 100 ng/ml; Qiagen, Australia) and processed for RNA extraction using the Plasma/Serum Circulating & Exosomal RNA Purification Mini Kit (Norgen Biotek, Cat. 51000) according to the manufacturer’s protocol. Extracted total RNA samples were analyzed with a Eukaryote Total RNA chip on an Agilent 2100 Bioanalyser (Agilent Technologies, United States) to confirm sufficient yield, quality and size of RNA. EV RNA sequencing libraries were then constructed using the QIAseq miRNA Library Kit (Qiagen) according to the manufacturer’s instructions. The yield and size of the resultant cDNA libraries were validated using a high-sensitivity DNA Assay on an Agilent 2100 Bioanalyser (Agilent Technologies). The cDNA libraries were then pooled at equal proportions for multiplexed sequencing on an Illumina NovaSeq6000 System at the Ramaciotti Centre for Genomics. The libraries were sequenced in a single-end 75 bp, high-output protocol to generate an output of up to 20 million reads-per-sample, with a quality control of >80% bases higher than Q30 at 1x100 bp. The demultiplexed FASTQ files of the sRNA transcript data were analysed using the Qiagen RNA Seq Portal. The sRNA transcripts were mapped to the sRNA human databases with default settings and Unique Molecular Index (UMI) counts of microRNA (miRNA) and piwi-interacting RNA (piRNA) species were generated. A filtering process was performed, retaining 907 transcripts that had less than 90% missing values across all samples. The transcripts were further filtered with the filterByExpr function and normalised by log2 transformation of counts per million (CPM) with the edgeR package [42], followed by batch correction with the ComBat function from sva R package [41]. 2.6 Batch effect correction of proteomic and transcriptomic data To identify protein and sRNA biomarkers for GBM and BMET, three comparisons were conducted; GBM vs. BMET, GBM vs. HC, and BMET vs. HC. The GBM cohort comprised of GBM cohorts A and B, and batch effects were evident across the two cohorts for both proteomic (Supplementary Figure 4) and transcriptomic data (Supplementary Figure 5) as shown by UMAP plots. To control for these cohort effects, batch correction was applied to the normalised protein and sRNA data using ComBat for the GBM vs. BMET and GBM vs. HC comparisons. However, batch correction was not applied to the BMET vs. HC comparison, as all samples originated from the same batch (Supplementary Table 1). The quantile-normalised, batch-corrected protein data is provided for GBM vs. BMET (Supplementary Table 4), GBM vs. HC (Supplementary Table 5) and BMET vs. HC (Supplementary Table 6). The log2CPM normalised, batch-corrected sRNA data is provided for GBM vs. BMET (Supplementary Table 7), GBM vs. HC (Supplementary Table 8) and BMET vs. HC (Supplementary Table 9). 2.7 Machine learning-based biomarker discovery approach Significant putative protein and sRNA biomarkers were identified through three comparative analyses: GBM vs. BMET, GBM vs. HC, and BMET vs. HC. Prior to analysis, data processing and batch correction were performed as described in Section 2.6. Processed data for each comparison was analysed using the FEMPipeline [43], a comprehensive machine learning (ML)-based framework for biomarker discovery. Briefly, this pipeline employs six repeats of 5-fold cross-validation, resulting in a total of 30 train-test splits. In each iteration, 80% of the data (4-folds) were used for training and 20% (1-fold) for testing, i.e. 80:20 train:test split. Across each iteration, the training data was subject to multiple feature selection methods, followed by training several classification models (Table 2), and the model performance was assessed using Area Under ROC Curve (AUC) on the test data (Supplementary Figure 7). Feature selection methods that resulted in the highest Mean AUC across the 30 different train-test splits were identified. Across these methods, the features that were common in at least 27, 28, 29 or 30 iterations were considered as ‘potential biomarker sets’. The omics data containing only these candidate biomarkers were re-analysed through the pipeline, and the most optimal biomarker set that had the highest Mean AUC were selected. The biomarker models that produced the highest AUCs are listed in Supplementary Table 10. Overall, this pipeline is designed to identify biomarkers that are strongly associated with clinical outcomes, demonstrate robustness to subsampling, and remain consistent across various feature selection methods [43]. The final biomarker panel was selected based on the highest Mean AUC, ensuring optimal predictive performance for stratifying clinical conditions while being independent of any specific classification threshold (i.e., the probability cutoff used to assign class labels), thereby providing a more objective and generalisable measure of model accuracy. Table 2. Summary of Feature Selection Methods and Classification Models Used in the FEMPipeline . The listed feature selection methods were used to identify features common in at least 27, 28, 29 or 30 iterations, and their performance was evaluated with classification models optimised for the FEMPipeline [43] . Feature Selection Methods No feature selection: all features included T-test Wilcoxon Test Ranger: selecting positive feature importance with impurity_corrected option MRMR (Minimum Redundancy Maximum Relevance) mRMR30: selecting 30 features mRMR50: selecting 50 features mRMR75: selecting 75 features mRMR100: selecting 100 features mRMR_perc50: selecting 50% of all features RF-RFE (Random Forest Recursive Feature Elimination) GA-RF (Genetic Algorithm with Random Forest optimisation) Classification Models Simple Logistic Regression L1-Regularised Logistic Regression L2-Regularised Logistic Regression Elastic Net Logistic Regression SVM with Sigmoid Kernel SVM with Radial Basis Function (RBF) Kernel Random Forest 2.8 Integrated differential expression and machine learning approach for multi-criteria biomarker discovery While the ML-based approach outlined in section 2.7 is designed to effectively identify robust and predictive biomarkers, we further introduced a multi-criteria selection strategy to ensure that the magnitude of abundance change is sufficiently large for clinical validation. By combining differential expression (DE) metrics with machine learning-based feature importance and stability, this approach prioritises biomarkers with statistical significance. To identify robust biomarker candidates, we applied this integrated approach to both transcriptomic and proteomic data across GBM, BMET, and HC cohorts. The workflow of the multi-criteria biomarker discovery approach is summarised in Figure 1. 2.8.1 Data Preprocessing and Normalisation For transcriptomic data, we applied the filterByExpr function to remove lowly expressed transcripts, followed by log-transformation of CPM. Proteomic data were normalised using quantile normalisation. All samples were jointly corrected for batch effects attributable to sample cohort differences using the ComBat method. 2.8.2. Cross-Validation Framework To ensure robustness and generalisability of biomarker selection, we employed six repeats of five-fold cross-validation, resulting in a total of 30 unique train-test splits (with an 80:20 train:test ratio). The following analytical steps were performed for each of these 30 subsets: 1) Differential Expression Analysis: DE analysis was performed on the training subset using the limma R package to compute log fold changes (logFC) and associated p-values for each protein/transcript. 2) Machine Learning (ML)-Based Feature Importance and Stability: The mean decrease in Gini Index was calculated using a Random Forest classifier trained on the same training data. Stability was assessed using the ranger() function from the ranger R package with the argument importance = ”impurity_corrected”. A protein or transcript was selected by this method if its variable importance score exceeded zero. 3) Metric Aggregation Across Splits: For each feature, the following metrics were computed across the 30 train-test iterations: • Mean logFC • Fisher’s combined p-value • Mean ML importance (mean of the Gini Index decrease) • Stability score (frequency of selection by Ranger) Additionally, features that had previously been identified as candidate biomarkers through ML-based approaches were flagged accordingly. 2.8.3 Biomarker Scoring and Ranking All candidate features were ranked within each metric. A combined score was then computed for each protein/transcript to prioritise features, with added weight for those previously identified as biomarkers. If previously identified as a biomarker: Combined score = rank(meanLogFC) + rank(Fisher’s p-value) + rank(ML importance) + rank(stability) + (2 × n) Here, n denotes the number of proteins or transcripts in the dataset and is assigned to markers that were also that were also identified in the prior ML-based approach (Section 2.7). If not previously identified: Combined score = rank(meanLogFC) + rank(Fisher’s p-value) + rank(ML importance) + rank(stability) 2.8.4 Recursive Feature Elimination We applied Recursive Feature Elimination (RFE) to refine the biomarker panel while maintaining predictive power and biological interpretability. For proteomics, the RFE process began with proteins in the top 25 th percentile of the combined score distribution due to the large feature set. For transcriptomics, RFE was applied to all features remaining after initial filtering. In each iteration, Random Forest models were used to evaluate performance on the 30 train-test splits. Features were removed only if their exclusion did not reduce the Mean AUC by more than a defined cutoff (0.01, 0.03, or 0.05). Furthermore, the final set was constrained to retain a maximum of 25 biomarkers and ensure that at least 25% had positive logFC values. The optimal biomarker panel was selected as the one achieving the highest Mean AUC across all RFE iterations and cutoff values while also meeting the minimum size requirement. An overview of the experimental workflow is shown in Figure 1. 