A deep learning-based pathological image proteomics analysis method and system

Abstract

The application discloses a kind of based on deep learning's pathological image proteomics analysis method and system, it is related to bioinformatics, the method includes: using laser microdissection and mass spectrometry constructs paired image-protein dataset;Adopt pre-training VGG16 convolutional neural network to extract image deep feature, fusion cell morphological feature constructs whole proteome regression mapping model;Through U-Net single cell recognition and iterative deconvolution algorithm output cell level protein expression matrix;Combining mixed semi-supervised learning strategy handles label scarcity problem;Based on protein expression matrix, dimension reduction clustering, spatial heterogeneity quantification (CV+SHI) and visual analysis are carried out.The application realizes the regression mapping of pathological image to whole proteome (>2800) for the first time, breaks through the flux bottleneck of traditional virtual staining, realizes true single cell resolution proteomics, and provides new means for precision medicine.

Description

Technical Field

[0001] This invention relates to bioinformatics, and more particularly to a method and system for proteomics analysis of pathological images based on deep learning. Background Technology

[0002] Proteins are key executors of life activities, and their expression levels and spatial distribution are crucial for understanding disease mechanisms, diagnosis, and treatment. Traditional protein detection methods have the following main limitations:

[0003] (1) High cost: Mass spectrometry analysis requires expensive instruments and equipment, with a single detection costing thousands to tens of thousands of yuan; Immunohistochemistry requires the preparation of specific antibodies for each protein;

[0004] (2) Limited throughput: Immunohistochemistry can only detect 1-3 proteins at a time; multiplex immunofluorescence can detect 10-50 proteins but is technically difficult;

[0005] (3) Insufficient spatial resolution: Traditional bulk proteomics loses spatial information; laser microdissection (LCM) has a resolution of 100-500μm, which cannot reach the single-cell level (10-30μm).

[0006] (4) Difficult to apply on a large scale: Clinical pathological diagnosis mainly relies on H&E stained sections, but H&E only provides morphological information and cannot directly reflect protein expression.

[0007] Existing deep learning-based virtual staining methods are mostly limited to a few predefined proteins and cannot be extended to the whole proteome level. The main technical bottlenecks include: difficulty in integrating multimodal data, scarcity of labeled data (high cost of acquiring paired image-protein data), limited resolution (mass spectrometry is usually performed at the ROI level), and insufficient model generalization ability. Therefore, there is an urgent need for a method that can predict whole proteome expression at low cost, high throughput, and single-cell resolution. Summary of the Invention

[0008] The purpose of this invention is to solve the problems in the prior art.

[0009] The technical solution adopted by this invention to solve its technical problem is: to provide a method for proteomics analysis of pathological images based on deep learning, comprising the following steps:

[0010] Whole slices of tissue samples were obtained by WSI and regions of interest (ROIs) were cut using laser microdissection. The protein expression matrix of each ROI was identified by mass spectrometry, and normalization and missing value imputation were performed to establish a one-to-one correspondence between ROI coordinates and protein expression data, forming a paired image-protein dataset.

[0011] We used a VGG16 convolutional neural network pre-trained on ImageNet as a feature extractor, fused cell morphology features, and constructed a regression mapping model from image to protein expression. We then trained the model using an image-protein dataset with hierarchical K-fold cross-validation to obtain a protein prediction model that maps image features to the protein expression space.

[0012] The pathological image to be predicted is used to identify single cells using a cell segmentation model, and the protein prediction model is used to predict proteins and perform iterative deconvolution to output a cell-level protein expression matrix.

[0013] Proteomics analysis based on the protein expression matrix includes dimensionality reduction clustering, spatial heterogeneity quantification, and visualization.

