A tumor subcellular protein analysis method based on two-dimensional to three-dimensional deep learning and targeted fluorescent nanoprobes

Abstract

The application discloses a tumor subcellular protein analysis method based on two-dimensional-to-three-dimensional deep learning and a targeted fluorescent nanoprobe, and relates to the field of biological information technology.The targeted fluorescent nanoprobe comprises a nanoparticle matrix, the nanoparticle matrix is internally loaded with a fluorescent marker, and the surface of the nanoparticle matrix is sequentially modified with a hydrophilic modification layer and a protein target recognition unit grafted on the surface of the hydrophilic modification layer.A 2D-to-3D deep learning analysis framework for a confocal nanoprobe-labeled fluorescence image is built, the three-dimensional structure of a cell is automatically predicted and generated from stacked two-dimensional images, and subcellular segmentation and fluorescent three-dimensional quantitative statistical analysis are simultaneously completed.The quantitative system deviation caused by two-dimensional slices is eliminated, and the three-dimensional distribution of a target protein is truly restored;full-automatic operation is adopted, manual reconstruction and labeling steps are omitted, the analysis flux and data repeatability are greatly improved, and tumor single-cell heterogeneity is accurately quantified.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method for analyzing tumor subcellular proteins based on two-dimensional to three-dimensional deep learning and targeted fluorescent nanoprobes. Background Technology

[0002] Aberrant expression, dysregulation of localization, and abnormal activation of subcellular proteins are important contributing factors to the occurrence, progression, and drug resistance of various malignant tumors, including lung cancer, breast cancer, and gastric stromal tumors. Tumor cells exhibit significant single-cell heterogeneity, with considerable differences in the activation levels and subcellular spatial distribution of target proteins across different cells. This is a key reason for poor response to targeted therapy and tumor recurrence. Therefore, achieving in-situ identification, spatial localization, and precise quantification of subcellular proteins at the single-cell level is of significant practical importance for tumor pathology research, targeted drug guidance, and efficacy monitoring.

[0003] Current protein detection and analysis technologies struggle to simultaneously achieve both single-cell resolution and subcellular three-dimensional characterization capabilities. Conventional detection methods often rely on cell lysate samples or tissue sections for analysis, reflecting only the overall level of the cell population and failing to resolve the true spatial distribution of proteins within individual cells. Fluorescence imaging, with its advantages of high resolution and in-situ visualization, has become the mainstream method for single-cell protein analysis, and various nanofluorescent probes are widely used for specific labeling of tumor targets. However, current mainstream analytical protocols are based on two-dimensional fluorescence section images for identification and quantification. Due to the limitations of imaging dimension, two-dimensional single-section data cannot reconstruct the complete three-dimensional structure of cells, resulting in significant deviations in quantitative results and making it difficult to accurately reflect the overall expression and spatial distribution characteristics of subcellular proteins.

[0004] Traditional small-molecule fluorescent probes generally suffer from drawbacks such as poor water solubility, weak cell penetration, and insufficient subcellular targeting, making it difficult to accurately identify target proteins within tumor cells. In contrast, nanoparticles, polymeric microspheres, and metal-organic frameworks, with their excellent structural designability, can achieve specific recognition of tumor proteins in different subcellular regions through surface modification and conjugation of targeting groups such as small-molecule inhibitors, antibodies, and aptamers. However, these nanoprobes still have limitations: the controllability of modification and intracellular delivery stability vary among different carriers, and most existing probes are designed for two-dimensional imaging scenarios, failing to fully adapt to the application requirements of three-dimensional observation.

[0005] To overcome the limitations of two-dimensional imaging, the industry often uses confocal microscopy to perform Z-axis layer-by-layer scanning, acquiring a continuous stack of two-dimensional images, and then reconstructing the three-dimensional morphology and fluorescence signal distribution of cells through image reconstruction. However, manually performing three-dimensional reconstruction, region segmentation, and data statistics is labor-intensive, inefficient, and prone to introducing human error, failing to meet the needs of high-throughput sample detection.

[0006] Deep learning technology offers a novel approach to intelligent image analysis. However, most existing image models are developed for single 2D images, making it difficult to handle continuously stacked image sequences and achieve automatic conversion of 2D raw data into 3D spatial parameters. To address these existing problems, those skilled in the art are dedicated to developing an integrated analysis system combining multifunctional targeted nanoprobes with 2D-to-3D deep learning algorithms. This system aims to effectively solve pain points such as biased 2D detection results and low efficiency of manual analysis, providing a novel technical solution for single-cell 3D detection of tumor subcellular proteins. Summary of the Invention

[0007] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is to develop a method for tumor subcellular protein analysis based on two-dimensional to three-dimensional deep learning and targeted fluorescent nanoprobes.

[0008] To achieve the above objectives, this invention provides a method for tumor subcellular protein analysis based on two-dimensional to three-dimensional deep learning and targeted fluorescent nanoprobes.

[0009] Furthermore, a subcellular targeted fluorescent nanoprobe for three-dimensional fluorescence imaging includes a nanoparticle matrix (nanoseeds), wherein the nanoparticle matrix is ​​loaded with a fluorescent marker, and its surface is sequentially modified with a hydrophilic modification layer and a protein targeting recognition unit grafted onto the surface of the hydrophilic modification layer, exhibiting excellent intracellular stability and subcellular targeting capability. The nanoparticles are silica, polystyrene-acrylic acid copolymer, or ZIF series metal-organic framework nanoparticles.

