The Challenge of NSCLC Subtyping: Non-small cell lung cancer (NSCLC) comprises two major subtypes - adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC) - with distinct molecular drivers, treatment responses, and prognoses. Accurate subtyping is critical for selecting appropriate targeted therapies, yet current tissue biopsy methods are invasive and may not capture tumor heterogeneity.
Exosomes as Liquid Biopsy Carriers: Exosomes are nanoscale vesicles (30-150 nm) shed by tumor cells into blood and other body fluids. They carry surface proteins, nucleic acids, and lipids that mirror the molecular identity of their parent tumor cells. Analyzing exosome content provides a liquid biopsy window into tumor heterogeneity without surgical intervention.
SERS for Molecular Fingerprinting: Surface-Enhanced Raman Scattering (SERS) amplifies the Raman spectral signal from molecules near gold nanostructures by factors of up to 10^10, enabling single-molecule sensitivity. When applied to exosomes, SERS generates complex spectral fingerprints encoding molecular composition that varies between cancer subtypes.
Integrated Platform: This study created a fully integrated platform coupling: (1) PACD probes - gold nanocages functionalized with an aggregation-inducing peptide and anti-CD9 antibody; (2) a microfluidic chip for efficient exosome capture; and (3) a multiscale 1D convolutional neural network for deep learning classification of the captured SERS spectra.
PACD Probe Construction: Gold nanocages (AuNCs) were synthesized and functionalized with a peptide that induces nanoparticle aggregation upon exosome binding (creating SERS hot spots for signal amplification), conjugated to anti-CD9 antibodies for specific exosome capture. CD9 is a tetraspanin protein universally expressed on exosome surfaces, ensuring broad capture efficiency.
Microfluidic Chip Architecture: The chip featured herringbone-patterned microchannels that create chaotic mixing to maximize probe-exosome contact. Channel dimensions were optimized to achieve 85% exosome trapping efficiency from cell-conditioned media. The captured exosome-PACD complexes were interrogated with a 785 nm laser for SERS spectral acquisition.
SERS Dataset: A total of 944 SERS spectra were collected from 4 cell lines: two LUAD lines (A549, PC-9) and two LUSC lines (H520, SK-MES-1). Each spectrum was pre-processed by background subtraction and normalization to remove laser noise and baseline drift before being input to the neural network.
Deep Learning Architecture: A multiscale 1D convolutional neural network processed SERS spectra as one-dimensional signals. Three parallel convolutional streams with different kernel sizes captured spectral features at different frequency resolutions. Grad-CAM (Gradient-weighted Class Activation Mapping) was applied to highlight the spectral regions most important for classification decisions, providing model interpretability.
Classification Accuracy: The deep learning model achieved 97.88% overall accuracy on the 944-spectrum dataset for distinguishing the four NSCLC cell lines. Individual pairwise AUCs exceeded 0.995 for all cell line comparisons, demonstrating excellent discrimination not just between LUAD and LUSC subtypes but also between different cell lines within the same subtype.
Handling Exosome Heterogeneity: A key finding was that SERS spectra from individual exosomes within the same cell line showed considerable variability (exosome heterogeneity), yet the CNN successfully classified the subtype despite this noise. The model learned ensemble features across multiple exosome spectra rather than relying on any single spectral peak.
Grad-CAM Explainability: Gradient-weighted activation maps highlighted specific Raman shift regions corresponding to proteins, lipids, and nucleic acids that differed between LUAD and LUSC exosomes. These highlighted spectral regions align with known biological differences between the two subtypes, validating that the model learned biologically meaningful features rather than artifacts.
Microfluidic Capture Validation: The 85% exosome trapping efficiency was confirmed by nanoparticle tracking analysis comparing particle counts before and after chip processing. Electron microscopy images confirmed intact exosome morphology after capture, verifying that the microfluidic process did not damage or alter the exosome content being measured.
Liquid Biopsy Alternative to Tissue Biopsy: Current NSCLC subtyping relies on tissue biopsies, which are invasive and sometimes unfeasible for patients with deep or inaccessible tumors. An exosome-based liquid biopsy platform could provide subtype information from a simple blood draw, removing procedural risk and enabling repeated sampling to monitor tumor evolution.
Capturing Tumor Heterogeneity: Unlike a single tissue biopsy that samples one tumor region, circulating exosomes integrate signals from multiple tumor sites and metastases. This systemic sampling may provide a more complete molecular picture, particularly in patients with heterogeneous tumors or multiple metastatic lesions.
Treatment Guidance: LUAD and LUSC differ critically in treatment recommendations - LUAD is more commonly associated with EGFR and ALK alterations targeted by specific drugs, while LUSC histology guides squamous-specific regimens. Rapid, non-invasive subtyping could accelerate time to appropriate therapy, particularly important for urgent clinical presentations.
Point-of-Care Potential: The microfluidic format is amenable to miniaturization into a point-of-care device. Future development of integrated chip-to-readout systems could enable rapid NSCLC subtyping in clinic without sending samples to a central laboratory, particularly valuable in resource-limited settings.
Cell Line to Patient Blood Samples: All 944 spectra were collected from in vitro cell line supernatants rather than patient blood samples. Transitioning to clinical blood samples introduces challenges including lower exosome concentrations, contamination from non-tumor exosomes, and patient-to-patient biological variability that cell lines cannot replicate.
Expanding Beyond Four Cell Lines: Classification among four cell lines is a proof-of-concept. Clinical deployment would require training on exosomes from a much larger and diverse cohort of patient-derived samples representing the full histological, molecular, and clinical spectrum of NSCLC.
Integration with Molecular Testing: Future platforms could combine SERS-based subtyping with concurrent mutation detection - for example, using specific Raman reporters linked to antibodies targeting EGFR mutations on exosome surfaces - creating a multi-analyte test that simultaneously determines subtype and actionable mutations.
Automated Data Processing Pipeline: Clinical implementation would require a fully automated pipeline from sample loading to classification output, including automated spectral preprocessing, quality control filtering, and standardized deep learning inference. Development of this end-to-end system remains an important engineering challenge.