2.9 Data visualisation and exploratory analysis The proteomes of six individual EV were compared using FunRich 3.0. Heatmaps were created using ComplexHeatmap R Package, and all other plots were created using ggplot R package. Area Under the Receiver Operating Curves (ROC) and Simple Linear Regressions were performed in GraphPad Prism (10.0.0). Orange Data Mining Software (3.36.2) was used to perform Principal Component Analyses (PCA), Linear Projections, and testing and scoring of highly ranked protein and transcript markers for discriminating BMET by primary tumour malignancy and site-of-origin. 2.10 Data availability The GBM spectral library is available in PeptideAtlas with the identifier PASS01487. The filtered, normalised and batch-corrected proteomic and transcriptomic data are provided in the Supplementary Tables 4-9. The RAW LC-MS/MS proteomic data has been deposited to the ProteomeXChange Consortium via the PRIDE partner repository with the data set identifier PXD062759. The raw transcriptomic data are accessible at NCBI Gene Expression Omnibus (GEO submission: GSE294371) 3. Results: 3.1 Characterisation of plasma-EVs for Optimised Proteomic and Transcriptomic Analyses Plasma-EVs were isolated by size exclusion chromatography and characterised by NTA, cryo-TEM and SWATH-MS in accordance with the minimum information for EV studies [44]. Using 500 µL plasma (n=4), six sequential 500 µL PBS fractions (F7-12) were collected. NTA quantification revealed a progressive increase in EV particle numbers across fractions, with 7.5x10 9 ± 3.95x10 8 EV particles eluting in the earliest fraction (F7) to 1.42x10 11 ± 7.65x10 9 EV particles in the later F12 fraction (Supplementary Figure 3A-1). All EV fractions displayed similar size distributions, with an average EV modal size of 101.2 ± 4.2 nm (Supplementary Figure 3A-2). Our established SWATH-MS approach in conjunction with targeted data extraction using a custom glioma spectral library comprised of 8662-proteins, was used for in-depth proteomic characterisation of each of the six individual plasma-EV fractions F7-12. This method identified a total of 3858 proteins across all six plasma-EV fractions (Supplementary Table 3). Unlike the EV particle numbers, the number of protein IDs increased sequentially across F7, F8 and F9, with 2732, 3098 and 3684 proteins, respectively, and then stabilised in F10, F11, F12 at 3675, 3639 and 3605 proteins, respectively (Supplementary Figure 3B). Among these, 2136 proteins, representing 55.3% of all identified plasma-EV proteins, were consistently present across all fractions (Supplementary Figure 3D). Notably, F7 exhibited the most distinct proteomic profile, with a lower overlap in protein identifications with all the other EV fractions (60.0-70.8%). In contrast, F9-F12 had the greatest similarity in their protein profile compositions, sharing over 90% of their protein identifications (Supplementary Figure 3D). The proteins identified in each of the 6 fractions are listed in Supplementary Table 3. To establish a strategy that uses a single plasma sample for tandem EV proteomic and transcriptomic analyses, EV fractions were surveyed for both EV-specific proteins [35] and common, high abundance serum protein contaminants by LC-MS/MS. The proteomes of all plasma-EV fractions 7-12 contained 76%-83% of the top-100 EV proteins as reported by Vesiclepedia [35]; the top-100 EV proteins are listed in Supplementary Table 3. These included canonical EV-marker proteins, such as CD9, CD63, PDCD6IP, ITGβ1 and HSP90AB1 which were present in all fractions (Supplementary Figure 3C-1). CD9 levels were highest in F7, while CD63 levels were highest in F10 and PDCD6IP levels were highest in F12. In stark contrast, common serum proteins such α-2 macroglobulin (A2M), albumin (ALB), apolipoprotein B-100 (APOB) and haptoglobin (HP) were found at markedly lower levels in early eluting fractions F7-9, compared to later eluting fractions F10-12 (Supplementary Figure 3C-2). To enhance proteomic and transcriptomic analyses, the early eluting plasma-EV fractions (F7-9), characterised by lower serum protein contamination and distinct proteomic profiles, were pooled for mass spectrometry-based proteomic analysis. The later eluting plasma-EV fractions, F10-12, which contained higher plasma-EV particle counts, were pooled to ensure sufficient RNA yields (≥ 1 ng) for sRNA next-generation-sequencing (Figure 1). Cryo-TEM imaging confirmed that both pools of EV fractions displayed similar vesicular morphologies, with heterogenous, multi-layered EV structures measuring 100-200 nm in size (Supplementary Figure 3E-1 and 3E-2). Performing this SEC-based fractionation twice (i.e., two sets of 500 µL plasma per patient using the same qEV original column) isolated plentiful yields of relatively pure plasma-EVs from 1 ml of blood-plasma for high-dimensional profiling of both the plasma-EV proteome and transcriptome. 3.2 Multi-omics Analysis of Plasma-EVs Our established proteomics and sRNA sequencing pipelines were applied to plasma-EV fractions F7-9 and F10-12, respectively, from GBM patients (n=26, Cohort A+B), BMET patients (n=21), and HCs (n=21) (Figure 2A). Using quantitative mass spectrometry and sRNA next-generation-sequencing, we confidently identified 4117 proteins and 907 transcripts across all groups for interrogative biomarker discovery (Figure 1). Protein identifications showed substantial overlap across the specimen cohorts (Figure 2B), as well as transcript identifications (Figure 2C). A total of 4092 proteins (99.4% of identified proteins) were shared across BMET, HC and GBM cohorts A and B (Figure 2B). Similarly, 849 sRNA transcripts (93.6% of identified transcripts) were common to GBM, BMET and HC specimens (Figure 2C). 3.3 Candidate plasma-EV biomarkers for distinguishing GBM and BMET patients pre-operatively To evaluate differences in the molecular profiles of GBM and BMET plasma-EVs, we performed pairwise comparisons of the shared proteins (Figure 2B) and transcripts (Figure 2C) between pre-operative GBM and BMET patient cohorts. DE analysis revealed 95 proteins with significantly elevated levels in GBM relative to BMET (fold-change, FC>|1.5|, p |1.5|, p <0.05). A volcano plot (Supplementary Figure 6A-1) illustrates these significant differences, including the top-10 proteins higher in GBM plasma-EVs (CISD1, EIF2A, EIF3I, H1-10, MESD, PRKAA2, RBM15, RPL15, TMEM259, and TTC17) and in BMET plasma-EVs (ACTL6A, ARFIP2, BRAT1, FIP1L1, GYPC, MT3, PAF1, RAB24, STX7 and ZWINT). Similarly, a pairwise comparison of the sRNA transcripts revealed 86 DE transcripts between GBM and BMET cohorts. Among these, seven transcripts had significantly higher levels in GBM relative to BMET (fold-change > 1.2, p 1.2, p miR-126-3p, miR-23a-3p, miR-223-3p, miR-199a-3p, miR-221-3p, miR-27a-3p, miR-146a-5p, miR-222-3p, piR-28764 and piR-28096, indicated in a volcano plot (Supplementary Figure 6A-2). The cross-validation biomarker discovery workflow was then implemented (Figure 1). Cross validation was repeated six times, generating 30 train-test data subsets (80% train, 20% test). For each data subset, DE analysis and feature selection were performed on the training data, and feature ranking by importance and RFE were applied to identify the best-performing markers (Figure 1). This process consistently selected 23 proteins, ACTL6A, APOC3, ARFIP2, BRAT1, CISD1, EIF3I, GPRIN1, MRE11, MT3, NIBAN2, PAF1, PLEKHO2, PRRC2C, RBM15, RNASEH2A, RNPEP, RPL10, RPL15, STBD1, STX6, TMEM259, UROS and ZWINT, as the strongest discriminators for distinguishing GBM from BMET, with a mean AUC of 0.99 (Figure 3A; Supplementary Table 11). Similarly, four miRNA species, let-7f-5p, miR-27b-3p, miR-182-5p and miR-190a-5p, were also identified as robust discriminators between GBM and BMET, achieving an average AUC of 0.912 (Figure 3A; Supplementary Table 12). The expression levels of these candidate protein (Figure 3B-1) and transcript (Figure 3C-1) biomarkers across GBM and BMET cohorts are visualised as box plots. The biomarkers with notably higher median expression in GBM plasma-EVs include STX6, RNPEP, RNASEH2A, GPRIN1, TMEM259, PLEKHO2, RPL15, EIF3I, RBM15, CISD1 (Figure 3B-1) and, let-7f-5p and miR-182-5p (Figure 3C-1), while those with higher expression in BMET include APOC3, ZWINT, UROS, PRRC2C, ACTL6A, BRAT1, PAF1, NIBAN2, RPL10, MRE11, MT3, ARFIP2 (Figure 3B-1), miR-27b-3p and miR-190a-5p (Figure 3C-1). UMAP plots based on the expression levels of these top 23 proteins (Figure 3B-3) and 4 miRNAs (Figure 3C-3) showed an improved separation of GBM and BMET samples compared to all plasma-EV proteins (Figure 3B-2) and transcripts (Figure 3C-2). These features were then cross-referenced with biomarkers identified by the ML-based approach to determine the most optimal panels for distinguishing GBM from BMET preoperatively. Using Ranger feature selection, the ML-based approach identified 9 proteomic biomarkers (mean AUC = 0.9622, Supplementary Figure 7A-1) and 12 transcriptomic biomarkers (mean AUC = 0.9043, Supplementary Figure 7A-2) that had high performance with a Random Forest classification model (see Supplementary Table 10 for biomarker lists). Plasma-EV biomarkers identified in both the ML-based approach and cross-validation DE analysis biomarker discovery workflow included NIBAN2, MT3, RBM15, APOC3, ACTL6A, GPRIN1, PRRC2C, RPL10, RNPEP, miR-27b-3p, miR-182-5p and miR-190a-5p (Figure 3A). 