SPECTRA-based Detection of Drug-Induced Hepatotoxicity through Extracellular Vesicle Analysis

Abstract

Extracellular vesicles (EV) are becoming crucial targets in liquid biopsy, diagnostics, and therapeutic applications, yet their nanoscale characterization remains challenging. In this context, the detection of drug-induced liver injury, i.e., hepatotoxicity, through EV molecular content remains unexplored. To this end, we present the SPECTRA - Surface Plasmonic Enhancement with Combined Transformative RAman and Interferometric Microscopy approach, which provides rapid, label-free EV content analysis under 30 minutes and requires only 1.3 microliters of sample. Using hepatic cultures as a model, our platform captures distinct and reproducible EV molecular changes in response to acetaminophen-induced hepatotoxicity. Across independent culture sets, a Gaussian Process Regression model predicted Acetaminophen (APAP) dose with median absolute error of 1.50 mM and minimal bias (Bland–Altman bias ~0.03 mM); cross-validated RMSE was 3.145 mM. These findings establish EVs as dynamic reporters of cellular drug responses and demonstrate use of SPECTRA for EV detection of hepatotoxicity.

Graphical abstract

Introduction

Extracellular vesicles (EVs) are lipid-bilayer-enclosed nanomaterials secreted by virtually all cell types that transport bioactive proteins¹, lipids¹, and nucleic acids^2,3, thereby mediating intercellular communication⁴. By encapsulating molecular signatures that mirror the physiological states of their parent cells, EVs have become central to nano-biomaterial research⁵, drug delivery⁶, medical imaging^7,8, biomedical diagnostics⁹ and therapeutic development^10,11. In particular, their molecular cargo can be harnessed for non-invasive monitoring of cellular stress and toxicity, making them ideal candidates for studying drug-induced hepatotoxicity^12–14.

Drug induced liver injury (DILI), i.e., hepatotoxicity, remains a leading cause of drug failure during development^15–18 and withdrawal post approval^19,20. Standard hepatotoxicity assessment methods, such as liver enzyme assays, histopathological evaluations, and transcriptomic analyses, remain indispensable but are hindered by their invasiveness, lengthy procedures, and limited diagnostic precision^21,22. Importantly, EVs have already been implicated in acetaminophen (APAP)-associated liver injury and related hepatotoxic mechanisms, supporting the premise that EVs can carry injury-linked molecular signatures^13,14. At the same time, EV populations are intrinsically heterogeneous and analytical choices in isolation/characterization can strongly influence downstream readouts²³; therefore, rigorous EV reporting and characterization consistent with community guidelines (e.g., MISEV2023) is essential when developing EV-based assays²⁴. Although EVs have been implicated in hepatotoxicity, robust frameworks that translate EV molecular signatures into rapid, quantitative, dose-resolved toxicity readouts remain limited^25–28.

Surface-enhanced Raman spectroscopy (SERS) is a label-free vibrational transduction technique that enhances Raman signals through plasmonic nanostructures, enabling sensitive molecular fingerprinting²⁹. While several studies, including our work³⁰ and those of others^31–35, have explored EV-based SERS sensing for diagnostic applications, the potential of EVs for measuring drug-induced effects, including label-free assessment and rapid classification of cytotoxicity, needs to be further studied. Many EV–SERS biosensing implementations improve selectivity by using immunoassay-style capture (e.g., antibody-based recognition) and related affinity strategies, which inherently enrich marker-defined EV subpopulations and presuppose knowledge of targets. ^33,36. Likewise, aptamer-affinity methods isolate EVs by binding predefined surface markers, enriching marker-defined subsets rather than providing an unbiased readout of the full EV population^37–39. Furthermore, many of the approaches typically focus on binary classification (e.g., healthy vs. diseased)^40,41 rather than quantitative dose-response assessment. Consequently, a label-free framework capable of unbiased, quantitative prediction of drug-induced toxicity grades remains a critical unmet need.

Our work in this article focused on a single type of cell and impact of drug exposure on EVs derived from that cell type. Here, we present SPECTRA (Surface Plasmonic Enhancement with Combined Transformative RAman and Interferometric Microscopy), a label-free framework for drug-induced hepatotoxicity detection using EV content. This framework provides quantitative prediction / dose-resolved regression of drug-induced hepatotoxicity using a SERS-based platform (Fig. 1) by analyzing EV content. Using EVs derived from 2D rat hepatocyte cultures exposed to acetaminophen (APAP), culture supernatants were collected before dosing (Day 0) and at 48 h post-dosage (Day 2), followed by EV isolation and characterization. Our method utilizes 1.3 μL of EV sample volume dried on a disposable gold nanopillar-structured SERS surface. Within 30 minutes, we quantified APAP-induced hepatotoxicity by capturing the molecular changes in EV cargo two days after drug exposure. Our regression model, trained on SERS spectral features, successfully predicted APAP doses across independent cultures with root-mean-square error (RMSE) values as low as 1.50 mM. This study demonstrates the potential of EV-SERS integration as a precise, non-invasive, and scalable approach for hepatotoxicity screening, offering a new avenue for improving preclinical drug development and toxicity prediction.

Figure 1: Experimental approach used in this study.

Hepatic cell cultures were prepared in vitro on plates for various doses and durations. Extracellular vesicles were isolated from the supernatants of these cultures and prepared on gold nanopillar SERS substrates. SERS spectra were compared across different days and doses. Concurrently, functional assay measurements were taken to assess cell viability as the dose increased. These viability results were then compared with the SERS intensities (Created in https://BioRender.com).

