Single-cell protein subcellular localization model based on weakly supervised multiple-instance learning

Abstract

The present application relates to a single-cell protein subcellular localization model based on weakly supervised multi-instance learning, and relates to the technical field of biological information. The model comprises a cell feature extractor, an image branch, a cell branch and a class-aware adaptive pruning module. The cell feature extractor is used to encode single-cell images into single-cell feature vectors. An image-level classifier in the image branch is used to output an image-level protein subcellular localization prediction result according to the image-level feature representation. A cell-level classifier in the cell branch is used to output a single-cell-level protein subcellular localization prediction result according to the single-cell feature vector. The class-aware adaptive pruning module generates single-cell pseudo labels through several label sources. The model of the present application can solve the problems of cell label noise and class long-tail distribution under weak supervision, thereby realizing more accurate and stable single-cell localization prediction and providing a powerful tool for single-cell heterogeneity analysis.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and specifically to a single-cell protein subcellular localization model based on weakly supervised multi-instance learning. Background Technology

[0002] Subcellular localization of proteins is a key spatial determinant of their biological functions. Precisely determining the spatial distribution of proteins in organelles or subcellular structures is crucial for elucidating protein function, analyzing disease mechanisms, and discovering drug targets. Traditional techniques mainly rely on experimental methods such as subcellular fractionation and green fluorescent protein fusion, but these biochemical methods have inherent limitations such as low throughput and high cost.

[0003] With the development of high-resolution microscopy and artificial intelligence, deep learning-based subcellular localization methods have become an important paradigm in proteomics research. However, existing deep learning methods are mainly developed for multi-cell immunofluorescence images. This is because currently widely used protein databases (such as the Human Protein Atlas, HPA) only provide protein or image-level subcellular localization annotations, i.e., a holistic localization label is given by manually evaluating all cells in each image. Therefore, existing methods usually treat image-level annotations directly as single-cell annotations for training and testing, failing to resolve differences in protein localization patterns between individual cells, which are crucial for studying protein function in different cell types and their association with diseases.

[0004] In recent years, with the rise of single-cell research, the study of protein subcellular localization at the single-cell level has begun to attract attention. The core challenge lies in how to construct a reliable cell-level classification model based on multi-cell images with only image-level annotations. In 2021, the HPA team launched a single-cell protein localization recognition competition on the Kaggle platform, further highlighting the importance and challenge of this problem. Existing solutions are mostly based on multi-instance learning frameworks, using image-level labels to supervise single-cell feature learning. For example, some studies use clustering and heuristic methods to generate single-cell pseudo-labels from weak supervision; others propose hybrid two-stream networks to fuse image-level and cell-level features to reduce false positives. In addition, some studies employ strategies such as self-supervised pre-training to improve feature quality.

[0005] Despite the impressive performance of these methods, single-cell analysis still faces two key challenges: first, the inherent bias between image-level labels and actual cell localization patterns leads to cell-level label noise; second, the long-tailed distribution of subcellular localization category data severely degrades the model's ability to identify tail categories. These challenges limit the predictive accuracy and robustness of existing models. For example, the winning solution in the 2021 Human Single-Cell Protein Subcellular Localization Competition hosted by the HPA team on the Kaggle platform exhibited significant performance fluctuations on two independent test sets.

[0006] Therefore, there is an urgent need for a single-cell protein subcellular localization method that can effectively overcome label noise and class imbalance problems and achieve high accuracy and robustness. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides a single-cell protein subcellular localization model based on weakly supervised multi-instance learning. This model can solve the problems of cell label noise and long-tailed category distribution under weak supervision, thereby achieving more accurate and stable single-cell localization prediction and providing a powerful tool for single-cell heterogeneity analysis.

[0008] This invention provides a single-cell protein subcellular localization model based on weakly supervised multi-instance learning, including a cell feature extractor, image branching, cell branching, and a class-aware adaptive pruning module;

[0009] The cell feature extractor is used to encode a single-cell image into a single-cell feature vector, which is input to the image branch, the class-aware adaptive pruning module, and the cell branch.

[0010] The image branch includes: a feature aggregation unit and an image-level classifier; the feature aggregation unit is used to perform max pooling aggregation on the single-cell feature vectors to obtain an image-level feature representation, which is then input into the image-level classifier; the image-level classifier is used to output an image-level protein subcellular localization prediction result based on the image-level feature representation.

