Method and device for constructing lung cancer auxiliary diagnosis model based on sEVs membrane protein and intramembrane miRNA

Abstract

The application provides a method for constructing a lung cancer auxiliary diagnosis model based on sEVs membrane proteins and intramembrane miRNAs, comprising the following steps: constructing a plasma sample library, wherein the ratio of normal control group, lung benign disease patients and lung cancer patients in the plasma sample is 1:1:1; for each plasma sample, obtain sEVs solution by separation and purification; obtain the expression amount of three miRNAs in the sEVs membrane from the sEVs solution by using an in-situ synchronous detection method based on membrane fusion delivery and DSN signal amplification; the three miRNAs are miR-21-5p, miR-375-3p and miR-451a nucleotide sequences as shown in SEQ. No. 1-3; obtain the expression amount of three membrane proteins of sEVs from the sEVs solution by using a synchronous detection method of tumor-derived sEVs membrane proteins based on a label-type surface-enhanced Raman spectrum; the three membrane proteins are GPC1, LYPD3 and NCAM1; the expression amount of the three miRNAs and the expression amount of the three membrane proteins of each plasma sample are taken as input characteristics and sent into a machine learning model for training to obtain a lung cancer auxiliary diagnosis model to assist lung cancer diagnosis.

Description

Technical Field

[0001] This invention relates to the field of molecular biology detection technology, and more specifically, to a method and apparatus for constructing an auxiliary diagnostic model for lung cancer based on sEVs membrane proteins and intramembrane miRNAs. Background Technology

[0002] Small extracellular vesicles (sEVs) are considered an ideal entry point for liquid biopsy due to their abundant content in body fluids, high stability, and the presence of numerous tumor-specific molecular targets. By screening for tumor-specific biomarkers, highly specific molecular targets can be provided for lung cancer screening and auxiliary diagnosis, thereby improving the sensitivity and specificity of screening and supporting early and accurate diagnosis.

[0003] Studies have shown that small RNAs in sEVs, including microRNAs (miRNAs), are considered potential biomarkers for various cancers. Patent application CN119736384A discloses the application of plasma small extracellular vesicle miRNA biomarkers in the preparation of a product for predicting the efficacy of lung cancer immunotherapy combined with chemotherapy. This involves obtaining the expression levels of miRNA biomarkers in plasma sEVs of subjects before undergoing immunotherapy combined with chemotherapy for lung adenocarcinoma. Based on the expression levels of these miRNA biomarkers in plasma sEVs, an efficacy prediction model is established using any one of the following analytical methods: minimum absolute contraction and selection operator, support vector machine, neural network, or random forest. In practice, besides using the expression levels of miRNAs on sEVs for efficacy prediction, the expression levels of membrane proteins are also used for lung cancer screening.

[0004] However, relying solely on the expression levels of miRNAs on sEVs or solely on sEV membrane proteins for lung cancer detection has significant limitations. When only sEVs miRNAs are detected, the same miRNA may be abnormally expressed in various cancers and even inflammatory diseases, leading to insufficient lung cancer specificity. Furthermore, the extraction, quantification, and selection of internal controls for sEVs miRNAs lack standardized procedures, resulting in poor comparability of results between different laboratories. Additionally, miRNAs are easily degraded during sample preservation and separation, affecting detection stability. On the other hand, when only sEV membrane proteins are detected, their expression levels are significantly affected by sEV heterogeneity and the tumor microenvironment, making it difficult to simultaneously achieve both sensitivity and specificity with a single or a few membrane protein biomarkers. Moreover, membrane protein detection often requires high-affinity antibodies or probes, which are costly and cannot provide information on tumor proliferation activity, invasiveness, or other biological behaviors. Therefore, detection of any single sEV biomarker is insufficient for reliable early diagnosis of lung cancer.

[0005] In order to solve the above problems, people have been seeking an ideal technological solution. Summary of the Invention

[0006] Therefore, it is necessary to provide a method and apparatus for constructing an auxiliary diagnostic model for lung cancer based on sEVs membrane proteins and intramembrane miRNAs to address the aforementioned technical problems.

[0007] To achieve the above objectives, the first aspect of the present invention provides a lung cancer-assisted plasma sample library based on sEVs membrane proteins and intramembrane miRNAs, wherein the ratio of normal control group, patients with benign lung disease and lung cancer patients in the plasma sample is 1:1:1. For each plasma sample, a sEVs solution was obtained through separation and purification; The expression levels of three miRNAs in the sEVs membrane were obtained from the sEVs solution using an in situ synchronous detection method based on membrane fusion delivery and DSN signal amplification; the three miRNAs were miR-21-5p, miR-375-3p and miR-451a nucleotide sequences as shown in SEQ. No. 1-3; The expression levels of three membrane proteins of tumor-derived sEVs were obtained from sEVs solution using a tag-based surface-enhanced Raman spectroscopy-based method for simultaneous detection of these membrane proteins. The three membrane proteins were GPC1, LYPD3, and NCAM1. The expression levels of three miRNAs and three membrane proteins from each plasma sample were used as input features and fed into a machine learning model for training to obtain a lung cancer auxiliary diagnostic model to assist in the diagnosis of lung cancer.

[0008] Specifically, the above scheme uses six sEVs molecular markers as input features, including three miRNAs (miR-21-5p, miR-375-3p, and miR-451a) and three membrane proteins (GPC1, LYPD3, and NCAM1). Among them, sEVs-miRNAs can reflect the post-transcriptional regulatory state of tumor cells and are closely related to key gene networks such as tumor immune escape and chemotherapy resistance. Membrane proteins can indicate the cellular origin of sEVs and the expression status of surface antigens and receptors, directly linking the tumor immune microenvironment and targeted therapy target characteristics. Through the above feature-level fusion, two types of heterogeneous data can be mapped to the same feature space, realizing multi-scale characterization of miRNAs and proteins, thereby improving the sensitivity and specificity of screening. Furthermore, machine learning is used to learn cross-modal associations to overcome the limitations of single modality and improve the predictive ability for lung cancer.

[0009] To achieve the above objectives, a second aspect of the present invention provides a method for the auxiliary diagnosis of lung cancer based on sEVs membrane proteins and intramembrane miRNAs, comprising the following steps: Using the lung cancer auxiliary diagnostic model construction method described in the first aspect, multiple models are constructed, and the model with the best performance is selected as the lung cancer auxiliary diagnostic model. Plasma samples were obtained from patients to be tested, and sEVs solution was obtained through separation and purification. The expression levels of three miRNAs in the sEVs membrane were obtained from the sEVs solution using an in situ synchronous detection method, and the expression levels of three membrane proteins of sEVs were obtained from the sEVs solution using a synchronous detection method. The expression levels of three miRNAs and three membrane proteins were used as input features and fed into a lung cancer auxiliary diagnostic model for identification to obtain lung cancer prediction results.

[0010] Furthermore, when the lung cancer prediction result is lung cancer, lung CT images of the patient to be tested are obtained, and CT imaging features are extracted from the lung CT images based on machine vision. The CT imaging features include the diameter of the CT image nodules, the location of the nodules, the morphological features of the nodules, and the density of the nodules. The expression levels of three miRNAs, three membrane proteins, and CT imaging features of the patients to be tested are used as input features and fed into a trained multimodal lung nodule benign and malignant differential diagnosis model to obtain lung cancer diagnosis results.

[0011] The above scheme uses six sEVs molecular markers (miR-21-5p, miR-375-3p, miR-451a, three miRNAs, and GPC1, LYPD3, and NCAM1, three membrane proteins) and CT image quantitative features (nodule size, nodule location, nodule morphology, and nodule density) as multimodal input features. Among them, sEVs-miRNAs can reflect the post-transcriptional regulatory state of tumor cells and are closely related to key gene networks such as tumor immune escape and chemotherapy resistance; membrane proteins can indicate the cellular origin of sEVs and the expression status of surface antigens and receptors, directly related to the tumor immune microenvironment and targeted therapy target characteristics; CT images provide information on the macroscopic morphology and anatomical structure of tumors, reflecting the spatiotemporal heterogeneity of tumors. A phased progressive fusion strategy is adopted: the first stage constructs a lung cancer screening model based on the six sEVs molecular markers; the second stage fuses the molecular markers and CT image quantitative features at the feature layer, maps them to a unified feature space, and inputs them into the classifier to achieve complementary integration of molecular functional state and anatomical structural information. This strategy retains the advantages of liquid biopsy being non-invasive and dynamically monitored, while also obtaining imaging anatomical localization and morphological information. By using machine learning to uncover cross-modal correlation patterns, it overcomes the limitations of single-modal approaches, ultimately improving the sensitivity of early lung cancer screening and the specificity of differentiating between benign and malignant lung nodules.

[0012] To achieve the above objectives, a third aspect of the present invention provides a lung cancer auxiliary diagnostic device based on sEVs membrane proteins and intramembrane miRNAs, comprising: The model library includes a lung cancer auxiliary diagnostic model and a multimodal lung nodule benign / malignant differential diagnostic model. The lung cancer auxiliary diagnostic model is constructed using the method described in the first aspect for constructing a lung cancer auxiliary diagnostic model based on sEVs membrane proteins and intracellular miRNAs, and is used to predict lung cancer based on the expression levels of three miRNAs and three membrane proteins. The multimodal lung nodule benign / malignant differential diagnostic model is used to diagnose lung cancer based on the expression levels of three miRNAs, three membrane proteins, and lung CT imaging features. The molecular data acquisition module is used to acquire the expression levels of three miRNAs and three membrane proteins in the plasma samples of the patients to be tested. The three miRNAs are miR-21-5p, miR-375-3p and miR-451a nucleotide sequences as shown in SEQ. No. 1-3; the three membrane proteins are GPC1, LYPD3 and NCAM1. The CT image acquisition module acquires lung CT images of the patient to be tested and extracts CT image features based on machine vision. The CT image features include nodule diameter, nodule location, nodule morphological features, and nodule density. The diagnostic module is used to take the expression levels of three miRNAs and three membrane proteins in the plasma sample of the patient to be tested as input features, call the lung cancer auxiliary diagnostic model for identification, and obtain a lung cancer prediction result; and when the lung cancer prediction result is lung cancer, it takes the expression levels of the three miRNAs, three membrane proteins and CT imaging features of the patient to be tested as input features, calls the multimodal lung nodule benign and malignant differential diagnosis model for identification, and obtains a lung cancer diagnosis result.