3.4 Putative diagnostic plasma-EV markers for GBM To identify putative plasma-EV biomarkers specific to GBM, a similar comparative and biomarker discovery analysis was performed against HCs. DE analysis found 210 proteins with significantly higher levels and 219 proteins with significantly reduced levels in GBM plasma-EVs relative to HCs (fold-change > 1.5, p <0.05, Supplementary Figure 6B-1). The top-10 proteins higher in GBM included CLIP3, CRABP1, EEF1E1, FRY, NUCB2, PCOLCE, RECQL, RENBP, RPS19 and STAT6, while the top-10 proteins with lower levels in GBM included ANKRD52, APOC2, ATP5F1D, C1QTNF4, IRF2BPL, LIN7C, MLH1, NAA30, SSR1 and UBE2K (Supplementary Figure 6B-1). Similarly, we found higher miR-190a-5p levels in GBM plasma-EVs relative to HCs, alongside 33 transcripts with significantly lower levels (fold-change > 1.2, p <0.05), with the top-10 lower transcripts including miR-342-3p, miR-10a-5p, miR-99b-5p, miR-126-5p, miR-125a-5p, miR-150-5p, miR-30a-5p, miR-10b-5p, miR-483-5p and miR-4433b-5p (Supplementary Figure 6B-2). As before, we applied our biomarker discovery approach to distil these findings into a concise panel of candidate biomarkers for GBM (Figure 1). RFE and feature ranking across each of the 30 train-test data subsets in a 5-fold cross-validation identified 23 protein markers for GBM, including ATP5F1D, CRABP1, APOC3, DEGS1, YWHAQ, PRRC2C, RPS19, CLIP3, RENBP, PON1, SDF4, NF2, PHB1, UBE2K, TRNT1, MLH1, FAF1, NUCB2, AEBP1, FAAH, STAT6, EIF4A3 and PTBP1. These 23 protein markers had a mean AUC of 0.978 for distinguishing GBM from HCs (Figure 4A; Supplementary Table 11). In addition, our biomarker discovery approach identified 11 transcripts, miR-342-3p, miR-10a-5p, miR-27b-3p, miR-190a-5p, miR-150-5p, miR-24-3p, miR-144-5p, let-7g-5p, let-7f-5p, miR-103a-3p and piR-28533; together, these 11 transcripts demonstrated a remarkable, near-perfect capacity for discriminating GBM patients from HCs with a mean AUC of 0.998 (Figure 4A; Supplementary Table 12). Our ML-based method also determined proteomic and transcriptomic biomarkers with strong discriminative power to distinguish GBM from HCs. The Ranger feature selection method identified a set of 6 proteomic biomarkers that had a mean AUC of 0.9815 when evaluated by a Random Forest model (Supplementary Figure 7B-1), while a total of 17 transcriptomic biomarkers were commonly selected by Ranger and mRMR_perc50 feature selection methods and had a mean AUC of 0.9833 with a Radial Kernal SVM model (Supplementary Figure 7B-2). The biomarkers corresponding to the highest performing models by AUC are listed in Supplementary Table 10. Numerous GBM plasma-EV candidate biomarkers determined by the biomarker discovery workflow overlapped with these ML-based biomarkers; ATP5F1D, CRABP1, APOC3, DEGS1, YWHAQ, PRRC2C, miR-342-3p, miR-10a-5p, miR-27b-3p and miR-190a-5p (Table 4A). Interestingly, GBM plasma-EVs exhibited consistently higher median expression levels for proteins, PON1, STAT6, NUGB2, CLIP3, TRNT1, RENBP, CRABP1 and RSP9 (Figure 4B-1), and while marginally higher median expression was observed for the sRNAs, let-7f-5p, let-7g-5p, miR-103a-3p, miR-144-5p and miR-190a-5p (Figure 4C-1), relative to HCs. In contrast, lower median expression levels in GBM plasma-EVs were found for APOC3, AEBP1, FAF1, EIF4A3, NF2, MLH1, PRRC2C, SDF4, YWHAQ, UBE2K, PHB1, PTB1, FAAH, DEGS1, ATP5F1D, miR-342-3p, miR-150-5p, miR-10a-5p, miR-24-3p and miR-27b-3p (Figure 4B-1, Figure 4C-1). UMAP dimensionality reduction of the selected protein (Figure 4B-3) and transcript (Figure 4C-3) biomarkers shows improved separation of GBM and HC patient specimens compared to the 4117 proteins (Figure 4B-2) and 907 transcripts (Figure 4C-2). 3.5 Putative diagnostic plasma-EV markers for BMET The same analyses were applied to determine BMET-specific biomarkers. DE analysis identified 193 proteins with significantly higher levels in BMET (fold-change > 1.5, p <0.05, Supplementary Figure 5C-1) and 164 proteins with significantly lower levels in BMET (fold-change < 1.5, p <0.05, Supplementary Figure 6C-1) relative to HC. The top-10 elevated proteins in BMET included CLSTN3, CRABP1, CYP51A1, DDX23, FRY, MRE11, OGA, PCOLCE, PISD, RECQL, while the top-10 reduced proteins in BMET included ATP5F1D, CD47, EHD2, EPB41L5, GABRB3, KARS1, MMP25, PAG1, TMEM259, UBE2K (Supplementary Figure 6C-1). Similarly, 37 transcripts exhibited significantly higher levels (fold-change > 1.2, p BMET relative to HC (fold-change < 1.2, p miR-20a-5p, miR-378a-3p, miR-1246, miR-223-5p, miR-1290, piR-32833, piR-23446, piR-32994, piR-28764, and the top-10 lower transcripts include let-7b-5p, let-7c-5p, miR-92a-3p, miR-483-5p, miR-183-5p, miR-432-5p, miR-342-3p, miR-382-5p, miR-30a-3p and piR-32194 (Supplementary Figure 6C-2). Our systematic biomarker discovery process (Figure 1) identified 19 proteins (DNAJC8, CAVIN1, GADD45GIP1, ACAT1, ATP1B1, ATP5F1D, FCN3, LRRC57, RPL14, PISD, DCTN3, HP, ATP11C, DDX23, OGA, SEC24D, TMEM165, COPS6 and NIBAN2) as robust diagnostic markers for BMET with an impressive mean AUC of 0.982 (Figure 5A; Supplementary Table 11). Similarly, 11 transcripts (miR-223-3p, miR-20a-5p, let-7b-5p, miR-142-3p, miR-378a-3p, miR-1246, miR-342-3p, miR-432-5p, miR-122-5p, miR-196b-5p and miR-126-5p) demonstrated near-perfect discriminatory power with mean AUC of 0.996 (Supplementary Table 12). Box plots highlight the expression differences of these markers in BMET and HC specimens (Figure 5B-1, 5C-1). Markers with elevated levels in the plasma-EVs of BMET patients include HP, GADD45GP1, LRRC57, DCTN3, DNAJC8, TMEM165, NIBAN2, PISD, ATP11C, OGA, DDX23, miR-142-3p, miR-223-3p, miR-1246, miR-20a-5p and miR-378a-3p (Figure 5B-1, Figure 3C-1), while markers with lower levels include CAVIN1, FCN3, RPL14, SEC24D, ACAT1, ATP5F1D, ATP1B1 and COPS6, let-7b-5p, miR-122-5p, miR-342-3p and miR-432-5p (Figure 5B-1, Figure 5C-1). Dimensionality reduction via UMAP plots exhibits the diagnostic potential of these biomarkers with the expression profiles of the 19 proteins (Figure 5B-3) and 11 transcript (Figure 5C-3) biomarkers providing enhanced separation of BMET and HC samples compared to the complete protein and sRNA profiles captured across BMET and HC plasma-EV cohorts (Figure 5B2, 5C-2). Application of the ML-based approach to the BMET vs. HC comparison yielded a biomarker set of nine proteins that had an AUC of 0.9779 (Supplementary Figure 7C-1) and a set of 10 transcripts that had an AUC of 0.9792 (Supplementary Figure 7C-2) using the mRMR feature selection method and an L2-regularised logistic regression model. The biomarkers corresponding to the highest performing models by AUC are listed in Supplementary Table 10. Cross-referencing these ML-based biomarker sets with markers identified by our biomarker discovery workflow found that DNAJC8, CAVIN1, GADD45GIP1, ACAT1, ATP1B1, miR-223-3p, miR-20a-5p, let-7b-5p, miR-142-3p, miR-378a-3p, miR-1246, miR-342-3p, were strong candidate biomarkers for BMET (Figure 5A). 3.6 Plasma-EV markers distinguish BMET by their primary malignancy The BMET cohort included plasma-EV samples derived from patients with a history of melanoma, adenocarcinoma, and various other carcinomas, such as non-small cell carcinoma, testicular embryonal carcinoma, and acinar cell carcinoma (Supplementary Table 1). Plasma-EV protein and sRNA profiles of BMET patients (Figure 6A) were analysed using multiple feature selection methods to identify markers that distinguish BMETs based on their primary tumour malignancy, with a particular focus on differentiating metastatic melanoma (MMEL, n=10) from other BMETs (OTHER, n=11). Among the 4117 plasma-EV proteins, an L1-regularised logistic regression model ranked 20 proteins as the most-informative. These were further evaluated using Random Forest, Gini Decrease and Relief F methods, identifying a total of seven proteins, GAL3ST4, TMT1B, MCCC2, C2orf76, PODXL, VGF, and ARF4, that were consistently highly ranked across all three methods (Figure 6B-1). Receiver Operating Characteristic (ROC) analysis confirmed their high sensitivity and specificity for distinguishing MMEL from OTHER BMET (Figure 6B-2), with strong discriminative performance metrics; GAL3ST4 (AUC = 0.95, p = 0.0004), TMT1B (AUC = 0.95, p = 0.0006), MCCC2 (AUC = 0.97, p = 0.0003), C2orf76 (AUC = 0.91, p = 0.0015), PODXL (AUC = 0.89, p = 0.0028), VGF (AUC = 0.87, p = 0.0043) and ARF4 (AUC = 0.95, p = 0.0006). PCA analysis based on the expression profiles of these seven proteins effectively separated MMEL samples from OTHER BMETs (Figure 6B-3), and a multivariate Random Forest model achieved an average classification accuracy of 93.0% (70%-30% train-test split, 100 iterations; Figure 6B-4). A similar approach was applied to identify proteins distinguishing BMETs originating from SKIN (n=10), LUNG (n=6) and OTHER (n=5) primary sites (Figure 6A). L1-Logistic Regression analysis selected 49 most-informative proteins, which were subsequently ranked by Random Forest, Gini Decrease, and ReliefF feature selection methods. Across all three methods, six proteins (ATP5F1A, ACACA, YARS2, MPDZ, VGF and PYCR2) consistently scored in the top-20 (Figure 6C-1). PCA analysis based on the levels of these six plasma-EV proteins enabled clear clustering of samples by primary tumour site (Figure 6C-2), with ACACA and ATP5F1A contributing to the separation of LUNG-derived BMETs, while MPDZ, PYCR2, and VGF were associated with SKIN-derived BMETs (Figure 6C-3). In contrast, plasma-EV sRNA transcripts exhibited weaker discriminatory power for distinguishing BMETs based on primary tumour malignancy and primary site (Supplementary Figure 8). To evaluate their discriminatory potential, the plasma-EV transcripts were ranked by L1-Logistic Regression and Random Forest models to differentiate MMEL from OTHER BMETS (Supplementary Figure 8A). Three transcripts (miR-122-5p, miR-30e-3p and miR-148a-3p) were consistently ranked in the top-30 features by both methods (Supplementary Figure 8A-1). However, PCA analysis did not reveal clear separation between MMEL and OTHER BMET samples (Supplementary Figure 8A-2). Using the same feature ranking approach, seven sRNA transcripts (miR-484, miR-148a-3p, miR-320b, miR-192-5p, miR-122-5p, piR-23444, and miR-155-5p) were highly-ranked for distinguishing BMETS by their primary site (Supplementary Figure 8B-1). Nonetheless, their expression profiles did not effectively cluster BMET samples by their SKIN-, LUNG- or OTHER- primary sites (Supplementary Figure 8B-2). 