[0014] Preferably, the VGG16 convolutional neural network includes convolutional blocks 1 to 5, with the parameters of convolutional blocks 1 to 3 frozen, and convolutional blocks 4 to 5 fine-tuned; the custom top-level network structure includes a Flatten layer, multiple fully connected layers with ReLU activation functions, a Dropout layer, and a batch normalization layer, with the number of neurons in the output layer equal to the number of target proteins N and the ReLU activation function ensuring non-negative output; the regression mapping model also includes a feature fusion module, which fuses image depth features with cellular geometry, texture, color, and spatial relationship morphological features before inputting them into the regression prediction module; the loss function uses the mean absolute percentage error (MAPE).

[0015] Preferably, the process of establishing the paired image-protein dataset includes the following steps:

[0016] ROIs are precisely cut using laser micro-cutting technology, with each ROI having an area ranging from 100×100μm² to 500×500μm².

[0017] Protein identification and quantification were performed using an ultra-high sensitivity mass spectrometer, with a database search FDR of <1%.

[0018] Protein expression data were normalized using iBAQ values, and missing values ​​were imputed using a normal distribution.

[0019] Based on ROI coordinate information, affine transformation image registration is performed on WSI, corresponding image patches are extracted and standardized to a uniform size, and a one-to-one correspondence is established with protein expression vectors.

[0020] Preferably, the cell segmentation model adopts a U-Net architecture, which includes multiple encoder layers and corresponding decoder layers, and the loss function is a combination of Dice Loss and Binary Cross-Entropy. The pathological WSI is identified globally using a sliding window scanning method. Post-processing includes binarization, morphological operations, watershed algorithm to separate adherent cells, and connected component analysis. The geometric features, texture, color, and spatial relationship features of each cell are extracted. The geometric features include centroid coordinates, bounding box, area, perimeter, and roundness.

[0021] Preferably, the iterative deconvolution includes the following steps:

[0022] Initialization: For each identified cell j within the ROI, set the initial predicted value equal to the average protein expression value of the ROI, P_ROI_measured; where the average protein expression value of the ROI, P_ROI_measured, is obtained from the actual measurement.

[0023] Forward propagation: Predict the protein expression P_j^(t) of each cell using a protein prediction model; where t represents the iteration step;

[0024] Aggregation: Calculate P_ROI^(t) = mean(P_j^(t)); where P_ROI^(t) represents the predicted average protein expression value of the ROI;

[0025] Error calculation and update: P_j^(t+1)=P_j^(t)+α×(P_ROI_measured-P_ROI^(t)), where α is the learning rate;

[0026] Constraints: Single-cell expression values ​​are non-negative, denoted as P_j^(t)≥0; and the coefficient of variation is limited by a biological prior threshold θ.

[0027] Repeat the initialization steps until the loss function converges or the maximum number of iterations is reached.

[0028] Preferably, the process of training an image-protein dataset using hierarchical K-fold cross-validation to obtain a protein prediction model that maps image features to the protein expression space includes the following four stages:

[0029] The first phase involves training an initial model using labeled data.

[0030] In the second stage, a label propagation algorithm is used to generate pseudo-labels for unlabeled data;

[0031] In the third stage, the model trained in the first stage is retrained by combining pseudo-labeled data and consistency regularization.

[0032] Fourth stage: Use a self-training strategy to select high-confidence prediction samples and iteratively optimize the model trained in the third stage.

[0033] Preferably, the dimensionality reduction clustering includes the following steps:

[0034] Dimensionality reduction can be achieved by using PCA to retain principal components that explain the pre-defined proportion of variance, or by using UMAP for nonlinear dimensionality reduction.

[0035] Clustering was performed, using the Leiden algorithm to construct a K-nearest neighbor graph and optimizing the modularity function to identify cell subpopulations;

[0036] We identified meta-programs for protein co-expression and interpreted the biological functions of each subpopulation through GO enrichment analysis and KEGG pathway analysis.

[0037] Preferably, the spatial heterogeneity quantification includes:

[0038] Calculate the spatial variation coefficient CV for each protein: CV = σ / μ, where σ is the standard deviation and μ is the mean.