[0010] Furthermore, a method for preparing the subcellular targeted fluorescent nanoprobe for three-dimensional fluorescence imaging includes the following steps: Step 1: Based on the matrix characteristics, select the appropriate method to bind the fluorescent label to the interior or surface of the nanoparticles, and stir continuously for 5-6 hours to prepare fluorescent functionalized nanoparticles; among them, silica and polystyrene-acrylic acid copolymer are fixed by covalent bonding, and ZIF material relies on its own porous cavity to embed fluorescent molecules through hydrophilic and hydrophobic interactions.

[0011] Step 2: Using polyethylene glycol as a hydrophilic linker, the protein targeting molecule is grafted to the end of the PEG molecular chain using an end-group active reagent. After solvent dispersion, constant temperature stirring, dialyzing and freeze drying, the targeting-PEG composite functional molecule is obtained. Step 3: Mix the targeted-PEG composite functional molecule with the fluorescent functionalized nanoparticles and react them. After stirring overnight, purify the mixture by centrifugation and repeated washing to obtain the targeted fluorescent nanoprobe.

[0012] Furthermore, the matrix is ​​polystyrene-acrylic acid copolymer nanoparticles, which undergo secondary surface modification using natural polysaccharides to enhance the biocompatibility and resistance to non-specific adsorption. The polysaccharides are selected from one or more of dextran, chitosan, hyaluronic acid, alginate, agarose, and pectin, and the modification is achieved through physical adsorption and supramolecular interactions.

[0013] Furthermore, the matrix is ​​a ZIF metal-organic framework particle, and its surface is modified using bioactive molecules to enhance its structural stability in physiological environments. The bioactive molecules are polysaccharides; these polysaccharides are selected from one or more of dextran, chitosan, hyaluronic acid, alginate, agarose, and pectin, and the modification is primarily through physical processes.

[0014] Furthermore, in step two, the protein-targeting molecule is a small molecule targeted drug or a monoclonal antibody; The small molecule targeted drug is gefitinib, erlotinib, imatinib, or dasatinib; The monoclonal antibody is either bevacizumab or trastuzumab.

[0015] Furthermore, by adjusting the synthesis reaction time, monomer ratio, and reaction temperature, the particle size of the nanoparticle matrix can be controlled within the range of 100~500nm, adapting to the labeling needs of different subcellular structures.

[0016] Furthermore, fluorescent markers can be flexibly selected based on imaging equipment, detection system, and solvent environment to match different fluorescence detection scenarios.

[0017] Furthermore, the binding modes of fluorescent labels to nanomaterials include three categories: covalent bonding, supramolecular interactions, and physical encapsulation.

[0018] Furthermore, in step two, the molecular weight of the hydrophilic linker is 3000, 5000, or 10000; and the end group of the polyethylene glycol linker is amino, carboxyl, or NHS active ester.

[0019] Furthermore, the fluorescent marker is selected from one or more of the following: fluorescein isothiocyanate, cyanine dyes, rhodamine, and Alexa Fluor series fluorescent reagents.

[0020] Furthermore, a method for analyzing tumor subcellular proteins utilizes the aforementioned targeted fluorescent nanoprobes, combined with confocal continuous imaging and intelligent algorithms, to achieve three-dimensional segmentation of the cell membrane, cytoplasm, and organelles. It also quantifies and statistically analyzes the positioning and fluorescence signal distribution of the nanoprobes, thereby effectively and comprehensively identifying protein mutations and abnormally activating proteins in tumor cells. Simultaneously, it enables batch analysis of heterogeneous tumor cells and clinical primary cells. Specifically, it includes the following steps: S1. The target fluorescent nanoprobe is used to incubate and label the cells to be tested; S2. The labeled cells are fixed and stained, and then the Z-axis is scanned layer by layer using a confocal microscope to obtain a two-dimensional image stack. S3. Input the two-dimensional image stack into a pre-trained deep learning model. The model extracts the image features of each two-dimensional slice layer by layer, reconstructs the three-dimensional structure of a single cell, and automatically segments the cell membrane, cytoplasm and nucleus regions in three-dimensional space. S4. Based on the segmented three-dimensional structure, calculate the co-localization relationship between the targeted fluorescent nanoprobe and the target protein, or calculate the three-dimensional volume ratio of the targeted fluorescent nanoprobe in the subcellular region. S5. Determine the expression level and mutation degree of the target protein based on the statistical results.

[0021] Furthermore, in step S1, the cells to be tested are experimental tumor cell models or clinically derived samples; The experimental tumor cell models include any one or more of the following: EGFR exon19 deletion mutant non-small cell lung cancer HCC827 cells, EGFR T790M drug-resistant mutant non-small cell lung cancer H1975 cells, c-KIT mutant gastric stromal tumor GIST-T1 cells, BCR-ABL fusion type chronic myeloid leukemia K562 cells, HER2 positive breast cancer SK-BR-3 cells, KRAS G12D mutant colorectal cancer HCT116 cells, BRAF V600E mutant melanoma A375 cells, and any one or more of the following as normal control cells: normal lung epithelial cells BEAS-2B, normal human breast epithelial cells MCF-10A, normal human colorectal epithelial cells NCM460, human gastric mucosal cells GES-1, and human umbilical vein endothelial cells HUVEC. The clinically sourced samples include surgically removed tumor tissue, malignant pleural effusion, malignant ascites, peripheral blood circulating tumor cells, and biopsy tissue. Two patient groups can be selected for the samples: one group consists of newly diagnosed patients who have not received any targeted drug intervention, used to detect baseline levels of protein mutations and abnormal activation; the other group consists of patients who have developed drug resistance and relapse after targeted therapy, used to analyze the expression changes and spatial distribution characteristics of drug resistance-related proteins.