Results

We used a SERS-based, label-free approach, to evaluate the changes in EV content secreted by rat primary hepatocytes under different APAP exposures, representing various levels of hepatotoxicity. Increasing doses of APAP led to a dose-dependent decline in hepatocyte viability, functionality, and morphological integrity over 2 days (Fig. 2). Cell viability measurements on day 2, based on a CCK8 assay, indicated an IC50 of 14.55 mM (Fig. 2a). Concurrent functional assays on day 2 showed significant reductions in albumin secretion (IC50: 12.6 mM, Fig. 2b) and urea excretion (Fig. 2c) with increasing APAP doses. Morphological analysis on day 2 corroborated these findings, revealing extensive cellular damage and culture disintegration at 40 mM (Supplementary Figure 1). These results collectively underscore the severe cytotoxic effects of APAP at high doses on primary rat hepatocytes.

Figure 2: Viability, functionality, and morphology of primary rat hepatocyte cultures with varying doses of Acetaminophen exposure.

a) The normalized viability of cell samples via CCK8 assay at different concentrations of acetaminophen (APAP) at the end of the experiment (after 2 days of incubation with APAP). The 2-day IC₅₀ was determined to be 14.55 mM. b) The normalized change of albumin concentration in cell culture medium at the endpoint of the experiment. The day 2 IC₅₀ was determined to be 12.6 mM. c) The normalized change in urea concentration in cell culture medium at the endpoint of the experiment.

Next, we isolated and concentrated EVs from hepatic culture supernatants using our ExoTIC platform⁴² as described in the methods section. NTA analysis revealed that the EVs primarily fall within the small EV size range (50 – 220 nm) (Fig. 3a). As seen in Fig. 3a, the Day 2 – 40 mM measurements have a large variance in concentration since the levels are close to limit of detection of the device⁴³. This variation does not affect the analysis of molecular signatures from the individual EVs using SERS. Still, we used a complementary technique, interferometric microscopy, and confirmed the presence of EVs in the low concentration range and quantified the decrease in the concentration in day 2 for 40 mM APAP (Fig. 3b). The tables summarizing the size profiles from NTA and interferometric microscope can be seen in Supplementary tables 2 and 3. We confirmed the identity of isolated EV samples using flow cytometry with a CD63 antibody, a tetraspanin marker. High CD63 positivity (>90%), in our samples, indicates that most analyzed particles are EVs (Fig. 3c,d). Extended version of the flow cytometry data can be seen in Supplementary Figure 2. Supplementary Figure 3 shows the raw images from Western blot analysis of EV-specific markers TSG101, and Flotillin-1 (Flot1), which are strongly expressed, supporting the presence of EVs. In addition to TSG101, an ESCRT complex protein, and Flotillin-1, a membrane-associated lipid raft protein, further confirm the EV identity. The absence of a Calnexin band, an endoplasmic reticulum marker, verifies the purity of the EV preparation by excluding cellular contamination. TEM further validated the morphology and size of the hepatic EVs (Fig. 3e).

Figure 3: Characterization hepatic EVs.

a) Nanoparticle tracking analysis (NTA) of EV size and concentration for Day 0 and Day 2 samples under 0 mM and 40 mM APAP exposure. Shaded areas represent standard errors. b) interferometric imaging analysis comparing EVs under 0 mM and 40 mM APAP exposure. Top: Day 0 samples; Bottom: Day 2 samples. c-d) Flow cytometry was utilized with and without CD63 antibodies to prove the presence of small EVs from Day 0 samples with 0 mM (c) and 40 mM (d) APAP dosage. e) Transmission electron microscopy (TEM) was used to show the particles have the expected small EV shapes within sizes smaller than 220.

To characterize the commercial nanopillar SERS substrates used for EV measurements, we first leveraged capillary-force–induced nanopillar leaning upon droplet deposition to generate plasmonic hot spots and enhance Raman scattering (Fig. 4a). We verified substrate morphology using representative SEM images provided for the manufacturer’s nanopillar platform (Fig. 4b). We then assessed chip-scale performance by Raman mapping over a 200 μm × 200 μm region to verify spatial signal uniformity (Fig. 4c) and by collecting control spectra (blank substrate and DI water–leaned pillars) alongside EV-deposited pillars to demonstrate EV-associated signal enhancement (Fig. 4d). The color scale in Fig. 4c from 0.10 (blue) to 0.11 (yellow) indicates that the variation in signal intensity across the sampled region is less than 1%, reflecting a high degree of spatial consistency. The substrate exhibits a theoretical enhancement factor ranging from 2.4 × 10⁶ to 2.1 × 10¹¹ depending on local hot-spot geometry⁴⁵, with our measurements confirming spatially uniform performance (CV < 1%, Fig. 4c). Finally, we quantified analytical performance by measuring the relationship between mean SERS intensity and EV concentration (NTA-derived) and annotating the calibration with LoD/LoQ guides (Fig. 4e). We found that the LoD and LoQ of the system are 67 and 307 particles/mL. Thus, we found that the safe concentration region for repeatable measurements is between 10³ and 10⁷ particles/mL.

Figure 4: SERS sensor characterization and analytical performance.

(a) Schematic illustrating nanopillar leaning after sample deposition, which promotes plasmonic “hot spot” formation and SERS enhancement. (b) Representative SEM images of the commercial SERS nanopillar substrate (images reproduced from the Silmeco website)⁴⁴. (c) Raman intensity map over a 200 μm × 200 μm region demonstrating spatial signal uniformity across the substrate. (d) Representative SERS spectra acquired on a blank substrate (blue), after DI-water deposition (orange), and after incubation with the selected EV sample (yellow). (e) Mean SERS intensity versus EV concentration (particles/mL; NTA-derived) with the fitted trend and overlaid limits of detection (LoD) and quantification (LoQ).

We then evaluated the dose-dependent effects of APAP exposure using label-free molecular vibrational analysis. Specifically, we used SERS data on EVs derived from culture supernatants derived from 0, 10, 20, and 40 mM APAP exposure across two independent sets (Set 1 and Set 2) to assess hepatotoxicity. To this end, we applied principal component analysis (PCA) and trained regression models to explore the spectral differences. Figures 5a and d present the average normalized spectra, highlighting reduced intensities with increasing drug concentration on Day 2, particularly in the 810–1600 cm⁻¹ range, while we observed no significant differences before dosing on Day 0, indicating dose-dependent spectral changes. The spectra shown in Fig. 4A were collected from multiple chips (two to four per condition) to ensure reproducibility across substrates. The coefficient of variation across chips was calculated to be 1.01% (see Methods section).

