Integration of label-free surface enhanced Raman spectroscopy (SERS) of extracellular vesicles (EVs) with Raman tagged labels to enhance ovarian cancer diagnostics

Abstract

We report a proof-of-concept diagnostic strategy that integrates multiplexed Raman-tagged antibody labeling with label-free surface-enhanced Raman spectroscopy (SERS) and machine learning (ML) to improve the detection of ovarian cancer via extracellular vesicles (EVs). EVs were isolated from patient plasma using size-exclusion chromatography and labeled with polyyne-based Raman tags targeting three ovarian cancer biomarkers: CA-125, HE4, and CA-19-9. Labeled and unlabeled EVs were deposited onto SERS-active substrates, and spectra were collected using a custom confocal Raman microscope. Incorporating the tag-derived signal into SERS analysis enhanced interpretability and added molecular specificity. We evaluated classification performance using various ML models applied to spectral datasets from a cohort of ovarian cancer patients and healthy controls. Combined use of the Raman tag and label-free regions improved classification accuracy compared to either modality alone. Notably, support vector machine (SVM) achieved over 95 % accuracy, sensitivity, and specificity. Compared to ELISA, our SERS platform demonstrated improved sensitivity in detecting EV-associated biomarkers from small sample volumes. This approach addresses a key limitation of SERS-based diagnostics by linking spectral features to known biomarkers, offering improved transparency and performance in ML-enabled liquid biopsy.

Keywords: Artificial intelligence; Biomarkers; Exosomes; Liquid biopsy.

Fig. 1.. SERS analysis of EVs isolated from patient biofluids.

a) Raman confocal imaging optical set-up for SERS analysis of EV samples deposited on SERS active substrates. Substrates are placed on the inverted confocal objective and spectra are acquired. Scale bar = 1 μm. b) Antibodies against cancer-related EV membrane proteins (CA125, CA19-9, and HE4) are covalently conjugated to three distinct Raman polyyne tags. EVs are first enriched by SEC directly from patient plasma (150 μL) and then co-incubated with three Raman tagged-antibody pairs. An additional SEC step removes unbound antibodies and samples (20 μL) are incubated on the SERS-active substrate for analysis.

Fig. 2.. EV characterization.

a) NTA is used to measure particle concentration and size distribution. b) A negative-stained TEM image shows the size and morphology of isolated particles. c) SP-IRIS demonstrates the presence of EVs captured in an antibody sandwich of tetraspanins CD9, CD63, and CD81, alongside mouse-IgG as a control. Particle counts from the fluorescently labeled detection antibodies are shown on the y-axis. d) Relative protein expression for the three chosen biomarkers. Isolated EVs were tested using sandwich ELISAs to test the level of surface expression for CA125, HE4, and CA19-9.

Fig. 3.. SERS analysis of Raman tagged antibodies for EV biomarker detection.

a) EVs are tagged with antibodies (CA125, CA19-9, and HE4) covalently conjugated with three different Raman-active tags (diyne, triyne, and tetryne), respectively. b) Distinct characteristic SERS peaks are visible for all three tags at 2125 cm⁻¹ (tetryne), 2166 cm⁻¹ (triyne), and 2210 cm⁻¹ (diyne). c) A mixture of all three tags reveals peaks are still discrete and readily distinguishable. d) Mean spectra of patients from cancer and control groups, average as solid lines and pale-shaded regions as standard deviation. Mean and standard deviation of the accumulated intensity of the Raman tag peaks associated with biomarkers of interest (from the green-shaded region) are shown at the e) patient level (errors bars represent RSD) and f) individual spectrum level grouped across cancer vs control. HE4-tetryne between 2120 and 2130 cm⁻¹, CA-19-9-triyne between 2161 and 2171 cm⁻¹, and CA-125-diyne between 2205 and 2215 cm⁻¹. The Raman intensities of the Raman tags associated with biomarkers of interest show significant differences between cancer and control groups. (**** represent p-values <0.0001).

Fig. 4.. Unsupervised learning of Raman tag signature region (cell silent region), cell signature region and full spectrum region.

a) The first 5 PCs of PCA based on Raman tag signature region. The Raman peaks marked with dotted lines in the orange shaded area from left to right are peaks at 2125 cm⁻¹ (HE4-tetryne), 2166 cm⁻¹ (CA-19-9-triyne), and 2210 cm⁻¹ (CA-125-diyne) b) The first 5 PCs of PCA based on EVs signature regions, the peaks marked with dotted lines in the blue shaded area, when observed from left to right, are as follows: at 1118 cm⁻¹ corresponding to C-N stretching in protein backbones and lipids, at 1267 cm⁻¹ attributed to NH₃ rocking and amide III vibrations, at 1493 cm⁻¹ linked to conjugated C=C vibrations in nucleic acids, at 1548 cm⁻¹ denoting protein vibrational modes (Amide II), and finally at 1597 cm⁻¹ representing vibrations in nucleic acids. c) The first 5 PCs of PCA based on full spectrum region, with aforementioned peaks are marked with dotted line in shaded area. d)-f) KDE plot of PC0 and PC1, and g)-i) confusion matrix of K means cluttering for Raman tag signature region (cell silent region), cell signature region and full spectrum region.

Fig. 5.. ML model training and performance.

a) Schematic of the workflow to test 9 different ML algorithms to classify patient samples, using either the tag, EV fingerprint, or full spectral information. The heatmaps display the average scores from all ML models for each spectrum obtained from patients, where orange denotes a control classification and blue indicates a cancer classification. b) Confusion matrices and c) individual ROC curves for all and an overarching mean curve showing the combined performance. d) The heatmap summarize the accuracy rates across ML models trained in cross-validation: samples from the cancer and control group are on the vertical and horizontal axis, respectively. Each cell represents the accuracy of a specific cancer-control test group pairing. e) A bar chart presenting the accuracy rate and associated standard deviation for models trained on distinct spectrum regions. f) This heatmap represents the performance metrics, including accuracy, sensitivity, and specificity, for ML models across each region.