An artificial intelligence-enhanced early ovarian cancer diagnosis biosensor

Abstract

In early cancer diagnosis, extracellular vesicles (EVs) are more advantageous than circulating tumor cells due to their smaller size, greater stability, and enhanced tissue penetration. These qualities lead to higher EV concentrations in body fluids, facilitating early detection. This study leverages surface-enhanced Raman scattering (SERS) for EV detection, employing a novel biosensor made with a molybdenum disulfide (MoS₂) composite film on silicon and demonstrating a lower limit of detection (LOD) and multi-marker synchronous quantitative testing performance compared to existing methodologies. This biosensor efficiently measures EV concentrations and precisely detects three specific proteins on ovarian cancer EVs simultaneously (CD63, CD24, and CA125). Using the ovarian cancer cell line HO8910, the sensor demonstrated a detection limit of 1.4 × 10⁴ particles per mL and a wide linear range of 3.4 × 10⁴ particles per mL to 3.4 × 10⁸ particles per mL. It also effectively discriminated between serum samples from healthy individuals and ovarian cancer patients at different stages. Additionally, machine learning was applied to analyze detection data, resulting in a diagnostic model with a 97.78% prediction accuracy. This highlights the sensor's potential in revolutionizing early cancer detection and establishing new diagnostic models.

---

1. Introduction

Cancer, as the most significant 'health killer' for humanity, poses a severe threat to human life and health. Early diagnosis and screening for cancer are of utmost urgency. Compared to traditional biomarkers, extracellular vesicles (EVs) are considered promising novel tumor markers.^1,2 EVs are nanoscale vesicles secreted by cells, containing various biomolecules such as proteins, nucleic acids, and lipids.³ The membrane surface of EVs also carries protein markers that are indicative of cancer-related information.⁴ EVs play an important role in the occurrence and development of diseases, and their concentration is one of the indicators reflecting the degree of tumor disease progression.⁵ Differential expression of EV membrane proteins is an important basis for distinguishing EVs from different sources and studying EV origins. Therefore, accurate characterization of tumor EVs, especially the high-specificity and high-sensitivity detection of membrane proteins, is crucial for early screening of tumor based on EVs.⁶ Currently, solid biopsy remains the gold standard for pathological diagnosis and is the basis for treating most cancers. However, as an invasive detection method, solid biopsy has certain risks of trauma and infection.⁷ Liquid biopsy, as an emerging non-invasive diagnostic technique, can diagnose diseases more quickly and accurately. Therefore, EVs in cancer patients' body fluids are considered a highly potential target for liquid biopsy.⁸ For example, Shin et al. achieved highly sensitive detection of lung cancer via liquid biopsy based on EV membrane proteins.⁹

Nanomaterials¹⁰ generally refer to ultrafine particle materials composed of nanoparticles with sizes ranging from 1 to 100 nm. They exhibit distinct properties compared to bulk materials. In recent years, researchers have incorporated nanomaterials into biosensors, particularly in the field of EV detection, to develop highly sensitive and rapid detection biosensors. Firstly, nanomaterials with high surface areas and porous structures, such as graphene¹¹ and molybdenum disulfide (MoS₂),¹² can provide more active binding sites for capturing antibodies. Combined with specific antibodies targeting EV membrane proteins, nanomaterials can be used for isolation, purification, and sensitive detection of EVs.¹³ Secondly, nanomaterials can be used for EV labeling. For example, by utilizing the surface modification of gold nanoparticles and specific antibodies recognizing EV membrane proteins, efficient probes¹⁴ for recognizing EV membrane proteins can be prepared to investigate their expression. Some two-dimensional nanomaterials have demonstrated excellent physicochemical properties and have shown great application prospects in biosensing, such as MoS₂ and graphene.^15,16

Traditional methods for EV detection, such as nanoparticle tracking analysis (NTA), dynamic light scattering (DLS), flow cytometry, and transmission electron microscopy (TEM),^17–19 suffer from limitations including complex sample preparation, cumbersome operation, high equipment cost, and limited sensitivity, making them less suitable for high-throughput, low-cost, and rapid early screening. Surface-enhanced Raman scattering spectroscopy (SERS), as a fast and non-destructive non-contact optical detection method, amplified Raman signals through “hotspots” formed by plasma^20,21 or electromagnetic enhancement of gold, silver and other particles to achieve highly sensitive detection. It has been widely used in quantitative detection of EVs and exploration of the expression of EV surface mask proteins, and a series of significant advances have been made.^22–24 Meanwhile, researchers have demonstrated innovation by developing a series of biosensors incorporating nanomaterials.^25–27 These advanced biosensors offer high-sensitivity measurements of EV concentrations and detailed analyses of EV membrane protein expression, marking significant progress towards early cancer diagnosis. Furthermore, the integration of machine learning and Raman spectroscopy has revolutionized the analysis and interpretation of molecular structural data. Machine learning²⁸ enhances data analysis efficiency, converting vast datasets into actionable insights, while Raman spectroscopy provides detailed, high-resolution spectral data.^29,30 Together, these technologies improve the accuracy of EV detection and analysis, facilitating the optimization of algorithms and contributing to the advancement of diagnostic methodologies.³¹

In this work, to achieve high-sensitivity detection of the EV concentration and multi-channel analysis of surface membrane proteins, we introduced MoS₂,³² a nanomaterial with a high specific surface area and excellent biocompatibility, and prepared a SERS biosensor capable of detecting multiple surface membrane proteins of ovarian cancer EVs simultaneously. The sensor was used to detect the expression of multiple membrane proteins on EVs from the ovarian cancer cell line HO8910, and the lowest detection limit reached 1.4 × 10⁴ particles per mL, with a linear range of 3.4 × 10⁴ particles per mL to 3.4 × 10⁸ particles per mL. Expression differences in surface membrane proteins among different cell lines were also evaluated. Furthermore, the sensor was used to detect EVs in serum samples from healthy individuals and early and late-stage ovarian cancer patients, and the expression levels of surface membrane proteins were statistically compared. Finally, machine learning was introduced to analyze and classify the detection data, and a cancer diagnostic classification model was established with an overall prediction accuracy of 97.78%, demonstrating the enormous potential of the sensor in early cancer diagnosis (Fig. 1).

Fig. 1 Illustrative diagram of a surface-enhanced Raman scattering biosensor based on the MoS₂ composite film for ovarian cancer detection and diagnosis.