[0039] Calculate the Shannon heterogeneity index SHI = -Σ_i(p_i × log2(p_i)), where p_i is the normalized protein expression percentage in region i;

[0040] Proteins with both CV and SHI values ​​exceeding a preset threshold are defined as high spatial heterogeneous proteins and considered as potential biomarker candidates.

[0041] Preferably, the visualization analysis includes:

[0042] Generate a heatmap of the spatial distribution of protein expression and overlay single-cell protein expression levels onto WSI;

[0043] Spatial distribution map of generated cells in the meta-program;

[0044] Output the prediction result matrix, with rows representing single-cell identifiers, columns representing protein names, and values ​​representing predicted expression levels, and save it as a CSV or Excel format;

[0045] Output a model performance report, including Pearson correlation coefficient, Spearman correlation coefficient, MAPE and AUC evaluation metrics for each protein.

[0046] This invention also provides a deep learning-based proteomics analysis system for pathological images, comprising:

[0047] The dataset acquisition module acquires whole slices (WSI) of tissue samples and uses laser microdissection technology to cut regions of interest (ROIs). The protein expression matrix of each ROI is obtained by mass spectrometry identification, normalization and missing value imputation are performed, and a one-to-one correspondence between ROI coordinates and protein expression data is established to form a paired image-protein dataset.

[0048] The prediction model training module uses a VGG16 convolutional neural network pre-trained on ImageNet as a feature extractor, integrates cell morphology features, and constructs a regression mapping model from image to protein expression. The model is trained using an image-protein dataset with hierarchical K-fold cross-validation to obtain a protein prediction model that maps image features to the protein expression space.

[0049] The protein training module uses a cell segmentation model to identify single cells in the pathological image to be predicted, and uses the protein prediction model to predict proteins and perform iterative deconvolution to output a cell-level protein expression matrix.

[0050] The protein analysis module performs proteomics analysis based on the protein expression matrix, including dimensionality reduction clustering, spatial heterogeneity quantification, and visualization.

[0051] The present invention has the following beneficial effects:

[0052] (1) Breakthrough in throughput bottleneck: For the first time, regression mapping from large-scale screening and other problematic pathological H&E images to the whole proteome (≥2,805 proteins) was achieved, breaking through the limitation of traditional virtual staining that only predicts a few proteins. The overall prediction AUC reached 0.718, and the AUC of 938 proteins exceeded 0.8.

[0053] (2) True single-cell resolution: The iterative deconvolution algorithm accurately decomposes regional protein expression to the single-cell level (cell diameter 15-35μm). After deconvolution, the Pearson correlation coefficient r between the aggregated value and the mass spectrometry reference value is >0.95, realizing "virtual protein staining", which is a true single-cell proteomics for tumor classification.

[0054] (3) Low cost and large-scale application: Only routine H&E staining slides are required (cost < 100 yuan / slide), no mass spectrometer or specific antibody is needed, the cost is reduced by about 100 times, and it can be directly deployed on the existing digital pathology infrastructure;

[0055] (4) Cross-cancer generalization: After fine-tuning with a small number of samples (target domain ≥ 10% of labeled data), it maintains good predictive performance in multiple cancer types such as lung cancer, breast cancer, and gastric cancer, and has universality;

[0056] (5) Biomarker discovery: By quantifying spatial heterogeneity using the CV+SHI dual-indicator approach, highly heterogeneous proteins can be systematically screened. Combined with clinical validation, this provides new technical targets for tumor subtyping and precision medicine. The invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the invention is not limited to the embodiments. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating a deep learning-based proteomics analysis method for pathological images according to an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of the AUC value distribution of the protein prediction model in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the AUC value distribution of the protein prediction model based on CRC data in TCGA, as an embodiment of the present invention.