[0022] Further, in step S2, the acquisition parameters of the two-dimensional image stack are: layer thickness 0.5~1μm, 20~100 two-dimensional slices are continuously acquired for each single cell sample; before inputting the two-dimensional image stack into the deep learning model, preprocessing such as background denoising, multi-fluorescence channel splitting, invalid edge slice removal and coarse screening of cell contours is automatically completed.

[0023] Furthermore, in step S3, the deep learning model is built based on nnU-Net, 3D-UNet, CNN convolutional neural network or general vision large model, and is trained using a manually labeled confocal 2D image stack-3D ground truth pairing dataset. After cross-validation optimization, the subcellular region segmentation accuracy is ≥90%, and the deviation between the 3D quantitative result and the gold standard manual detection is ≤10%.

[0024] The model first extracts features frame-by-frame from each preprocessed 2D slice, then completes spatial information association through a cross-slice 3D feature fusion module, outputting voxel-level single-cell 3D morphology to fully restore the cell's true spatial structure, eliminate sampling bias from a single 2D slice, and simultaneously completes 3D segmentation of subcellular regions, classifying multiple subcellular regions such as cell membrane, cytoplasm, nucleus, endoplasmic reticulum, and mitochondria. It also distinguishes between target probe-specific fluorescence signals, cell autofluorescence, background noise, and residual signals from free probes, significantly improving the accuracy of subcellular localization. Based on the 3D reconstruction and segmentation results, the model can automatically perform multi-dimensional quantitative statistics, outputting indicators including: total fluorescence intensity, fluorescence signal volume ratio, signal spatial distribution uniformity, and relative expression level of target proteins in each subcellular region. Simultaneously, it statistically analyzes the number and proportion of high-expression, low-expression, and negative cells within the entire sample, completing the automatic classification of tumor cell subpopulations.

[0025] In a preferred embodiment 1 of the present invention, the preparation process of the gefitinib-modified EGFR-targeting silica fluorescent nanoprobe is described in detail.

[0026] In a preferred embodiment 2 of the present invention, the preparation process of the dasatinib-modified BCR-ABL-targeted silica fluorescent nanoprobe is described in detail.

[0027] In a preferred embodiment 3 of the present invention, the preparation process of the imatinib-modified c-KIT-targeted silica fluorescent nanoprobe is described in detail.

[0028] In a preferred embodiment 4 of the present invention, the preparation process of the hyaluronic acid modified EGFR-targeting ZIF-8 fluorescent nanoprobe is described in detail.

[0029] In a preferred embodiment 5 of the present invention, the preparation process of the hyaluronic acid modified BCR-ABL targeting PS-co-PAA fluorescent nanoprobe is described in detail.

[0030] In a preferred embodiment 6 of the present invention, the preparation process of the chitosan-modified HER2-targeting PS-co-PAA fluorescent nanoprobe is described in detail.

[0031] In the preferred embodiment 7 of the present invention, the experimental process for verifying the targeting specificity of EGFR-targeting fluorescent nanoprobes in EGFR-positive lung cancer cells is described in detail.

[0032] In a preferred embodiment 8 of the present invention, the experimental process for verifying the targeting specificity of the c-KIT-targeted fluorescent nanoprobe in c-KIT-positive gastric stromal tumor cells is described in detail.

[0033] In a preferred embodiment 9 of the present invention, the experimental process for verifying the targeting specificity of the BCR-ABL-targeted fluorescent nanoprobe in BCR-ABL-positive chronic myeloid leukemia cells is described in detail.

[0034] In a preferred embodiment 10 of the present invention, the process of constructing a two-dimensional to three-dimensional deep learning model is described in detail.

[0035] In a preferred embodiment 11 of the present invention, the process of analyzing the three-dimensional mention ratio of membrane mutation targets (EGFR, c-KIT) based on a two-dimensional to three-dimensional deep learning model is described in detail.

[0036] In a preferred embodiment 12 of the present invention, the process of analyzing the proportion of three-dimensional colocalization of intracellular mutation targets (BCR-ABL) based on a two-dimensional to three-dimensional deep learning model is described in detail.

[0037] In a preferred embodiment 13 of the present invention, the three-dimensional quantitative analysis process of standard tumor cell line probes is described in detail.

[0038] In a preferred embodiment 14 of the present invention, the process of quantitative analysis of subcellular proteins in clinical tumor samples based on a two-dimensional to three-dimensional deep learning model is described in detail.

[0039] The principle of this invention is as follows: This invention selects easily modifiable and stable substrates such as silica, polymer microspheres, and MOF as nanoparticle matrices. Through precise and controllable surface covalent modification, it can directionally couple highly specific targeting units such as small molecule inhibitors, monoclonal antibodies, and aptamers. Simultaneously, it modifies hydrophilic polymers to enhance the stability of the physiological environment, regulates the probe particle size and surface functional group arrangement, and precisely binds to abnormally activated proteins in the cell membrane, cytoplasm, and organelles.

[0040] This invention trains a neural network using a confocal 2D stacked-paired 3D ground truth dataset with annotations; the network fuses the image features of each 2D slice layer by layer, maps and reconstructs the complete 3D morphology of a single cell, automatically divides subcellular regions such as cell membrane and cytoplasm, and statistically analyzes the fluorescence signal intensity and volume ratio in 3D space.