2. Experimental section

2.1. Reagents and instruments

Reagents including silicon wafers and MoS₂ powder were sourced from ZHONGXINQIHENG and Sigma-Aldrich, respectively, with a purity of 99%. Sodium deoxycholate and sodium chloride (NaCl), each with purities of 99% and 99.5%, were obtained from Alfa Aesar and the Chengdu Kelong Chemical Reagent Factory. Poly(diallyldimethylammonium chloride) (PDDA) and 5,5′-dithiobis(2-nitrobenzoic acid) (DTNB) were acquired from Shanghai Aladdin, with concentrations of 35% and 97.74%, respectively. Additional compounds, including acetone and isopropanol, were obtained from Sinopharm Chemical Reagent, both at analytical grade. Chloroauric acid (HAuCl4, 48–50% Au basis) was purchased from Macklin, while sodium citrate (99%) was obtained from Alfa Asia. Argon gas was supplied by Nanjing Special Gas Co., Ltd. Further reagents such as 1-(3-dimethylaminopropyl)-3-ethylcarbodiimide hydrochloride (EDC) and N-hydroxysuccinimide (NHS) were sourced from Shanghai Aladdin, with a purity of 98%. Bovine serum albumin (BSA, 96%) and anti-AFP antibodies were supplied by Beijing Bioss. Phosphate-buffered saline (PBS) was procured from Langeco Technology Co., Ltd. Ovarian cancer cell lines HO8910 and SKOV3 were supplied by Nanjing Zhongda Hospital, while anti-EpCAM, anti-CD63, anti-CD24, anti-CD9, anti-CA125, and anti-EGFR antibodies were sourced from Beijing Bioss. TFMBA (≥98%) and 4-MBN (≥95%) were obtained from Macklin and Dibai, respectively. Anhydrous ethanol, of analytical grade, was provided by Sinopharm Chemical Reagent. All reagents were used as received, ensuring high-quality experimental results.

The experimental instruments used include a cell crusher (1500YC) from Nanjing Emmanuel and an ultrapure water system (UPT-II-10T) from Sichuan Youpu. Weighing measurements were conducted using an electronic balance (ME104E) from Mettler Toledo (Switzerland). Cleaning procedures were carried out using an ultrasonic cleaner (KQ-400DE) from Kunshan Ultrasonic Instruments, and high-speed centrifugation was performed with a Heraeus Multifuge ×1 (Thermo Fisher, USA). For surface treatments, a plasma gun from Femto Science was employed. UV-vis spectroscopy was conducted using a UV-1600PC from Shanghai Mapada (MAPADA), while confocal Raman spectroscopy was performed with an Alpha 300 Raman-AFM from Witec (Germany). High-resolution imaging was achieved with a transmission electron microscope (Talos F200X) and a scanning electron microscope (FEI Inspect F50), both from Thermo Fisher. Cell culture was performed using DMEM medium sourced from GIBCO, and filtration was carried out with 0.22 μm filters from Merck Millipore Ltd. Ultracentrifuge tubes were provided by Beckman Coulter, and nanoparticle tracking analysis was conducted using a ZetaView PMX 110.

2.2. Preparation of MoS₂ dispersion solution and PDDA solution

Weigh 2.625 g of MoS₂ powder and 1.125 g of sodium cholate powder. Add them together into a 100 mL beaker containing 50 mL of deionized water. Stir the mixture rapidly with a glass rod. Clean and dry the ultrasonic disruptor probe with deionized water. Place the beaker with the MoS₂ mixture solution in a larger beaker containing a certain amount of water, which can dissipate the large amount of heat generated during ultrasonic treatment more effectively. Cover the top of the small beaker with a sealing film, leaving only a small hole in the middle for the ultrasonic probe to be placed, in order to prevent excessive evaporation of the solution during ultrasonic treatment. Place the ultrasonic probe deep into the solution. Set the power parameter of the disruptor probe to 750 W, the amplitude to 75%, the ultrasonic interval time to 1 second, and the ultrasonic time to 18 hours. Set the alarm temperature of the solution during ultrasonic treatment to 70 °C. After ultrasonic treatment, let the solution stand for 12 hours. After standing, take 30 mL of the upper solution, divide it into two equal parts, and transfer them to two centrifuge tubes using a pipette. Place the two centrifuge tubes symmetrically in a centrifuge and set the centrifugation speed to 10 000 rpm for 30 minutes. After centrifugation, carefully remove and discard the upper clear liquid using a pipette. Then, add 1 mL of a 1 mg mL⁻¹ polyvinylpyrrolidone solution to the centrifuge tube and mix it thoroughly by shaking. Place the tube in an ultrasonic cleaner and treat it with 90% power for 10 minutes to obtain a MoS₂ dispersion with an ink-like appearance. Store the prepared MoS₂ dispersion solution in a refrigerator at 0–4 °C for later use. Dissolve a 35% (mass fraction) PDDA solution in a 0.1 M NaCl solution to obtain a final PDDA solution with a mass fraction of 0.1%.

2.3. Preparation of the MoS₂ composite film

A rectangular area of 4 mm × 6 mm was defined on the Si substrate, and the silicon wafer was ultrasonically cleaned with acetone, isopropyl alcohol and deionized water for 10 minutes each. A 120 W argon plasma gun was used to pretreat the surface of the silicon wafer for 1 minute to make the silicon wafer negatively charged. Then, part of the silicon wafer was immersed in the PDDA solution for 5 minutes and rinsed with deionized water to remove excess PDDA. After that, the silicon wafer was immersed in the MoS₂ dispersion solution for 5 minutes and rinsed with deionized water to remove excess MoS₂. This assembly process was repeated several times to obtain a MoS₂ composite film on the silicon wafer surface.

2.4. Preparation of gold nanoparticle solution

Take a large beaker and add 200 mL of deionized water. Heat the solution to boiling and add 1.36 mL of a 50 mg mL⁻¹ HAuCl4 solution. Stir the solution with a magnetic stirrer at a speed of 1000 rpm. After the solution comes back to a boil, quickly add 20 mL of a 10 mg mL⁻¹ sodium citrate solution. Continue heating and stirring. When the solution turns deep red, continue boiling for 15 minutes to ensure complete reaction of the remaining reagents. After the reaction is complete, stop heating and stirring, and allow the solution to cool to room temperature. Seal the beaker with the sealing film and store it in a refrigerator at 0–4 °C for later use.