[0060] Figure 4 This is the ROC curve of the U-Net cell segmentation model in an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram of the structure of a pathological image proteomics analysis system based on deep learning, according to an embodiment of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0063] See Figure 1 The diagram illustrates the steps of a deep learning-based proteomics analysis method for pathological images according to an embodiment of the present invention, including the following steps:

[0064] S101, obtain whole slices of tissue samples (WSI) and cut the region of interest (ROI) using laser microdissection technology. Based on the one-to-one correspondence of mass spectrometry protein expression data, a paired image-protein dataset is formed.

[0065] S102 uses a VGG16 convolutional neural network pre-trained on ImageNet as a feature extractor, integrates cell morphology features, and constructs a regression mapping model from image to protein expression. It is trained using an image-protein dataset with hierarchical K-fold cross-validation to obtain a protein prediction model that maps image features to the protein expression space.

[0066] S103, The cell segmentation model is used to identify single cells in the pathological image to be predicted, and the protein prediction model is used to predict proteins and perform iterative deconvolution to output a cell-level protein expression matrix.

[0067] S104, Perform proteomics analysis based on the protein expression matrix, including dimensionality reduction clustering, spatial heterogeneity quantification, and visualization.

[0068] Using 50 patients with stage II / III colorectal cancer (CRC) as subjects, the complete process of the method of the present invention is described in detail.

[0069] Specifically, the data acquisition and pairing dataset construction in S101 includes:

[0070] Sample processing: Fresh tumor tissue was cryopreserved at -80°C within 30 minutes postoperatively for LCM and mass spectrometry analysis. Adjacent tissue sections were FFPE-treated for H&E staining and pathological image acquisition. FFPE sections (4 μm) were H&E stained and scanned using an AperioAT2 scanner (20×, 0.5 μm / pixel) to obtain WSI. A Leica LMD7 laser microdissection system was used to cut 800 ROIs from 50 samples (an average of 16 ROIs per sample, with each ROI area approximately 200×200 μm²).

[0071] Ultra-high sensitivity mass spectrometry (timsTOF Pro) was used for analysis, and the MaxQuant database was searched (FDR < 1%). On average, 1,850 ± 320 proteins were identified per ROI. After pooling, 2,805 proteins detected in ≥80% of the samples were selected for subsequent analysis. iBAQ normalization was performed, and missing values ​​were imputed using a normal distribution (randomly selected from a normal distribution centered at 1.8 times the mean and 0.3 times the mean).

[0072] Based on LCM coordinate information, affine transformation registration is performed on WSI (using 3-5 pairs of morphological markers, with a registration error of <50μm). The corresponding image patch is extracted and resized to 128×128 pixels to establish a one-to-one correspondence with the protein expression vector.

[0073] Specifically, in S102, a VGG16 convolutional neural network (including convolutional blocks 1 to 5) pre-trained on ImageNet is used. The first 15 layers (i.e., convolutional blocks 1 to 3, extracting general low-level visual features such as edges and textures) are frozen, and convolutional blocks 4 to 5 are fine-tuned to adapt to the specific semantic features of pathological images. The custom top-level network includes a Flatten layer, multiple fully connected layers with ReLU activation functions, a Dropout layer, and a batch normalization layer. The connection order is as follows: Flatten layer → fully connected layer (256, ReLU activation function) → Dropout layer (0.5) → batch normalization layer → fully connected layer (128, ReLU) → Dropout layer (0.5) → batch normalization layer → fully connected layer (64, ReLU) → fully connected layer (2805, ReLU). The loss function used was MAPE, the Adam optimizer (initial learning rate 0.001), hierarchical 5-fold cross-validation, and an early stopping strategy (patience = 30 epochs). Performance (5-fold average): Pearson r = 0.68 ± 0.05, Spearman ρ = 0.71 ± 0.04, MAPE = 32.5 ± 4.2%, AUC = 0.72 ± 0.03 (see...) Figure 2 (As shown); Proteins with AUC>0.8: 938 (33.4%); High-performance proteins include cell cycle-related proteins (Ki-67 / PCNA / CyclinD1, AUC=0.85-0.92) and epithelial markers (E-cadherin / Cytokeratin, AUC=0.82-0.88).