[0011] The class-aware adaptive pruning module generates single-cell pseudo-labels through several label sources, including a first label source, a second label source, and a third label source. The first label source includes the protein subcellular localization prediction results output by the image-level classifier for each single cell. The second label source includes external single-cell pseudo-labels generated based on the single-cell feature vector and external clustering and heuristic algorithms. The third label source includes image-level protein subcellular localization annotations.

[0012] The cell branch includes: a cell-level classifier; the cell-level classifier is used to output single-cell-level protein subcellular localization prediction results based on the single-cell feature vector.

[0013] The model comprises an image branch and a cell branch. The image branch employs a multi-instance learning framework to extract global features from multi-cell immunofluorescence protein images and output image-level protein subcellular localization predictions. A class-aware adaptive pruning module generates pseudo-labels for single-cell immunofluorescence protein images, while the cell branch outputs single-cell-level protein subcellular localization predictions. This model, through its dual-branch architecture combined with the class-aware adaptive pruning module, effectively utilizes and purifies cell-level signals from weak supervision, significantly reducing the impact of label noise. Trained using only weak image-level annotations, it effectively overcomes the problems of single-cell label noise and long-tailed data distribution. It achieves optimal prediction accuracy and generalization stability in public benchmark tests, providing a precise tool for analyzing single-cell protein localization heterogeneity.

[0014] The protein subcellular localization prediction results output by the image-level classifiers of each single cell mentioned above refer to the protein subcellular localization prediction results of all single cells in the current multi-cell immunofluorescence image obtained by directly inputting the single cell feature vector into the trained image branch image-level classifier.

[0015] In one embodiment, the image-level classifier includes a fully connected layer.

[0016] In one embodiment, the single-cell image is obtained by segmentation of a multi-cell immunofluorescence image, and the image-level protein subcellular localization annotation is derived from the multi-cell immunofluorescence image;

[0017] The class-aware adaptive pruning module generates single-cell pseudo-labels through several label sources, including the following steps: performing dynamic threshold filtering on the confidence scores of the first label source and the second label source, weightedly fusing the filtered second label source with the third label source, and then weightedly summing it with the filtered first label source to obtain single-cell pseudo-labels.

[0018] In one embodiment, the image branch employs a multiple instance learning framework, and the cell-level classifier includes a fully connected layer.

[0019] In one embodiment, the training loss of the single-cell protein subcellular localization model total Includes: Image branch classification loss img Loss of cell branching cell Consistency regularization loss con ;

[0020] The image branch classification loss img Loss of cell branching cell Consistency regularization loss con The weight allocation is as follows:

[0021]

[0022] in, and Let the loss weight hyperparameter be set as follows: =0.1, =0.05.

[0023] In one embodiment, the image branch classification loss img Loss of cell branching cell All use asymmetric loss functions.

[0024] This model employs asymmetric loss to address class imbalance and combines it with consistency regularization to enhance robustness. By combining asymmetric loss and consistency regularization strategies, the model effectively mitigates class imbalance and enhances feature robustness.

[0025] In one embodiment, the image branch classification loss and the cell branch classification loss are calculated using an asymmetric loss function, which is shown below:

[0026]

[0027] in, For positive sample loss, For negative sample loss, The focusing parameters are set for positive samples. The focusing parameters set for negative samples. A hard threshold set for positive samples. A hard threshold is set for negative samples. For positive sample prediction, For negative sample prediction.

[0028] The aforementioned asymmetric loss function is decoupled into positive sample loss. With negative sample loss To address the imbalance between positive and negative samples, different focusing parameters are set for the positive and negative samples respectively. and and different hard thresholds and Simultaneously, model predictions are decoupled into positive and negative sample predictions. and .

[0029] In one embodiment, the positive sample prediction The negative sample prediction As shown below:

[0030]

[0031] in, p This represents the prediction of a single-cell protein subcellular localization model for a sample, when... or If the prediction fails, then the sample prediction will fail.

[0032] The above-mentioned single-cell protein subcellular localization model uses the training loss described above. total During the optimization process, the hyperparameters of the negative samples are first fixed at their default values. =0.05 , =2 Then, through hyperparameter search, the hyperparameters for positive samples are set to... =0.01 , =0 .