[0041] Technical effects: 1. The modular universal targeted fluorescent nanoprobe system established in this invention is compatible with various nanomaterials and targeted recognition molecules, and is specifically designed to meet the labeling and detection requirements of three-dimensional stacked imaging. It significantly improves probe targeting specificity and subcellular localization accuracy, greatly reduces non-specific adsorption, and exhibits stronger intracellular stability, supporting long-term continuous stacked image acquisition. It provides high signal-to-noise ratio raw fluorescent images for backend 2D-to-3D deep learning analysis. The entire detection platform can be flexibly adapted to multiple cancer types and multiple subcellular target detection scenarios, offering greater scalability and versatility.

[0042] 2. The 2D-to-3D deep learning analysis framework for fluorescent images labeled with confocal nanoprobes, constructed in this invention, automatically predicts and generates three-dimensional cell structures from stacked two-dimensional images, simultaneously completing subcellular segmentation and three-dimensional quantitative statistical analysis of fluorescence. This eliminates the quantitative system bias caused by two-dimensional slices, accurately restoring the three-dimensional distribution of target proteins; the entire process is automated, eliminating manual reconstruction and annotation steps, significantly improving analysis throughput and data repeatability, and accurately quantifying tumor single-cell heterogeneity. Attached Figure Description

[0043] Figure 1 This invention combines 2D to 3D deep learning to achieve subcellular protein labeling and analysis.

[0044] Flowchart.

[0045] Figure 2 This is the verification result of the three-dimensional segmentation model in Example 10 on the segmentation accuracy of various cell structures.

[0046] Figure 3 This is a comparison of the three-dimensional binding ratios of probes in different positive and negative cell lines in Example 13. Detailed Implementation

[0047] The following describes several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0048] Example 1: Preparation of EGFR-targeting fluorescent nanoprobes modified with gefitinib

[0049] 1. Weigh gefitinib and NHS-PEG-silane coupling agent (molecular weight 5000) at a molar ratio of 1.2:1, add 20 mL of dimethyl sulfoxide (DMSO), and stir at room temperature for 24 h to complete the coupling reaction; after purification by dialysis bag with a molecular weight cutoff of 1000 Da, the product is freeze-dried to obtain the dried product, namely gefitinib-PEG-silane coupling agent.

[0050] 2. Mix 9.8 mg of fluorescein isothiocyanate (FITC) with 1.4 mL of 3-aminopropyltrimethoxysilane (APTMS), add 20 mL of anhydrous ethanol, stir in the dark for 24 h to obtain the FITC-APTMS fluorescent intermediate.

[0051] 3. Add 1 mL of tetraethyl orthosilicate (TEOS), 5 mL of 28 wt% ammonia, and 50 mL of anhydrous ethanol to a reaction flask and stir at room temperature for 3 h to obtain a dispersion of silica nanoparticles. Add FITC-APTMS at a nanoparticle to fluorescein mass ratio of 500:1 and stir at room temperature in the dark for 5 h. Then add 10 mL of an ethanol solution of gefitinib-PEG-silane coupling agent with a concentration of 1 mg / mL and stir at room temperature in the dark for 24 h. After the reaction is completed, wash the nanoparticles with anhydrous ethanol and deionized water alternately by centrifugation 3-4 times to obtain gefitinib-modified EGFR-targeting silica fluorescent nanoprobes.

[0052] Example 2: Preparation of dasatinib-modified BCR-ABL-targeting silica fluorescent nanoprobes

[0053] 1. Weigh dasatinib and NHS-PEG-silane coupling agent (molecular weight 5000) at a molar ratio of 1.15:1, add 14 mL of N,N-dimethylformamide (DMF), and stir at room temperature for 24 h; the product is dialyzed and lyophilized to obtain the dried product of dasatinib-PEG-silane coupling agent.

[0054] 2. Mix 10.7 mg FITC with 1.5 mL APTMS, add 25 mL anhydrous ethanol, and stir in the dark for 24 h to obtain FITC-APTMS.

[0055] 3. Add 4 mL of TEOS, 15 mL of 28 wt% ammonia, and 150 mL of anhydrous ethanol to a reaction flask and stir at room temperature for 3 h to obtain a dispersion of silica nanoparticles. Add FITC-APTMS at a nanoparticle to fluorescein mass ratio of 700:1 and stir at room temperature in the dark for 6 h. Then add 5 mL of 2 mg / mL dasatinib-PEG-silane coupling agent ethanol solution and stir at room temperature in the dark for 24 h. Centrifuge, wash, and purify to obtain dasatinib-modified BCR-ABL targeted silica fluorescent nanoprobe.

[0056] Example 3: Preparation of hyaluronic acid-modified EGFR-targeting ZIF-8 fluorescent nanoprobes

[0057] 1. Weigh imatinib and COOH-PEG-silane coupling agent (molecular weight 5000) at a molar ratio of 1.12:1, add 20 mL of DMSO, and stir at room temperature for 24 h; the product is dialyzed and lyophilized to obtain the dried imatinib-PEG-silane coupling agent product.

[0058] 2. Mix 15 mg of rhodamine fluorescein with 2 mL of APTMS, add 50 mL of anhydrous ethanol, and stir in the dark for 24 h to obtain rhodamine-APTMS.