[0074] Specifically, external validation was performed on the TCGA-COAD and TCGA-READ datasets. The validation results can be found in [link to relevant documentation]. Figure 3 As shown: the AUC of protein abundance predicted by H&E image is 0.702, which verifies the generalization ability of the model.

[0075] Specifically, the S103 U-Net single-cell recognition model (4 layers encoder, 4 layers decoder, 512 channels bottleneck layer) was trained on data of approximately 15,000 cells manually labeled in 10 representative WSIs (loss: Dice Loss + Binary Cross-Entropy; Adam, learning rate 0.001, batch=8, epochs=100). Validation set: Dice coefficient = 0.92, IoU = 0.86, overall AUC = 0.94, see [link / reference]. Figure 4As shown. For all WSI sliding window scans (1000×1000 pixels, stride 800 pixels), post-processing included binarization (threshold 0.5) → morphological opening operation (3×3) → morphological closing operation (5×5) → Watershed separation of adherent cells → connected component analysis (filtered area <100 or >5,000 pixels). A total of 1,245,678 cells (diameter 15-35 μm) were identified from 50 samples.

[0076] Specifically, the iterative deconvolution in S103 to single-cell resolution is set to approximately 100-200 cells per ROI, and the average protein expression P_ROI_measured in the ROI is measured by mass spectrometry. The deconvolution process is as follows:

[0077] ① Initialize P_j^(0) = P_ROI_measured;

[0078] ②Predict P_j^(t) using a deep learning model;

[0079] ③ Aggregate P_ROI^(t) = mean(P_j^(t));

[0080] ④ Error Δ = P_ROI_measured - P_ROI^(t);

[0081] ⑤ Update P_j^(t+1) = P_j^(t) + α × Δ; constrain P_j^(t) ≥ 0 and CV < θ;

[0082] Repeat steps 2-5 until convergence.

[0083] Validation (random 10 ROIs): The aggregated value after deconvolution and the mass spectrometry measurement value Pearson r = 0.95 ± 0.02, proving the effectiveness of the algorithm. The final output is a single-cell proteomics matrix of 1,245,678 cells × 2,805 proteins.

[0084] Specifically, in S103, the hybrid semi-supervised learning extracts an additional 8,000 unlabeled ROIs from WSI. The label propagation algorithm is: Y^(t+1)=α×S×Y^(t)+(1-α)×Y^(0), where α=0.8-0.9, and S=normalized cosine similarity matrix. Consistency regularization is applied, with the total loss L_total=L_supervised+λ×L_consistency (L_consistency=||f(x)-f(T(x))||²), where L_supervised represents the supervised loss. Self-training: Samples with confidence > 0.9 are iteratively added to the training set. Results: The test set AUC improved from 0.72 to 0.76 (a relative improvement of 5.6%), with particularly significant improvements for proteins with inconspicuous morphological features (average 8-12%).

[0085] Specifically, the proteomics analysis of S106 includes:

[0086] (1) Dimensionality reduction clustering: PCA (first 50 principal components, explaining 90.3% of variance) → UMAP → Leiden clustering identifies 12 co-expressed meta-programs (including: MP1 proliferating tumors / MP2 quiescent tumors / MP3 EMT cells / MP4 CD8+T / MP5 CD4+T / MP6 Treg / MP7 M1 macrophages / MP8 M2 macrophages / MP9 B cells / MP10 fibroblasts / MP11 endothelial cells / MP12 necrosis and apoptosis).

[0087] (2) Spatial heterogeneity quantification: 187 highly heterogeneous proteins were screened with CV>0.4 and SHI>1.5. Top proteins: SRSF6 (CV=0.52, SHI=1.89) and FXYD3 (CV=0.48, SHI=1.76).