[0033] In one embodiment, the consistency regularization loss con This is the sum of image branch differences, cell branch differences, and single-cell differences.

[0034] The consistency regularization loss con The calculation method includes: applying weak data augmentation and strong data augmentation to the single-cell image to generate weak data augmentation view and strong data augmentation view respectively; and processing the weak data augmentation view and strong data augmentation view through the cell feature extractor and inputting them into the image branch, the cell branch and the image-level classifier respectively.

[0035] The image branch outputs image-level protein subcellular localization prediction results for weak data and image-level protein subcellular localization prediction results for strong data, and the difference between the two is the image branch difference.

[0036] The difference between the single-cell level protein subcellular localization prediction results with weak data and the single-cell level protein subcellular localization prediction results with strong data output by the cell branch is the cell branch difference.

[0037] The difference between the image-level classifier outputting the protein subcellular localization prediction results for each single cell with weak data and the image-level classifier outputting the protein subcellular localization prediction results for each single cell with strong data is the single-cell difference.

[0038] Understandably, the above-mentioned consistency regularization loss con This is used to constrain the model's predictions for different augmented views, thereby maintaining consistency.

[0039] The present invention also provides a method for constructing the single-cell protein subcellular localization model, comprising the following steps: acquiring multi-cell immunofluorescence images with image-level protein subcellular localization annotations, performing cell segmentation to obtain a dataset of single-cell images, inputting the single-cell images into the cell feature extractor to obtain single-cell feature vectors, and inputting the single-cell feature vectors into the image branch to obtain image-level protein subcellular localization prediction results;

[0040] The single-cell feature vector is input into the image-level classifier to obtain the protein subcellular localization prediction results output by the image-level classifier for each single cell;

[0041] The protein subcellular localization prediction results output by the image-level classifier of each single cell, the external single-cell pseudo-label, and the image-level protein subcellular localization annotation are input into the class-aware adaptive pruning module to obtain the single-cell pseudo-label;

[0042] The training of the image branch is supervised by the image-level protein subcellular localization annotation, and the training of the cell branch is supervised by the single-cell pseudo-label. The optimization yields the single-cell protein subcellular localization model.

[0043] Understandably, the multicellular immunofluorescence images with image-level protein subcellular localization annotations are from publicly available databases, including at least one of the following: The Human Protein Atlas online database, training data published in the Kaggle 2021 challenge, and the OpenCell online image database.

[0044] In one embodiment, the optimization includes: employing the training loss. total The single-cell protein subcellular localization model was optimized.

[0045] In one embodiment, the optimization includes: incorporating image branch classification loss. img Loss of cell branching cell Consistency regularization loss con The aforementioned single-cell protein subcellular localization model was jointly optimized.

[0046] This invention also provides a method for predicting subcellular protein localization in a single cell, comprising the following steps: segmenting a multi-cell immunofluorescence image to be predicted into single-cell images; inputting the single-cell images into the single-cell protein subcellular localization model; the image branch outputting the image-level protein subcellular localization prediction result, while the cell branch outputting the single-cell-level protein subcellular localization prediction result; and multiplying the image-level protein subcellular localization prediction result with the single-cell-level protein subcellular localization prediction result to obtain the single-cell protein subcellular localization prediction result.

[0047] Preferably, the prediction method further includes: a cell integrity detection step, wherein after multiplying the image-level protein subcellular localization prediction result with the single-cell-level protein subcellular localization prediction result to obtain a single-cell protein subcellular localization prediction result, the integrity of the single-cell image is evaluated and weights are assigned to it, and the weights are multiplied with the single-cell protein subcellular localization prediction result to obtain an adjusted single-cell protein subcellular localization prediction result.

[0048] The aforementioned weights are used to adjust the final prediction score of the cell, thereby reducing the unreliability of predictions for edge or incomplete cells.

