P05.21.A MACHINE LEARNING MODELS DETECT MULTIANALYTE BLOOD ‘FINGERPRINTS FOR ACCURATE GLIOBLASTOMA SURVEILLANCE

Abstract BACKGROUND Glioblastoma (GBM) is the most aggressive adult primary brain tumour, with high recurrence rates and poor survival despite standard therapy. Distinguishing true tumour progression from treatment-related effects (e.g., pseudoprogression) is a persistent clinical challenge. Circulating biomarkers offering early, reliable, and noninvasive insights into tumour dynamics could transform patient monitoring and guide timely treatment decisions. Extracellular vesicles (EVs) are submicron (30-1000 nm) particles secreted by all cells, including GBM cells, containing selectively packaged molecular cargo. EVs cross the blood-brain barrier and are easily isolated from the circulation, making them ideal candidates for liquid biopsy. We developed a novel EV-based liquid biopsy platform to identify multianalyte biomarker signatures—‘blood fingerprints’—to classify tumour burden, recurrence, and treatment resistance in GBM patients. MATERIAL AND METHODS We analysed 104 serial plasma samples (1 mL) from 50 IDH-wildtype GBM patients at three clinical time points: pre-surgery (PreOP, n=27), post-surgery (PostOP, n=49), and at pathologically confirmed recurrence (REC, n=28). EVs were isolated using standardised protocols and analysed using complementary high-throughput proteomics and small-RNA sequencing platforms. Following data preprocessing (filtering and normalisation), datasets were input into a 30-iteration cross-validation pipeline (80% train, 20% test split) incorporating multiple feature selection strategies and classification algorithms. Top-performing protein and sRNA biomarkers were identified and used to build multianalyte predictive models via ensemble stacking machine learning. RESULTS Our pipeline generated 4117 proteins and 272 sRNA transcripts common to all plasma samples. Three multianalyte blood fingerprints were developed to classify tumour burden (PreOP vs. PostOP), recurrence (PostOP vs. REC), and treatment resistance (PreOP vs. REC). Each model combined the most informative protein and sRNA features. All models achieved 100% training accuracy and demonstrated robust cross-validation performance, with test accuracies ranging from 89.9% to 93.4%. CONCLUSION Our study demonstrates the feasibility and clinical potential of an EV-based multianalyte liquid biopsy for noninvasive monitoring of GBM tumours. This approach could complement current imaging techniques, providing clinicians with blood-based tools to differentiate recurrence from treatment-related effects and monitor treatment response. Validation is ongoing using plasma samples and clinicopathologic data from the VERTU clinical trial (ACTRN12615000407594) and the Sydney Brain Tumour Bank. These efforts aim to advance our liquid biopsy strategy toward clinical implementation, ultimately improving GBM patient management.