4. Discussion: 4.1 Circulatory EVs as Practical Diagnostic Tools for Brain Cancers GBM and BMET comprise an array of distinct tumour entities, with diverse aetiology, pathophysiology, clinical behaviour and treatment approaches. However, they often exhibit overlapping features on conventional neuroimaging, making pre-surgical differentiation challenging, with misdiagnosis rates exceeding 40% [45, 46]. Definitive diagnoses rely on brain tissue biopsy, but this approach carries significant limitations, including surgical risks, patient reluctance, tumour accessibility issues, sampling errors, and insufficient tissue to capture tumour heterogeneity. Additionally, tissue-based diagnostics can be complex, time-consuming, and may delay critical treatment decisions. A major clinical concern is the reliance on imaging techniques, such as MRI, CT and PET scans, to diagnose BMETs, especially in patients with a known history of systemic malignancy. When a new brain lesion is detected in patients with known malignancy, it is often presumed to be metastatic disease and treated accordingly [47]. However, this assumption can lead to misdiagnoses and poor outcomes, especially in rare cases of synchronous tumour pathologies [47]. For instance, a case report described a patient with non-small cell lung cancer and a brain lesion initially presumed to be a BMET based on imaging alone. The lesion was treated with stereotactic radiosurgery, but continued progression of the brain lesion prompted a delayed biopsy, and ultimately revealed a GBM diagnosis [47]. Although synchronous malignancies are uncommon, they highlight the urgent need for accurate, minimally invasive diagnostic tools to distinguish BMET and GBM pre-operatively. This is particularly crucial for BMET patients who could undergo effective non-surgical treatment for their intracranial metastases [47]. In this study, we demonstrate that sampling circulating EVs is a valuable diagnostic strategy for distinguishing GBM from BMET. We developed a plasma-EV isolation method for high-dimensional proteomic and sRNA profiling for biomarker discovery (Figure 1). By integrating ML-methods, we identified biomarkers capable of classifying both GBM (Figure 4) and BMET (Figure 5) and differentiating between the two entities (Figure 3), with high discriminatory power with AUC values greater than 0.90. Notably, these high-performing biomarker models were developed using as little as 500 μL of plasma for EV proteomic analysis and up to 1 ml plasma for EV transcriptomic analysis, demonstrating the feasibly of detecting biomarkers from minimal blood volumes. In contrast, biomarker discoveries related to circulating tumour cells (CTCs) and free circulating tumour DNA (ctDNA) typically require much larger plasma volumes for optimal detection. Studies have shown that CTC detection rates increase significantly from 13% to 47% when increasing plasma volumes from 7.5 mL to 30 mL, while ctDNA detection improves from 66.6% to 100% [48, 49]. The large sample volumes or multiple blood draws currently required for CTC and ctDNA approaches for a brain tumour liquid biopsy is not feasible for critically ill patients, especially when frequent assessments of other blood parameters are necessary. In this context, the ability to accurately detect EV biomarkers using smaller plasma volumes offers significant advantages, including reduced patient discomfort and potential side effects, compatibility with frequent routine sampling, and cost-effectiveness. 4.2 Machine Learning Models Provide an Accurate Distinction Between GBM and BMET This pilot study is the first to investigate molecular biomarker panels from circulatory EVs to distinguish GBM from BMET. Our biomarker discovery approach integrated differential expression analysis with ML-based selection of high-performing diagnostic markers. We have identified highly accurate putative protein and sRNA biomarker panels for diagnosing and differentiating GBM and BMET. Cross-validation revealed robust protein biomarkers with impressive performance metrics for diagnosing GBM (AUC = 0.978) and BMET (AUC = 0.982), and for distinguishing them (AUC = 0.990). Similarly, the sRNA biomarkers demonstrated excellent performance for diagnosing GBM (AUC = 0.998) and BMET (AUC = 0.996), and for distinguishing the two entities (AUC = 0.912). These results are comparable to recent advancements in radiographic imaging-based deep learning, which reported classification accuracies ranging from 83.5% to 95.6% in validation cohorts [50, 51] and perfusion-weighted imaging achieving an AUC of 0.92 [52]. Previous studies have also explored unbiased ML models for biomarkers distinguishing GBM and BMET using plasma proteomics. However, these studies did not report any overlap with the circulatory-EV biomarkers identified in this study. For example, one study identified a panel of eight-proteins (PRL, SNC, MACF1, ACAT2, VWF, GSN, FCGBP and F10) that differentiated gliomas from BMETs with 74.6% accuracy. Similarly, BMET-specific proteins were included COL1A1, VWF, GSN, F10, DES, NCAM1, SNG521, DST and CD44, and achieved a classification accuracy of 92.3% [45], while glioma-specific proteins included COL1A1, PRL, VWF, HBD, SF3B1, NME3, RRBP1 and CSRP1, and had a cumulative accuracy of 78.5% [45]. Furthermore, integrating ML-models with Raman spectroscopy has demonstrated high classification accuracy for detecting brain tumour tissue, with this approach capable of accurately classifying GBM patients (91%), BMET (97%) and meningiomas (96%) [53]. Recently the implementation of artificial neural networks on molecular spectral profiles of sera-EVs determined by surface-enhanced Raman Spectroscopy, facilitated accurate distinction between BMETs and GBM with 97% accuracy [54]. The spectral profiles also successfully predicted the primary tumour of origin for BMETs, achieving 94% for breast cancer and 100% for lung cancer. These findings align with our results and emphasise the potential of EV-based blood tests as reliable, non-invasive tools for distinguishing BMET and GBM [54]. 4.3 Key Molecular Markers for GBM and BMET The cross-validation methodology identified proteomic and sRNA biomarkers in plasma EVs that effectively distinguish GBM from BMET. In our analysis, 23 protein markers demonstrated strong discriminatory power with a mean AUC of 0.99, while four key sRNA species had a mean AUC of 0.912. Notably, several of these markers, NIBAN2, MT3, RBM15, APOC3, ACTL6A, GPRIN1, PRRC2C, RPL10, RNPEP, miR-27b-3p, miR-182-5p and miR-190a-5p, exhibited high diagnostic performance in ML-based analysis (Figure 3A). Many of the protein biomarkers elevated in BMET plasma-EVs (Figure 3B-1, Figure 3C-1) have well-established roles in cancer progression and metastasis. For example, NIBAN2 is a downstream target of the MAP kinase (Erk1/2) signalling cascade and is known to promote melanoma metastasis and suppress apoptosis [55, 56], and was also a BMET-specific plasma-EV marker (Figure 5A). Similarly, APOC3 is linked to lymph node metastasis in oral squamous cell carcinoma through its interaction with ENO1 [57], a protein that is 11 times more expressed in uveal melanoma-EVs compared to normal choroidal melanocytes and serves as a biomarker for melanoma and non-small cell lung carcinoma [58]. PRRC2C regulates tumour progression and is associated with shorter survival in liver cancer and also promotes metastasis through EMT-related pathways [59]. Interestingly, APOC3 and PRRC2C were also part of the plasma-EV biomarker signature for GBM (Figure 4A), albeit at lower levels in GBM relative to HCs. Conversely, markers such as MT3, ACTL6A, and miR-190a-5p have reported roles in both GBM and brain metastasis. MT3, a CNS-selective isoform of metallothionein heavy-metal binding proteins that is abundant in astrocytes, is not only implicated with the metastasis of solid tumours [60], but also correlates with GBM treatment resistance and reduced survival [61, 62]. ACTL6A expression is often higher in metastatic tumours compared to primary tumours, suggesting a key role in cancer migration [63], however its high expression is also associated with poor survival outcomes for GBM [64], promoting GBM tumour migration and invasion [65]. An intriguing finding is that related to miR-190a-5p, a known tumour suppressor, where low levels play important roles in tumour metastasis via interaction with VEGF-mediated tumour angiogenesis, and also glioma malignancy [67, 68]. Contrary to this, our study showed elevated levels of miR-190a-5p in both BMET plasma-EVs relative to GBM, and even in GBM relative to HCs. This may reflect selective export of miR-190a-5p from tumour cells via EVs to mitigate its tumour suppressive effects, or a compensatory anti-tumour function by normal cells of the body [69]. In contrast, protein biomarkers more abundant in GBM plasma-EVs (Figure 3B-1, Figure 3C-1) have been linked to cancer stemness, migration and invasion, and have been previously reported as biomarkers for GBM or glioma. RBM15, a key component of the m6A methylation ‘writer’ complex, plays key roles in epitranscriptome regulation and is part of a prognostic risk signature for glioma, and promotes tumour growth in basal-like breast cancer [70-73]. Likewise, GPRIN1 regulates glioma progression and seizure development [74], and plays key roles in lung cancer proliferation, migration and epithelial-to-mesenchymal transition (EMT) [75], while RPL10 was previously identified as a GBM urinary-EV biomarker [76]. miR-182-5p promotes glioma tumorigenesis [77], and has been shown to be elevated in EVs secreted by hypoxic GBM cells, which directly targets Kruppel-like factor 2 and 4, leading to accumulation of VEGFR, promotion of angiogenesis, vascular permeability and tumour transendothelial migration, with evaluated circulating miR-182-5p correlated with poor patient prognosis [78]. It is also overexpressed in triple-negative breast cancer and has been associated with lymph node metastasis, suggesting that it may have potential relevance in the context of brain metastases [66]. In contrast, let-7f, is a tumour suppressor that inhibits glioma cell proliferation, migration, and invasion, with high let-7f expression negatively associated with glioma grade [79]. Furthermore, lower let-7f-5p levels are also associated with increased propensity for metastasis in human gastric cancer [80], a trend that was observed in BMET plasma-EVs relative to GBM (Figure 3C-1), while elevated RNPEP is found in high-metastatic hepatocellular carcinoma exosomes, and drives tumour stemness and EMT through NF-κB signaling [81]. 