2.5. Preparation of Raman tags

First, concentrate the prepared gold nanoparticle solution to a final concentration of approximately 1 nM. Then, take 1 mL of the gold nanoparticle solution and add 50 μL of a 1 mM DTNB ethanol solution. Mix the solution well and incubate it in a refrigerator for 4 hours to allow DTNB to spontaneously adsorb onto the surface of the gold nanoparticles through Au–S covalent bonding. Centrifuge the mixture at 7500 rpm for 10 minutes, and dissolve the precipitate obtained after centrifugation in 1 mL of a 0.1 mM PBS solution. Add 10 μL of a 0.1 mg mL⁻¹ recognition antibody solution to 1 mL of the Au@DTNB solution and incubate it in a refrigerator for 4 hours. Then, centrifuge the mixture at 7500 rpm for 10 minutes, remove the supernatant, and dissolve the bottom red precipitate in 200 mL of 0.1% BSA solution. Occasionally, some black precipitate may also be observed in addition to the red precipitate. The formation of this black precipitate is mainly due to the high concentrations of Na⁺ and Mg²⁺ ions in the solution, which cause aggregation of gold nanoparticles.³³ Store the final solution in a refrigerator at 0–4 °C overnight for later use. This immune recognition probe can be stored stably in the refrigerator for several months.

2.6. Sandwich-type immunosensor structure

On the surface of the prepared MoS₂ film, add 25 μL of a 0.2 M EDC/NHS PBS solution and let it stand at room temperature for 1 hour to activate the carboxyl groups on the surface of the MoS₂ film. After activation, wash the film surface 2–3 times with deionized water and dry it under a slow stream of nitrogen gas. On the activated MoS₂ film surface, add 25 μL of a 10 μg mL⁻¹ anti-EpCAM PBS solution and then incubate it in a refrigerator at 0–4 °C for 12 hours. After incubation, remove the film, wash the surface 2–3 times with deionized water, and dry it under a slow stream of nitrogen gas. Prepare a 1% BSA solution and add 25 μL of the BSA solution to the MoS₂ film surface. Close the film at room temperature for 1 hour to fully block the unreacted active groups on the film surface, preventing non-specific adsorption. After blocking, wash the film surface 2–3 times with deionized water and dry it under a slow stream of nitrogen gas. Add 25 μL of the EV solution to the MoS₂ film surface and incubate in a refrigerator at 0–4 °C for 4 hours. After incubation, wash the film surface 2–3 times with deionized water and dry it under a slow stream of nitrogen gas. Add 25 μL of the prepared AuNP@DTNB@Anti-CD63 composite immune recognition probe to the MoS₂ film surface and incubate in a refrigerator at 0–4 °C for 2 hours. After incubation, wash the film surface 2–3 times with deionized water and dry it under a slow stream of nitrogen gas. At this point, a sandwich structure is obtained, which can be stored at room temperature until Raman detection. The serum samples used in the experiment were provided by Zhongda Hospital, Affiliated to Southeast University, and informed written consent was obtained from participants before the study began. The study was approved by the IEC of Clinical Research of Zhongda Hospital, Affiliated to Southeast University, No. 2020ZDSYLL307-P01/2021.02.03.

2.7. Raman detection

Raman detection, shown in Fig. S1, was performed at an excitation wavelength of 532 nm and a laser power of 1 mW and the integration time was 2 seconds.³⁴

Raman detection, shown in Fig. 3e, was performed at an excitation wavelength of 633 nm. A 50 μm × 50 μm area on the film was selected for surface scanning. The laser power was set to 1 mW. The scanned area was divided into 10 × 10 small regions, and the integration time for each region was 2 seconds.

All other Raman detection experiments were performed at an excitation wavelength of 633 nm, and 4–5 scanning areas of 40 μm × 40 μm were randomly selected on the MoS₂ composite films for surface scanning processing. The laser power was 1 mW. The scanning range was divided into 20 × 20 small regions, and the integration time of each small region was 2 seconds.

2.8. NTA detection

First, the sample chamber was cleaned with deionized water, followed by calibration of the nanoparticle tracking analyzer using 100 nm polystyrene microspheres. The chamber was then rinsed with PBS. A 10 μL EV solution in PBS was diluted 20 times with standard PBS solution and loaded into the instrument. NTA was performed under experimental conditions of an excitation wavelength of 488 nm, a temperature of 25 °C, a pH of 7, and a conductivity of 11 μS cm⁻¹. A total of seven measurement points were set, with one measurement taken at each point.

2.9. Statistical analysis

To determine the sensor sensitivity, test results indicated that the sensor response was nearly zero at an EV concentration of 3.7 × 10³ particles per mL or below, suggesting that EVs at this concentration or lower could not be detected. Consequently, the detection range of the sensor was established to be 3.7 × 10⁴ particles per mL to 3.7 × 10⁸ particles per mL. Sensitivity is defined as the slope of the linear region of the calibration curve, obtained by linear fitting. The limit of detection (LOD) was calculated using the equation LOD = 3σ/s, where σ represents the standard deviation obtained from ten repeated measurements of the blank group and s denotes the slope of the calibration curve.³⁵ The specific results are summarized as follows:

-For protein CD63, the linear equation is y = 1555.45x − 6463.47, R² = 0.9979, with a LOD of 1.4 × 10⁴ particles per mL;

-For protein CD24, the linear equation is y = 1419.05x − 5982.65, R² = 0.9950, with a LOD of 1.6 × 10⁴ particles per mL;

-For protein CA125, the linear equation is y = 816.89x − 3471.91, R² = 0.9974, with a LOD of 1.8 × 10⁴ particles per mL.

Based on the lowest protein detection limit, the detection limit of this SERS biosensor for EVs from ovarian cancer cell line HO8910 is determined to be 1.4 × 10⁴ particles per mL.

Multivariate statistical analysis was performed using MATLAB, including principal component analysis (PCA). For cancer staging detection, a support vector machine (SVM) model was implemented in MATLAB with the deep learning toolbox. Spectral data from healthy, early-stage cancer, and late-stage cancer groups were divided into training (3000 groups) and testing (150 groups) sets. The SVM model, trained with the “fitcecoc” function with a linear kernel, was validated using test set accuracy, the confusion matrix, the receiver operating characteristic (ROC) curve, and the area under the curve (AUC).