Figure 5: SERS analysis of EVs from hepatic cell cultures.

a,d) Dose-response spectra for days 0 and 2 (Sets 1 and 2). b,e) PCA score maps for dose responses on days 0 and 2 (Sets 1 and 2). c) Gaussian Process Regression (GPR) model calibrated with 80% of Set 1 day 2 data. Top: Predictions for Set 1 day 0. Bottom: Test set predictions for Set 1 day 2. f) Predictions for Set 2 day 0 (Top) and day 2 (Bottom) using the Set 1 day 2 model.

To assess whether spectral analysis of EV content can distinguish differences in APAP exposure, we performed principal component analysis (PCA) (Figs. 5b and e). On Day 2, PCA revealed well-separated clusters corresponding to concentration levels in both sets. In contrast, day 0 data showed no visible clustering (See Supplementary table 1 for overlap analysis). This demonstrated that spectral differences induced by hepatotoxicity became more pronounced on Day 2, highlighting the progressive effect of hepatotoxicity and the improved ability to distinguish concentrations over time.

Next, we created a regression model using the Set1 – Day 2 dataset by applying Gaussian Process Regression (GPR)⁴⁶ analysis (Figs. 5c and f). The GPR model demonstrated strong predictive performance on the test set for Set 2 – Day 2 dataset, with good agreement between actual and predicted concentrations, as reflected by RMSE values (Table 1). To validate the model’s prediction ability, we analyzed Day 0 data as a negative control, where the model consistently failed to predict concentrations accurately across both sets, showing substantial deviations regardless of concentration levels. We found the model’s accuracy by performing RMSE, R², mean absolute error (MAE), 5-fold cross-validated RMSE calculations and plotting Bland-Altman graph (Supplementary figure 4). According to these calculations our model’s error parameters are R²=0.952, overall MAE = 1.495 mM, and from Bland-Altman graph, prediction bias = 0.03 mM. A more robust error prediction was also performed using 5-fold cross-validated RMSE which was found to be 3.145 mM.

Table 1:

Calculated RMSE values of the regression analysis.

	Set 1 RMSE Values (mM)		Set 2 RMSE Values (mM)
APAP dose (mM)	Day 0	Day 2	Day 0	Day 2
0 mM	8.40	1.50	5.89	7.02
10 mM	7.87	5.77	8.97	5.90
20 mM	16.57	4.74	18.36	11.94
40 mM	26.07	2.32	25.07	7.65

We also sought to identify SERS wavenumbers that show the highest correlation with the traditional features of hepatic cell viability across the APAP dose range. Specifically, we calculated the Pearson correlation between the normalized SERS intensities for each wavenumber with CCK-8, albumin and urea measurements and rank-ordered them. Furthermore, we conducted a corresponding Spearman rank correlation analysis to compare outcomes with and without urea. These results are presented in Supplementary Tables 4–7. Figure 6, where in each panel the x-axis represents the SERS wavenumber intensity and y-axis represents the normalized value of different traditional features, illustrates the correlation of select SERS wavenumbers which showed the highest correlations (>0.95 with albumin) and are biochemically relevant. The six highly correlated SERS intensities were around 739 cm⁻¹ (Fig. 6a), 959.8 cm⁻¹ (Fig. 6b), 1249.8 cm⁻¹ (Fig. 6c), 1525.3 cm⁻¹ (Fig. 6d), 1576.2 cm⁻¹ (Fig. 6e), and 1602.1 cm⁻¹ (Fig. 6f) which showed high linearity against both the CCK-8 and albumin signals. A detailed matrix of all correlations among the SERS intensity and traditional hepatotoxicity readings is tabulated in Table 2. We note that albumin and CCK-8 measurements are highly correlated with each other and albumin is considered one of the most reliable viability markers for hepatic cultures⁴⁷. Urea measurements, on the other hand lag these two both temporally and in a dose-dependent manner ⁴⁸, which is also evident in our results (Fig. 2 and Table 1). Table 3 presents the tentative peak assignments for the selected SERS wavenumbers and includes the corresponding molecular vibrational bonds and the most probable macromolecular associations retrieved from the literature. In addition, we scanned whole spectral range to calculate the distribution of the correlation coefficients per wavenumbers and found that 36.3% of them have coefficients less then 0.7, while only 6.4% exceeded r > 0.95 (Supplementary figure 6). The bimodal distribution indicates that high correlations (Fig. 6) are not artifacts of overall intensity scaling but reflect specific molecular vibrations.

Figure 6:

SERS intensity vs viability test results using functional assays on albumin, CCK8 and urea. The relations of these analytes with SERS peak intensities at a) 739, b) 960, c) 1250, d) 1525, e) 1576, f) 1602 cm⁻¹ are displayed in the figure.

Table 2:

Correlation coefficients between normalized viability scores via albumin, CCK8 and urea and SERS intensities at 739, 960, 1250, 1525, 1576 and 1602 cm^-1.