4.4 Circulating-EV Profiles Predict Primary Malignancy Melanoma, a neural ectodermal tumour, is notorious for its capacity to metastasise to the brain. Melanoma-EVs play key roles in preparing a pre-metastatic niche, by altering the host tissue to make it more hospitable for the metastasising tumour [58, 82]. Here, we report a key novel finding that plasma-EV profiles can stratify BMET patients by their primary malignancy. Seven proteins (GAL3ST4, TMT1B, MCCC2, C2orf76, PODXL, VGF, and ARF4) exhibited strong discriminatory power (AUC = 0.87–0.97) in differentiating MMEL from ‘other’ BMET types (Figure 6B-2, B-3). Similarly, a subset of six proteins (ATP5F1A, ACACA, YARS2, MPDZ, VGF and PYCR2) enabled classification of BMET origin by primary tumour site (skin, lung, or other). Notably, proteomic markers demonstrated superior performance compared to the sRNA signatures for classification of BMET specimens, suggesting that protein-based EV biomarkers may provide greater diagnostic specificity. Intriguingly, multiple plasma-EV proteins that could differentiate MMEL from ‘other’ BMETs are implicated in tumour progression and migration. MCCC2 plays important roles in cancers metabolism, promoting cell proliferation, migration and invasion [83], suggesting that it could also support melanoma metastasis to the brain by modulating energy metabolism. C2orf76 is associated with increased cancer aggressiveness and contributes to malignant phenotypes [84]. Elevated PODXL levels promotes stemness and EMT in melanoma, and is associated with increased migratory capacity [85], while elevated VGF is reported in melanoma metastasis [86], and linked with neurogenesis and neurite outgrowth. It is a secreted mediator for metastatic breast cancer tropism to the brain [87], thereby potentially facilitating melanoma cells’ adaptation within the brain microenvironment. These findings have critical implications for cases where the primary tumour is unknown. In these cases, plasma-EV profiling could streamline this process, offering a fast, blood-based method for determining cancer-of-origin. This is particularly valuable in clinical settings where identifying the primary site of metastasis is essential for guiding treatment decisions. Moreover, the ability to distinguish MMEL from other BMETs with high accuracy provides a promising diagnostic tool for earlier detection of MMEL, and improved patient stratification for the development of personalised treatment strategies. 4.5 Study Limitations and Future Directions A rigorous cross-validation approach was employed in this pilot study to identify highly accurate biomarker panels for GBM and BMET. However, multiple limitations need to be addressed to ensure the generalisability, reproducibility and clinical utility of any biomarker panels resolved here. The primary limiting factor of this investigation was the relatively small sample sizes of 26 GBM and 21 BMET plasma-EVs. While sufficient for a pilot study, they necessitate further validation in future investigations using large, independent cohorts of blood sampled across multiple collection sites. However, despite the small sample sizes, the GBM-specific biomarkers demonstrate robustness as they were consistently resolved across two independent GBM cohorts (Figure 2B) using plasma samples that were collected and processed at different sites and under varied conditions. Future studies should continue to incorporate broad collection of bloods, particularly specimens collected during clinical trials, to validate identified plasma-EV biomarker panels and confirm their pre-operative diagnostic accuracy, as well as assess their clinical impact on treatment-decisions and clinical outcomes. Furthermore, the successful translation and adoption of plasma-EV biomarkers into routine clinical practice will require the standardisation of various pre-analytical variables, including EV isolation, storage and blood processing protocols. Finally, employing this biomarker discovery workflow on longitudinal cohorts holds promising potential for deciphering biomarkers that may facilitate effective post-treatment surveillance of both GBM and BMET patients. 4.6. Conclusion This pilot study identified biomarker panels associated with circulating-EVs that may facilitate the development of highly accurate, minimally invasive blood tests for distinguishing GBM from BMET prior to surgery. These blood-based biomarkers could have significant potential to complement neuroimaging and provide additional diagnostic clarity in cases where BMETs and GBM exhibit similar imaging features. Notably, circulatory-EV proteomes can also distinguish BMET patients based on their primary malignancy with remarkable sensitivity, which may improve patient stratification for neoadjuvant treatments. Further validation of the biomarker panels described in large external cohorts is warranted to establish their clinical utility and facilitate their integration into standard brain cancer diagnostic workflows. Acknowledgements We thank the patients who contributed their blood samples and associated information to the Sydney Brain Tumour Bank and GlioNET Observational Study at Royal Prince Alfred Hospital (RPA) and Chris O’Brien Lifehouse. The Sydney Brain Tumour Bank is supported by funding from RPA Brainstorm brain cancer research charity, NSW Office of Health and Medical Research, Australian government Medical Research Future Fund (Australian Brain Cancer Mission) as well as philanthropic funding to the Brain Cancer Research Program at the Chris O’Brien Lifehouse. We are incredibly grateful to all staff supporting Sydney Brain Tumour Bank operations, in particular Mary Lordan, Jane Raftesath, Amy Lonergan, Kristine Deang and Matin Ramezani. Blood samples and information was also accessed from the Hunter Cancer Biobank, which supported by the Mark Hughes Foundation at the Hunter Medical Research Institute. Our research was enabled by access to Sydney University facilities, including Sydney Mass Spectrometry, and Sydney Microscopy and Microanalysis, A special thanks to Adam Costin from Sydney Microscopy and Microanalysis for capturing the cryo-transmission electron microscopy images. Funding This work was supported by grants from NSW Office of Health and Medical Research, Tour de Cure, Mark Hughes Foundation, James N Kirby Foundation, BF Foundation, Cure My Brain and Cure Brain Cancer Foundation. Author contributions Specific contributions are as follows: Research study conception, data interpretation, and manuscript preparation by SMH and KLA. Clinical sample collection from consented participants and clinical annotations by VG, BS and HWS. LS and MEB performed neuropathological assessments and case characterisations. The plasma-EV isolation method was optimised by SMH, and executed by AT and SMH. SMH performed the EV characterisation studies. SMH completed the proteome preparations, mass spectrometry data acquisition and data alignment. AT prepared the transcriptomes for next-generation-sequencing at Ramaciotti Centre for Genomics. AV and FV performed data processing, statistical analyses, machine learning modelling, and data visualisation under the supervision of FV. All authors have reviewed the manuscript and approve the final version. Corresponding author Correspondence to Kimberley L. Alexander. Competing interests The authors declare no competing interests. Ethics approval and consent This research was approved by The University of Sydney Human Research Ethics Committee under protocol 2019/705 and performed in accordance with the Declaration of Helsinki. Supplementary Materials Supplementary Tables_1-12 Supplementary Figures_1-8 Figure Legends Figure 1 – Strategised methodological workflow for in-depth proteomic and transcriptomic profiling of plasma-EVs and identification of putative biomarker sets . EVs were isolated from 1 ml platelet-depleted plasma into six fractions (F7-F12) using commercially available size exclusion chromatography columns (qEVoriginal, Izon). Fractions 7-9, containing fewer soluble protein contaminants, were pooled for proteomic analysis. Fractions 10-12 comprise the highest EV yields and were pooled for small RNA sequencing. Pooled F7-9 EV proteomes were extracted and prepared for data-independent acquisition mass spectrometry (DIA-MS or SWATH-MS), followed by a targeted data extraction strategy that aligns the MS data to a glioma-specific protein library, comprised of spectral coordinates for 8662 proteins. Pooled F10-12 EVs were pre-treated with RNaseA to remove contamination by extravesicular RNA species, followed by extraction of total RNA. Using the QiaSeq miRNA library kit, small RNA (sRNA) libraries were generated for next-generation-sequencing at 20 million reads per sample. The sequenced sRNA transcripts were mapped to human databases. A total of 4117 proteins and 907 sRNA transcripts were confidently identified for interrogative biomarker discovery, after data preprocessing (filtering, imputation, normalisation and batch effect correction). The multi-criteria biomarker discovery approach, summarised in the blue box, was comprised of a cross-validation framework of 30 train-test (80%-20%) data subsets generated by six repeats of a 5-fold cross validation. The cross-validation method is combined with inter-cohort differential expression (DE) analyses and, machine learning-based feature importance and stability assessment, to identify biomarkers with statistical significance. Details of the multi-criteria biomarker discovery approach can be found in Methods Sections 2.6, 2.7 and 2.8. Figure 2 –Summary of cohorts and multi-omics data used for plasma-EV biomarker discovery. A) Summary table of study cohorts for differential expression (DE) analysis and interrogative biomarker discovery. The groups are comprised of samples from patients with GBM, ID-wildtype (Cohorts A and B), Brain Metastasis (BMET) and Healthy Controls (HC). The Venn diagrams display the overlap of identified proteins and transcripts across the sample cohorts. B) Protein counts for each group were determined after filtering to include only proteins present in at least 50% of samples within each respective group. A total of 4,092 proteins were found to be shared across all the GBM, BMET and HC cohorts. C) Transcripts were filtered to retain those with fewer than 90% zeros across all samples. A transcript was considered present in a category if at least one sample in that category exhibited non-zero expression. Based on this criterion, 849 transcripts were common