[0013] This invention has outstanding substantive features and significant progress compared to the prior art, specifically: This invention uses the expression levels of three miRNAs and three membrane proteins as input features to train a machine learning model, thereby obtaining a lung cancer auxiliary diagnostic model. Compared with traditional lung cancer prediction based solely on miRNAs, the fusion of multimodal features, through multi-scale miRNA-protein characterization, significantly improves the accuracy, robustness, and clinical interpretability of early lung cancer screening and efficacy prediction, representing the current development direction of precision medicine.

[0014] Furthermore, a SAMA system was constructed to achieve efficient, specific capture and gentle release of sEVs. On the other hand, a high-sensitivity detection strategy based on membrane fusion delivery and DSN signal amplification was used for the in-situ simultaneous quantitative detection of three low-abundance miRNAs within sEVs. This method integrates miRNA release, specific recognition, and signal amplification into the same reaction space by constructing a multifunctional cationic liposome.

[0015] Furthermore, a sandwich sensing structure was established with magnetic trapping beads, tumor-derived sEVs, and multifunctional SERS probes as the core, enabling simultaneous, high-throughput quantitative detection of three sEV membrane surface proteins, GPC1, LYPD3, and NCAM1. This breaks through the limitations of traditional methods that require individual detection, and can acquire expression data of three biomarkers in a single reaction, significantly improving detection throughput and sample utilization.

[0016] Finally, based on the multimodal input features composed of three miRNAs, three membrane proteins, and CT imaging features, the machine learning model was trained to obtain a lung cancer auxiliary diagnostic model. Through multimodal complementarity of macroscopic image localization and molecular subtyping, the accuracy, robustness, and clinical interpretability of early lung cancer screening were further significantly improved. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method for constructing the lung cancer auxiliary diagnostic model of the present invention.

[0018] Figure 2 This is the confusion matrix of the training set for five-fold cross-validation.

[0019] Figure 3 This is the confusion matrix of the test set for five-fold cross-validation.

[0020] Figure 4 The ROC curves are for the classification results of six machine learning models on the test set.

[0021] Figure 5 This is a schematic diagram of the principle of three miRNA in situ synchronous detection methods based on membrane fusion delivery and DSN signal amplification according to the present invention.

[0022] Figure 6 This is a schematic diagram of the SAMA system of the present invention for the separation and release of sEVs.

[0023] Figure 7 This is a comparison between the in situ synchronous detection method of three miRNAs of the present invention and RT-qPCR; where the horizontal axis is the relative expression level of RT-qPCR, and the vertical axis is the difference in fluorescence signal intensity (FL−FL0) measured by the established method, that is, the fluorescence intensity of the target channel (FL) - the background fluorescence intensity (FL0).

[0024] Figure 8 This is a heatmap showing the expression of three miRNAs in plasma sEVs in three population groups. Figure 8 A is miR-21-5p; Figure 8 B is miR-375-3p; Figure 8 C represents miR-451a; NC represents normal controls; BLD represents benign lung disease; LC represents lung cancer patients.

[0025] Figure 9 This is a schematic diagram of the experimental principle for the simultaneous detection of three membrane proteins (sEVs) based on tag-based SERS.

[0026] Figure 10 Correlation analysis between SERS and ELISA methods. Figure 10 A: GPC1; Figure 10 B: LYPD3; Figure 10 C: NCAM1; the horizontal axis represents the ELISA test results, and the vertical axis represents the SERS system test results.

[0027] Figure 11 These are the detection results of GPC1, LYPD3, and NCAM1 in plasma sEVs. In the figure, A: GPC1; B: LYPD3; C: NCAM1; NC: normal control; BLD: patients with benign lung disease; LC: patients with lung cancer.

[0028] Figure 12 The expression levels of three proteins in plasma sEVs were compared between the lung cancer group and the control group.

[0029] In the sequence list: SEQ. No. 8 is the aptamer Apt used in the embodiments of the present invention. EpCAM The DNA sequence; SEQ No. 9 is the aptamer Apt used in the embodiments of the present invention. PD-L1 The DNA sequence; SEQ No. 10 is the aptamer Apt used in the embodiments of the present invention. EGFR DNA sequence.

[0030] Figure 13 This is a flowchart illustrating a method for auxiliary diagnosis of lung cancer.

[0031] Figure 14 This is a schematic diagram of another method for auxiliary diagnosis of lung cancer.

[0032] Figure 15 This is the confusion matrix of the training set for five-fold cross-validation.

[0033] Figure 16 This is the confusion matrix of the test set for five-fold cross-validation.

[0034] Figure 17 The ROC curves are for the classification results of six machine learning models on the test set. Detailed Implementation

[0035] SAMA: A targeted separation system based on the coupling of streptavidin-coated magnetic beads (SA@MBs) and aptamers (Apt).

[0036] SERS: A detection system based on aptamer-antibody dual-recognition tag-based surface-enhanced Raman scattering (SERS).

[0037] AuNPs: Gold nanoparticles.

[0038] The technical solution of the present invention will be further described in detail below through specific embodiments. Example 1

[0039] This embodiment proposes a method for constructing an auxiliary diagnostic model for lung cancer based on sEVs membrane proteins and intramembrane miRNAs, such as... Figure 1 As shown, it includes the following steps: (1) Construct a plasma sample bank in which the ratio of normal control group, patients with benign lung disease and lung cancer patients in plasma samples is 1:1:1.

[0040] It is important to note that plasma samples should be collected in the morning on an empty stomach. Collect 5 mL of blood from the antecubital vein into an EDTA anticoagulant tube and gently invert to mix 5 times. Within 2 hours of collection, centrifuge at 3000×g for 5 min at 4°C to separate the supernatant plasma. Filter the plasma through a 0.22 μm filter membrane, aliquot, and store at -80°C for later use.

[0041] Specifically, stratified random sampling was used to divide the plasma sample bank into a training set and an independent test set in a 7:3 ratio. Stratified sampling ensures that the distribution of samples of each category in the training set and the test set is consistent with the overall distribution, avoiding class imbalance. The training set is used only for model building and hyperparameter optimization, while the independent test set is not involved in model training or hyperparameter tuning and is only used for final performance evaluation to objectively reflect the model's generalization ability on unknown data.

[0042] Specifically, regarding the sample size required for model construction, the study follows the Events per Variable (EPV) principle for reliability assessment. EPV is a key statistical indicator for measuring the stability of a predictive model and assessing the risk of overfitting. To ensure the reliability of model parameter estimation, the ratio of the number of outcome events to the number of predictor variables is typically required to be at least 10:1. In this embodiment, the EPV value of the training set must meet the statistical evaluation criterion of EPV ≥ 10.

[0043] It is understandable that before constructing various machine learning algorithms, it is necessary to perform normalization preprocessing using the range standardization method for feature data of different dimensions, and uniformly map all data to the interval [0, 1], so as to accelerate the convergence speed of machine learning algorithms and improve the overall performance of the model.

[0044] (2) For each plasma sample, sEVs solution was obtained by separation and purification; the expression levels of three miRNAs in the sEVs membrane were obtained from the sEVs solution by in situ synchronous detection method, and the expression levels of three membrane proteins of sEVs were obtained from the sEVs solution by synchronous detection method; the three miRNAs were miR-21-5p, miR-375-3p and miR-451a nucleotide sequences as in SEQ.No.1-3; the three membrane proteins were GPC1, LYPD3 and NCAM1.

[0045] Specifically, sEVs extraction from plasma samples was performed using a commercially available sEVs isolation kit. Plasma was removed from the freezer at -80°C, thawed in a 25°C water bath, and immediately placed on ice. 200 μL of plasma was centrifuged at 2000 × g for 20 min at room temperature, then centrifuged again at 10000 × g for 20 min. 100 μL of sterile PBS buffer was added to the pretreated supernatant and mixed thoroughly. Then, 60 μL of a dedicated coprecipitation reagent was added and mixed thoroughly. The mixture was incubated at 4°C for 30 min. After incubation, the mixture was centrifuged at 4°C and 10000 × g for 5 min. The precipitate was then rapidly centrifuged at 10000 × g for 30 s at room temperature to remove residual liquid. The precipitate was resuspended in 100 μL of pre-cooled PBS and stored at -80°C for later use.

[0046] Then, a high-sensitivity detection strategy based on membrane fusion delivery and DSN signal amplification was employed for the in-situ simultaneous quantitative detection of three low-abundance miRNAs within sEVs. This method integrates miRNA release, specific recognition, and signal amplification within the same reaction space by constructing multifunctional cationic liposomes.

[0047] Furthermore, a SERS-based method for the simultaneous detection of tumor-derived sEVs membrane proteins was employed to achieve highly sensitive simultaneous quantitative detection of sEVs membrane surface proteins (GPC1, LYPD3, NCAM1).

[0048] The expression levels of three miRNAs and three membrane proteins from each plasma sample were used as input features and fed into a machine learning model for training to obtain a lung cancer auxiliary diagnostic model to assist in the diagnosis of lung cancer.

[0049] The following are the specific steps for training a lung cancer auxiliary diagnostic model based on machine learning.

[0050] The expression levels of miR-21-5p, miR-375-3p and miR-451a in 450 plasma samples (including lung cancer patients, patients with benign lung diseases and healthy controls, 150 in each group) were quantitatively detected using an in situ synchronous detection method based on membrane fusion delivery and DSN signal amplification. The expression levels of tumor-derived sEVs membrane proteins GPC1, LYPD3, and NCAM1 in 450 plasma samples (including lung cancer patients, patients with benign lung diseases, and healthy controls, 150 in each group) were quantitatively detected using a tag-based surface-enhanced Raman spectroscopy-based method.

[0051] Then, based on the PyCharm programming platform, using the Python 3.x programming language and the scikit-learn machine learning library, six machine learning models were built: decision tree, random forest, adaptive boosting algorithm, logistic regression, support vector machine, and artificial neural network. These are all common models and will not be discussed in detail here.

[0052] It is understandable that during the model training and hyperparameter tuning phases, a 5-fold cross-validation grid search strategy is adopted: the training set is split into 5 mutually exclusive subsets, and 4 subsets are used as training data and 1 subset as validation data in turn. The preset hyperparameter combinations of each model (such as the maximum depth of the decision tree, the number of decision trees in the random forest, the kernel function parameters of the SVM, etc.) are traversed, and the average classification accuracy of the 5-fold validation is used as the evaluation index to select the optimal hyperparameter combination of each model. Based on the complete training set and test set data, the final version of each model is obtained by training with the optimal hyperparameter combination.