Set1	Albumin	CCK8	Urea	739	959	1250	1525	1576	1602
Albumin	1.000	0.957	0.768	0.992	1.000	0.972	0.977	0.998	0.998
CCK8		1.000	0.879	0.981	0.957	0.864	0.894	0.940	0.960
Urea			1.000	0.844	0.771	0.662	0.746	0.753	0.798
739				1.000	0.992	0.943	0.964	0.986	0.996
959					1.000	0.972	0.978	0.998	0.999
1250						1.000	0.992	0.984	0.970
1525							1.000	0.988	0.982
1576								1.000	0.997
1602									1.000
Set2	Albumin	CCK8	Urea	739	959	1250	1525	1576	1602
Albumin	1.000	0.957	0.768	0.955	0.991	0.976	0.980	0.956	0.984
CCK8		1.000	0.879	0.994	0.981	0.986	0.986	0.988	0.986
Urea			1.000	0.914	0.846	0.888	0.879	0.921	0.868
739				1.000	0.985	0.995	0.993	0.998	0.992
959					1.000	0.996	0.998	0.986	0.999
1250						1.000	1.000	0.997	0.999
1525							1.000	0.995	1.000
1576								1.000	0.993
1602									1.000

Table 3:

Tentative band assignments of the selected Raman bands.

Wavenumber	Bond Vibration	Assignment	Reference
739	Pyridine ring breathing	RNA, observed in miR-122	58
959	v(PO43-)	Ribose	49
1185	Ring δip(C-H)	Saccharides, lipids, nucleic acids	50,51
1250	Amide III	Proteins	52
1525	Vip(C=C)	C, G, RNA assignment	53,54
1576	Ring breathing modes	A,G	55
1602	Pyridine ring stretching	RNA, observed in miR-122	56,57

Discussion

In this study, we demonstrated the potential of SPECTRA, a SERS-based EV analysis platform, to quantify drug-induced hepatotoxicity through dose-dependent molecular signatures. SERS intensities between 810–1600 cm⁻¹ consistently decreased with increasing APAP exposure, and a Gaussian Process Regression (GPR) model achieved RMSE values as low as 1.50 mM across independent datasets. The observed decrease in SERS intensity with increasing APAP dose could arise from either reduced EV secretion or molecular changes within individual EVs. Several lines of evidence indicate the latter: (1) our dilution calibration (Fig. 4e) confirmed linear SERS response across the working range, (2) dose-dependent spectral trends persisted after NTA normalization (Supplementary Figure 5), demonstrating that molecular content changes occur independent of EV abundance, and (3) normalization to the internal silicon substrate peak (642 cm⁻¹) corrects for local EV number variations at each measurement spot with high reproducibility (CV = 1.01%).

A correlation analysis identified six wavenumbers (739, 960, 1250, 1525, 1576, 1602 cm⁻¹) that strongly correlated with albumin and CCK8 viability assays (r > 0.94), predominantly reflecting nucleic acid and protein vibrational modes^49–57. This interpretation is consistent with prior reports identifying cellular and EV-derived miRNAs, including miR-122⁵⁸, miR-192⁵⁹, miR-193⁶⁰, as characteristic features of drug-induced liver injury. Protein-associated bands such as amide III further support molecular cargo changes beyond nucleic acids. The strong correlation with functional markers (albumin, CCK8) supports the biological relevance of the observed spectral changes. However, the current study was limited to a monoculture system with a single hepatotoxicant and only two biological replicates; these constraints necessitate cautious interpretation and broader validation.

In our study, due to the high dilution (1:10,000) of the samples, some measurements lacked EVs and they were excluded via a cluster analysis by filtering out the uncharacteristic spectral signature. To minimize batch effects, all measurements used substrates from a single batch and uniformity of the substrates were confirmed (See Fig. 4). Future translation will require validation across multiple batches, supported by machine learning–based correction or normalization methods ^61,62. The limited number of APAP doses introduced some non-linearity in regression predictions, but the overall MAE (1.495 mM) and minimal bias (0.03 mM) support robustness.

Translation to real-world settings will require validation across hepatotoxic compounds with distinct mechanisms, robustness testing in co-culture and organoid systems to better capture EV diversity, and enrichment of liver-specific EV subpopulations from complex biofluids. These steps will help determine whether the spectral features identified here are conserved across injury pathways and can support reliable, liver-specific toxicity readouts. While the label-free SERS platform enables rapid analysis without immunoaffinity labeling, extending this approach to heterogeneous in vivo EV populations will necessitate careful consideration of mixed cellular origins and spectral complexity. Future studies incorporating co-culture and organoid models, additional hepatotoxicants, and surface-marker–based EV enrichment will be important for evaluating specificity in clinically relevant contexts. Overall, this monoculture, single-compound study establishes foundational feasibility and highlights the utility of SERS–EV integration as a preclinical hepatotoxicity screening approach, providing a basis for broader evaluation across diverse biological systems.

Methods

Cell Culture and Analysis Materials

Two different cell culture media were used during culture. First, primary rat hepatocyte culture media (C + H) was prepared with high glucose (4.5 g/L) Dulbecco’s modified eagle’s medium (DMEM; Life Technologies, CA, United States) and was supplemented with 10% fetal bovine serum (FBS, Sigma, St. Louis, MO, United States), 2% penicillin-streptomycin, 7.5 μg/mL hydrocortisone, 20 ng/mL epidermal growth factor (EGF) and 14 ng/mL glucagon, was used for seeding and initial overnight maintenance. Hepatocyte Culture media, as defined for the rest of the protocol, is the HCM Hepatocyte Culture Medium BulletKit including HBM Basal Medium (CC-3199) and HCM Singlequot Supplements (CC-4182) (0.5 mL transferrin, 0.5 mL ascorbic acid, 0.5 mL HEGF, 0.5 mL insulin, 0.5 mL hydrocortisone, 10.5 mL BSA (Fatty Acid Free), and 0.5 mL GA-100). HCM Hepatocyte Culture Medium BulletKit including HBM Basal Medium (CC-3199) and HCM Singlequot Supplements (CC-4182) were purchased from Lonza Bioscience (Walkersville, MD). Hepatocyte Culture media was used after the first day of culture and for the duration of the experiment.