[0088] (3) Biomarker validation (taking SRSF6 as an example): Kaplan-Meier analysis of 35 independent cohorts of CRC showed that the median overall survival in the high SRSF6 group was 38.2 months vs. the low SRSF6 group was 62.5 months (log-rank p=0.018); Cox HR=2.31 (95% CI: 1.15-4.63, p=0.019); CD8+ T cell density in the SRSF6 high expression region was 15±8 cells / mm², and in the low SRSF6 region it was 89±22 cells / mm² (p<0.001); functional experiments proved that SRSF6 enhances immune escape by regulating PD-L1 selective splicing.

[0089] The validation was performed using lung cancer proteomics analysis as an example. Thirty non-small cell lung cancer (NSCLC) samples (20 adenocarcinoma and 10 squamous cell carcinoma), with 12-18 ROIs per sample (480 in total), were analyzed by mass spectrometry, identifying 3,215 proteins. A random forest classifier (1000 trees) for adenocarcinoma vs. squamous cell carcinoma was trained based on predicted single-cell protein expression (3,215 dimensions): accuracy = 92%, AUC = 0.96. Important marker proteins: adenocarcinoma (TTF-1, NapsinA, CEACAM5), squamous cell carcinoma (p63, CK5 / 6, SOX2).

[0090] The generalization ability of the embodiments of the present invention across cancer types was evaluated. Source domain: colorectal cancer (50 samples); target domain: lung cancer (30 samples), breast cancer (20 samples), gastric cancer (15 samples). Zero-sample transfer: lung cancer AUC=0.62, breast cancer AUC=0.58, gastric cancer AUC=0.60 (performance decrease of 11-17%). Small-sample fine-tuning (10% of the target domain data, fine-tuning only the top layer network): lung cancer AUC=0.68, breast cancer AUC=0.65, gastric cancer AUC=0.66, demonstrating that the present invention has good cross-cancer generalization.

[0091] See Figure 5 The diagram shown is a schematic representation of a deep learning-based pathological image proteomics analysis system according to an embodiment of the present invention, comprising:

[0092] The dataset acquisition module 501 acquires whole slices of tissue samples (WSI) and uses laser microdissection technology to cut regions of interest (ROI). The protein expression matrix of each ROI is obtained by mass spectrometry identification, normalization and missing value imputation are performed, and a one-to-one correspondence between ROI coordinates and protein expression data is established to form a paired image-protein dataset.

[0093] The prediction model training module 502 uses a VGG16 convolutional neural network pre-trained on ImageNet as a feature extractor, integrates cell morphology features, and constructs a regression mapping model from image to protein expression. It is trained using an image-protein dataset with hierarchical K-fold cross-validation to obtain a protein prediction model that maps image features to the protein expression space.

[0094] The protein training module 503 uses a cell segmentation model to identify single cells in the pathological image to be predicted, and uses the protein prediction model to predict proteins and perform iterative deconvolution to output a cell-level protein expression matrix.

[0095] The protein analysis module 504 performs proteomics analysis based on the protein expression matrix, including dimensionality reduction clustering, spatial heterogeneity quantification, and visualization.

[0096] The present invention has the following industrial application prospects: (1) Clinical pathological diagnosis: assisting pathologists in quickly assessing the molecular characteristics of tumors; (2) Precision medicine: predicting patients' response to targeted / immunotherapy; (3) Drug development: assessing the impact of drugs on the tumor microenvironment; (4) Biomarker discovery: screening candidate biomarkers of highly heterogeneous proteins; (5) Digital pathology platform: serving as the core analysis module of AI-assisted diagnostic system.