[0053] Figure 2 A to Figure 2 E shows the confusion matrix of the five-fold cross-validation training set. In the five folds, 108, 107, 108, 109, and 108 out of 120 lung cancer patients were correctly predicted as lung cancer patients; 96, 95, 96, 94, and 94 out of 120 patients with benign lung diseases were correctly predicted as benign lung diseases; and 112, 108, 110, 108, and 106 out of 120 healthy individuals were correctly predicted as healthy individuals. The accuracy of the training set for each fold is 87.78%, 86.11%, 87.22%, 86.39%, and 85.56%.

[0054] Figure 3 A~ Figure 3E shows the confusion matrix of the five-fold cross-validation test set. The confusion matrix of the test set reflects the model's classification performance on three types of samples: lung cancer, benign lung diseases, and normal population. As shown in Table 1, the model's average sensitivity for lung cancer, benign lung diseases, and normal population is 80.00%, 89.33%, and 90.00%, respectively. The overall classification accuracy of the five-fold test set is 87.78%, 85.56%, 86.67%, 85.56%, and 87.78%, respectively, with an average accuracy of 86.67%. The sensitivity for lung cancer is 76.67%–83.33%, and the sensitivity for benign lung diseases and normal population is above 80%. The AUC value is 0.858–0.883, and the 95% confidence interval is above 0.778, indicating that the model's generalization performance is stable and there is no obvious overfitting or underfitting.

[0055] Table 1. Classification performance of the five-fold cross-validation test set.

[0056] Accuracy was used to evaluate the tri-class classification performance of each model. Benign lung diseases and the normal population were grouped into a "non-lung cancer group," while lung cancer patients were grouped into a "lung cancer group." Sensitivity, specificity, positive predictive value, negative predictive value, and AUC were used to evaluate the binary classification performance of the models. The results are shown in Table 2. Figure 4 As shown, the RF model has the best predictive performance, with an accuracy of 86.67%, a sensitivity of 86.67%, a specificity of 93.10%, a positive predictive value of 86.67%, a negative predictive value of 93.10%, and an AUC of 0.901 (95% CI: 0.854-0.968).

[0057] Table 2. Classification performance of six models integrating sEVs multi-omics and CT radiomics on the test set.

[0058] As shown in the table above, the RF model has the best overall performance in terms of accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and AUC, and is therefore the optimal model.

[0059] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps. Example 2

[0060] This embodiment differs from Embodiment 1 in that it provides specific steps for obtaining the expression levels of three miRNAs within the sEVs membrane from the sEVs solution using an in-situ simultaneous detection method, as follows: Figure 5 As shown.

[0061] The specific steps are as follows: Purification of target sEVs: Based on the SA@MBs and Apt coupled target separation system, namely the SAMA system, the sEVs solution is purified to obtain the purified target sEVs.

[0062] Specifically, such as Figure 6 As shown, the core working principle of the SAMA system is as follows: (1) Preparation of SAMA magnetic beads: Three desulfurized biotin-labeled nucleic acid aptamers, Apt, were synthesized. The three nucleic acid aptamers target EpCAM, PD-L1, and EGFR of sEVs, respectively, and were named Apt. EpCAM Apt PD-L1 and Apt EGFR Utilizing the affinity between SA@MBs and desulfurized biotin, the three nucleic acid aptamers Apt were coupled to the surface of streptavidin magnetic beads, resulting in Apt-coated beads. EpCAM Apt PD-L1 and Apt EGFR SAMA magnetic beads.

[0063] The specific steps include: taking 600 μL of SA@MBs (2 mg / mL), magnetically separating, and washing three times with PBST. Adding 1 mL of solution containing 400 nmol / L Apt... EpCAM / Apt PD-L1 / Apt EGFR The aptamer solution was reacted with shaking at room temperature for 1 h. Unbound aptamers were removed by magnetic separation, washed three times with PBST, resuspended in PBS, and stored at 4 °C protected from light. It should be used within one week.

[0064] (2) Capture of sEVs: SAMA magnetic beads were used to capture sEVs, and the SAMA-sEVs complex was obtained after magnetic separation and purification. Specifically, the specific recognition of EpCAM, PD-L1 and EGFR on the surface of the sEVs membrane by aptamers was used to achieve precise capture of target sEVs. The magnetic bead-sEVs complex was rapidly separated under the action of an external magnetic field, and unbound and non-specifically adsorbed components were washed away.

[0065] (3) Release of sEVs: The captured SAMA-sEVs complex was mixed with free biotin and incubated by shaking to elute and release sEVs from the SAMA-sEVs complex, thus obtaining purified target sEVs.

[0066] It is understandable that the extremely high affinity between free biotin and streptavidin (stronger than that between desulfurized biotin and streptavidin, approximately 10 times that of desulfurized biotin) can be utilized. 4 By adding an excess of free biotin to competitively replace the desulfurized biotin aptamer-sEVs complex bound to the surface of magnetic beads, intact sEVs are gently released under neutral pH 7.4 conditions, thus preserving their morphology and biological activity to the greatest extent.

[0067] Preparation of Lip@AuNFs@DSN: The Lip@AuNFs@DSN comprises cationic liposomes and gold nanoflares and DSN enzymes encapsulated within the cationic liposomes.

[0068] The specific preparation process is as follows: 1) Preparation of gold nanoparticles (AuNPs) using sodium citrate reduction. Specifically, the synthesis steps for AuNPs are as follows: 10 mL of 100 mmol / L chloroauric acid was added to 90 mL of Milli-Q water and magnetically stirred to dissolve, resulting in a final concentration of 1 mmol / L. A round-bottom flask connected to a condenser was heated in an oil bath at 120 ℃ until stable reflux was achieved. 10 mL of 38.8 mmol / L sodium citrate was quickly added, and the reaction was sealed. The solution color changed from pale yellow to colorless, then grayish-black, and finally wine-red (approximately 1 min), indicating the formation of AuNPs. The reaction was continued for 15–20 min, then heating was stopped, and the mixture was allowed to cool naturally to room temperature with stirring. The solution was filtered through a 0.22 μm filter and stored at 4 ℃ protected from light. The absorbance at 520 nm was measured using a UV-Vis spectrophotometer from 450 to 700 nm, and the molar extinction coefficient ε = 2.7 × 10⁻⁶ was determined. 8 L·mol -1 ·cm -1 Calculate the concentration.

[0069] 2) Three fluorescently labeled DNA probes were coupled to the surface of the AuNPs to obtain gold nanoflares; the gold nanoflares consist of gold nanoparticles and three DNA probes cDNA21, cDNA375, and cDNA451 with nucleotide sequences such as SEQ.No.12-14 coupled to the surface of the gold nanoparticles. DNA probe cDNA21 is labeled with the FAM fluorescent group, DNA probe cDNA375 is labeled with the TAMRA fluorescent group, and DNA probe cDNA451 is labeled with the Cy5 fluorescent group. Specifically, the synthesis steps of gold nanoflares are as follows: 60 μL of 10 μmol / L cDNA is mixed with 2 μL of 0.2 mol / L TTCEP and incubated at a constant temperature for 30 min to reduce thiol groups. 300 μL of AuNPs is added, mixed at room temperature, and allowed to stand for 5 min; then incubated at -20 ℃ in the dark for 2 h. Thaw at 4 ℃ for 30 min, adding 15 μL of 2.5 mol / L NaCl every 30 min to a final concentration of 500 mmol / L. Incubate overnight at 4 ℃ in the dark to allow for complete cDNA assembly. Centrifuge at 12000 ×g for 15 min, discard the supernatant; resuspend in PBS, centrifuge again, and wash twice. Resuspend in 300 μL PBS and store at 4 ℃ in the dark.

[0070] 3) The gold nanoflares and DSN enzyme were dissolved in a reaction buffer for preparing cationic liposomes by thin-film hydration. The cationic liposomes were prepared by thin-film hydration using the reaction buffer to obtain the Lip@AuNFs@DSN.

[0071] The synthesis steps of cationic liposomes are as follows: 6.0 mg of 1,2-dioleoyl-3-trimethylammonium-propane (DOTAP), 4.9 mg of cholesterol, 3.9 mg of 1,2-distearoyl-sn-glycero-3-phosphoethanolamine-N-polyethylene glycol (DSPE-PEG), and 6.3 mg of 1,2-dipalmitoyl-sn-glycero-3-phosphocholine (DPPC) were weighed and dissolved in 8 mL of chloroform and 2 mL of methanol. After sonication for 10 min, the solution was transferred to a 250 mL oval flask. The organic solvent was removed by rotary evaporation under reduced pressure at 37 °C (150 r / min) to form a lipid film. Further reduced pressure evaporation for 30 min was then performed to remove residual solvent. Add 2 mL of reaction buffer, wash the membrane by rotating at 45 r / min under normal pressure, and hydrate at 37 ℃ for 2 h. Disperse the liposomes by squeezing them back and forth 10 times using a liposome extruder into 2 mL of reaction buffer, and store at 4 ℃ for later use. Dissolve the pre-prepared AuNFs and DSN enzyme in the reaction buffer, replacing the blank buffer in the hydration step above, and perform hydration and squeezing in the same manner to obtain cationic liposomes (Lip@AuNFs@DSN) encapsulated with AuNFs and DSN enzyme. The remaining steps are the same as for the preparation of blank liposomes.

[0072] Fluorescence intensity detection: The purified target sEVs were mixed with the Lip@AuNFs@DSN to undergo membrane fusion, and the fluorescence intensity of each channel was measured simultaneously. A fluorescence intensity-sEVs standard curve was found to obtain the expression levels of the three miRNAs.

[0073] Figure 5 The principle of the in-situ synchronous detection method is shown below: (1) Construction and membrane fusion delivery of multifunctional liposomes: Multifunctional liposomes (Lip@AuNFs@DSN) loaded with fluorescently labeled DNA probes and DSN enzyme were prepared. The three DNA probes targeted different miRNAs (labeled with FAM, TAMRA, and Cy5 fluorescent groups, respectively); the initial fluorescence of the cDNA probe coupled to the surface of the gold nanoparticles was effectively suppressed by the quenching effect of the adjacent gold nanoparticles. When Lip@AuNFs@DSN fused with sEVs, the miRNA inside the sEVs was released into the liposome cavity, and the detection reagent entered the common reaction space formed by the fusion, thereby initiating target recognition and signal generation.