12 well tissue culture plates were purchased from CellTreat Scientific Products (Pepperell, MA, United States). GenClone 25–508 Dulbecco’s phosphate-buffered saline (PBS) was purchased from Genesee Scientific (Research Triangle Park, NC, United States). Trypan blue was purchased from Sigma Aldrich, United States. Geltrex^™ LDEV-Free Reduced Growth Factor Basement Membrane Matrix (A1413202), and 12 and 96 well plates were purchased from Invitrogen by Thermo Fisher Scientific (Carlsbad, CA, United States). Cell counting kit-8 (CCK8) was purchased from ApexBio (ab228554; Boston, MA, United States). The urea assay kit was purchased from Stanbio Laboratory (0580–250). Tween 20 (P2287), albumin from rat serum (A6414), o-phenylenediamine dihydrochloride (OPD) tablets (10 mg substrate per tablet) (P8287), hydrogen peroxide (216763), sulfuric acid (320501), dimethyl sulfoxide (DMSO) (472301), and acetaminophen (A7085) were purchased from Sigma Aldrich (St. Louis, MO, United States). Sheep anti-rat albumin HRP conjμgated (A110–134P) (1 mL at 1 mg/mL) was purchased from Fortis Life Sciences (Waltham, MA, United States).

Primary Rat Hepatocyte (PRH) Isolation

Primary rat hepatocytes (PRH) were freshly isolated from 10 to 12 weeks-old adult female (180–200 g) Lewis rats (Charles River Laboratories, United States). The Cell Resource Core (CRC) performed the isolation according to protocol #2011N000111, as approved by the Institutional Animal Care and Use Committee (IACUC) at Massachusetts General Hospital (MGH). After complete digestion of the liver, a yield of 400–600 million primary rat hepatocytes with greater than 85% viability was given for use. Cell count and viability were provided by quantification with a hemocytometer count after trypan blue staining.

2D Rat Hepatocyte Monolayer Culture and Acetaminophen Dosage

Prior to cell seeding, 12-well tissue culture plates were coated with 500 μL 1% Geltrex in PBS, and incubated at 37°C in a humidified environment containing 5% CO₂. Next, the remaining solution was aspirated and the well was washed with PBS. The PRH cell suspension was diluted to 560,000 viable cells/mL in C+H medium and 1 mL was added to each well. After introduction of the hepatocyte suspension to the wells, the plate was shaken gently in an alternating horizontal and vertical pattern to uniformly distribute the cells within the wells. The plates were then incubated for 45 minutes at 37°C with 5% CO₂. Following the initial incubation, the medium was aspirated and each well washed with warm PBS to remove dead or non-adherent cells. Finally, 1 mL of C+H medium was added. After 24 hours, the media was aspirated and cells were washed with PBS. For the second 24 hours of culture, 1 mL of Hepatocyte Culture medium with 2% Geltrex by volume (2% Geltrex-media) was added to each well.

The 24 hours after the introduction of 2% Geltrex-media, media from each well was collected and stored at −80°C. These samples were later unfrozen for analysis and divided by volume based on experimental usage. The 1 mL samples were split into two parts, a 600 μL aliquot for exosome analysis and 400 μL for functionality assay completion. Following collection of media, a CCK8 assay was completed.

Following absorbance readings, the media from each well was aspirated and washed with PBS. The well plate was then split into four groups and introduced to HCM-2% Geltrex media with either control (no added acetaminophen), 10mM, 20 mM, or 40 mM acetaminophen, diluted from a 2M stock of acetaminophen in DMSO. 48 hours after initial dosage (Day 2), the collection and CCK8 dosage was repeated, and the cells were dosed with acetaminophen-media again.

CCK8 Assay and Brightfield Imaging

After collection, each well was washed with PBS before 500 μL of Hepatocyte Culture media was added for use in the CCK-8 viability assay. 50 μL of CCK-8 were then added to 8 of the 12 wells and incubated for 1 hour at 37°C with 5% CO₂. After incubation, triplicate samples of 80 μL were pipetted into a 96-well plate and their absorbance values were read on a microplate reader (Molecular Devices, SpectraMax^® ID3) at 450 nm.

Following CCK8 assay and dosing with acetaminophen-media each day, each well was imaged using an EVOS^™ M5000 microscope. Morphological changes were used to qualitatively validate the other viability and functional assay results.

Albumin Quantification

Albumin secretion in PRH cultures was determined using an enzyme-linked immunosorbent assay (ELISA) protocol. The following solutions were prepared prior to beginning the assay: substrate buffer (5.1 g citric acid and 7.29 g sodium phosphate dibasic in 1 L distilled water), 8N sulfuric acid (22.2 mL concentrated sulfuric acid in 77.8 mL DI water), and PBS-Tween 20 solution (0.05 v/v% Tween 20 in PBS).

96 well plates were coated with 100 μL of rat albumin at a concentration of 50 μg/mL. The plate was then covered with an adhesive plate sealer, and incubated overnight at 4°C. After incubation, plates were washed four times with PBS-Tween solution. After washing, 50 μL of sample or standard were added in triplicate. Standards were created by serially diluting 100 μg/mL albumin and added to each individual plate. Following introduction of samples/standards, 50 μL of antibody solution, stock antibodies diluted 1:10000 in PBS-Tween, was added to each well. The plate was then covered with an adhesive plate sealer and incubated for 90 minutes at 37°C.

At the end of the incubation, the plate was washed again with PBS-Tween solution four times. Subsequently, 100 μL of OPD development solution, a mixture of one pill of OPD, 25 mL substrate buffer, and 10 μL of 30% H2O2, was introduced via multichannel pipette at set intervals. After 5 minutes, the reaction was quenched with the addition of 50 μL of 8N sulfuric acid at the same interval as previously before measuring the difference between 490 and 650 nm on a microplate reader (Molecular Devices, SpectraMax^® ID3).