3. Results and discussion

3.1. Preparation of biosensors

MoS₂, as a novel nanomaterial, demonstrates distinctive advantages in the field of biosensors: its stable structure enables facile carboxyl group modification for antibody immobilization, while its excellent biocompatibility preserves biomolecular activity, ensuring reliable detection outcomes. Furthermore, MoS₂ exhibits two characteristic Raman peaks near 400 cm⁻¹, a spectral region that avoids overlap with most Raman-active reporter molecules. When combined with gold–silver nanoparticles, the inherent Raman enhancement effect of MoS₂ synergizes with plasmonic electronic coupling effects from noble metal components, collectively achieving substantial sensitivity improvements in detection systems.^36,37

MoS₂ dispersion was prepared via a liquid-phase exfoliation method, followed by surface modification with sodium cholate to introduce carboxyl groups for antibody conjugation. Characterization, including zeta potential analysis, spectroscopic characterization, and TEM, confirmed the successful preparation of negatively charged MoS₂ nanosheets (Fig. S1). Subsequently, the surface of the silicon wafer after ultrasonic cleaning was charged by argon plasma pretreatment, then immersed in PDDA and MoS₂ dispersion solution in turn, and repeatedly assembled to form a MoS₂ composite film (Fig. 2a). Moreover, by observing the surface morphology of the MoS₂ film using SEM (Fig. 2b), we identified the presence of wrinkles and protrusions on the surface. Through a layer-by-layer self-assembly process, the two-dimensional nanomaterials formed a three-dimensional structure on the silicon substrate, effectively increasing the specific surface area of the detection substrate. This provides additional active sites for antibody connection, thereby enhancing the sensitivity of the sensor. Then, the Raman enhancement effect of the sensor was evaluated using DTNB as a Raman reporter (Fig. 2c), and we found that the 10-layer MoS₂ showed significantly better enhancement than the 5-layer MoS₂, while the improvement beyond 10 layers was minimal. Considering the fabrication time and cost, 10 layer MoS₂ was selected for the composite film. With fewer layers, MoS₂ and PDDA cannot form a complete film, resulting in thinner or even uncovered regions, leading to weaker overall Raman enhancement. As the layer count increases, the composite film becomes more complete, significantly enhancing the Raman effect, though this effect does not change significantly at higher layer counts.

Fig. 2 Fabrication of the biosensor and characterization of the MoS₂ composite film and the immunorecognition probe of gold nanoparticles: (a) fabrication process of the biosensor. (b) SEM image of the 10 layer MoS₂ composite film, and the insets on the top right corner show the physical image. (c) Raman enhancement effect with different MoS₂ layers. (d) TEM image of the gold nanoparticle. (e) Schematic diagram of the gold nanoparticle immunocomplex probe. (f) Visible absorption spectra of gold nanoparticle immunocomposite probes.

We obtained a stable, wine-red gold nanoparticle solution using a thermal reduction method (Fig. S2). Based on TEM analysis (Fig. 2d), we statistically analysed 50 synthesized gold nanoparticles and determined an average particle diameter of approximately 12.9 nm. The overall morphology and size distribution of the nanoparticles are shown in Fig. S3. This particle size is suitable for capturing EVs in the diameter range of 40–200 nm and facilitates subsequent SERS analysis. Subsequently, a two-step method was employed, in which Raman reporters (DTNB) and recognition antibodies (anti-CD63) were sequentially added to the gold nanoparticle solution to modify the surface, yielding gold nanoparticle immunorecognition probes (Fig. 2e). The modification process was characterized via UV-vis absorption spectroscopy³⁸ (Fig. 2f). The UV-vis spectra at different stages of modification showed that, upon the addition of Raman reporters to the gold nanoparticle surface, the overall particle size increased,³⁹ causing the redshift of the absorption peak from 525 nm to 530 nm. When the recognition antibodies were adsorbed onto the modified surface, the absorption peak of Au further redshifted to 536 nm and the full width at half maximum (FWHM) increased by 9 nm, which is consistent with the size-dependent behavior of the UV-vis absorption peak of gold nanoparticles.⁴⁰

3.2. Performance of biosensors for the detection of EVs in ovarian cancer

As an important part of EV⁴¹ detection, EV surface membrane proteins can mediate the interaction between EVs and target cells. EpCAM,⁴² a glycosylated membrane protein, is present in different types of tumors, such as ovarian cancer, lung cancer, and pancreatic cancer, and its expression level is high. It is an effective tumor marker. The expression of EpCAM is mainly related to the proliferation of epithelial cells during tumor development and can effectively distinguish EVs derived from tumor cells and normal cells. Therefore, EpCAM protein was used as the capture protein, and the anti-EpCAM antibody was immobilized on the surface of the MoS₂ composite film to capture and immobilize EVs. CD63⁴³ is a transmembrane protein on the surface of EVs from different cell sources and is a reliable tumor marker. Studies have shown that the expression of CD63 protein is also related to tumor progression. CD24⁴⁴ is a cell surface protein connected by glycosylphosphatidylinositol and is expressed in various solid tumors. Studies have shown that the expression of CD24 protein can affect cancer metastasis and CD24 protein is a diagnostic protein marker for poor prognosis of ovarian cancer and other cancers. CA125⁴⁵ is a commonly used protein marker for ovarian cancer diagnosis and is expressed on the surface of ovarian cancer EVs. To effectively reflect the expression of surface membrane proteins of ovarian cancer EVs, Raman reporters were used to label corresponding protein markers. The detection signal of Raman reporters was used to reflect the presence and expression level of EV membrane proteins (Fig. 3a). Three Raman reporters, DTNB⁴⁶ (1328 cm⁻¹), TFMBA⁴⁷ (1627 cm⁻¹), and 4-MBN⁴⁸ (2223 cm⁻¹), were selected based on the principle that their characteristic peak signals do not overlap.

Fig. 3 Surface signal enhancement: (a) Raman signal of the Raman reporters, including DTNB, TFMBA and 4-MBN. (b) The detection process of the biosensor. (c) Raman enhancement from different combinations of DNTB with Au and MoS₂, verifying the SERS effectiveness. (d) Raman spectra of six comparison tests, verifying the detection principle. (e) Surface scan image of Raman reporters.