[0074] (2) DSN enzyme-triggered signal conversion and cyclic amplification: The released miRNA hybridizes with the corresponding DNA probe to form a DNA-RNA double strand. The DSN enzyme specifically cleaves the DNA strand in the double strand, allowing the fluorescent group to escape the quenching effect of AuNFs and restoring the fluorescence signal. After cleavage, the miRNA remains intact and can rehybridize with new probes, initiating multiple rounds of cyclic cleavage to achieve cascade amplification of the signal of a single miRNA molecule.

[0075] (3) Multiplexing detection: After the three cDNA probes are cleaved by DSN enzyme, they release three distinguishable fluorescent signals. By simultaneously measuring the intensity of each fluorescent channel, the in situ synchronous quantitative analysis of the three miRNAs can be achieved.

[0076] It should be noted that the main reagents and materials used in the purification of target sEVs and the preparation of Lip@AuNFs@DSN, as well as all nucleic acid sequences (including DNA probes, aptamers, and miRNA mimics, etc.), were purchased from Sangon Biotech Co., Ltd. (see Table 3 for details). All water used was ultrapure water prepared using the Milli-Q ultrapure water system (Millipore, USA). The instruments and equipment used were commonly used in this field.

[0077] Table 3 Nucleic acid names and base sequences

[0078] The preparation methods for the main buffer solutions and reagents involved are as follows: RPMI-1640 / DMEM complete medium: 5 mL of inactivated fetal bovine serum was mixed with 45 mL of basal medium containing penicillin-streptomycin antibiotics.

[0079] 0.01 mol / L phosphate buffer (PBS, pH 7.4): Dissolve commercially available PBS powder in 2 L Milli-Q ultrapure water according to the instructions; autoclave at 121 °C and 0.1 MPa for 40 min, cool to room temperature, and store at 4 °C.

[0080] Phosphate-Tween buffer PBST (pH 7.4): Slowly add 250 μL Tween-20 to 500 mL of 0.01 mol / L PBS (pH 7.4) and vortex for 1 min; store at 4 ℃.

[0081] TBST: 100 mL of 10×TBST concentrate was diluted to 1 L with Milli-Q ultrapure water and mixed well.

[0082] Hybridization buffer (pH 7.4): Dissolve 5.844 g NaCl, 2.033 g MgCl2·6H2O, and 2.423 g Tris in 950 mL Milli-Q water, adjust the pH to 7.4 with 0.1 mol / L HCl, bring the volume to 1 L, and autoclave.

[0083] 0.01 mol / L morpholine ethanesulfonic acid buffer (MES, pH 6.0): Weigh 1.066 g MES and dissolve it in 400 mL of Milli-Q ultrapure water; adjust the pH to 6.0 with 1 mol / L potassium hydroxide, bring the volume to 500 mL, and store at 4 ℃ protected from light.

[0084] 10 mmol / L Tris(2-Carboxyethyl)phosphine buffer (TCEP): Weigh 2.8665 g TCEP, dissolve in Milli-Q ultrapure water, and bring the volume to 1 L; aliquot and store at 4 ℃ for short-term storage, or freeze at -20 ℃.

[0085] It is important to note that a fluorescence intensity-sEVs standard curve needs to be established before finding the fluorescence intensity-sEVs standard curve. The steps for establishing a fluorescence intensity-sEVs standard curve include: The sEVs standard was serially diluted with PBS to 1.0 × 10⁻⁶. 4 ~5.0×10 8 Particle / μL series concentrations; The target separation system based on streptavidin-modified magnetic beads coupled with aptamers was used to purify the gradient dilution solution of sEVs standard to obtain purified target sEVs. The optimal conditions were 1.0 U of DSN enzyme, 120 min of reaction time, and 37 °C. Under these conditions, the target sEVs purified by each gradient dilution solution were mixed with Lip@AuNFs@DSN at a concentration of 10 × 10⁷ particles / μL to induce membrane fusion, and the fluorescence intensity of each channel was measured simultaneously. A standard curve of fluorescence intensity-sEVs was established using linear regression with the logarithm of sEVs concentration on the x-axis and net fluorescence intensity on the y-axis. The regression equation and R² were recorded. Wherein, net fluorescence intensity = FL - FL0, FL is the fluorescence intensity of the target channel, and FL0 is the background fluorescence intensity.

[0086] Under optimal conditions, the Lip@AuNFs@DSN system is suitable for 1.0 × 10⁻⁶ Ω·cm. 4 ~5.0×10 7The detection performance of the three target miRNAs in sEVs within the particle / μL range was good, and ΔFL showed a linear relationship with lgCsEVs. Figure 14 As shown. Where, miR-21-5p: Y=17292.63X1-19076.63, R 2 =0.9993. miR-375-3p:=15184.12X2-12247.81,R 2 =0.9977;miR-451a:Y=12893.67X 3 -17850.23, R 2 =0.9992.

[0087] Y is ΔFL, X 1 X 2 X 3 These represent the commonly used logarithms for the three miRNA concentrations. The system demonstrates good quantitative detection capabilities over a dynamic range of approximately four orders of magnitude, meeting the needs for simultaneous analysis of miRNAs of varying abundances in complex samples.

[0088] To evaluate the quantitative accuracy of the constructed detection system, using RT-qPCR, the gold standard for miRNA quantification, as a reference, a correlation model was established between the fluorescence signal increment ΔFL = FL - FL0 and the relative expression level of RT-qPCR in this method by detecting the expression levels of miR-21-5p, miR-375-3p, and miR-451a in plasma sEVs. Figure 7 As shown in the results, all three target miRNAs exhibited high linear correlations between this method and RT-qPCR: miR-21-5p correlation coefficient r = 0.988 (P < 0.001), miR-375-3p correlation coefficient r = 0.982 (P < 0.001), and miR-451a correlation coefficient r = 0.985 (P < 0.001). Compared with RT-qPCR, this system eliminates the need for sEV lysis and RNA extraction, allowing for the simultaneous detection of three target miRNAs with shorter detection time and simpler procedures. The high consistency between the two methods (r > 0.98) confirms that this method possesses quantitative accuracy and reliability comparable to the gold standard method for miRNA quantification in complex plasma samples, meeting the needs of clinical trace miRNA analysis.

[0089] When the detection system constructed above was used to test actual samples, a total of 450 subjects were included, including lung cancer patients (n=150), patients with benign lung diseases (n=150), and normal controls (n=150). The pathological types and TNM stages of the lung cancer patients are shown in Table 4. The pathological types of lung cancer patients included squamous cell carcinoma, adenocarcinoma, and small cell lung cancer. Among the lung cancer patients selected in this example, 53.33% had adenocarcinoma, 22.67% had squamous cell carcinoma, and 24.0% had small cell lung cancer, and the samples were evenly distributed between stages I / II (52.67%) and III / IV (47.33%).

[0090] Table 4. Clinicopathological types and TNM staging of lung cancer patients

[0091] Analysis of miRNA expression levels in three groups of plasma sEVs: The expression levels of miR-21-5p, miR-375-3p and miR-451a in plasma sEVs of 450 subjects were detected using the in situ synchronous detection method established in this embodiment.

[0092] like Figure 8 As shown in Table 5, the three miRNAs exhibited differential expression patterns among the three groups, with significant differences between groups (all P < 0.0001).

[0093] Table 5. Description and analysis results of plasma samples.

[0094] Therefore, the method for detecting the expression levels of three miRNAs within the sEVs membrane provided in this embodiment can achieve efficient capture, gentle release, and simultaneous in-situ detection of internal miRNAs from sEVs, realizing a breakthrough in the entire process from efficient enrichment of sEVs to multiplex miRNA detection. Methodological evaluation confirms that this technology system has a low detection limit, a recovery rate as high as 96.5%~104.8%, and good reproducibility (RSD<8%), providing reliable technical support for the accurate detection of low-abundance biomarkers and providing highly specific and clinically relevant molecular evidence for lung cancer screening and auxiliary diagnosis. Example 3

[0095] The difference between this embodiment and Embodiment 2 is that it further provides specific steps for obtaining the expression levels of three sEV membrane proteins from the sEV solution using a method for simultaneous detection of tumor-derived sEV membrane proteins based on tag-type surface-enhanced Raman spectroscopy (SERS).

[0096] like Figure 9As shown, its core working principle is as follows: Immunocapture magnetic beads (AgMNPs-Apt) are prepared to capture tumor-derived sEVs. Three SERS probes that target the sEV membrane surface proteins GPC1, LYPD3, and NCAM1 are synthesized. The sEVs are captured by the immunocapture magnetic beads (AgMNPs-Apt) and recognized by the SERS probes to form a sandwich structure. Finally, Raman signals are collected and calibrated with MBN as an internal standard to achieve simultaneous quantification of the three membrane proteins GPC1, LYPD3, and NCAM1.

[0097] The specific steps include: (1) Preparation of immunocapture magnetic beads (AgMNPs-Apt): Fe3O4 nanoparticles were used as magnetic cores, and AuNPs were electrostatically adsorbed and loaded to form a Fe3O4-AuNPs composite structure. The internal standard molecule 4-mercaptobenzonitrile (MBN, Raman characteristic peak 2226 cm⁻¹) was used. -1 (For Raman signal calibration) The Fe3O4-AuNPs composite structure was modified with Au-S bonds to encapsulate a silver shell, resulting in silver magnetic nanoparticles (AgMNPs). Three DNA sequences, aptamers Apt (SEQ No. 8-10) as shown in Table 2, were simultaneously coupled to the silver magnetic nanoparticles via Ag-S bonds. EpCAM Apt PD-L1 Apt EGFR Immunocapture magnetic beads AgMNPs-Apt were prepared.

[0098] Specifically, the preparation steps of Fe3O4 magnetic nanoparticles are as follows: Fe3O4 nanoparticles are synthesized using a modified solvothermal method. The specific steps are as follows: 2.0 g FeCl3·6H2O, 4.0 g sodium acetate, and 1.0 g polyethylene glycol are fully dissolved in 60 mL of ethylene glycol. After the mixture is sonicated until completely dissolved, it is transferred to a 100 mL high-pressure reactor and reacted at 180 ℃ for 8 h. After the reaction system cools naturally to room temperature, the generated Fe3O4 product is collected using a magnetic rack. The obtained product is washed three times with ultrapure water and anhydrous ethanol to remove residual impurities, and finally dried overnight in a vacuum oven at 60 ℃ for later use.