[0059] 3. Add 1 mL TEOS, 4 mL 28 wt% ammonia, and 50 mL anhydrous ethanol to a reaction flask and stir at room temperature for 5 h to obtain a dispersion of silica nanoparticles; add Rhodamine-APTMS at a nanoparticle to fluorescein mass ratio of 1000:1 and stir at room temperature in the dark for 4 h, then add 5 mL of 2 mg / mL imatinib-PEG-silane coupling agent ethanol solution and stir at room temperature in the dark for 24 h; centrifuge, wash and purify to obtain imatinib-modified c-KIT targeted silica fluorescent nanoprobe.

[0060] Example 4: Preparation of hyaluronic acid-modified EGFR-targeting ZIF-8 fluorescent nanoprobes

[0061] 1. Prepare 0.15 g / mL zinc nitrate hexahydrate aqueous solution, 2.85 g / mL dimethylimidazole aqueous solution, and 0.5 mg / mL FITC DMSO solution respectively.

[0062] 2. Mix 1 mL of zinc nitrate hexahydrate solution, 5.7 mL of FITC-DMSO solution, and 5 mL of deionized water, and stir at room temperature for 30 min. Adjust the FITC concentration according to the mass ratio of nanoparticles to fluorescein of 500:1. Add 10 mL of dimethylimidazole solution and continue stirring at room temperature for 1 h. Collect the product by centrifugation at 14000 rpm for 20 min, and wash it three times with deionized water to obtain FITC-labeled ZIF-8 nanoparticles.

[0063] 3. Prepare 2 mg / mL hyaluronic acid aqueous solution and 2 mg / mL gefitinib ethanol solution respectively, and mix them in equal volumes to obtain the targeting modification solution; mix 2 mg / mL FITC-ZIF-8 aqueous solution with the targeting modification solution at a volume ratio of 5:1, and shake at room temperature for 12 h to obtain hyaluronic acid modified EGFR targeting ZIF-8 fluorescent nanoprobe.

[0064] Example 5: Preparation of hyaluronic acid-modified BCR-ABL-targeting PS-co-PAA fluorescent nanoprobes

[0065] 1. Add 200 μL of acrylic acid and 45 mL of deionized water to the reactor, stir to dissolve, then add 920 μL of styrene and purge with nitrogen for 30 min to remove oxygen; add 0.03 g of potassium persulfate (dissolved in 10 mL of deionized water), heat to 70 °C, and stir under nitrogen protection for 10 h; centrifuge and wash to obtain PS-co-PAA nanoparticles.

[0066] 2. Disperse PS-co-PAA particles in MES buffer at pH 5.5 to prepare a 4 mg / mL solution; take 50 mg of this solution and add 6.4 mg of EDC sequentially. The nanoparticles were activated with HCl for 10 min, then with 18.2 mg sulfo-NHS for 3 h, and then FITC was added and reacted in the dark for 24 h. The amount of FITC was adjusted according to the mass ratio of nanoparticles to fluorescein of 800:1. After centrifugation and washing, FITC-labeled PS-co-PAA particles were obtained.

[0067] 3. Immerse the fluorescent particles in 5 mL of 5 mg / mL hyaluronic acid aqueous solution, vortex mix for 1 h, and centrifuge to obtain hyaluronic acid modified particles; then immerse them in 2 mL of 5 mg / mL diamino PEG solution, add EDC / NHS activation system and react overnight.

[0068] 4. The particles were dispersed in an EDC / NHS activated aqueous solution containing 0.5 mg / mL dasatinib, reacted overnight, and then centrifuged and washed to obtain hyaluronic acid modified BCR-ABL targeting PS-co-PAA fluorescent nanoprobe.

[0069] Example 6: Preparation of chitosan-modified HER2-targeting PS-co-PAA fluorescent nanoprobes

[0070] 1. Add 250 μL of acrylic acid and 50 mL of deionized water to the reactor, stir to dissolve, then add 1050 μL of styrene and purge with nitrogen for 30 min to remove oxygen; add 0.04 g of potassium persulfate, heat to 70 °C, and stir under nitrogen protection for 12 h; centrifuge and wash to obtain PS-co-PAA nanoparticles.

[0071] 2. Disperse the particles in MES buffer at pH 5.5 to prepare a 5 mg / mL solution; take 50 mg of this solution and add 6.8 mg of EDC sequentially. After activation with HCl and 19.4 mg sulfo-NHS, rhodamine fluorescein was added and reacted in the dark for 24 h. The amount of fluorescein was adjusted according to the mass ratio of nanoparticles to fluorescein of 800:1. Rhodamine-labeled PS-co-PAA particles were obtained by centrifugation and washing.

[0072] 3. Immerse the fluorescent particles in 5 mL of 5 mg / mL chitosan aqueous solution, vortex mix for 3 h, and centrifuge to obtain chitosan-modified particles; then immerse them in 2 mL of 5 mg / mL diamino-terminated PEG solution, add EDC / NHS activation system and react overnight.

[0073] 4. The particles were dispersed in an EDC / NHS activated aqueous solution containing 0.5 mg / mL trastuzumab, reacted overnight, and then centrifuged and washed to obtain chitosan-modified HER2-targeting PS-co-PAA fluorescent nanoprobes.