[0049] This invention also provides the application of the single-cell protein subcellular localization model in single-cell protein subcellular localization heterogeneity analysis and / or protein function analysis of single-cell immunofluorescence images.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] This invention presents a single-cell protein subcellular localization model based on weakly supervised multi-instance learning. This model addresses the problems of cell label noise and long-tailed category distribution under weak supervision, thereby achieving more accurate and stable single-cell localization prediction and providing a powerful tool for single-cell heterogeneity analysis. The model can be trained using only image-level annotations and exhibits excellent performance, solving the core challenge of scarce single-cell annotation data. The trained model of this invention can perform accurate subcellular localization classification of a large number of single cells, thereby achieving a quantitative description of protein localization heterogeneity in cell populations. This provides a precise and generalizable computational tool for studying functional differences of proteins at the single-cell level, cell state differentiation, and disease-related abnormal expression. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the structure of the single-cell protein subcellular localization model in the embodiment;

[0053] Figure 2 This is a visualized single-cell-level prediction result obtained by performing protein subcellular localization on immunofluorescent protein images using a single-cell protein subcellular localization model. Detailed Implementation

[0054] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0056] Unless otherwise specified, all reagents, materials, and equipment used in this embodiment are commercially available; unless otherwise specified, all test methods are conventional test methods in this field.

[0057] Example

[0058] I. Constructing the dataset.

[0059] The training set uses multi-cell immunofluorescent protein images with image-level annotations (in this embodiment, the immunofluorescent protein images are sourced from the Human Protein Atlas online database and training data published in the Kaggle 2021 challenge). The test set, containing expert-annotated single-cell labels, is used from the aforementioned challenge (in the challenge, the two test sets are named the public test set and the private test set, respectively; technicians can access the public and private test sets by clicking the publicly available competition link on the Kaggle platform). The detailed information for the training and test sets is shown in the table below.

[0060] In this embodiment, the image-level annotations described above are descriptions of the immunofluorescent protein images, not descriptions of subcellular structures such as the cell nucleus and endoplasmic reticulum. The image-level annotations refer to the location markers of the protein as a whole within one or more subcellular structures in the immunofluorescent protein image, while the single-cell-level truth labels refer to the different subcellular structures in which the protein in each single cell of the immunofluorescent protein image is located. These are divided into 19 categories (these 19 categories of subcellular structures are the subcellular structures specified in the Kaggle 2021 challenge, divided into 18 subcellular structures plus one negative category, which can be identified by technicians using the Kaggle platform). Each of the above immunofluorescent protein images contains the following four channels: green channel (target protein), blue channel (cell nucleus), red channel (microtubules), and yellow channel (endoplasmic reticulum).

[0061] Table 1. Detailed information on the training and testing sets of immunofluorescent protein images.

[0062]

[0063] II. Constructing a single-cell protein subcellular localization model based on weakly supervised multi-instance learning.

[0064] This embodiment constructs a weakly supervised multiple instance learning model, trained on single-cell image data with an image size of 224×224. This weakly supervised multiple instance learning model consists of four parts: a cell feature extractor extracts single-cell image features; an image branch provides image-level protein subcellular localization prediction; a class-aware adaptive pruning module dynamically generates single-cell pseudo-labels; and a cell branch provides single-cell-level protein subcellular localization prediction (e.g., ...). Figure 1 (As shown).

[0065] 1. Cell feature extractor extracts features from single-cell images.

[0066] Cell Segmentator (https: / / github.com / CellProfiling / HPA-Cell-Segmentation / ), a cell segmentation tool provided by the aforementioned online human protein atlas database, was used to segment each multi-cell immunofluorescence protein image, yielding an average of approximately 15 single-cell images per image. All single-cell images were resampled to 224×224 pixels. Subsequently, a deep neural network pre-trained on the ImageNet dataset (J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, “Imagenet: A large-scale hierarchical image database”, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 248-255. 2009.) was used as the cell feature extractor. Each single-cell image, after passing through the cell feature extractor, was encoded into a 2048-dimensional feature vector, which was used as input to the subsequent model.

[0067] 2. Image branch provides image-level protein subcellular localization prediction.

[0068] The image branch employs a multi-instance learning framework. In this framework, a multi-cell immunofluorescence protein image is considered a "packet," and each single cell within it is considered an "instance." All single-cell feature vectors belonging to the same image, extracted in step 1 of this embodiment, are input into this branch. First, these features are aggregated using a max-pooling operation to obtain an image-level global feature representation. Then, this global feature is fed into a fully connected layer (i.e., an image-level classifier) ​​for image-level protein subcellular localization classification, outputting predicted scores corresponding to 19 protein subcellular categories. During the training phase, the classification loss is calculated using image-level ground truth annotation labels to simultaneously fine-tune the weights of the cell feature extractor and the parameters of the fully connected layer.