The sensor detection process is shown in Fig. 3b. To systematically evaluate the SERS enhancement effect, equal volumes of DTNB solution or prepared Au@DTNB tags were, respectively, applied onto clean silicon substrates, MoS₂-modified silicon substrates, and Au@DTNB tag-modified MoS₂ substrates, and Raman measurements were performed under identical experimental conditions. As shown in Fig. 3c, the Raman signal intensity of DTNB was significantly enhanced on both the MoS₂ film and gold nanoparticles. When both components were present, the Raman signal was further amplified, confirming the synergistic enhancement effect and the feasibility of the SERS sensor. Furthermore, upon formation of the sensor's sandwich-type structure, the close proximity between the immunorecognition probe and the MoS₂ substrate further strengthened the Raman signal, thereby enhancing the sensitivity of the sensor. Furthermore, upon formation of the sensor's sandwich-type structure, the proximity between the immunorecognition probe of gold nanoparticles and the MoS₂ substrate notably strengthened the Raman signal, thereby enhancing the sensor sensitivity. Before detecting ovarian cancer EVs, the sandwich-type immunosensor structure of the sensor was verified, and six control groups were set up to detect each group of sensors by Raman spectroscopy. The results showed that only when a complete sandwich-type immunosensor structure was formed, a signal response could be obtained (Fig. 3d).

The sensor surface was subsequently subjected to area scanning, with specific parameters provided in the Experimental section. Three spectral windows centered at 1328, 1627, and 2223 cm⁻¹ (each spanning 50 cm⁻¹) were selected, corresponding to the principal Raman peaks of DTNB, TFMBA, and 4-MBN, which were employed to label the exosomal membrane proteins CD63, CD24, and CA125, respectively. The signals corresponding to these proteins were visualized in red, green, and blue channels (Fig. 3e). Prominent signal intensities were observed within each spectral window, indicating successful EV capture and effective immunorecognition of the targeted membrane protein markers.

The bright regions in the scanned images of the three proteins notably overlapped, consistent with the multi-channel detection, suggesting that the EVs in this area feature multiple surface membrane proteins labeled by the immunorecognition probe.

Then, EVs from the ovarian cancer cell line HO8910 were selected as detection targets to evaluate the analytical performance of surface protein marker detection. Initial characterization of EV samples by NTA demonstrated that the particle size distribution showed predominant nanoparticles in the 100–175 nm range (Fig. S4), consistent with typical EV dimensions. Quantitative analysis revealed a sample concentration of 3.7 × 10⁸ particles per mL. Based on these findings, the EVs were serially diluted across a concentration gradient spanning 3.7 × 10² to 3.7 × 10⁸ particles per mL to establish detection systems at varying concentrations. The EpCAM antibody was used as the capture antibody to capture EVs, and the Raman reporter DTNB-CD63, TFMBA-CD24, and 4-MBN-CA125 immunodetection probes were used to label the surface membrane proteins of EVs. After removing the background signal, the corresponding Raman signal waterfall plot at each concentration was obtained (Fig. 4a–c). At extremely low concentrations, the signal response of the sensor to ovarian cancer EVs was nearly zero, while at higher concentrations, obvious Raman characteristic peaks of the Raman reporters DTNB, TFMBA, and 4-MBN at 1328 cm⁻¹, 1627 cm⁻¹, and 2223 cm⁻¹, respectively, were observed, and the signal intensity of the peaks increased with the increase of the EV concentration, showing a certain linear relationship. Leveraging the relationship between peak signal strengths and EV concentrations, we ascertained the detection limit of the sensor for EVs from the HO8910 cell line using linear regression. According to the linear curves (Fig. 4d–f), the linear detection ranges of the sensor for the three protein markers were determined to be 3.7 × 10⁴ particles per mL to 3.7 × 10⁸ particles per mL. The detection limit of the sensor for CD63 protein was determined to be 1.4 × 10⁴ particles per mL; for CD24 protein, it was 1.6 × 10⁴ particles per mL; and for CD63 protein, it was 1.8 × 10⁴ particles per mL. Therefore, based on the lowest detection limit, the detection limit of the sensor for EVs from the HO8910 cell line was determined to be 1.4 × 10⁴ particles per mL. We compared the detection results obtained with our developed sensor with the LOD reported in the existing literature (Tables S1 and S2). It was found that our sensor not only enables the simultaneous detection of multiple surface proteins but also possesses a significantly lower LOD, underscoring the pronounced advantages of this biosensor.

Fig. 4 Detection of single membrane proteins in EVs of ovarian cancer. (a)–(c) Raman images of CD63, CD24, and CA125 proteins at exosomal concentrations from 3.7 × 10² particles per mL to 3.7 × 10⁸ particles per mL. (d)–(f) Linear curve of CD63, CD24, and CA125 proteins at exosomal concentrations from 3.7 × 10² particles per mL to 3.7 × 10⁸ particles per mL. The error bars represent the standard deviation of three repetitive measurements of each sample.

To explore the sensor's ability to simultaneously detect multiple membrane proteins on EVs and detect EVs from different ovarian cancer cell lines and to validate its feasibility in ovarian cancer diagnosis, we conducted experiments on EVs derived from the ovarian cancer cell lines HO8910 (a concentration of 3.7 × 10⁷ particles per mL) and SKOV3 (a concentration of 5.2 × 10⁷ particles per mL). We used Anti-EpCAM as the capture probe and Anti-CD24/TFMBA/Au, Anti-CD63/DTNB/Au, and Anti-CA125/4-MBN/Au as composite immunorecognition probes. These probes were mixed and simultaneously used to recognize and label the surface membrane proteins of EVs. The specific procedures followed the same protocol described in the previous section. After incubation, Raman spectroscopy was performed for detection (Fig. 5a). The results clearly showed the characteristic peaks of the corresponding Raman reporters for the same three proteins in both cell lines. This validated the sensor's capability to simultaneously identify and characterize multiple membrane protein expressions on EVs.

Fig. 5 Other properties of biosensors. (a) Comparison of protein expression in ovarian cancer cell lines HO8910 and SKOV3. (b) Signal repeatability of biosensors. (c) Stability of biosensors. The error bars represent the standard deviation of three repetitive measurements of each sample.