[0074] Example 7: Validation experiment of the targeting specificity of EGFR-targeting fluorescent nanoprobes in EGFR-positive lung cancer cells

[0075] 1. EGFR-negative human lung squamous cell carcinoma cells NCI-H520 and EGFR-mutant non-small cell lung cancer cells HCC827 were mixed and seeded in confocal culture dishes. When the cell adhesion reached 70%~80%, the old culture medium was removed, and 1.5 mL of fresh culture medium containing 30 μg / mL of gefitinib-modified EGFR-targeting fluorescent nanoprobe prepared in Example 1 was added to each dish. The mixture was incubated at 37°C for 6 h.

[0076] 2. After removing the culture medium, the sample was washed three times by centrifugation with PBS buffer at 500g. It was then fixed with 4% paraformaldehyde at room temperature for 20 min, and stained with cell membrane dyes and DAPI nuclear dyes. Z-axis layer-by-layer scanning was performed using a laser confocal microscope with a layer thickness of 0.8 μm. Fifty consecutive two-dimensional fluorescent sections were acquired from each sample to form a two-dimensional image stack for subsequent three-dimensional analysis. Before inputting the data into the model, the original images underwent automatic preprocessing steps including background denoising, multi-fluorescence channel splitting, removal of invalid edge sections, and coarse screening of cell contours.

[0077] 3. Imaging results showed that the probe fluorescence signal of HCC827 cells was mainly enriched in the cell membrane region, while the probe fluorescence signal of NCI-H520 cells had no specific distribution and was diffused inside the cell.

[0078] Example 8: Validation experiment of the targeting specificity of c-KIT-targeted fluorescent nanoprobes in c-KIT-positive gastric stromal tumor cells.

[0079] 1. Mix c-KIT negative human gastric mucosal cells GES-1 and c-KIT mutant gastric stromal tumor cells GIST-T1 and seed them in a confocal culture dish. When the cell adhesion reaches 70%~80%, add culture medium containing 30μg / mL c-KIT targeted fluorescent nanoprobe and incubate at 37℃ for 6h.

[0080] 2. After washing, fixing, and staining with PBS, the samples were scanned along the Z-axis using a confocal microscope with a layer thickness of 0.8 μm. Fifty two-dimensional slices were continuously collected from each sample to form a two-dimensional image stack.

[0081] 3. Imaging results showed that the probe fluorescence signal of GIST-T1 cells was specifically enriched in the cell membrane, while the probe signal of GES-1 cells was diffusely distributed in the cytoplasm.

[0082] Example 9: Validation experiment on the targeting specificity of BCR-ABL-targeted fluorescent nanoprobes in BCR-ABL-positive chronic myeloid leukemia cells.

[0083] 1. Mix BCR-ABL negative human breast cancer cells MCF-7 and BCR-ABL positive chronic myeloid leukemia cells K562 and seed them in a confocal culture dish. Add culture medium containing 30 μg / mL BCR-ABL targeted fluorescent nanoprobes and incubate at 37°C for 6 h.

[0084] 2. After washing and fixing with PBS, cells were permeated with 0.1% Triton X-100 solution, incubated overnight at 4°C with BCR-ABL primary antibody, and immunofluorescence staining was completed the next day with secondary fluorescent antibody. Cell nuclei were labeled with DAPI. Confocal microscopy was used for Z-axis scanning at a slice thickness of 0.8 μm, and 60 consecutive two-dimensional slices were acquired from each sample to form a two-dimensional image stack. Before inputting the image into the model, the original images underwent automatic preprocessing steps including background denoising, multi-fluorescence channel splitting, removal of invalid edge slices, and coarse screening of cell contours.

[0085] 3. Imaging results showed that the probe fluorescence signal of K562 cells highly overlapped with the immunofluorescence signal of BCR-ABL protein, while the probe signal of MCF-7 cells did not have specific co-localization.

[0086] Example 10: Construction of a 2D to 3D Deep Learning Model

[0087] 1. Paired Dataset Construction: A stack of confocal Z-axis 2D images labeled with different tumor cell lines and different targeted fluorescent probes was collected (a continuous 2D tomographic image sequence obtained by scanning layer by layer along the cell Z-axis with a fixed step size, completely covering the 3D spatial range of the cell). Five types of target regions were manually labeled in the spatial coordinate system using a 3D image annotation tool: cell nucleus, cell membrane, cytoplasm, target subcellular organelles, and signal regions of targeted fluorescent nanoprobes. The annotation results were strictly spatially registered with the original images, and finally, a paired dataset with a one-to-one correspondence between "2D image stack - 3D ground truth annotation" was constructed.

[0088] 2. Model Building and Training: The paired datasets were randomly divided into training and test sets in a 7:3 ratio to ensure a uniform distribution of different sample types across both datasets. 30% of the training set was then split off as a validation set. A 2D-to-3D deep learning model was built based on nnU-Net (an adaptive medical image segmentation network framework that automatically configures network structure and training strategies according to data characteristics). A new cross-slice 3D feature fusion module was added (linking spatial features of adjacent slices through 3D convolution and attention mechanisms to enhance Z-axis structural continuity). The model inputs continuous 2D slices, first extracting image features frame-by-frame through a 2D feature extraction branch, then completing spatial information association through the aforementioned fusion module, finally outputting voxel-level (smallest volume unit level in 3D space) 3D reconstruction and multi-class segmentation results. Training employed the AdamW optimizer and cosine annealing learning rate strategy, using a weighted combination of Dice loss and cross-entropy loss as the loss function, combined with data augmentation such as random flipping and brightness perturbation to improve model generalization ability. During training, an early stopping mechanism was implemented based on the validation set accuracy to save the optimal model parameters.