Urea Quantification

Urea excretion of PRH cultures was quantified using a Stanbio urea BUN assay kit following the manufacturer’s protocol. The urea assay reagent was prepared by mixing one part BUN color reagent and two parts BUN acid reagent. Next, 10 μL of either standard or sample are pipetted in duplicate into a 96 well plate, and combined with 150 μL of urea assay reagent. The plate was then covered with a plate sealer and incubated for 90 minutes at 60°C. After removal from the incubator, the plate was cooled at room temperature for 10 minutes. The plate was finally read on a microplate reader (Molecular Devices, SpectraMax^® ID3) at 540 nm.

Extracellular Vesicle Isolation

All EV isolation steps were performed in a cold room to prevent gelation, which could reduce the number of samples and lead to larger size distributions due to aggregation. Supernatants from triplicate samples were pooled to increase the yield of EVs from each sample. The total volume was brought to 5 mL by diluting with PBS. The diluted sample was then processed using our ExoTIC platform⁴² to select nanoparticles within the 50 to 220 nm size range. After the initial isolation, a second wash step was conducted with 5 mL of PBS. This platform is now well established for yield and intact EV isolation outcomes in various applications including cancers ^63–66, cardiovascular diseases^67–69, and immunology ⁷⁰.

Extracellular Vesicle Characterization

We characterized EVs using transmission electron microscopy (TEM), nanoparticle tracking analysis (NTA), flow cytometry, interferometric microscopy and western blotting. For TEM, 10 μl EVs were loaded onto the grid (Electron Microscopy Sciences – EM Sciences FCF-300-CU), and incubated with 1% uranyl acetate before imaging with a JEOL JEM 1400 120 kV TEM microscope (JEOL Inc., USA) equipped with a Ultrascan digital high-resolution camera (Gatan Inc., Pleasanton, CA). NTA NanoSight equipped with 430 nm laser (NS300, Malvern Panalytical, UK) quantified and profiled EVs in terms size and concentrations. Bead-assisted flow cytometry was performed to further characterize EVs, using 4% w/v, 4 μm aldehyde-sulfate latex beads (A37304, ThermoFisher, Carlsbad, CA, United States), and staining with the EV tetraspanin marker CD63 (#556019, BD Pharmingen, San Jose, CA, USA) antibody. Western blot analysis of EVs was performed using Calnexin (#PA1–30197, Invitrogen, Carlsbad, CA, United States), Flotillin-1 (#A6220, Abclonal, Woburn, MA, USA), TSG101 4A10 (#NB200–112, Novus Bio, Centennial, CO, USA) and CD63 (#556019, BD Pharmingen, San Jose, CA, USA) antibodies, according to the insights provided in MISEV 2023 guidelines ²⁴.

Interferometric Imaging:

To analyze the concentration and size distribution of Extracellular Vesicles (EVs), a custom-built interferometric microscope is employed^71,72. Our interferometric imaging enables the visualization of nanoscale particles by leveraging the interference between the scattered light from the nanoparticles and the background light reflected from the sample substrate. This interference generates a signal on the camera plane that correlates with the particle size. In this study, EVs are immobilized on a silicon (Si) substrate coated with a 100 nm-thick SiO₂ thin film.

The substrate is illuminated in a widefield configuration using a 530 nm LED light source (M530L4, Thorlabs, Newton, NJ, USA), which is focused onto the back focal plane of a 50x objective lens (LMPlanFL, Nikon Corporation, Tokyo, Japan). The reflected and scattered light are collected through the same objective lens and subsequently imaged onto a CMOS camera (acA2440, Basler AG, Ahrensburg, Germany). For each measurement, a sample spot of 0.2 μL is deposited using a standard pipette and allowed to dry for 15 minutes to immobilize the particles.

The sample spot is then scanned, and z-stacks of interferometric images are acquired using a motorized sample stage. The best-focused images are selected based on the particle intensity in the z-stack. Particle detection and counting are performed using a custom MATLAB script. The total particle count in the interferometric images is converted into a concentration (particles/mL) by accounting for the sample volume.

Surface Enhanced Raman Spectroscopy

Surface-enhanced Raman spectroscopy (SERS) experiments were conducted using a custom-built setup with a 785 nm, 100 mW diode laser (CrystaLaser, Reno, NV, USA). The laser output power was reduced to 5 mW using a half-wave plate and a polarized beam splitter. A 10x microscope objective (0.25 NA) focused the beam on the sample. The Stokes-shifted backscattered photons were separated using a dichroic mirror (LP805, Thorlabs, Newton, NJ, USA) and coupled into a 0.22 NA multimode fiber (M105L02S-B, Thorlabs, Newton, NJ, USA) connected to the Raman spectrometer (HyperNova, StellarNet, USA) with a thermoelectrically cooled (−60°C) CCD camera (Andor Technology, Belfast, Northern Ireland).

Isolated EV samples were diluted 10⁴ times (see dilution response of SERS substrate in Fig. 4) with distilled water and drop-casted on Au nanopillar-coated SERS substrates (Silmeco, Copenhagen, Denmark). The fabrication parameters and characteristics of these surfaces have been earlier reported ⁴⁵. A 1.3 μL drop of each sample was placed on a hydrophobic surface and dried for 15 minutes. The particles attached to gold nanopillars were measured with an exposure time of 1 second. To avoid overlooking sparsely distributed EVs in lower-concentration samples (particularly the 40 mM APAP day 2 group), we expanded our scanning strategy to cover a larger surface area. Specifically, we performed a 31 × 31 grid, resulting in 961 total measurement points, each separated by a 5 μm step size. This approach maximized coverage of the sample surface and ensured that lower-density EV populations were captured and included in subsequent data analysis.

Blank measurements of distilled water on the SERS substrate were performed before each experimental session to establish baseline signals and ensure no contamination was present. These blank measurements showed minimal spectral features (Fig. 4d), confirming that the observed spectral signatures are EV-derived rather than from the suspension medium.