across all GBM, BMET and HC sample cohorts. Figure 3 – The best-performing plasma-EV biomarkers for distinguishing GBM and BMET patients pre-operatively. A) Summary table outlining the best-performing plasma-EV proteins and transcripts for distinguishing GBM from BMET patients preoperatively. The table includes the number of features, their recursive feature elimination cut-offs and mean area under the curve (AUC). A total of 23 proteins had a mean AUC of 0.990, while 4 miRNA transcripts had a mean AUC of 0.912. Protein and miRNA markers denoted by an asterisk were also identified as high-performing biomarkers for discriminating GBM and BMET in a ML-based approach (see Methods section 2.7). Box-plots of the B-1) 23 best-performing protein biomarkers and C-1) 4 best-performing transcript biomarkers, illustrate the differences in expression levels between GBM and BMET cohorts. The middle line represents the median and points above and below the boxes represent outliers. UMAP plots depict the data dimensionality of the B-2) 4117 proteins and C-2) 907 transcripts, and show an increased separation between GBM and BMET cohorts based on expression levels of the B-3) 23 candidate biomarker proteins, and C-3) 4 candidate miRNA biomarkers. Figure 4 – The best-performing plasma-EV markers for diagnosing GBM from HC. A) Summary table outlining the best-performing proteins and transcripts for distinguishing GBM patients from HCs using pre-operative plasma-EV specimens. The table details the number of features and their identities, their recursive feature elimination cut-offs, and mean area under the curve (AUC). Protein and miRNA markers denoted by an asterisk were also identified in a separate ML-based analysis as high-performing biomarkers for discriminating GBM and HC. Box-plots of the B-1) 23 best-performing protein biomarkers and C-1) 11 best-performing transcript biomarkers, illustrate the difference in their expression levels between GBM and HC cohorts. The middle line represents the median and points above and below the boxes represent outliers. UMAP dimensionality reduction illustrates minimal clustering of GBM and HC based on the expression levels of B-2) 4117 proteins and C-2) 907 transcripts, and exhibit and improved separation between GBM and HC samples based on expression levels of the B-3) 23 candidate biomarker proteins, and C-3) 11 candidate miRNA biomarkers. Figure 5 – The best-performing plasma-EV markers for diagnosing BMET from HC. A) Summary table outlining the best-performing proteins and transcripts for distinguishing BMET from HC patients using pre-operative plasma-EV specimens. The table outlines the number of features and their identities, as well as their recursive feature elimination cut-offs and mean area under the curve (AUC). Protein and miRNA markers denoted by an asterisk were also identified in a prior ML-based approach as high-performing biomarkers for discriminating BMET and HC. Box-plots of the B-1) 19 best-performing protein biomarkers and C-1) 11 best-performing transcript biomarkers, illustrate the difference in their expression levels between BMET and HC cohorts. The middle line represents the median, and points above and below the boxes represent outliers. UMAP dimensionality reduction illustrates negligible clustering of BMET and HC based on the expression levels of B-2) 4117 proteins and C-2) 907 transcripts, and exhibit and improved separation between BMET and HC samples based on expression levels of the B-3) 19 candidate biomarker proteins, and C-3) 11 candidate miRNA biomarkers. Figure 6 – Top-ranked plasma-EV proteins for distinguishing brain metastases (BMET) based on their primary tumour type. A) Plasma-EV samples from metastatic brain tumour patients (BMET; n=21) were annotated by their primary tumour type (Metastatic melanoma (MMEL, n=10) or other metastases (OTHER, n=11)), and their primary site-of-origin (SKIN (n=10), LUNG (n=6), or OTHER (n=5)). B-1) An L1-regularised logistic regression model (C=1) was used to rank 4117 proteins by importance for distinguishing MMEL from OTHER BMETs. The model assigned nonzero coefficients (β) to 20 most-informative proteins (ranked 1-20). The 20 most-informative proteins were further ranked using Random Forest, Gini Decrease and ReliefF methods, and seven proteins (GAL3ST4, TMT1B, MCCC2, C2orf76, PODXL, VGF, and ARF4) were consistently highly-ranked across the three methods. B-2) The top-seven proteins exhibited high sensitivity and specificity for distinguishing MMEL from OTHER BMETs with Area Under the Receiver Operating Characteristic Curve (AUC-ROC) values ranging from 0.87-0.97 (0.0006 < p < 0.0028). B-3) The expression profiles of the top-seven proteins also enabled distinct clustering of MMEL and OTHER samples in a Principal Component Analysis (PCA); PC1 = 59% and PC2 = 14%. B-4) A multivariate Random Forest model using the top-seven proteins (stratified 70% training, 30% testing split, 100 iterations) achieved high-performance for classifying MMEL, with average AUC = 0.990, classification accuracy = 93.0%, F1-score = 92.2%, precision = 88.4%, recall = 96.3%, and specificity = 90.5%. C-1) An L1-regularised logistic regression model (C=1) was used to rank the 4117 plasma-EV proteins by importance for distinguishing BMETs originating from SKIN, LUNG and OTHER primary sites-of-origin. The model assigned nonzero coefficients (β) to 49 most informative proteins, which were then ranked by three methods (Random Forest, Gini Decrease, and ReliefF). A total of six proteins (ATP5F1A, ACACA, YARS2, MPDZ, VGF and PYCR2) were consistently ranked in the top-20 across the three feature selection methods for distinguishing BMET by primary tumour site. C-2) PCA analysis of the top-six proteins enabled clear clustering of samples by primary tumour site; PC1 = 40% and PC2 = 22%. C-3) Linear projection of the PCA showed that ACACA and ATP5F1A contributed to the separation of LUNG-derived METs, while MPDZ, PYCR2, and VGF distinguished SKIN-derived BMETs. Supplementary Figure 1 – Simple Linear Regression and R2 values comparing the abundance of proteins identified in technical replicates. Four technical LC-MS/MS analyses of a GBM plasma-EV specimen (Int Ctrl) were performed in SWATH mode on a TripleTOF®6600 on separate days (day 1, 3, 5, 7) to ensure repeatability and reproducibility of our MS quantitation approach. Normalised protein abundances were plotted as scatter plots for all replicate pairs and the Pearson’s correlation coefficient ( r 2 ) was calculated in GraphPad Prism (10.0.0). Stable, reproducible, and repeatable MS/MS quantitation was observed between injections ( r 2 > 0.9697). Supplementary Figure 2 – Simple Linear Regression and R2 values comparing the abundance of proteins identified in technical replicates. Four technical LC-MS/MS analyses of a GBM plasma-EV specimen (SB) were performed in DIA-mode on an Orbitrap Eclipse Tribid on separate days (day 1, 3, 5, 7) to ensure repeatability and reproducibility of our MS quantitation approach. Normalised protein abundances were plotted as scatter plots for all replicate pairs and the Pearson’s correlation coefficient ( r 2 ) was calculated in GraphPad Prism (10.0.0). Stable, reproducible, and repeatable MS/MS quantitation was observed between injections ( r 2 > 0.9218). Supplementary Figure 3 – Characterisation of qEV isolated plasma-EVs. Nanoparticle tracking analysis (NTA) was performed to determine A-1) the average number of particles and, A-2) modal population size (± standard error of the mean of triplicate NTA readings; n=4) per fraction of EVs isolated from 0.5 ml plasma. SWATH-MS measurements of the plasma-EV fractions (n=1) determined the B) number of protein identifications across the plasma-EV fractions and top-100 EV proteins as reported by Vesiclepedia, as well as the abundance of C-1) EV-marker proteins (CD9, CD63, Programmed Cell Death 6 Interacting Protein (PDCD6IP), Integrin subunit β1 (ITGB1), Heat Shock Protein 90 α family class β member 1 (HSP90AB1)) and C-2) common highly-abundant soluble protein contaminants (α-2 macroglobulin (A2M), albumin (ALB), apolipoprotein B-100 (APOB) and haptoglobin (HP)). D) Table shows numbers (and percentages) of overlapping protein species across the different plasma-EV fractions. Cryo-transmission electron microscopy images of pooled EV fractions E-1) F7-9 and E-2) F10-12 plasma EVs. Supplementary Figure 4 – UMAP plots of quantile normalised proteomics data before and after batch correction. Supplementary Figure 5 – UMAP plots of log2-CPM normalised transcriptomics data before and after batch correction. Supplementary Figure 6 – Volcano plots of differentially expressed proteins and transcripts in comparisons of GBM, BMET, and HC samples. For proteomics, the differential expression (DE) analysis was performed using limma on quantile normalised and Combat batch-corrected data, comparing A-1) GBM vs. BMET, B-1) GBM vs. HC and C-1) BMET vs. HC. Significantly different protein levels are defined by fold-change >1.5 and p-value <0.05. For transcriptomics, the DE analysis was conducted with limma on filtered(using filterByExpr), normalised by log transformation of counts per million (CPM) and Combat batch-corrected data, comparing A-2) GBM vs. BMET, B-2) GBM vs. HC and C-2) BMET vs. HC. Significantly different transcript levels have a fold-change >1.2 and p-value <0.05. The top-10 upregulated and top-10 downregulated proteins and transcripts, ranked by fold-change, are labelled in each plot. Colour-codes denote proteins that are significantly high in GBM (blue), BMET (green) and HC (red). Supplementary Figure 7 – Heatmap showing Mean AUC performance of markers determined by ML-based biomarker discovery. Heatmaps illustrating the Mean AUC results for candidate proteomic and transcriptomic biomarkers across three classification comparisons: GBM vs . HC, GBM vs. BMET, and BMET vs . HC. The rows represent the classification models, while the columns indicate the biomarker sets, labelled based on the feature selection method(s) used, the number of iterations in which the biomarker set was consistently selected, and the comparison performed. The heatmaps display the Mean AUC values for each biomarker set: A1, A2) GBM vs. BMET, B1, B2) GBM