[0099] The preparation steps of AgMNPs were as follows: A silver shell was grown on the surface of Fe3O4-AuNPs-MBN by chemical reduction. 14 mg AgNO3 and 300 mg polyvinylpyrrolidone were dissolved in 200 mL of ultrapure water, ultrasonically dispersed, and then 10 mg Fe3O4-AuNPs were added. Then, 300 μL of ammonia and 150 μL of formaldehyde were added sequentially, and the reaction was continued with ultrasonication for 5 min. The Ag-encapsulated magnetic nanoparticles (AgMNPs) were collected using magnetic separation and washed five times with ultrapure water to completely remove unreacted reagents.

[0100] The preparation steps of Fe3O4-AuNPs-MBN are as follows: Gold nanoparticles (AuNPs) are prepared by sodium borohydride reduction: 10 mL of 5 mmol / L chloroauric acid (HAuCl4) is added to 180 mL of ultrapure water and magnetically stirred until homogeneous; 10 mL of 5 mmol / L trisodium citrate is added, and after stirring for 5 min, 5 mL of freshly prepared 0.1 mol / L sodium borohydride is rapidly added dropwise, and the mixture is stirred at room temperature for 4 h until the solution turns a stable orange-red color. Fe3O4-AuNPs are prepared by electrostatic adsorption: Fe3O4 dispersion and AuNPs solution are mixed and sonicated, and after standing, the solid is collected by magnetic separation; the assembly is repeated until the supernatant turns a light orange-red color, indicating that AuNPs on the Fe3O4 surface have reached saturation adsorption. The product is washed three times with ultrapure water and twice with ethanol, and the solvent is evaporated at room temperature.

[0101] Mix 400 μL of 10 μmol / mL MBN ethanol solution with 3 mL of 10 mg / mL Fe3O4-AuNPs ethanol dispersion and sonicate for 5 min; freeze at -20 ℃ for 30 min to promote MBN adsorption. After returning to room temperature, perform magnetic separation, wash three times with ethanol to remove unbound MBN, resuspend in 3 mL of anhydrous ethanol, and store at 4 ℃ for later use.

[0102] The preparation steps of AgMNPs-Apt are as follows: Take 50 μL of 10 μmol / L specific aptamer (Apt) EGFR Apt PD-L1 and Apt EpCAM The product was mixed with 5 μL of 10 mmol / L TCEP solution and incubated at 25 °C for 45 min to reduce disulfide bonds (denoted as Pre solution). Subsequently, 200 μL of 10 mg / mL AgMNPs was added to the Pre solution and incubated overnight at 25 °C in the dark. After the reaction, the product was collected by magnetic separation, and the supernatant (Post solution) and subsequent wash solutions (Wash1, Wash2) were collected. The final product was resuspended in PBS buffer and stored at 4 °C in the dark for later use.

[0103] (2) Preparation of SERS probes (porous core-shell gold nanorods): Gold nanorods (AuNRs) were prepared by seed growth and modified with Au-S bonds with three Raman reporter molecules: 2-nitrobenzoic acid (DTNB, Raman characteristic peak at 1332 cm⁻¹). -1 ), 2-Mercaptopyridine (MPY, Raman characteristic peak 1002 cm⁻¹) -1 ), 2-Nitrotoluene (NT, Raman characteristic peak at 1375 cm⁻¹) -1 Ag-Au alloy shells were deposited on the surface of AuNRs by co-reduction of HAuCl4 and AgNO3 with ascorbic acid; the Ag components were selectively etched with Fe(NO3)3 to form porous core-shell AuNRs (Au@AuNRs). Finally, three SERS probes were constructed by coupling with specific antibodies targeting GPC1, LYPD3, and NCAM1.

[0104] Specifically, the preparation steps for AuNRs are as follows: Seed solution preparation: Mix 2.5 mL of 0.2 mol / L CTAB with 2.5 mL of 0.5 mmol / L HAuCl4, incubate at 29 ℃ in the dark, and rapidly add 0.3 mL of ice-cold 0.01 mol / L sodium borohydride while stirring at 1000 r / min. Stir for 1 min and let stand at 29 ℃ for 1 h. Growth solution preparation: Mix 20 mL of 0.2 mol / L CTAB with 20 mL of 1 mmol / L HAuCl4 in a 29 ℃ water bath in the dark. Add 0.2 mL of 0.01 mol / L AgNO3, 0.38 mL of 2 mol / L HCl, and 0.33 mL of 0.1 mol / L ascorbic acid (AA) sequentially. Add 130 μL of the seed solution to the growth solution and let stand overnight at 29 ℃ in the dark. Centrifuge at 10000 r / min for 14 min, repeat 3 times to remove excess surfactant, and resuspend the precipitate in 5 mmol / L CTAB.

[0105] Synthesis of AuNRs modified with Raman reporter molecules: 1 mL of 0.01 mol / L 2-nitrobenzoic acid (DTNB, Raman characteristic peak 1332 cm⁻¹) was used. -1 ), 2-Mercaptopyridine (MPY, Raman characteristic peak 1002 cm⁻¹) -1 ), 2-Nitrotoluene (NT, Raman characteristic peak at 1375 cm⁻¹) -1The mixture was reacted with 1 mL of 0.01 mol / L TCEP at room temperature for 40 min. 100 μL of the mixture was then incubated with 10 mL of AuNRs at room temperature with shaking for 4 h. The mixture was centrifuged at 9000 r / min for 14 min, repeated 3 times, and resuspended in 0.1 mol / L CTAB.

[0106] Encapsulation of Au-Ag alloy shells: 2 mL of AuNRs labeled with Raman molecules were added to 10 mL of 1% polyvinylpyrrolidone solution, followed by the dropwise addition of 20 μL of 10 mmol / L HAuCl4 and 140 μL of 10 mmol / L AgNO3. Under stirring in a 29 °C water bath, 300 μL of 0.1 mol / L AA and 750 μL of 0.1 mol / L NaOH were rapidly added to induce shell growth. The reaction was allowed to proceed for 15 min, followed by centrifugation at 9000 r / min for 14 min, washing three times with 0.5 mmol / L CTAB, and storage at 4 °C protected from light.

[0107] Preparation of porous shell-structured AuNRs: Porous structures were prepared using a chemical etching method. 2 mL of Au-Ag@AuNRs were mixed with 150 μL of 10 mmol / L Fe(NO3)3 and shaken at room temperature for 30 min. The mixture was then centrifuged at 6000 r / min for 5 min, washed twice with 0.5 mmol / L CTAAB, and resuspended in 0.5 mmol / L CTAAB.

[0108] Antibody conjugation to specific SERS tags: Surface activation: An equal volume of 0.01 mol / L 3-mercaptoundecanoic acid (MUA) was mixed with 0.01 mol / L TCEP and reacted for 40 min. 10 μL of the mixture was added to 200 μL of porous AuNRs and incubated overnight at 4 °C. Antibody conjugation: After centrifugation and purification, the antibody was resuspended in pH 6.5 MES buffer. EDC and NHS (10 mmol / L each, 100 μL each) were added for activation for 2 h. Then, 25 μL of 100 μg / mL specific antibody (GPC1, LYPD3, or NCAM1) was added, and the antibody was incubated overnight at 4 °C. Blocking and storage: Blocking with 1% BSA for 1 h, centrifuging, and resuspending in 200 μL PBS, then stored at 4 °C for later use.

[0109] (3) Capture of sEVs and assembly of the sandwich complex: AgMNPs-Apt was co-incubated with the sEVs solution to achieve the separation and enrichment of tumor-derived sEVs through aptamer-membrane protein specific binding. Three SERS probes were added sequentially to the co-incubation system, and the AgMNPs-Apt-sEVs-SERS probe sandwich structure was formed by antibody-antigen recognition assembly. Free probes were removed by magnetic separation, and the purified complex was obtained after washing.

[0110] (4) Raman signal detection and quantitative analysis: After the complex was dropped onto a glass slide and dried and fixed, Raman spectroscopy was collected. Under the same optical and acquisition parameters, the Raman intensity of each probe was collected, and the corresponding Raman intensity-sEVs standard curve was found to obtain the expression levels of the three membrane proteins.

[0111] The specific Raman signal detection steps for the three signal probes included: diluting the SERS probes for GPC1, LYPD3, and NCAM1 to a uniform concentration of 100 μg / mL with PBS buffer, and vortexing the samples for 30 s to ensure uniform dispersion. Detection was performed using a Raman spectrometer equipped with a 633 nm excitation source. The parameters were set as follows: laser power 10 mW, spectral acquisition range 800–1600 cm⁻¹. -1 Integration time 5 s, spectral resolution 2 cm -1 Before each test, the silicon wafer characteristic peak (520 cm⁻¹) was used. -1 Wavelength calibration was performed. 2.5 μL of each probe dilution solution was added to a clean glass slide, dried at room temperature, and then placed on the sample stage. Raman spectra of each probe were acquired sequentially under the same optical and acquisition parameters. Each sample was measured independently three times, and the raw spectral data were recorded for subsequent analysis.

[0112] It should be noted that the main chemical reagents, biological materials, and nucleic acid sequences involved in the nucleic acid experiments used in this embodiment were all synthesized and provided by Sangon Biotech Co., Ltd. For details of the specific nucleic acid sequences and membrane protein amino acid sequences, please refer to Tables 3 and 6.

[0113] Table 6. Amino acid sequence information of membrane proteins

[0114] The preparation methods for the main buffer solutions and reagents involved are as follows: 0.01 mol / L phosphate buffer (PBS, pH 7.4): See Example 2.

[0115] 10 mmol / L Tris(2-carboxyethyl)phosphine buffer (TCEP): See Example 2.

[0116] 0.01 mol / L morpholine ethanesulfonic acid buffer (MES, pH 6.0): See Example 2.

[0117] PBST, pH 7.4: See Example 2.

[0118] PBS-CTAB buffer: Weigh 91.1 mg of hexadecyltrimethylammonium bromide (CTAB), add 50 mL of 0.01 mol / L PBS (pH 7.4), and sonicate until clear; store at 4 ℃ protected from light, and reconstitute by sonication for 5 min before use.

[0119] 1-(3-Dimethylaminopropyl)-3-ethylcarbodiimide (EDC) solution: Weigh 10 mg of EDC hydrochloride and dissolve it rapidly in 1 mL of 0.01 mol / L MES (pH 6.0), vortex for 1 min; prepare fresh before use and leave at room temperature for no more than 2 h.

[0120] Blocking solution: Weigh 0.11 g bovine serum albumin (BSA, purity ≥98%), dissolve in 10 mL PBS, vortex to mix, and incubate at 4°C for ≥1 h until completely dissolved.