[0089] 3. Model Accuracy Validation and Optimization: The model parameters are iteratively optimized using the 3D Intersection over Union (IoU) as the core evaluation metric, which is the ratio of the intersection volume of the model's predicted 3D region to the union volume of the manually labeled ground truth region, used to quantify the spatial overlap of 3D segmentation. After training, the model is validated on a test set, comparing the output with manually labeled 3D ground truth. The average segmentation accuracy is required to be ≥90%, and the deviation between the 3D quantitative results and the human gold standard should be ≤10%. If these standards are not met, the fusion module structure, loss function weights, or training strategy are adjusted, iteratively optimizing until the accuracy requirements are met. After model optimization, accuracy validation is performed on a test set. The segmentation accuracy results for various structures are shown below. Figure 2 As can be seen, the segmentation accuracy of all categories reaches over 90%, meeting the preset accuracy standard.

[0090] Example 11: Three-dimensional volume proportion analysis of membrane mutation targets (EGFR, c-KIT) based on a two-dimensional to three-dimensional deep learning model

[0091] The confocal 2D images collected in Examples 7 and 8 are stacked and input into the trained 2D-to-3D deep learning model, and the following analysis is automatically performed: 1. Extract two-dimensional slice features layer by layer to complete the reconstruction of the three-dimensional morphology of single cells; 2. Automatic segmentation of cell membrane, cytoplasm, and nucleus regions in three-dimensional space, distinguishing specific probe fluorescence from background noise; 3. Using 90% of the three-dimensional volume as the threshold: if the fluorescent signal of the targeted nanoprobe accounts for ≥90% of the three-dimensional volume in the cell membrane region, it is judged as membrane-specific binding (positive result), and the rest is judged as non-specific binding (negative result).

[0092] 4. Based on the three-dimensional reconstruction and segmentation results, the model automatically completes multi-dimensional quantitative statistics, and the output indicators include: total fluorescence intensity, fluorescence signal volume ratio, signal spatial distribution uniformity, and relative expression level of target protein in each subcellular region.

[0093] 5. Calculate the three-dimensional volume percentage of membrane-binding probes in each cell, determine the overall percentage of positive cells in the sample, and classify the degree of abnormal protein activation in subcellular cells (high expression / medium expression / low expression) based on the percentage of individual cells. Simultaneously, count the number and percentage of high-expressing, low-expressing, and negative cells within the entire sample to complete the automatic classification of tumor cell subpopulations.

[0094] Example 12: Three-dimensional co-localization ratio analysis of intracellular mutation target (BCR-ABL) based on 2D-to-3D deep learning model

[0095] The confocal 2D image stack acquired in Example 9 is input into the 2D-to-3D deep learning model, and the following analysis is automatically performed: 1. Extract two-dimensional slice features layer by layer to complete the three-dimensional morphology reconstruction of cells and segment subcellular regions; 2. Match the probe fluorescence signal with the immunofluorescence-labeled target protein signal in three-dimensional space, and calculate the three-dimensional colocalization coefficient between the two; 3. Using 90% of the three-dimensional colocalization volume as the threshold: if the three-dimensional colocalization volume of the probe and the target protein is ≥90%, it is judged as target-specific binding (positive result), and the rest is judged as negative result.

[0096] 4. Calculate the proportion of probe-protein three-dimensional colocalization in each cell, and quantitatively analyze the expression level and mutation degree of intracellular target proteins.

[0097] Example 13: Three-dimensional quantitative analysis of standard tumor cell line probes

[0098] Using the three-dimensional analysis methods of Examples 11 and 12, batch quantification of standard tumor cell line samples from Examples 7-9 was performed, and the results are as follows: Figure 3 As shown: 1. The mean three-dimensional binding rate of membrane probes in EGFR mutant HCC827 cells was significantly higher than that in negative NCI-H520 cells; 2. The mean three-dimensional binding rate of membrane probes in c-KIT mutant GIST-T1 cells was significantly higher than that in negative GES-1 cells; 3. The mean three-dimensional colocalization rate of the probe-immunofluorescent protein in BCR-ABL positive K562 cells was significantly higher than that in negative MCF-7 cells.

[0099] Three-dimensional quantitative results showed that within the same positive cell population, the probe binding rate of individual cells differed significantly, and they could be divided into three subpopulations of high, medium and low expression, which directly reflected the single-cell heterogeneity of tumor cells.

[0100] Example 14: Quantitative analysis of subcellular proteins in clinical tumor samples based on a 2D-to-3D deep learning model

[0101] Clinical samples from patients who have not received targeted drug therapy were collected, including pleural effusion from non-small cell lung cancer and surgical tissue from gastric stromal tumors. Primary tumor cells were obtained by enzymatic digestion, cleavage red blood cell separation, and magnetic bead sorting. After inoculation and culture, corresponding targeted fluorescent nanoprobes were added and incubated for 6 hours. After fixation and staining, confocal Z-axis two-dimensional image stacks were acquired.

[0102] A 2D-to-3D deep learning model was used to complete 3D reconstruction and quantitative analysis. 1. For samples targeting membrane mutations (EGFR, c-KIT), the overall mutation status of the patient is determined by the mean three-dimensional binding rate of the membrane probe, and tumor heterogeneity and mutation degree are analyzed by the binding rate distribution of individual cells. 2. For intracellular mutation target (BCR-ABL) samples, the patient's mutation status was determined by the mean of probe-protein three-dimensional colocalization rate, and the expression level of mutant protein was quantitatively analyzed.