vs. HC, and C1, C2) BMET vs . HC. The best proteomic biomarker set for A1) GBM vs. BMET achieved an AUC of 0.9622 and was identified using a Random Forest model with the Ranger feature selection method across 27 iterations. For B1) GBM vs. HC, the best proteomic set had an AUC of 0.9815, determined by a Random Forest model with the Ranger feature selection method across 27 iterations. The second highest-performing proteomic biomarker set for C1) BMET vs. HC achieved an AUC of 0.9779 and was identified using an L2-regularized logistic regression model with the mRMR100 feature selection method across 28 iterations. Despite achieving the second-best mean AUC, it was chosen for further analysis as it was comprised of fewer features than the top-ranked biomarker set (50 features were commonly selected by t-test, Wilcoxon test and mRMR_perc50 methods (t_w_mp50_30_METvsHC) and had a mean AUC of 1.0000 with an L2 Regularised logistic regression model. For transcriptomic biomarkers, the best-performing set for A2) GBM vs. MET achieved an AUC of 0.9043, determined using a Random Forest model with the Ranger feature selection method across 30 iterations. For B2) GBM vs. HC, the highest-performing set reached an AUC of 0.9833, identified using a Radial Kernel SVM model with the Ranger feature selection method across 30 iterations. Finally, for C2) BMET vs. HC, the top transcriptomic biomarker set achieved an AUC of 0.9792, determined using an L2-regularized logistic regression model with the mRMR30 feature selection method, across 29 and 30 iterations. Supplementary Figure 8 – Top-ranked plasma-EV sRNA transcripts for distinguishing brain metastases (BMET) by their primary tumour malignancy. A-1) The top overlapping transcripts ranked by importance in both an L1-regularised logistic regression and random forest model, for discriminating between MMEL and OTHER BMETs (miR-122-5p, miR-30e-3p, miR-148a-3p). B-2) The expression levels of the 3 transcripts enabled modest separation of MMEL specimens from OTHER BMETs in a principal component analysis (PCA; x-axis PC1 = 38% and y-axis PC2 = 33%). B-1) The top overlapping transcripts ranked by importance by both a L1-regularised logistic regression and random forest model, for distinguishing BMETs by primary tumour site (SKIN, LUNG, and OTHER). B-2) In a PCA analysis, the expression levels of the top transcripts allowed separation of some samples by their tumour site-of-origin, with most SKIN and LUNG samples forming clusters away from each other (x-axis PC1 = 28%, y-axis PC2 = 24%). References 1. Gavrilovic, I.T. and J.B. Posner, Brain metastases: epidemiology and pathophysiology. Journal of Neuro-Oncology, 2005. 75 (1): p. 5-14.2. Schouten, L.J., et al., Incidence of brain metastases in a cohort of patients with carcinoma of the breast, colon, kidney, and lung and melanoma. Cancer, 2002. 94 (10): p. 2698-2705.3. Cagney, D.N., et al., Incidence and prognosis of patients with brain metastases at diagnosis of systemic malignancy: a population-based study. Neuro-Oncology, 2017. 19 (11): p. 1511-1521.4. Sloan, A.E., C.J. Nock, and D.B. Einstein, Diagnosis and treatment of melanoma brain metastasis: a literature review. Cancer Control, 2009. 16 (3): p. 248-55.5. Fordham, A.-J., et al. Differentiating Glioblastomas from Solitary Brain Metastases: An Update on the Current Literature of Advanced Imaging Modalities . Cancers, 2021. 13 , DOI: 10.3390/cancers13122960.6. Yousaf, T., G. Dervenoulas, and M. Politis, Chapter Two - Advances in MRI Methodology , in International Review of Neurobiology , M. Politis, Editor. 2018, Academic Press. p. 31-76.7. Schiff, D., Single brain metastasis. Current Treatment Options in Neurology, 2001. 3 (1): p. 89-99.8. Hassaneen, W., et al., Multiple craniotomies in the management of multifocal and multicentric glioblastoma. Clinical article. J Neurosurg, 2011. 114 (3): p. 576-84.9. Kalita, O., et al., Glioblastoma multiforme in patients with history of extracranial cancer: Case series. Clinical Neurology and Neurosurgery, 2016. 144 : p. 39-43.10. Maluf, F.C., et al., High-grade gliomas in patients with prior systemic malignancies. Cancer, 2002. 94 (12): p. 3219-3224.11. Lee, E.J., et al., Potential role of advanced MRI techniques for the peritumoural region in differentiating glioblastoma multiforme and solitary metastatic lesions. Clin Radiol, 2013. 68 (12): p. e689-97.12. Jung, B.Y., et al., Differentiation between Glioblastoma and Solitary Metastasis: Morphologic Assessment by Conventional Brain MR Imaging and Diffusion-Weighted Imaging. Investig Magn Reson Imaging, 2021. 25 (1): p. 23-34.13. Sunwoo, L., et al., Differentiation of Glioblastoma from Brain Metastasis: Qualitative and Quantitative Analysis Using Arterial Spin Labeling MR Imaging. PLoS One, 2016. 11 (11): p. e0166662.14. Law, M., et al., High-Grade Gliomas and Solitary Metastases: Differentiation by Using Perfusion and Proton Spectroscopic MR Imaging. Radiology, 2002. 222 (3): p. 715-721.15. Kadota, Y., et al., Differentiation between glioblastoma and solitary brain metastasis using neurite orientation dispersion and density imaging. Journal of Neuroradiology, 2020. 47 (3): p. 197-202.16. Liu, Y., et al., Image-Based Differentiation of Intracranial Metastasis From Glioblastoma Using Automated Machine Learning. Frontiers in Neuroscience, 2022. 16 .17. Suh, C.H., et al., Perfusion MRI as a diagnostic biomarker for differentiating glioma from brain metastasis: a systematic review and meta-analysis. European Radiology, 2018. 28 (9): p. 3819-3831.18. Bauer, A.H., et al., Differentiation of solitary brain metastasis from glioblastoma multiforme: a predictive multiparametric approach using combined MR diffusion and perfusion. Neuroradiology, 2015. 57 (7): p. 697-703.19. Server, A., et al., Diagnostic examination performance by using microvascular leakage, cerebral blood volume, and blood flow derived from 3-T dynamic susceptibility-weighted contrast-enhanced perfusion MR imaging in the differentiation of glioblastoma multiforme and brain metastasis. Neuroradiology, 2011. 53 (5): p. 319-30.20. Mangla, R., et al., Percentage Signal Recovery Derived from MR Dynamic Susceptibility Contrast Imaging Is Useful to Differentiate Common Enhancing Malignant Lesions of the Brain. American Journal of Neuroradiology, 2011. 32 (6): p. 1004-1010.21. Neska-Matuszewska, M., et al., Differentiation of glioblastoma multiforme, metastases and primary central nervous system lymphomas using multiparametric perfusion and diffusion MR imaging of a tumor core and a peritumoral zone—Searching for a practical approach. PLOS ONE, 2018. 13 (1): p. e0191341.22. Dong, F., et al., Differentiation of supratentorial single brain metastasis and glioblastoma by using peri-enhancing oedema region–derived radiomic features and multiple classifiers. European Radiology, 2020. 30 (5): p. 3015-3022.23. Théry, C., et al., Minimal information for studies of extracellular vesicles 2018 (MISEV2018): a position statement of the International Society for Extracellular Vesicles and update of the MISEV2014 guidelines. Journal of Extracellular Vesicles, 2018. 7 (1).24. Xavier, C.P.R., et al., The Role of Extracellular Vesicles in the Hallmarks of Cancer and Drug Resistance. Cells, 2020. 9 (5).25. Bao, Q., et al., Tumor-Derived Extracellular Vesicles Regulate Cancer Progression in the Tumor Microenvironment. Front Mol Biosci, 2021. 8 : p. 796385.26. Hallal, S., et al., Extracellular Vesicles from Neurosurgical Aspirates Identifies Chaperonin Containing TCP1 Subunit 6A as a Potential Glioblastoma Biomarker with Prognostic Significance. Proteomics, 2019. 19 (1-2): p. e1800157.27. Hallal, S., et al., Deep Sequencing of Small RNAs from Neurosurgical Extracellular Vesicles Substantiates miR-486-3p as a Circulating Biomarker that Distinguishes Glioblastoma from Lower-Grade Astrocytoma Patients. Int J Mol Sci, 2020. 21 (14).28. Akers, J.C., et al., MiR-21 in the extracellular vesicles (EVs) of cerebrospinal fluid (CSF): a platform for glioblastoma biomarker development. PLoS One, 2013. 8 (10): p. e78115.29. Street, J.M., et al., Identification and proteomic profiling of exosomes in human cerebrospinal fluid. J Transl Med, 2012. 10 : p. 5.30. Ebrahimkhani, S., et al., Deep sequencing of circulating exosomal microRNA allows non-invasive glioblastoma diagnosis. NPJ Precis Oncol, 2018. 2 : p. 28.31. Hallal, S., et al., A Comprehensive Proteomic SWATH-MS Workflow for Profiling Blood Extracellular Vesicles: A New Avenue for Glioma Tumour Surveillance. International journal of molecular sciences, 2020. 21 (13): p. 4754.32. Hallal, S.M., et al., Glioblastoma biomarkers in urinary extracellular vesicles reveal the potential for a ’liquid gold’ biopsy. Br J Cancer, 2024.33. Hallal, S.M., et al., Size matters: Biomolecular compositions of small and large extracellular vesicles in the urine of glioblastoma patients. Journal of Extracellular Biology, 2024. 3 (11): p. e70021.34. Hallal, S., et al., Deep Sequencing of Small RNAs from Neurosurgical Extracellular Vesicles Substantiates miR-486-3p as a Circulating Biomarker that Distinguishes Glioblastoma from Lower-Grade Astrocytoma Patients. International Journal of Molecular Sciences, 2020. 21 (14): p. 4954.35. Hallal, S., et al., Extracellular Vesicles Released by Glioblastoma Cells Stimulate Normal Astrocytes to Acquire a Tumor-Supportive Phenotype Via p53 and MYC Signaling Pathways. Molecular Neurobiology, 2018. 56 (6): p. 4566-4581.36. Mallawaaratchy, D.M., et al., Comprehensive proteome profiling of glioblastoma-derived extracellular vesicles identifies markers for more aggressive disease. Journal of Neuro-Oncology, 2017. 131 (2): p. 233-244.37. Nanou, A., et al., Tumour-derived extracellular vesicles in blood of metastatic cancer patients associate with overall survival. British Journal of Cancer, 2020. 122 (6): p. 801-811.38. Mastronarde, D.N., Automated electron microscope tomography using robust prediction of specimen movements. Journal of Structural Biology, 2005. 152 (1): p. 36-51.39. Choi, M., et al., MSstats: an R package for statistical analysis of quantitative mass spectrometry-based proteomic experiments. (1367-4811 (Electronic)).40. Stekhoven, D.J. and P. Bühlmann, MissForest—non-parametric missing value imputation for mixed-type data. Bioinformatics, 2012. 