Furthermore, to minimize the error between experimental results in multiple detections, the repeatability of the sensor's Raman signal was validated. Using DTNB to label the membrane protein CD63 on EVs at a concentration of 3.7 × 10⁷ particles per mL from the HO8910 cell line, Raman spectra were obtained from 30 random regions to assess signal uniformity across different areas of the same sensor film. The intensities of two Raman peaks at 1328 cm⁻¹ and 1560 cm⁻¹ were statistically analyzed. The results (Fig. 5b) showed that the relative standard deviations (RSD) of the Raman signal intensities at 1328 cm⁻¹ and 1560 cm⁻¹ were 4.45% and 8.04%, respectively, validating the good reproducibility of the biosensor's response. Additionally, we stored 40 prepared sensors for two weeks and recorded the Raman peak intensity at 1328 cm⁻¹ of the DTNB Raman reporter by selecting five sensors every two days (Fig. 5c). After four weeks, the intensity had decreased to 79.3% of the original value. These results indicate that the biosensor has good reproducibility and stability. It is worth noting that, although EVs are inherently heterogeneous in terms of size and surface marker expression, the average SERS signals obtained from multiple measurements remain stable and reproducible.⁴⁷ Therefore, we consider that such heterogeneity does not significantly affect the evaluation of signal uniformity in our experiments.

3.3. Performance of the sensor in detecting EVs in clinical serum samples

To validate the performance of the SERS biosensor in clinical detection and obtain a dataset for training an early ovarian cancer diagnosis model, we applied the sensor to detect EVs in serum samples from ovarian cancer patients and healthy females. The aim was to explore the biosensor's capability to diagnose ovarian cancer by detecting the expression of various surface membrane proteins on EVs. Given that the concentration of EVs in serum samples is approximately 10⁸ particles per mL^49,50 and that previous studies have indicated a higher EV concentration in serum from cancer patients compared to healthy individuals,⁵¹ the serum samples were diluted 100 times with PBS before actual detection. Subsequently, anti-EpCAM was used as the capture antibody to immobilize EVs, and the Raman reporters DTNB, TFMBA, and 4-MBN were used to label CD63, CD24, and CA125 proteins on the EV surface. The Raman signal intensities at 1328 cm⁻¹, 1627 cm⁻¹, and 2223 cm⁻¹ were analyzed.

We examined EVs in a total of 30 serum samples, including 10 samples from healthy individuals, 10 samples from early-stage ovarian cancer patients (stage I/II), and 10 samples from late-stage ovarian cancer patients (stage III/IV), and plotted the results (Fig. 6a and b). The figures show that only background signals were observed in healthy samples, and as EpCAM expression in normal serum EVs is relatively low, theoretically, EVs secreted by normal cells would not be captured by the capture probe immobilized on the composite membrane, resulting in no corresponding Raman signals. Although a few samples might exhibit nonspecific adsorption of immune recognition probes or low EpCAM expression due to inflammation, leading to their capture by the probe, the overall signals corresponding to the three membrane proteins in patient samples were significantly higher than those in healthy samples. For early- and late-stage ovarian cancer patients, the expression levels of CD63 showed minimal differences, as CD63 is an exosome-specific protein expressed in almost all EVs, regardless of the cell origin or cancer stage.⁵² In contrast, the expression levels of CD24 and CA125 increased as the disease progressed. Furthermore, from the bar graph, it can be observed that even within the same group, the signal intensities of the same membrane protein varied due to differences in membrane protein expression on the EV surface, which makes it difficult to accurately classify serum samples using traditional data analysis methods.

Fig. 6 Multi-marker approach for cancer staging detection. (a) Statistics and (b) box plot statistics of CD63, CD24 and CA125 protein expression in healthy, patients with early ovarian cancer and patients with advanced ovarian cancer. (c) Results of principal component analysis. (d) Average spectra of each sample. (e) SVM model. (f) Confusion matrix results. (g) ROC curve diagram.

Therefore, machine learning was introduced to achieve the visualization of multidimensional data,⁵³ more intuitively demonstrate the expression differences of surface membrane proteins of EVs in serum samples between healthy individuals and ovarian cancer patients, and establish an ovarian cancer diagnosis and discrimination platform. Initially, we attempted to distinguish healthy individuals from ovarian cancer patients using Raman signal features of individual reporters. For the 30 sample sets, we extracted the Raman signal features corresponding to each reporter and divided the dataset into a training set of 20 samples and a test set of 10 samples. The model was trained using support vector machines (SVM) and validated on the test set. Ultimately, the classification accuracies for the models based on CD63, CD24, and CA125 were 65%, 85%, and 75%, respectively (Fig. S5), indicating that the diagnostic accuracy of a single biomarker remains suboptimal. Subsequently, we further explored a combined analysis using the Raman signal characteristics of these three biomarkers. In our analysis, the machine learning methods include principal component analysis (PCA)⁹ and SVM.⁵⁴ PCA is utilized to reduce the dimensions of serum sample data collected from healthy individuals, as well as patients with early and late-stage ovarian cancer, thereby intuitively highlighting the characteristic differences between the groups. Furthermore, the dataset was expanded by data augmentation techniques, in order to train a SVM model with enhanced accuracy, aiming to improve the efficiency of early ovarian cancer diagnosis. To elucidate the distinctive features among the three groups more intuitively, we subjected the serum sample data to dimensionality reduction through principal component analysis (PCA). The PCA-processed results, presented in Fig. 6c, reveal that the combined variance ratio of the two principal factors exceeded 90%, satisfying the criteria for the new lower-dimensional space. Moreover, in this transformed coordinate system, there are distinct differences between healthy individuals and patients with early- and late-stage ovarian cancer. PC1 is strongly associated with the separation of the three sample groups, and it typically corresponds to overall intensity, which is consistent with our previous analyses. These distinctions underscore the potential of our biosensor for clinical applications, as it effectively discriminates between the different states of ovarian cancer.

The SVM algorithm is a supervised learning model based on the principle of maximum margin separation. Its core concept involves identifying the optimal hyperplane that effectively separates data points of different categories. The support vectors determine the final classification boundary, enabling SVM to excel in classification tasks. Given that we only have detection results for 30 samples (10 samples each from healthy, stage I/II, and stage III/IV groups), such a small dataset can lead to overfitting and poor generalization when training the model. To address this issue, we expanded the dataset by adding Gaussian noise to the original data. Specifically, 25 sets of data were augmented 120-fold, resulting in a total of 3000 datasets for the training set, while 5 sets of data were augmented 30-fold, resulting in 150 datasets for the test set. The average spectrum of the expanded dataset is shown in Fig. 6d. The model was trained using the augmented training set, and the detailed procedure can be found in the Methods section. Upon completion of the training, the test set was used to evaluate the model's predictions. The confusion matrix of the prediction results is shown in Fig. 6f, and the overall prediction accuracy of the model is 97.78%. This high accuracy rate is indicative of the model's robust capability to discriminate between cancer stages in subjects. To address the potential risk of overfitting due to data augmentation, we also performed 10-fold cross-validation using only the original dataset (Fig. S6). The model maintained a high classification accuracy (93.3%) after cross-validation, indicating that data augmentation did not result in artificially inflated accuracy. This demonstrates that the model's discriminative power is indeed grounded in the genuine differences present in the original data. Moreover, the receiver operating characteristic (ROC) curves were plotted (Fig. 6g), and the area under the curve (AUC) was calculated. The three ROC curves were plotted using a one-vs-rest strategy, where each class (healthy, early-stage, and late-stage) was treated as the positive class and the other two classes as negative. All ROC curves and AUC values were derived from the same model. The AUC values for healthy individuals, early-stage ovarian cancer patients, and late-stage ovarian cancer patients were 1, 0.9988, and 0.9988, respectively. These AUC outcomes demonstrate that the ovarian cancer diagnosis and discrimination platform we developed exhibits considerable efficacy. Upon comparing the prediction accuracy with the existing literature (Table S1), it is found that our model not only effectively predicts different stages of cancer in patients but also demonstrates superior accuracy. Therefore, it holds promise as a powerful auxiliary tool for cancer diagnosis.