[0097] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for proteomics analysis of pathological images based on deep learning, characterized in that, Includes the following steps: Whole slices of tissue samples were obtained by WSI and regions of interest (ROIs) were cut using laser microdissection. The protein expression matrix of each ROI was identified by mass spectrometry, and normalization and missing value imputation were performed to establish a one-to-one correspondence between ROI coordinates and protein expression data, forming a paired image-protein dataset. We used a VGG16 convolutional neural network pre-trained on ImageNet as a feature extractor, fused cell morphology features, and constructed a regression mapping model from image to protein expression. We then trained the model using an image-protein dataset with hierarchical K-fold cross-validation to obtain a protein prediction model that maps image features to the protein expression space. The pathological image to be predicted is used to identify single cells using a cell segmentation model, and the protein prediction model is used to predict proteins and perform iterative deconvolution to output a cell-level protein expression matrix. Proteomics analysis was performed based on the cell-level protein expression matrix, including dimensionality reduction clustering, spatial heterogeneity quantification, and visualization. The iterative deconvolution includes the following steps: Initialization: For each identified cell j within the ROI, set the initial predicted value equal to the average protein expression value of the ROI, P_ROI_measured; where the average protein expression value of the ROI, P_ROI_measured, is obtained from the actual measurement. Forward propagation: Predict the protein expression P_j^(t) of each cell using a protein prediction model; where t represents the iteration step; Aggregation: Calculate P_ROI^(t) = mean(P_j^(t)); where P_ROI^(t) represents the predicted average protein expression value of the ROI; Error calculation and update: P_j^(t+1)=P_j^(t)+α×(P_ROI_measured-P_ROI^(t)), where α is the learning rate; Constraints: Single-cell expression values ​​are non-negative, denoted as P_j^(t)≥0; and the coefficient of variation is limited by a biological prior threshold θ. Repeat the initialization steps until the loss function converges or the maximum number of iterations is reached.

2. The method for proteomics analysis of pathological images based on deep learning according to claim 1, characterized in that, The VGG16 convolutional neural network includes convolutional blocks 1 to 5, with the parameters of convolutional blocks 1 to 3 frozen, and convolutional blocks 4 to 5 fine-tuned. The custom top-level network structure includes a Flatten layer, multiple fully connected layers with ReLU activation functions, a Dropout layer, and a batch normalization layer. The number of neurons in the output layer is equal to the number of target proteins N, and the ReLU activation function is used to ensure that the output is non-negative. The regression mapping model also includes a feature fusion module, which fuses image depth features with cellular geometry, texture, color, and spatial relationship morphological features and then inputs them into the regression prediction module. The loss function used is the mean absolute percentage error (MAPE).

3. The method for proteomics analysis of pathological images based on deep learning according to claim 1, characterized in that, The process of creating the paired image-protein dataset includes the following steps: ROIs are precisely cut using laser micro-cutting technology, with each ROI having an area ranging from 100×100μm² to 500×500μm². Protein identification and quantification were performed using an ultra-high sensitivity mass spectrometer, with a database search FDR of <1%. Protein expression data were normalized using iBAQ values, and missing values ​​were imputed using a normal distribution. Based on ROI coordinate information, affine transformation image registration is performed on WSI, corresponding image patches are extracted and standardized to a uniform size, and a one-to-one correspondence is established with protein expression vectors.

4. The method for proteomics analysis of pathological images based on deep learning according to claim 1, characterized in that, The cell segmentation model adopts the U-Net architecture, which includes multiple encoder layers and corresponding decoder layers. The loss function is a combination of DiceLoss and Binary Cross-Entropy. A sliding window scanning method is used to perform global single-cell recognition on pathological WSI. Post-processing includes binarization, morphological operations, watershed algorithm to separate adherent cells, and connected component analysis. Geometric features, texture, color, and spatial relationship features of each cell are extracted. The geometric features include centroid coordinates, bounding box, area, perimeter, and roundness.

5. The method for proteomics analysis of pathological images based on deep learning according to claim 1, characterized in that, The protein prediction model, which maps image features to the protein expression space, is trained using an image-protein dataset through hierarchical K-fold cross-validation. This process includes the following four stages: The first phase involves training an initial model using labeled data. In the second stage, a label propagation algorithm is used to generate pseudo-labels for unlabeled data; In the third stage, the model trained in the first stage is retrained by combining pseudo-labeled data and consistency regularization. Fourth stage: Use a self-training strategy to select high-confidence prediction samples and iteratively optimize the model trained in the third stage.