28 (1): p. 112-118.41. Johnson, W.E., C. Li, and A. Rabinovic, Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics, 2007. 8 (1): p. 118-127.42. Robinson, M.D., G.K. McCarthy Dj Fau - Smyth, and G.K. Smyth, edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics, 2010. 26 (1): p. 139.43. Vijayan, A., et al., Blood-based transcriptomic signature panel identification for cancer diagnosis: benchmarking of feature extraction methods. Briefings in Bioinformatics, 2022. 23 (5): p. bbac315.44. Welsh, J.A., et al., Minimal information for studies of extracellular vesicles (MISEV2023): From basic to advanced approaches. J Extracell Vesicles, 2024. 13 (2): p. e12404.45. Yang, S., et al., Multidimensional Proteomic Landscape Reveals Distinct Activated Pathways Between Human Brain Tumors. Advanced Science, 2025. 12 (7): p. 2410142.46. Yan, P.-F., et al., Accuracy of conventional MRI for preoperative diagnosis of intracranial tumors: A retrospective cohort study of 762 cases. International Journal of Surgery, 2016. 36 : p. 109-117.47. Carmicheal, J., et al., Synchronous double primary malignancy of non-small cell lung cancer and glioblastoma: A case report and literature review. Case Reports in Oncology, 2025.48. Alba-Bernal, A., et al., Increased blood draws for ultrasensitive ctDNA and CTCs detection in early breast cancer patients. npj Breast Cancer, 2024. 10 (1): p. 36.49. Yang, Y.P., T.M. Giret, and R.J. Cote, Circulating Tumor Cells from Enumeration to Analysis: Current Challenges and Future Opportunities. Cancers (Basel), 2021. 13 (11).50. Shin, I., et al., Development and Validation of a Deep Learning–Based Model to Distinguish Glioblastoma from Solitary Brain Metastasis Using Conventional MR Images. American Journal of Neuroradiology, 2021. 42 (5): p. 838.51. Bae, S., et al., Robust performance of deep learning for distinguishing glioblastoma from single brain metastasis using radiomic features: model development and validation. Scientific Reports, 2020. 10 (1): p. 12110.52. Knutsson, M., et al., Differentiation between glioblastoma and solitary brain metastases using perfusion and amide proton transfer weighted MRI. Frontiers in Neuroscience, 2025. 19 .53. Ember, K., et al., In situ brain tumor detection using a Raman spectroscopy system—results of a multicenter study. Scientific Reports, 2024. 14 (1): p. 13309.54. Premachandran, S., et al., Detection of brain metastases from blood using Brain nanoMET sensor: Extracellular vesicles as a dynamic marker for metastatic brain tumors. Biosensors and Bioelectronics, 2025. 269 : p. 116968.55. Schmidlin, C.J., et al., FAM129B-dependent activation of NRF2 promotes an invasive phenotype in BRAF mutant melanoma cells. Molecular Carcinogenesis, 2021. 60 (5): p. 331-341.56. Old, W.M., et al., Functional Proteomics Identifies Targets of Phosphorylation by B-Raf Signaling in Melanoma. Molecular Cell, 2009. 34 (1): p. 115-131.57. Wang, J., et al. ENO1 Binds to ApoC3 and Impairs the Proliferation of T Cells via IL-8/STAT3 Pathway in OSCC . International Journal of Molecular Sciences, 2022. 23 , DOI: 10.3390/ijms232112777.58. Tsering, T., et al. Uveal Melanoma-Derived Extracellular Vesicles Display Transforming Potential and Carry Protein Cargo Involved in Metastatic Niche Preparation . Cancers, 2020. 12 , DOI: 10.3390/cancers12102923.59. Zhang, K., J. Xu, and R. Chen, Silencing proline-rich coiled-coil 2C inhibit the proliferation and metastasis of liver cancer cells. Journal of Gastrointestinal Oncology, 2023. 14 (1): p. 287-301.60. Tao, Y.-F., et al., Metallothionein III (MT3) is a putative tumor suppressor gene that is frequently inactivated in pediatric acute myeloid leukemia by promoter hypermethylation. Journal of Translational Medicine, 2014. 12 (1): p. 182.61. Santangelo, R., E. Rizzarelli, and A. Copani, Role for Metallothionein-3 in the Resistance of Human U87 Glioblastoma Cells to Temozolomide. ACS Omega, 2020. 5 (29): p. 17900-17907.62. Mehrian-Shai, R., et al., High metallothionein predicts poor survival in glioblastoma multiforme. BMC Medical Genomics, 2015. 8 (1): p. 68.63. Chuu, C.-P., et al., ACTL6A is a novel oncogene and prognostic biomarker for prostate cancer. Annals of Oncology, 2019. 30 : p. vi106.64. Ji, J., et al., Actin like-6A promotes glioma progression through stabilization of transcriptional regulators YAP/TAZ. Cell Death & Disease, 2018. 9 (5): p. 517.65. Chen, X., et al., ACTL6A knockdown inhibits cell migration by suppressing the AKT signaling pathway and enhances the sensitivity of glioma cells to temozolomide. Exp Ther Med, 2021. 21 (2): p. 175.66. Paszek, S., et al., Dysregulation of microRNAs in triple-negative breast cancer. Ginekologia Polska = Polish Gynecology, 2017. 88 (10): p. 530.67. Wu, D., et al., Exostosin1 as a novel prognostic and predictive biomarker for squamous cell lung carcinoma: A study based on bioinformatics analysis. Cancer Medicine, 2021. 10 (8): p. 2787-2801.68. Xue, W., et al., PVT1 regulates the malignant behaviors of human glioma cells by targeting miR-190a-5p and miR-488-3p. Biochimica et Biophysica Acta (BBA) - Molecular Basis of Disease, 2018. 1864 (5, Part A): p. 1783-1794.69. Marar, C., B. Starich, and D. Wirtz, Extracellular vesicles in immunomodulation and tumor progression. Nat Immunol, 2021. 22 (5): p. 560-570.70. Tűzesi, Á., et al. Understanding the Epitranscriptome for Avant-Garde Brain Tumour Diagnostics . Cancers, 2023. 15 , DOI: 10.3390/cancers15041232.71. Guan, S., et al., A Risk Signature Consisting of Eight m6A Methylation Regulators Predicts the Prognosis of Glioma. Cellular and Molecular Neurobiology, 2022. 42 (8): p. 2733-2743.72. Hu, B., et al., Identification of RBM15 as a prognostic biomarker in prostate cancer involving the regulation of prognostic m6A-related lncRNAs. European Journal of Medical Research, 2024. 29 (1): p. 411.73. Park, S.H., et al., The m6A writer RBM15 drives the growth of triple-negative breast cancer cells through the stimulation of serine and glycine metabolism. Experimental & Molecular Medicine, 2024. 56 (6): p. 1373-1387.74. Mortazavi, A., et al., Differential gene expression underlying epileptogenicity in patients with gliomas. Neuro-Oncology Advances, 2024. 6 (1): p. vdae103.75. Zhuang, M.Q., et al., G protein regulated inducer of neurite outgrowth 1 is a potential marker for lung cancer prognosis. J Biol Regul Homeost Agents, 2020. 34 (3): p. 853-864.76. Hallal, S.M., et al., Glioblastoma biomarkers in urinary extracellular vesicles reveal the potential for a ‘liquid gold’ biopsy. British Journal of Cancer, 2024.77. Xue, J., et al., miR-182-5p Induced by STAT3 Activation Promotes Glioma Tumorigenesis. Cancer Research, 2016. 76 (14): p. 4293-4304.78. Li, J., et al., Hypoxic Cancer-Secreted Exosomal miR-182-5p Promotes Glioblastoma Angiogenesis by Targeting Kruppel-like Factor 2 and 4. Mol Cancer Res, 2020. 18 (8): p. 1218-1231.79. Yan, S., et al., Let-7f Inhibits Glioma Cell Proliferation, Migration, and Invasion by Targeting Periostin. Journal of Cellular Biochemistry, 2015. 116 (8): p. 1680-1692.80. Liang, S., et al., MicroRNA Let-7f Inhibits Tumor Invasion and Metastasis by Targeting MYH9 in Human Gastric Cancer. PLOS ONE, 2011. 6 (4): p. e18409.81. Chen, Y., et al., Tumor exosomal RNPEP promotes lung metastasis of liver cancer via inducing cancer-associated fibroblast activation. Cancer Science, 2025. 116 (3): p. 792-807.82. Klein, A., et al., Astrocytes facilitate melanoma brain metastasis via secretion of IL-23. The Journal of Pathology, 2015. 236 (1): p. 116-127.83. Liu, Y., Z. Yuan, and C. Song, Methylcrotonoyl-Coa Carboxylase 2 Overexpression Predicts An Unfavorable Prognosis and Promotes Cell Proliferation in Breast Cancer. Biomarkers in Medicine, 2019. 13 (6): p. 427-436.84. Ershov, P., et al. Uncharacterized Proteins CxORFx: Subinteractome Analysis and Prognostic Significance in Cancers . International Journal of Molecular Sciences, 2023. 24 , DOI: 10.3390/ijms241210190.85. Wenzina, J., et al., Inhibition of p38/MK2 Signaling Prevents Vascular Invasion of Melanoma. Journal of Investigative Dermatology, 2020. 140 (4): p. 878-890.e5.86. Ouyang, S., et al., Neuropeptide Precursor VGF Promotes Liver Metastatic Colonization of Gαq Mutant Uveal Melanoma by Facilitating Tumor Microenvironment via Paracrine Loops. Advanced Science, 2024. 11 (46): p. 2407967.87. Carvalho, R., et al., Nerve growth factor inducible (VGF) is a secreted mediator for metastatic breast cancer tropism to the brain. The Journal of Pathology, 2024. 264 (2): p. 132-147. Information & Authors Information Version history V1 Version 1 14 April 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords brain metastasis extracellular vesicles glioblastoma liquid biopsy ml modelling Authors Affiliations Susannah M. Hallal Chris O'Brien Lifehouse View all articles by this author Ágota Tűzesi Royal Prince Alfred Hospital View all articles by this author Abhishek Vijayan University of New South Wales View all articles by this author Vineet Gorolay Royal North Shore Hospital View all articles by this author Brindha Shivalingam Chris O'Brien Lifehouse View all articles by this author Hao-Wen Sim University of New South Wales View all articles by this author Michael E. Buckland Royal Prince Alfred Hospital View all articles by this author Laveniya Satgunaseelan Royal Prince Alfred Hospital View all articles by this author Fatemeh Vafaee University of New South Wales View all articles by this author Kimberley Alexander 0000-0002-7239-039X [email protected] Chris O'Brien Lifehouse View all articles by this author Metrics & Citations Metrics Article Usage 361 views 212 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Susannah M. Hallal, Ágota Tűzesi, Abhishek Vijayan, et al. Circulating extracellular vesicle-based multianalyte biomarker signatures accurately distinguish glioblastoma from brain metastasis patients before surgery. Authorea . 14 April 2025. DOI: https://doi.org/10.22541/au.174465021.16864881/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.174465021.16864881/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fea98e358190db4',t:'MTc3OTI3MjMwNQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();