[0121] It is important to note that a Raman intensity-sEVs standard curve needs to be established before searching for the standard curve. The steps for establishing a Raman intensity-sEVs standard curve include: The sEVs standard was serially diluted 1.0 × 10⁻⁶ with PBS. 3 1.0×10 4 1.0×10 5 1.0×10 6 1.0×10 7 1.0×10 8 1.0×10 9 particles / μL.

[0122] With Apt EpCAM +Apt EGFR +Apt PD-L1Using three aptamers, with magnetic beads at a volume of 20 μL, incubation time of 50 min, probe volume of 25 μL, and incubation time of 30 min as optimal conditions, AgMNPs-Apt were mixed with serially diluted solutions of each SEV standard to form a co-incubation system. The separation and enrichment of SEVs were achieved through aptamer-membrane protein specific binding. Three SERS probes were added sequentially to the co-incubation system, and an AgMNPs-Apt-sEVs-SERS probe sandwich structure was formed by antibody-antigen recognition assembly. Free probes were removed by magnetic separation, and the purified complex was obtained after washing. The Raman intensities of the three SERS probes were collected and recorded under the same optical and acquisition parameters. like Figure 10 The above is based on the logarithm of sEVs concentration (lgC). sEVs The x-axis represents particles / μL, and the normalized Raman intensity (I / I) represents the particle size distribution. 2226 The vertical axis is used to establish three standard curves for Raman intensity-sEVs using linear regression, and the regression equations and correlation coefficients R are recorded. 2 .

[0123] GPC1 at 1×10 3 ~1×10 8 The system exhibits a linear response in the particle / μL range, with a regression equation of Y = 0.2795X - 0.5017. R 2 =0.9796.

[0124] The linear range of LYPD3 is 1×10 4 ~1×10 9 The particle / μL regression equation is Y=0.1886X-0.4284. R 2 =0.9762.

[0125] The linear range of NCAM1 is 1×10. 4 ~1×10 9 The regression equation for particles / μL is Y = 0.2047X - 0.4726. R 2 =0.9617. The superimposed results of the three-channel calibration curves show that each target has a good linear fit, and the difference in slope reflects the response sensitivity characteristics of different protein antigens.

[0126] Furthermore, to evaluate the quantitative accuracy of the detection system in actual biological samples, a comparative experiment was conducted using ELISA as the standard. Parallel tests were performed on the same group of plasma samples using ELISA as the reference method. Figure 11As shown, linear regression analysis revealed a significant positive correlation between the protein concentrations measured by SERS and ELISA, with high correlation coefficients for GPC1, LYPD3, and NCAM1. r The values ​​were 0.984, 0.964, and 0.965 respectively (all... P <0.001). Further comparison of the two methods in terms of detection steps and throughput: ELISA requires multiple steps including coating, blocking, sample addition, incubation, washing, color development, and termination, with single-target detection taking 4–6 hours; the SERS platform, through an aptamer-antibody dual recognition strategy, can achieve simultaneous detection of three targets in the same reaction system, with a total time of approximately 2 hours, and the magnetic separation step simplifies the washing process. This comparison shows that the SERS sensing system, while maintaining a high correlation with the gold standard method, has the advantages of simple operation and parallel detection of multiple targets.

[0127] Using the constructed SERS method, the expression levels of sEVs membrane proteins GPC1, LYPD3, and NCAM1 in 450 plasma samples (including lung cancer patients, patients with benign lung diseases, and healthy controls, 150 cases in each group) described in Example 2 were quantitatively detected, and the expression differences among the three groups were analyzed. The expression heatmaps of the three proteins in plasma sEVs of different groups are shown below. Figure 12 As shown.

[0128] The results showed that the expression levels of all three proteins were higher in the lung cancer group than in the benign lung disease group and the healthy control group. GPC1 expression showed a gradually increasing trend from the healthy control group to the benign lung disease group and then to the lung cancer group; LYPD3 and NCAM1 expression was highest in the lung cancer group, followed by the benign lung disease group, and lowest in the healthy control group. Statistical analysis indicated that the overall expression differences of the three proteins among the three groups were statistically significant. P <0.0001). Example 4

[0129] Based on the same inventive concept, this application also provides a method for the auxiliary diagnosis of lung cancer based on sEVs membrane proteins and intramembrane miRNAs, including the following steps: Using the lung cancer auxiliary diagnostic model construction method described in Example 1, 2, or 3, multiple models are constructed, and the model with the best performance is selected as the lung cancer auxiliary diagnostic model. Taking Table 2 in Example 1 as an example, the RF model showed the best predictive performance, with an accuracy of 90.00%, sensitivity of 86.67%, specificity of 93.33%, positive predictive value of 92.86%, negative predictive value of 87.50%, and AUC of 0.911 (95% CI: 0.828–0.972). Therefore, the RF model was selected as the auxiliary diagnostic model for lung cancer.

[0130] Plasma samples were obtained from the patients to be tested, and sEVs solutions were obtained through separation and purification. The expression levels of three miRNAs within the sEVs membrane were obtained from the sEVs solution using an in-situ simultaneous detection method, as were the expression levels of three membrane proteins from the sEVs solution using a simultaneous detection method. The three miRNAs were the nucleotide sequences miR-21-5p, miR-375-3p, and miR-451a as shown in SEQ. No. 1-3; the three membrane proteins were GPC1, LYPD3, and NCAM1. The expression levels of three miRNAs and three membrane proteins were used as input features and fed into a lung cancer auxiliary diagnostic model for identification, thereby obtaining lung cancer prediction results to assist in lung cancer diagnosis.

[0131] Furthermore, in this embodiment, the RF model was used to predict high-risk individuals for lung cancer in 170 coke oven workers, and 15 high-risk individuals were finally selected. Detailed information is shown in Table 7.

[0132] Due to the occupational requirements of coke oven workers, their age tends to be younger, therefore the age differences among the four groups were statistically significant (P<0.05), while there were no statistically significant differences in gender composition (P>0.05). Based on the median occupational exposure history, the 170 coke oven workers were divided into a short-to-medium-term exposure group (occupational exposure duration of 5–14 years) and a long-term exposure group (occupational exposure duration of 15–24 years), and there were statistically significant differences in age distribution between the two groups (P<0.05).

[0133] Analysis of the data revealed that individuals predicted to be at high risk for lung cancer generally had a long history of occupational exposure, all of whom had been engaged in a single type of work for a long time. The main occupational hazards were exposure to dust, coal dust, and benzene, among which 8 had more than 15 years of work experience.

[0134] Table 7. Basic information on high-risk individuals identified from coke oven workers. Example 5

[0135] The difference between this embodiment and embodiment 4 is that, when lung cancer was determined in embodiment 1, the benign or malignant nature of the lung nodules was further identified in order to ultimately determine whether it was lung cancer.

[0136] It is understandable that even though combined detection of miRNA expression levels and membrane proteins on sEVs can improve the accuracy of lung cancer identification from both gene regulation and protein function perspectives, key limitations remain. This bimodal analysis based on sEVs essentially falls under the category of liquid biopsy, reflecting only molecular evidence of tumor presence but failing to provide spatial information such as the tumor's anatomical location, precise size, morphological characteristics (e.g., lobulation, spiculation), density (solid / ground-glass component), and whether it invades adjacent structures (e.g., pleura, blood vessels). These imaging features are precisely what determine the TNM staging, resectability, and treatment strategy selection for lung cancer. Therefore, the introduction of CT imaging for fusion screening is irreplaceable: CT can provide three-dimensional localization and morphological assessment of suspicious lung lesions, while sEV molecular detection can simultaneously provide functional clues about the degree of tumor malignancy. The fusion of these two approaches avoids the problem of overdiagnosis of benign nodules based solely on imaging and compensates for the blind spot of liquid biopsy's inability to locate lesions, thus forming a complementary closed loop of molecular early warning + imaging confirmation, significantly improving the specificity, sensitivity, and clinical decision-making value of early lung cancer screening.

[0137] Specifically, the steps for identifying benign or malignant pulmonary nodules are as follows: When the lung cancer prediction result is lung cancer, lung CT images of the patient to be tested are obtained, and CT imaging features are extracted from the lung CT images based on machine vision. The CT imaging features include the diameter of the CT image nodules, the location of the nodules, the morphological features of the nodules, and the density of the nodules.

[0138] Specifically, in order to improve the quantitative analysis of lung cancer differential diagnosis and enhance the fitting efficiency of machine learning models, this embodiment standardized the assignment of values ​​to CT imaging features and related clinical indicators. The qualitative semantic descriptions were transformed into numerical variables through feature engineering, as shown in Table 8, to ensure consistency and operability in the feature extraction, model training and classification process.

[0139] Table 8. Variable assignments for CT imaging features

[0140] Then, the expression levels of three miRNAs, three membrane proteins, and CT imaging features of the patients to be tested are used as input features and fed into a trained multimodal lung nodule benign and malignant differential diagnosis model to obtain lung cancer diagnosis results.

[0141] Furthermore, this embodiment provides specific steps for training a lung cancer auxiliary diagnostic model based on a machine learning model.

[0142] A patient sample bank was constructed, with a 1:1 ratio of patients with benign lung diseases to patients with lung cancer. Each patient sample included a plasma sample and a lung CT image sample.

[0143] This embodiment still uses the ratio of patients with benign lung diseases and lung cancer patients as described in Example 1, which is 1:1, with 150 cases in each group. The plasma samples used are the 300 plasma samples described in Example 1.

[0144] The expression levels of miR-21-5p, miR-375-3p, and miR-451a in 300 plasma samples described in Example 1 were quantitatively detected using the in-situ synchronous detection method based on membrane fusion delivery and DSN signal amplification described in Example 2. The expression levels of sEVs membrane proteins GPC1, LYPD3, and NCAM1 in 300 plasma samples described in Example 1 were quantitatively detected using the simultaneous detection method of tumor-derived sEVs membrane proteins based on tag-type surface-enhanced Raman spectroscopy as described in Example 3.

[0145] Then, the CT imaging features corresponding to these 300 plasma samples were obtained, and used as input features along with three miRNAs, three membrane proteins, and CT imaging features for subsequent model training.

[0146] Then, based on the PyCharm programming platform, using the Python 3.x programming language and the scikitlearn machine learning library, six machine learning models were constructed: decision tree, random forest, adaptive augmentation algorithm, logistic regression, support vector machine and artificial neural network.