4. Conclusions

We successfully developed a highly sensitive and specific SERS biosensor for the early diagnosis of ovarian cancer. Although the biosensor prepared in this study did not respond to detection at extremely low concentrations of EVs, it showed high sensitivity and a wide linear range in EV concentration detection. Moreover, the SERS biosensor could simultaneously label multiple surface membrane proteins of EVs and characterize their expression levels, achieving multi-channel detection of EVs and further improving the sensitivity of EV detection. It demonstrated enormous potential in EV detection and early cancer diagnosis. Then, we utilized machine learning technology and a SERS biosensor to evaluate the detection outcomes of a single protein marker with the objective of distinguishing between Raman spectra of healthy women and ovarian cancer patients. However, the classification results did not meet our expectations. Subsequently, leveraging the multi-channel detection capability of the biosensor, we integrated the detection outcomes of three protein markers and developed a prediction model using the SVM methodology. The model successfully achieved accurate classification of Raman spectral features in healthy women, early-stage ovarian cancer patients, and advanced-stage ovarian cancer patients, thereby demonstrating exceptional sensitivity, specificity, and accuracy. The combination of SERS sensors and machine learning has great potential for application in biomedical diagnostics, offering promising directions for non-invasive cancer detection and early diagnosis.

Author contributions

T. H., Z. W., X. L., Z. W. and J. Z. conceived the idea. T. H., H. X., and Z. W. planned and performed the experiments, took all the photographs and draw parts of figures. S. J. wrote the analysis algorithm. T. H., H. X, S. J. and Z. W. contributed to the data analysis and drafted the manuscript. T. H., H. X., S. J., Y. Q., X. L., J. Z., Z. W. and Z. N. edited the manuscript. All authors discussed the results and commented on the manuscript.

Conflicts of interest

There are no conflicts to declare.

Data availability

The data and materials that support the findings of the study are available from the corresponding author upon reasonable request.

Acknowledgements

This work was supported by the National Natural Science Foundation of China (grant no. 52305596, 52375564), the Natural Science Foundation of Jiangsu Province (grant no. BK20231424), the Open Project of Key Laboratory of Silk Products for Medical and Health in Textile Industry (grant no. SDHY2129), and the Research Fund for Advanced Ocean Institute of Southeast University (Key Program, grant no. KP202406).