Statistical Analysis

in vitro Essay Data Analysis

Graphpad Prism 10 (Boston, MA) was used to analyze and graph data. All data is normalized to its corresponding well in the plate along the time course of the experiment. For each day of measurement, each well was also normalized to the average viability of the control wells. All data is presented as the mean ± standard deviation of the collected wells. IC₅₀ of CCK8 and albumin graphs was determined by fitting the data to a non-linear regression model to create an inhibition dose response curve (concentration of inhibitor versus normalized response) with variable slope. Specifically, Graphpad Prism fit data to the following model: Y=100/(1+(IC₅₀ /X)ĤillSlope), where Y is the normalized viability, X is the concentration of the inhibitor, in this case acetaminophen. IC₅₀ and Hill Slope are calculated and output as a results of least squares regression.

SERS Spectral Preprocessing

All SERS spectra were first read from individual text files and organized so that each sample’s measurements were stored as columns of a common matrix. Next, pixel indices were converted to wavenumbers by fitting the positions of known Raman peaks to a polynomial and applying the resulting polynomial across all pixel channels using the known peak positions of polystyrene⁷³. This ensured consistent mapping of each signal channel to its true Raman shift.

To reduce extraneous noise or non-informative regions, the data were cropped to a specified range of pixel (or wavenumber) indices, removing the edges where the signal was weak or prone to artifact. A reference substrate spectrum was then identified by calculating the area under the curve (AUC) of each individual spectrum and selecting the one with the minimum overall AUC. This reference was used in the subsequent baseline-correction step.

Baseline correction followed the asymmetric least squares (ALS) approach. Each spectrum was processed by estimating a baseline with defined smoothness and asymmetry parameters, which was then subtracted from the raw intensities. The corrected spectra were further smoothed using a Savitzky-Golay filter, chosen to preserve spectral features while reducing high-frequency noise. The smoothed spectra were normalized to the peak at 642 cm⁻¹ which arises from the Si substrate’s second order phonon vibrations (2LA)⁷⁴ and gives a signature independent of the EV cargo that serves as an internal standard commonly used by others⁶¹.

Statistical power was limited by the availability of primary hepatocyte isolations from independent animals (n=2 biological replicates). Each biological replicate included technical replicates (triplicate wells per condition) and multiple SERS chips (2–4 per condition) to assess within-experiment variability. Cross-validation between Set 1 and Set 2 demonstrates reproducibility of the dose-response model across independent experiments.

Due to the high dilution, some measurement sites did not contain extracellular vesicles (EVs) and only recorded surface responses. These blank measurements were filtered out during preprocessing using k-means cluster analysis. The measurements were divided into five groups, and the groups whose signatures did not contain the predefined EV peaks were eliminated.

Exploratory Analysis and Linear Regression on Raman Dataset

We used the spectral range from 810–1610 cm⁻¹ for subsequent analyses and excluded wavenumbers near the baseline region to avoid the high variability typically observed in those segments, thereby focusing on the more stable and informative portions of the spectra. Principal Component Analysis (PCA) was employed to reduce the dimensionality of the Raman spectral data and identify patterns corresponding to different concentration levels. Preprocessed spectral data, normalized to minimize variations due to instrumental factors, served as input for PCA. The analysis focused on the first two principal components (PC1 and PC2), which explained the majority of the variance in the datasets. PCA was applied separately to data from Day 0 and Day 2 for both sets to assess differences in clustering over time. Scatter plots of PC1 versus PC2 were used to visualize the separation between concentration groups, with the percentage of variance explained by each component indicated in the plots. All PCA computations and visualizations were performed using the built-in functions in Matlab (MathWorks, Natick, MA, USA).

The baseline-corrected Raman spectra were used to construct a regression model utilizing Gaussian Process Regression (GPR) to predict concentration levels. The dataset comprised spectra for four different concentrations: 0 mM, 10 mM, 20 mM, and 40 mM. For each concentration group, balanced number of spectra were randomly selected to form the training set.

The spectra were split into predictor variables (intensity values) and response variables (concentration levels). The linear regression model was trained using the built-in fitlm function in MATLAB which performs the linear regression operation with several options. Predictions were made using the same dataset, and the RMSE was calculated. The process was iterated until convergence was achieved. The final RMSE and residuals were calculated and plotted, with predictions visualized using box plots to compare predicted concentrations against actual concentrations. All data visualizations were made using the built-in tools in Matlab. The exported plots were arranged and finalized in Inkscape (Version 1.3, available at https://inkscape.org).

Limit of detection and quantification calculation

Spectral calibration was performed using a third-order polynomial fit to align pixel indices with known wavenumber standards. A calibration curve was established by plotting the peak intensity against the exosome concentration (verified by NTA) on a double-logarithmic scale. A power law regression model ($y=a{x}^b$) was fitted to the data within the linear dynamic range (dilution factors 10⁴ to 10⁸).

The Limit of Detection (LoD) and Limit of Quantification (LoQ) were determined based on the statistical variability of the blank sample. The intensity thresholds were defined as the mean intensity of the blank plus three times its standard deviation $\left({\mu }_{\mathit{blank}}+3{\sigma }_{\mathit{blank}}\right)$ for LoD, and ten times the standard deviation $\left({\mu }_{\mathit{blank}}+10{\sigma }_{\mathit{blank}}\right)$) for LoQ. These threshold intensities were then converted into concentration values (particles/mL) using the inverse of the power law regression equation:

$$x={\left(\frac{y}{a}\right)}^{\frac{1}{b}}$$

where $\mathrm{y}$ is the intensity threshold, $\mathrm{x}$ is the calculated concentration, and $\mathrm{a}$ and $\mathrm{b}$ are the fitting parameters derived from the regression model.

Correlation Matrix, Error and Reproducibility Calculations

The correlation matrix was calculated by determining the Pearson correlation coefficient between each pair of variables. This coefficient, $r_{xy}$, is given by the equation:

$$r_{xy}=\frac{\sum \left(x_i-\bar{x}\right)\left(y_i-\bar{y}\right)}{\sqrt{\sum {\left(x_i-\bar{x}\right)}^2\sum {\left(y_i-\bar{y}\right)}^2}}$$

where $x_i$ and $y_i$ are individual data points, and $\bar{x}$ and $\bar{y}$ are the sample means.