[0147] Similarly, a 5-fold cross-validation grid search strategy is adopted during the model training and hyperparameter tuning phases. Figure 15 A to Figure 15 E shows the confusion matrix of the training set in the five-fold cross-validation. The accuracies of each fold of the training set are 90.00%, 90.83%, 91.67%, 90.83%, and 90.00%, respectively. In the five-fold cross-validation, the confusion matrix of the test set (...) Figure 16 A~ Figure 16E) reflects the model's classification performance for lung cancer and benign lung diseases. As shown in Table 9, the model's average sensitivity for lung cancer and benign lung diseases is 96.00% and 83.33%, respectively, and the overall classification accuracy on the five-fold test set is 88.33%, 88.33%, 90.00%, 91.67%, and 88.33%, respectively, with an average accuracy of 89.33%. The sensitivities for lung cancer were 96.67%, 96.67%, 96.67%, 96.67%, and 93.33%, with an average sensitivity of 96.00%. The specificities for benign lung diseases were 80.00%, 80.00%, 83.33%, 86.67%, and 83.33%, with an average specificity of 82.67%. The AUC values ​​for each segment were between 0.890 and 0.913, and the 95% confidence intervals remained within the range of 0.812 to 0.996, indicating that the model's generalization performance for binary classification of lung cancer and benign lung diseases was stable, and no obvious overfitting or underfitting was observed.

[0148] Table 9. Classification performance on the five-fold cross-validation test set.

[0149] To comprehensively evaluate the clinical application value of each model, core clinical screening indicators were calculated based on the confusion matrix, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and AUC. AUC was calculated based on ROC curves, with a value ranging from 0.5 to 1.0. The closer the value is to 1.0, the better the model's discriminative efficacy; a value less than 0.5 indicates no clinical diagnostic value.

[0150] The core clinical screening indicators were calculated and ROC curves were plotted using an independent test set that was not involved in model training and parameter tuning. This quantified the clinical diagnostic efficacy of each model to select the best-performing lung cancer auxiliary diagnostic model. At the same time, the differences in core indicators (accuracy, AUC) of the best model between the training set and the independent test set were compared. Combined with the average performance of the 5-fold cross-validation within the training set and the test set, the model was analyzed to determine whether there was overfitting or underfitting, thereby verifying the stability and generalization ability of the model.

[0151] Accuracy was used to evaluate the binary classification performance of each model, while sensitivity, specificity, positive predictive value, negative predictive value, and AUC were used to evaluate model performance. The results are shown in Table 10. Figure 17 As shown.

[0152] Table 10. Classification performance of six models integrating sEVs multi-omics and CT radiomics on the test set.

[0153] It is understood that in this embodiment, the fusion of plasma sEVs miRNA, sEVs membrane proteins, and CT imaging features achieves complementary integration of molecular and imaging information. Based on this, a machine learning model is constructed, which can significantly improve the efficiency of lung cancer screening and treatment prediction. The advantages of this fusion strategy are mainly reflected in the following aspects: First, it achieves a combination of molecular level and anatomical localization.

[0154] Both miRNAs and membrane proteins originate from tumor-derived sEVs. sEVs-miRNAs reflect abnormalities in the post-transcriptional regulatory network of tumor cells (e.g., upregulation of miR-21-5p and downregulation of miR-375-3p), while membrane proteins indicate cell origin and pathological subtype specificity (GPC1 enriched in squamous cell carcinoma, LYPD3 in adenocarcinoma, and NCAM1 in small cell lung cancer). CT images provide macroscopic spatial information such as nodule size, location, morphology, density, and margin features. The fusion allows for dynamic monitoring of molecular marker changes through liquid biopsy and precise lesion localization through imaging, overcoming the dual limitations of single liquid biopsy lacking anatomical localization and single images failing to reflect the essence of molecular biology.

[0155] Second, improve the accuracy of early diagnosis and differential diagnosis.

[0156] In clinical practice, early-stage non-solid nodules and inflammatory pseudotumors often present atypically on CT scans, leading to false positives. However, the miRNAs and membrane proteins of tumor-derived sEVs can provide functional evidence of malignancy. The sEVs six-marker initial screening model and the CT image-fused lung nodule identification model constructed in this study were significantly superior to the single-modality model, achieving a complementary relationship between high molecular sensitivity and high imaging specificity.

[0157] Third, it reduces the interference of tumor spatial heterogeneity and non-tumor vesicles.

[0158] This study utilizes a targeted enrichment strategy to specifically capture tumor-derived sEVs, which can reduce background interference from non-tumor-derived vesicles such as circulating platelets and immune cells to some extent. However, due to the spatial heterogeneity of tumors, single-molecule detection may still result in signal bias. Combining CT imaging can clarify the presence, size, location, and morphological characteristics of solid lesions, and conversely assist in screening for tumor-specific molecular signals highly correlated with imaging phenotypes, further eliminating false positive interference. Simultaneously, molecular markers can help elucidate the biological nature of atypical lesions on imaging, forming a mutually corroborating and complementary multimodal diagnostic system with imaging features.

[0159] Fourth, it meets the clinical needs for precise, non-invasive, and high-frequency monitoring.

[0160] Liquid biopsy of sEVs offers the advantages of being non-invasive and allowing for repeated sampling, making it suitable for continuous dynamic monitoring before and after treatment. CT imaging, as the standard means of baseline assessment and efficacy confirmation in lung cancer, provides crucial imaging evidence. Combining the two can reduce unnecessary frequent CT radiation exposure and enable image review at key points, facilitating the implementation of individualized follow-up and management strategies based on risk stratification.

[0161] In summary, the fusion of multimodal data from miRNA, membrane proteins, and CT images breaks through the limitations of single-modal information dimensions, achieving complementary verification from microscopic molecular features to macroscopic imaging structures. This can provide a more accurate, reliable, and clinically interpretable solution for early lung cancer screening, benign and malignant differentiation, efficacy prediction, and dynamic management throughout the entire process. Example 6

[0162] This embodiment provides a lung cancer auxiliary diagnostic device based on sEVs membrane proteins and intramembrane miRNAs, including: The model library includes a lung cancer auxiliary diagnostic model and a multimodal lung nodule benign / malignant differential diagnostic model. The lung cancer auxiliary diagnostic model is constructed using the method described in the first aspect for constructing a lung cancer auxiliary diagnostic model based on sEVs membrane proteins and intracellular miRNAs, and is used to predict lung cancer based on the expression levels of three miRNAs and three membrane proteins. The multimodal lung nodule benign / malignant differential diagnostic model is used to diagnose lung cancer based on the expression levels of three miRNAs, three membrane proteins, and lung CT imaging features. The molecular data acquisition module is used to acquire the expression levels of three miRNAs and three membrane proteins in the plasma samples of the patients to be tested. The three miRNAs are miR-21-5p, miR-375-3p and miR-451a nucleotide sequences as shown in SEQ. No. 1-3; the three membrane proteins are GPC1, LYPD3 and NCAM1. The CT image acquisition module acquires lung CT images of the patient to be tested and extracts CT image features based on machine vision. The CT image features include nodule diameter, nodule location, nodule morphological features, and nodule density. The diagnostic module is used to take the expression levels of three miRNAs and three membrane proteins in the plasma sample of the patient to be tested as input features, call the lung cancer auxiliary diagnostic model for identification, and obtain a lung cancer prediction result; and when the lung cancer prediction result is lung cancer, it takes the expression levels of the three miRNAs, three membrane proteins and CT imaging features of the patient to be tested as input features, calls the multimodal lung nodule benign and malignant differential diagnosis model for identification, and obtains a lung cancer diagnosis result.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for constructing an auxiliary diagnostic model for lung cancer based on sEVs membrane proteins and intramembrane miRNAs, characterized in that, Includes the following steps: A plasma sample bank was constructed, in which the ratio of normal control group, patients with benign lung disease and lung cancer patients in the plasma samples was 1:1:1; For each plasma sample, a sEVs solution was obtained through separation and purification. The expression levels of three miRNAs in the sEVs membrane were obtained from the sEVs solution using an in situ synchronous detection method based on membrane fusion delivery and DSN signal amplification; the three miRNAs were miR-21-5p, miR-375-3p and miR-451a nucleotide sequences as shown in SEQ. No. 1-3; The expression levels of three membrane proteins of tumor-derived sEVs were obtained from sEVs solution using a tag-based surface-enhanced Raman spectroscopy-based method for simultaneous detection of these membrane proteins. The three membrane proteins were GPC1, LYPD3, and NCAM1. The expression levels of three miRNAs and three membrane proteins from each plasma sample were used as input features and fed into a machine learning model for training to obtain a lung cancer auxiliary diagnostic model to assist in the diagnosis of lung cancer.

2. The method for constructing a lung cancer auxiliary diagnostic model based on sEVs membrane proteins and intramembrane miRNAs according to claim 1, characterized in that, The expression levels of three miRNAs within the sEVs membrane were obtained from sEVs solution using an in situ synchronous detection method based on membrane fusion delivery and DSN signal amplification, including: Purification of target sEVs: The sEVs solution was purified using a targeted separation system based on streptavidin-modified magnetic beads coupled with aptamers to obtain purified target sEVs. Preparation of Lip@AuNFs@DSN: Gold nanoparticles were prepared using the sodium citrate reduction method; three fluorescently labeled DNA probes were coupled to the surface of the gold nanoparticles to obtain gold nanoflares; the gold nanoflares comprised gold nanoparticles and three DNA probes cDNA21, cDNA375, and cDNA451 with nucleotide sequences such as SEQ. No. 12-14 coupled to the surface of the gold nanoparticles, wherein cDNA21 was labeled with a FAM fluorescent group, cDNA375 with a TAMRA fluorescent group, and cDNA451 with a Cy5 fluorescent group; the gold nanoflares and DSN enzyme were dissolved in a reaction buffer for preparing cationic liposomes by thin-film hydration, and cationic liposomes were prepared by thin-film hydration using the reaction buffer to obtain the Lip@AuNFs@DSN, wherein the Lip@AuNFs@DSN comprised cationic liposomes and gold nanoflares and DSN enzyme encapsulated within the cationic liposomes; Fluorescence intensity detection: The purified target sEVs were mixed with the Lip@AuNFs@DSN to undergo membrane fusion, and the fluorescence intensity of each channel was measured simultaneously. A fluorescence intensity-sEVs standard curve was found to obtain the expression levels of the three miRNAs.