References

1. R. Xu, A. Rai, M. Chen, W. Suwakulsiri, D. W. Greening and R. J. Simpson, Nat. Rev. Clin. Oncol., 2018, 15, 617–638.
2. M. Nawaz, G. Camussi, H. Valadi, I. Nazarenko, K. Ekström, X. Wang, S. Principe, N. Shah, N. M. Ashraf and F. Fatima, Nat. Rev. Urol., 2014, 11, 688–701.
3. S. Bruno, G. Chiabotto, E. Favaro, M. C. Deregibus and G. Camussi, Am. J. Physiol.: Cell Physiol., 2019, 317, C303–C313.
4. V. R. Minciacchi, M. R. Freeman and D. Di Vizio, Semin. Cell Dev. Biol., 2015, 41–51.
5. S. Ludwig, T. Floros, M.-N. Theodoraki, C. S. Hong, E. K. Jackson, S. Lang and T. L. Whiteside, Clin. Cancer Res., 2017, 23, 4843–4854.
6. S. Ferguson, K. S. Yang and R. Weissleder, Trends Mol. Med., 2022, 28, 681–692.
7. E. Heitzer, I. S. Haque, C. E. S. Roberts and M. R. Speicher, Nat. Rev. Genet., 2018, 20, 71–88.
8. D. R. Santiago-Dieppa, J. Steinberg, D. Gonda, V. J. Cheung, B. S. Carter and C. C. Chen, Expert Rev. Mol. Diagn., 2014, 14, 819–825.
9. H. Shin, H. Jeong, J. Park, S. Hong and Y. Choi, ACS Sens., 2018, 3, 2637–2643.
10. J. Shi, P. W. Kantoff, R. Wooster and O. C. Farokhzad, Nat. Rev. Cancer, 2017, 17, 20–37.
11. T. Sattar, Top. Curr. Chem., 2019, 377, 1–45.
12. H. K. Neupane and N. P. Adhikari, Int. J. Comput. Mater. Sci. Eng., 2021, 10, 2150009.
13. S.-L. Chen, C.-Y. Chen, J. C.-H. Hsieh, Z.-Y. Yu, S.-J. Cheng, K. Y. Hsieh, J.-W. Yang, P. V. Kumar, S.-F. Lin and G.-Y. Chen, Nanomaterials, 2019, 9, 1725.
14. M. Shi, X. Wang, Y. Wu, K. Yuan, Z. Li, H.-M. Meng and Z. Li, Sens. Actuators, B, 2022, 355, 131315.
15. A. Tripathi and J. Bonilla-Cruz, ACS Appl. Nano Mater., 2023, 6, 5042–5074.
16. P. W. Dempsey, C.-M. Sandu, R. Gonzalezirias, S. Hantula, O. Covarrubias-Zambrano, S. H. Bossmann, A. S. Nagji, N. K. Veeramachaneni, N. O. Ermerak and D. Kocakaya, Commun. Med., 2024, 4, 37.
17. A. Albanese and W. C. Chan, ACS Nano, 2011, 5, 5478–5489.
18. D. Jin, P. Xi, B. Wang, L. Zhang, J. Enderlein and A. M. van Oijen, Nat. Methods, 2018, 15, 415–423.
19. R. Xu, Particuology, 2015, 18, 11–21.
20. X. Zhao, X. Liu, D. Chen, G. Shi, G. Li, X. Tang, X. Zhu, M. Li, L. Yao, Y. Wei, W. Song, Z. Sun, X. Fan, Z. Zhou, T. Qiu and Q. Hao, Nat. Commun., 2024, 15, 5855.
21. L. Zhao, Y. Wang, S. Jin, N. An, M. Yan, X. Zhang, Z. Hong and S. Yang, Nat. Synth., 2024, 3, 867–877.
22. J. Li, Y. Li, P. Li, Y. Zhang, L. Du, Y. Wang, C. Zhang and C. Wang, Acta Biomater., 2022, 144, 1–14.
23. J. Wang, A. Wuethrich, A. A. I. Sina, R. E. Lane, L. L. Lin, Y. Wang, J. S. Cebon, A. Behren and M. Trau, Sci. Adv., 2020, 6, eaax3223.
24. J. Li, Y. Li, S. Chen, W. Duan, X. Kong, Y. Wang, L. Zhou, P. Li, C. Zhang, L. Du and C. Wang, Small Methods, 2022, 6, 2200154.
25. L. Zhang, C. Gu, J. Wen, G. Liu, H. Liu and L. Li, Anal. Bioanal. Chem., 2021, 413, 83–102.
26. L. Xu, N. Shoaie, F. Jahanpeyma, J. Zhao, M. Azimzadeh and K. T. Al, Biosens. Bioelectron., 2020, 161, 112222.
27. X. Zhu, H. Chen, Y. Zhou, J. Wu, S. Ramakrishna, X. Peng, H. S. Nanda and Y. Zhou, Curr. Opin. Biomed. Eng., 2021, 18, 100280.
28. J. Zhu, Z. Ren and C. Lee, ACS Nano, 2020, 15, 894–903.
29. C.-S. Ho, N. Jean, C. A. Hogan, L. Blackmon, S. S. Jeffrey, M. Holodniy, N. Banaei, A. A. Saleh, S. Ermon and J. Dionne, Nat. Commun., 2019, 10, 1–8.
30. L. Huang, H. Sun, L. Sun, K. Shi, Y. Chen, X. Ren, Y. Ge, D. Jiang, X. Liu and W. Knoll, Nat. Commun., 2023, 14, 48.
31. J. Q. Li, P. V. Dukes, W. Lee, M. Sarkis and T. Vo-Dinh, J. Raman Spectrosc., 2022, 53, 2044–2057.
32. T. Hu, M. Zhang, Z. Wang, K. Chen, X. Li and Z. Ni, Microchem. J., 2020, 158, 105209.
33. B. Wang, Q. Zhu, D. Liao and C. Yu, J. Mater. Chem., 2011, 21, 4821–4826.
34. S. Park, H. Kim, M. S. Kim, G. H. Han and J. Kim, Opt. Express, 2016, 24, 21551–21559.
35. A. Shrivastava, Drug Discovery Ther., 2011, 2, 21–25.
36. H. Kim, J. Im, K. Nam, G. H. Han, J. Y. Park, S. Yoo, M. Haddadnezhad, S. Park, W. Park, J. S. Ahn, D. Park, M. S. Jeong and S. Choi, Sci. Rep., 2022, 12, 22252.
37. J. Li, W. Zhang, H. Lei and B. Li, Nano Res., 2018, 11, 2181–2189.
38. C. Song, L. Min, N. Zhou, Y. Yang, S. Su, W. Huang and L. Wang, ACS Appl. Mater. Interfaces, 2014, 6, 21842–21850.
39. A. Dhollander, E. Mathieu, H. Jans, G. Vande Velde, T. Stakenborg, P. Van Dorpe, U. Himmelreich and L. Lagae, Int. J. Nanomed., 2016, 11, 3703–3714.
40. X. Zhang, S. Liu, X. Song, H. Wang, J. Wang, Y. Wang, J. Huang and J. Yu, ACS Sens., 2019, 4, 2140–2149.
41. S. Zhou, T. Hu, G. Han, Y. Wu, X. Hua, J. Su, W. Jin, Y. Mou, X. Mou and Q. Li, Small, 2020, 16, 2004492.
42. X. Gong, H. Chi, D. F. Strohmer, A. T. Teichmann, Z. Xia and Q. Wang, Front. Immunol., 2023, 13, 1089410.
43. X. Wang, L. Tian, J. Lu and I. O.-L. Ng, Oncogenesis, 2022, 11, 54.
44. H. Xiong, Z. Huang, Z. Yang, Q. Lin, B. Yang, X. Fang, B. Liu, H. Chen and J. Kong, Small, 2021, 17, 2007971.
45. N. Li, G. Lin, Y. Zhang, Q. Zhang and H. Zhang, Discover Oncol., 2022, 13, 33.
46. D. Song, R. Yang, S. Fang, Y. Liu, F. Long and A. Zhu, Microchim. Acta, 2018, 185, 1–10.
47. W. Zhang, L. Jiang, R. J. Diefenbach, D. H. Campbell, B. J. Walsh, N. H. Packer and Y. Wang, ACS Sens., 2020, 5, 764–771.
48. W.-S. Zhang, Y.-N. Wang and Z.-R. Xu, Anal. Chim. Acta, 2020, 1094, 106–112.
49. M. R. Fernando, C. Jiang, G. D. Krzyzanowski and W. L. Ryan, PLoS One, 2017, 12, e0183915.
50. W. Zhang, X. Ou and X. Wu, Int. J. Oncol., 2019, 54, 1719–1733.
51. Y. Matsumoto, M. Kano, Y. Akutsu, N. Hanari, I. Hoshino, K. Murakami, A. Usui, H. Suito, M. Takahashi and R. Otsuka, Oncol. Rep., 2016, 36, 2535–2543.
52. S. Gurung, D. Perocheau, L. Touramanidou and J. Baruteau, Cell Commun. Signaling, 2021, 19, 47.
53. J. Zhu, M. Cho, Y. Li, T. He, J. Ahn, J. Park, T.-L. Ren, C. Lee and I. Park, Nano Energy, 2021, 86, 106035.
54. C. Chang and C. Lin, ACM Trans. Intell. Syst. Technol., 2011, 2, 1–27.

---