Regression error parameters overall RMSE, Per-concentration RMSE, Overall MAE and coefficient of determination (R²) was calculated using the equation below:

I. Overall RMSE

$$RMS{E}_{\mathit{all}}=\sqrt{\frac{\sum _{i=1}^N{\left({\mathit{Predicted}}_i-{\mathit{Actual}}_i\right)}^2}{N}}$$

II. Per-concentration RMSE

$$RMS{E}_c=\sqrt{\frac{\sum _{i=1}^N{\left({\mathit{Predicted}}_i-{\mathit{Acual}}_i\right)}^2}{N_c}}$$

III. Overall MAE

$$MA{E}_{all}=\frac{1}{N}\sum _{i=1}^N\left|y_i-{\hat{y}}_i\right|$$

IV. Coefficient of determination (R²)

$$R^2=1-\frac{\sum _{i=1}^N{\left(y_i-{\hat{y}}_i\right)}^2}{\sum _{i=1}^N{\left(y_i-{\bar{y}}_i\right)}^2}$$

Reproducibility score was calculated using the coefficient of variation (CV) along the chips via the formula below:

$$\mathrm{C}\mathrm{V}=\frac{\mathrm{standard}\mathrm{deviation}}{\mathrm{mean}}\times 100$$

All calculations were performed using Matlab’s Statistics Toolbox.

Highlights.

• Novel quantitative dose-resolved EV-SERS platform for hepatotoxicity screening

• SPECTRA platform allows for rapid, label-free extracellular vesicle (EV) analysis for detection of hepatotoxicity.

• Delivers result within <30 minutes from as little as 1.3 μL sample.

• Precisely estimates acetaminophen (APAP) dose with RMSE down to 1.5 mM.

Biographies

Biographies

Ugur Parlatan, PhD

Basic Life Research Scientist, Bio-Acoustic MEMS in Medicine (BAMM) Lab, Canary Center at Stanford, Department of Radiology, Stanford University. He received his PhD in Physics in 2017 from Istanbul Technical University. His research interests include Raman spectroscopy, optical tweezers, spectroscopic imaging, extracellular vesicle molecular fingerprinting, data-driven diagnostics, and applying optical techniques for disease detection.

Luke Boudreau

Research Technician at the Center for Engineering in Medicine and Surgery, Massachusetts General Hospital / Harvard Medical School. He received his Bachelor of Science, double major in Chemical Engineering and Biochemistry, in May 2022. His interests lie in microfluidic systems, translational medicine, and engineering tools for diagnostics.

Hulya Torun

Postdoctoral Researcher, Bio-Acoustic MEMS in Medicine (BAMM) Lab, Canary Center at Stanford University. She received her PhD from Koc University (Istanbul, Turkey) in Biomedical Sciences and Engineering in 2024. Her work focuses on extracellular vesicle biology, nanotechnology-based sensing, imaging, and optical methods for biological systems.

Letao Fan

Postdoctoral researcher at Massachusetts General Hospital / Harvard Medical School, Center for Engineering in Medicine & Surgery. He earned his PhD in Experimental Therapeutics and Pharmacology from University of Mississippi Medical Center. His research focuses on modeling, organ-on-chip systems, and translational tissue engineering.

Ugur Aygun, PhD

Postdoctoral Scholar, Radiology, Stanford University (Canary Center / BAMM Lab). He received his PhD in Electrical Engineering from Koç University in 2020. He holds two Bachelor’s degrees from Istanbul Technical University (Electronics Engineering and Physics) obtained in 2013. His research interests include optical biosensors, optical microscopy, computational imaging, spectroscopy, and development of novel imaging techniques for biomedical applications.

Ayse Aslihan Gokaltun

Research Scientist, Center for Engineering in Medicine & Surgery, Massachusetts General Hospital / Harvard Medical School. She earned her MS and PhD from Hacettepe University Chemical Engineering Department in 2009 and 2014. Her research focuses on biomaterials, organ-on-chip platforms, EVs, tissue modeling, and translational medicine.

Demir Akin, PhD

Senior Researcher, BAMM Lab, Stanford University. He earned his PhD from Purdue University Comparative Pathobiology. His focus includes Pathobiology-Virology, bioMEMS, computational biology, nanomedicine, microfluidics, and diagnostics leveraging extracellular vesicles and spectroscopy.

O. Berk Usta, PhD

Assistant Professor, Center for Engineering in Medicine & Surgery, Massachusetts General Hospital / Harvard Medical School. He took his B.S. in Chemical Engineering from Bogazici Unversity in 2000 and his PhD in Chemical Engineering from University of Florida in 2007. He has been a postdoc in Harvard Medical School Surgery department between 2009 and 2013. His research interests include microfluidic device development, cell-based therapies, EVs, and diagnostic platforms.

Utkan Demirci, PhD

Professor of Radiology (with tenure) and by courtesy of Electrical Engineering, at Stanford University School of Medicine. Interim Division Chief and Director of the Canary Center for Early Cancer Detection. He earned his B.S. in Electrical Engineering, from University of Michigan, Ann Arbor in 1999, his M.S. in Electrical Engineering from Stanford in 2001, a second M.S. in Management Science & Engineering in 2005, and his Ph.D. in Electrical Engineering also in 2005, from Stanford University. His research spans micro/nanoscale technologies, microfluidics, extracellular vesicle diagnostics, point-of-care devices, global health, and early detection of disease.

Data availability

All data supporting the findings of this study are available within the paper and its Supplementary Information. The custom code used to generate Figures 4 and 5 has been deposited on Zenodo with the identifier https://doi.org/10.5281/zenodo.14768753⁷⁵