3. The method for constructing a lung cancer auxiliary diagnostic model based on sEVs membrane proteins and intramembrane miRNAs according to claim 2, characterized in that, A targeted separation system based on streptavidin-modified magnetic beads coupled with aptamers was used to purify sEVs solutions, including: SAMA magnetic bead preparation: Three dethiobiotin-labeled nucleic acid aptamers were synthesized, targeting EpCAM, PD-L1, and EGFR of small extracellular vesicles, respectively. Utilizing the affinity between streptavidin magnetic beads and dethiobiotin, the three nucleic acid aptamers were coupled to the surface of streptavidin magnetic beads, yielding SAMA magnetic beads with dethiobiotin-labeled aptamers. EpCAM Apt PD-L1 and Apt EGFR SAMA magnetic beads; sEVs capture: SAMA magnetic beads were used to capture sEVs, and the SAMA-sEVs complex was obtained after magnetic separation and purification. Release of sEVs: The captured SAMA-sEVs complex was mixed with free biotin and incubated by shaking to elute and release sEVs from the SAMA-sEVs complex, thus obtaining purified target sEVs.

4. A method for constructing a lung cancer auxiliary diagnostic model based on sEVs membrane proteins and intramembrane miRNAs according to claim 2 or 3, characterized in that, The steps for establishing the fluorescence intensity-sEVs standard curve are as follows: The sEVs standard was serially diluted with PBS to 1.0 × 10⁻⁶. 4 ~5.0×10 8 Particle / μL series concentrations; The target separation system based on streptavidin-modified magnetic beads coupled with aptamers was used to purify the gradient dilution solution of sEVs standard to obtain purified target sEVs. The optimal conditions were 1.0 U of DSN enzyme, 120 min of reaction time, and 37 °C. Under these conditions, the target sEVs purified by each gradient dilution solution were mixed with Lip@AuNFs@DSN at a concentration of 10 × 10⁷ particles / μL to induce membrane fusion, and the fluorescence intensity of each channel was measured simultaneously. A standard curve of fluorescence intensity versus sEVs was established by using linear regression with the logarithm of sEVs concentration on the x-axis and net fluorescence intensity on the y-axis. The regression equation and R² were recorded. Wherein, net fluorescence intensity = FL - FL0, FL is the fluorescence intensity of the target channel, and FL0 is the background fluorescence intensity.

5. The method for constructing a lung cancer auxiliary diagnostic model based on sEVs membrane proteins and intramembrane miRNAs according to claim 1, characterized in that, Three membrane proteins of sEVs were obtained from the sEVs solution using a simultaneous detection method, including: Preparation of AgMNPs-Apt magnetic beads: Fe3O4 nanoparticles were used as the magnetic core, and gold nanoparticles were loaded via electrostatic adsorption to form a Fe3O4-AuNPs composite structure. The Fe3O4-AuNPs composite structure was modified with the internal standard molecule 4-mercaptobenzonitrile via Au-S bonds. Then, a silver shell was encapsulated to obtain silver magnetic nanoparticles. Finally, three DNA sequences, namely the aptamers Apt (SEQ No. 8-10), were simultaneously coupled to the silver magnetic nanoparticles via Ag-S bonds. EpCAM Apt PD-L1 Apt EGFR The AgMNPs-Apt were obtained. SERS probe preparation: Gold nanorods were prepared using a seed growth method. Three Raman reporter molecules—0.01 mol / L 2-nitrobenzoic acid, 0.01 mol / L 2-mercaptopyridine, and 0.01 mol / L 2-nitrotoluene—were used to modify the corresponding gold nanorods via Au-S bonds. Ag-Au alloy shells were deposited on the surfaces of the three gold nanorods by co-reduction of HAuCl4 and AgNO3 with ascorbic acid. The Ag components were selectively etched using Fe(NO3)3 to form porous core-shell gold nanorods. Finally, specific antibodies targeting membrane protein markers GPC1, LYPD3, and NCAM1 were coupled to the porous core-shell gold nanorods to obtain three SERS probes. Capture of sEVs and assembly of sandwich complex: The AgMNPs-Apt solution was mixed with sEVs solution to form a co-incubation system. The sEVs were separated and enriched by aptamer-membrane protein specific binding. Three SERS probes were added to the co-incubation system in sequence. The AgMNPs-Apt-sEVs-SERS probe sandwich structure was assembled by antibody-antigen recognition. The free probes were removed by magnetic separation. After washing, the purified complex was obtained. Raman signal detection: Under the same optical and acquisition parameters, the Raman intensity of each probe was acquired, and the corresponding Raman intensity-sEVs standard curve was found to obtain the expression levels of the three membrane proteins.

6. The method for constructing a lung cancer auxiliary diagnostic model based on sEVs membrane proteins and intramembrane miRNAs according to claim 5, characterized in that, The steps for establishing the Raman intensity-sEVs standard curve are as follows: The sEVs standard was serially diluted 1.0 × 10⁻⁶ with PBS. 3 1.0×10 4 1.0×10 5 1.0×10 6 1.0×10 7 1.0×10 8 1.0×10 9 particles / μL; With Apt EpCAM +Apt EGFR +Apt PD-L1 Using three aptamers, with magnetic beads at a volume of 20 μL, incubation time of 50 min, probe volume of 25 μL, and incubation time of 30 min as optimal conditions, AgMNPs-Apt were mixed with serially diluted solutions of each SEV standard to form a co-incubation system. The separation and enrichment of SEVs were achieved through aptamer-membrane protein specific binding. Three SERS probes were added sequentially to the co-incubation system, and an AgMNPs-Apt-sEVs-SERS probe sandwich structure was formed by antibody-antigen recognition assembly. Free probes were removed by magnetic separation, and the purified complex was obtained after washing. The Raman intensities of the three SERS probes were collected and recorded under the same optical and acquisition parameters. Plotting the logarithm of sEVs concentration on the x-axis and the normalized characteristic peak intensity on the y-axis, three Raman intensity-sEVs standard curves were established using linear regression. The regression equations and correlation coefficients R were recorded. 2 .

7. A method for auxiliary diagnosis of lung cancer based on sEVs membrane proteins and intramembrane miRNAs, characterized in that, Includes the following steps: Using the method for constructing the lung cancer auxiliary diagnostic model according to any one of claims 1-6, multiple models are constructed, and the model with the best performance is selected as the lung cancer auxiliary diagnostic model; Obtain plasma samples from the patients to be tested, and obtain sEVs solution through separation and purification; The expression levels of three miRNAs in the sEVs membrane were obtained from the sEVs solution using an in situ synchronous detection method, and the expression levels of three membrane proteins in sEVs were obtained from the sEVs solution using a synchronous detection method. The expression levels of three miRNAs and three membrane proteins were used as input features and fed into a lung cancer auxiliary diagnostic model for identification to obtain lung cancer prediction results.

8. The method for auxiliary diagnosis of lung cancer based on sEVs membrane proteins and intramembrane miRNAs according to claim 7, characterized in that, When the lung cancer prediction result is lung cancer, lung CT images of the patient to be tested are obtained, and CT imaging features are extracted from the lung CT images based on machine vision. The CT imaging features include the diameter of the CT image nodules, the location of the nodules, the morphological features of the nodules, and the density of the nodules. The expression levels of three miRNAs, three membrane proteins, and CT imaging features of the patients to be tested are used as input features and fed into a trained multimodal lung nodule benign and malignant differential diagnosis model to obtain lung cancer diagnosis results.

9. The method for auxiliary diagnosis of lung cancer based on sEVs membrane proteins and intramembrane miRNAs according to claim 8, characterized in that, The training steps for the multimodal pulmonary nodule benign-malignant differential diagnosis model are as follows: A patient sample bank was constructed, with a 1:1 ratio of patients with benign lung diseases to patients with lung cancer. Each patient sample included a plasma sample and a lung CT image sample. For each plasma sample, sEVs solution was obtained through separation and purification. The expression levels of three miRNAs within the sEVs membrane were obtained from the sEVs solution using an in-situ simultaneous detection method based on membrane fusion delivery and DSN signal amplification. The three miRNAs were the nucleotide sequences miR-21-5p, miR-375-3p, and miR-451a, as shown in SEQ. No. 1-3. The expression levels of three sEVs membrane proteins were obtained from the sEVs solution using a simultaneous detection method based on tag-type surface-enhanced Raman spectroscopy for tumor-derived sEVs membrane proteins. The three membrane proteins were GPC1, LYPD3, and NCAM1. For each lung CT image sample, CT imaging features are extracted based on machine vision. These features include nodule diameter, nodule location, nodule morphology, and nodule density. The expression levels of three miRNAs, three membrane proteins, and corresponding CT imaging features of each patient sample were used as input features and fed into a machine learning model for training to obtain a multimodal differential diagnostic model for benign and malignant pulmonary nodules.

10. A lung cancer auxiliary diagnostic device based on sEVs membrane proteins and intramembrane miRNAs, characterized in that, include: The model library includes a lung cancer auxiliary diagnostic model and a multimodal lung nodule benign / malignant differential diagnostic model. The lung cancer auxiliary diagnostic model is constructed using the method described in any one of claims 1-6 for constructing a lung cancer auxiliary diagnostic model based on sEVs membrane proteins and intracellular miRNAs, and is used to predict lung cancer based on the expression levels of three miRNAs and three membrane proteins. The multimodal lung nodule benign / malignant differential diagnostic model is used to diagnose lung cancer based on the expression levels of three miRNAs, three membrane proteins, and lung CT imaging features. The molecular data acquisition module is used to acquire the expression levels of three miRNAs and three membrane proteins in the plasma samples of the patients to be tested. The three miRNAs are miR-21-5p, miR-375-3p and miR-451a nucleotide sequences as shown in SEQ. No. 1-3; the three membrane proteins are GPC1, LYPD3 and NCAM1. The CT image acquisition module acquires lung CT images of the patient to be tested and extracts CT image features based on machine vision. The CT image features include nodule diameter, nodule location, nodule morphological features, and nodule density. The diagnostic module is used to take the expression levels of three miRNAs and three membrane proteins in the plasma sample of the patient to be tested as input features, call the lung cancer auxiliary diagnostic model for identification, and obtain a lung cancer prediction result; and when the lung cancer prediction result is lung cancer, it takes the expression levels of the three miRNAs, three membrane proteins and CT imaging features of the patient to be tested as input features, calls the multimodal lung nodule benign and malignant differential diagnosis model for identification, and obtains a lung cancer diagnosis result.

Images