6. The method for proteomics analysis of pathological images based on deep learning according to claim 1, characterized in that, The dimensionality reduction clustering includes the following steps: Dimensionality reduction can be achieved by using PCA to retain principal components that explain the pre-defined proportion of variance, or by using UMAP for nonlinear dimensionality reduction. Clustering was performed, using the Leiden algorithm to construct a K-nearest neighbor graph and optimizing the modularity function to identify cell subpopulations; We identified meta-programs for protein co-expression and interpreted the biological functions of each subpopulation through GO enrichment analysis and KEGG pathway analysis.

7. The method for proteomics analysis of pathological images based on deep learning according to claim 1, characterized in that, The spatial heterogeneity quantification includes: Calculate the spatial variation coefficient CV for each protein: CV = σ / μ, where σ is the standard deviation and μ is the mean. Calculate the Shannon heterogeneity index SHI = -Σ_i(p_i × log2(p_i)), where p_i is the normalized protein expression percentage in region i; Proteins with both CV and SHI values ​​exceeding a preset threshold are defined as high spatial heterogeneous proteins and considered as potential biomarker candidates.

8. The method for proteomics analysis of pathological images based on deep learning according to claim 1, characterized in that, The visualization analysis includes: Generate a heatmap of the spatial distribution of protein expression and overlay single-cell protein expression levels onto WSI; Spatial distribution map of generated cells in the meta-program; Output the prediction result matrix, with rows representing single-cell identifiers, columns representing protein names, and values ​​representing predicted expression levels, and save it as a CSV or Excel format; Output a model performance report, including Pearson correlation coefficient, Spearman correlation coefficient, MAPE and AUC evaluation metrics for each protein.

9. A deep learning-based proteomics analysis system for pathological images, characterized in that, include: The dataset acquisition module acquires whole slices (WSI) of tissue samples and uses laser microdissection technology to cut regions of interest (ROIs). The protein expression matrix of each ROI is obtained by mass spectrometry identification, normalization and missing value imputation are performed, and a one-to-one correspondence between ROI coordinates and protein expression data is established to form a paired image-protein dataset. The prediction model training module uses a VGG16 convolutional neural network pre-trained on ImageNet as a feature extractor, integrates cell morphology features, and constructs a regression mapping model from image to protein expression. The model is trained using an image-protein dataset with hierarchical K-fold cross-validation to obtain a protein prediction model that maps image features to the protein expression space. The protein training module uses a cell segmentation model to identify single cells in the pathological image to be predicted, and uses the protein prediction model to predict proteins and perform iterative deconvolution to output a cell-level protein expression matrix. The protein analysis module performs proteomics analysis based on the cell-level protein expression matrix, including dimensionality reduction clustering, spatial heterogeneity quantification, and visualization. The iterative deconvolution includes the following steps: Initialization: For each identified cell j within the ROI, set the initial predicted value equal to the average protein expression value of the ROI, P_ROI_measured; where the average protein expression value of the ROI, P_ROI_measured, is obtained from the actual measurement. Forward propagation: Predict the protein expression P_j^(t) of each cell using a protein prediction model; where t represents the iteration step; Aggregation: Calculate P_ROI^(t) = mean(P_j^(t)); where P_ROI^(t) represents the predicted average protein expression value of the ROI; Error calculation and update: P_j^(t+1)=P_j^(t)+α×(P_ROI_measured-P_ROI^(t)), where α is the learning rate; Constraints: Single-cell expression values ​​are non-negative, denoted as P_j^(t)≥0; and the coefficient of variation is limited by a biological prior threshold θ. Repeat the initialization steps until the loss function converges or the maximum number of iterations is reached.

Images