Machine learning-driven microfluidic deglycosylation for ultra-sensitive detection of PD-L1 extracellular vesicles

초록

Reliable detection of programmed death-ligand 1 (PD-L1) remains challenging due to heterogeneity and the masking of epitopes by glycosylation. Herein, we introduce a machine learning (ML)-driven microfluidic chip that performs on-chip deglycosylation of tumor-derived extracellular vesicles (EVs). This process unmasks PD-L1positive EVs (PD-L1 EVs), enabling improved immunodetection. The ML-powered microfluidic approach achieved a limit-of-detection (LOD) of 103 EVs for PD-L1 EVs, significantly surpassing yields from a conventional microfluidic approach (105 EVs) and microcentrifuge tube (106 EVs). The platform was optimized using a hybrid machine learning algorithm, the reTesla Stacking Ensemble (rTSE), maximizing deglycosylation efficiency and PD-L1 signal. In a clinical study involving 45 plasma samples from 30 lung cancer patients and 15 healthy donors, the platform detected PD-L1 EVs and distinguished cancer patients from healthy controls with 96.7 % sensitivity and 100 % specificity. These results underscore the clinical potential of this integrated microfluidic and rTSE model for sensitive and noninvasive PD-L1 biomarker detection.

키워드