Claims

1. A subcellular targeted fluorescent nanoprobe for three-dimensional fluorescence imaging, characterized in that, The invention includes a nanoparticle matrix, wherein a fluorescent marker is loaded inside the nanoparticle matrix, and its surface is sequentially modified with a hydrophilic modification layer and a protein targeting recognition unit grafted onto the surface of the hydrophilic modification layer. The nanoparticles are silica, polystyrene-acrylic acid copolymer, or ZIF series metal-organic framework nanoparticles.

2. A method for preparing a subcellular targeted fluorescent nanoprobe for three-dimensional fluorescence imaging as described in claim 1, characterized in that, Includes the following steps: Step 1: Based on the matrix characteristics, fluorescent markers are incorporated into the interior or surface of nanoparticles, and the mixture is stirred continuously for 5-6 hours to prepare fluorescent functionalized nanoparticles. Step 2: Using polyethylene glycol as a hydrophilic linker, the protein targeting molecule is grafted to the end of the PEG molecular chain using an end-group active reagent. After solvent dispersion, constant temperature stirring, dialyzing and freeze drying, the targeting-PEG composite functional molecule is obtained. Step 3: Mix the targeted-PEG composite functional molecule with the fluorescent functionalized nanoparticles and react them. After stirring overnight, purify the mixture by centrifugation and repeated washing to obtain the targeted fluorescent nanoprobe.

3. The preparation method according to claim 2, wherein the matrix is ​​polystyrene-acrylic acid copolymer nanoparticles, the surface of which is modified with polysaccharides; the polysaccharides are selected from one or more of dextran, chitosan, hyaluronic acid, alginate, agarose, and pectin.

4. The preparation method according to claim 2, wherein the matrix is ​​ZIF metal-organic framework particles, the surface of which is modified with bioactive molecules; the bioactive molecules are polysaccharides; the polysaccharides are selected from one or more of dextran, chitosan, hyaluronic acid, alginate, agarose, and pectin.

5. The preparation method according to claim 2, wherein in step two, the protein targeting molecule is a small molecule targeting drug or a monoclonal antibody; The small molecule targeted drug is gefitinib, erlotinib, imatinib, or dasatinib; The monoclonal antibody is either bevacizumab or trastuzumab.

6. The preparation method according to claim 2, wherein in step two, the molecular weight of the hydrophilic linker is 3000, 5000 or 10000; and the end group of the polyethylene glycol linker is amino, carboxyl or NHS active ester.

7. A method for analyzing tumor subcellular proteins, characterized in that, Includes the following steps: S1. The target fluorescent nanoprobe as described in claim 1 is used to incubate and label the cells to be tested; S2. The labeled cells are fixed and stained, and then the Z-axis is scanned layer by layer using a confocal microscope to obtain a two-dimensional image stack. S3. Input the two-dimensional image stack into a pre-trained deep learning model. The model extracts the image features of each two-dimensional slice layer by layer, reconstructs the three-dimensional structure of a single cell, and automatically segments the cell membrane, cytoplasm and nucleus regions in three-dimensional space. S4. Based on the segmented three-dimensional structure, calculate the co-localization relationship between the targeted fluorescent nanoprobe and the target protein, or calculate the three-dimensional volume ratio of the targeted fluorescent nanoprobe in the subcellular region. S5. Determine the expression level and mutation degree of the target protein based on the statistical results.

8. The tumor subcellular protein analysis method according to claim 7, wherein in step S1, the cell to be tested is an experimental tumor cell model or a clinically derived sample; The experimental tumor cell models include any one or more of the following: EGFR exon19 deletion mutant non-small cell lung cancer HCC827 cells, EGFR T790M drug-resistant mutant non-small cell lung cancer H1975 cells, c-KIT mutant gastric stromal tumor GIST-T1 cells, BCR-ABL fusion chronic myeloid leukemia K562 cells, HER2 positive breast cancer SK-BR-3 cells, KRAS G12D mutant colorectal cancer HCT116 cells, and BRAF V600E mutant melanoma A375 cells; and any one or more of the following as normal control cells: normal lung epithelial cells BEAS-2B, normal human breast epithelial cells MCF-10A, normal human colorectal epithelial cells NCM460, human gastric mucosal cells GES-1, and human umbilical vein endothelial cells HUVEC. The clinically sourced samples include any one or more of the following: surgically removed tumor tissue, malignant pleural effusion, malignant ascites, circulating tumor cells in peripheral blood, and biopsy tissue.

9. The tumor subcellular protein analysis method as described in claim 7, wherein in step S2, the acquisition parameters of the two-dimensional image stack are: layer thickness 0.5~1μm, 20~100 two-dimensional slices are continuously acquired for each single cell sample; before inputting the two-dimensional image stack into the deep learning model, preprocessing is automatically completed, including background denoising, multi-fluorescence channel splitting, invalid edge slice removal, and coarse screening of cell contours.

10. The tumor subcellular protein analysis method as described in claim 7, wherein in step S3, the deep learning model is built based on nnU-Net, 3D-UNet, CNN convolutional neural network or a general visual large model, and is trained using a manually labeled confocal 2D image stack-3D ground truth pairing dataset; the model first extracts features frame by frame from each preprocessed 2D slice, and then completes spatial information association through a cross-slice 3D feature fusion module, outputting voxel-level single-cell 3D morphology, and simultaneously completing 3D segmentation of subcellular regions.

Images