[0069] 3. The class-aware adaptive pruning module dynamically generates single-cell pseudo-labels.

[0070] This module aims to generate high-quality pseudo-supervisory labels for single cells in each training cycle. Its input integrates signals from three sources: first, internal cell prediction, where the feature vector of the current single cell is input into the fully connected layer of the trained image branch in step 2, and the output serves as the initial prediction score for that cell; second, external pseudo-labels, based on the single cell feature vector, using pseudo-label scores generated by clustering and heuristic algorithms (in this embodiment, the specific generation method of external pseudo-labels is detailed in XL Zhu, LX Bao, MQ Xue, and YY Xu, “Automatic recognition of protein subcellular location patterns in single cells from immunofluorescence images based on deep learning”, Briefings in Bioinformatics, vol. 24, Jan. 2023.); and finally, image-level strong supervision, the image-level annotation label (one-hot vector) of the original multicellular immunofluorescence protein image to which the cell belongs.

[0071] The module processing flow is as follows: First, for both internal cell predictions and external pseudo-labels, the average score within the current training batch is calculated by category, and this average score is used as a dynamically updated threshold to filter out low-confidence portions, thus retaining reliable supervision signals. Then, the filtered external pseudo-labels (weights) are first processed... =0.8) and image-level labels are weighted and fused together, and then the result is combined with the filtered internal cell prediction (weight = 0.8). = The protein is weighted and summed to generate a pseudo-label for each single cell. This pseudo-label is not a discrete category, but rather a quantitative confidence value reflecting the probability of protein localization in each subcellular segment. The weights are... (1) is a function that varies with the training cycle. t A dynamically changing value.

[0072] (1)

[0073] here T 1 is set to 5. T Set 2 to 10.

[0074] 4. Cell branching provides single-cell level prediction of protein subcellular localization.

[0075] The cell branch equips each single-cell feature vector with an independent fully connected layer to map the features to cell-level protein subcellular localization predictions. The training supervision signal for this branch comes entirely from the single-cell pseudo-labels dynamically generated by the class-aware adaptive pruning module in step 3 during each training epoch. The parameters of this fully connected layer and the cell feature extractor are optimized by minimizing the difference between the cell prediction results and the pseudo-labels.

[0076] 5. Overall setup for model training.

[0077] The overall setup for model training is as follows: Image branch classification loss img With loss of cell branching cell All use asymmetric loss functions.

[0078] The total loss function is ,in con The weights of each loss combination are for the consistency regularization loss based on feature perturbation. Set it to 0.1. Set it to 0.05.

[0079] The asymmetric loss function described above is shown below:

[0080]

[0081] in, For positive sample loss, For negative sample loss, The focusing parameters are set for positive samples. The focusing parameters set for negative samples. A hard threshold set for positive samples. A hard threshold is set for negative samples. For positive sample prediction, For negative sample prediction.

[0082] The above positive sample prediction The above negative sample prediction As shown below:

[0083]

[0084] in, p This represents the prediction of a single-cell protein subcellular localization model for a sample, when... or If the prediction fails, then the sample prediction will fail.

[0085] The above consistency regularization loss conThe specific calculation method is as follows:

[0086] Weak data augmentation and strong data augmentation are applied to single-cell images to generate weak data augmentation views and strong data augmentation views, respectively. After the weak data augmentation views and strong data augmentation views are processed by the cell feature extractor, they are input into the image branch, cell branch, and image-level classifier, respectively.

[0087] The image branch outputs image-level protein subcellular localization prediction results for weak data and image-level protein subcellular localization prediction results for strong data; the difference between the two is the image branch difference.

[0088] The difference between the single-cell level protein subcellular localization prediction results with weak data and the single-cell level protein subcellular localization prediction results with strong data is the cell branch difference.

[0089] The difference between the image-level classifier outputs the protein subcellular localization prediction results for each single cell with weak data and the image-level classifier outputs the protein subcellular localization prediction results for each single cell with strong data is the single-cell difference.

[0090] The sum of the image branch differences, cell branch differences, and single-cell differences mentioned above constitutes the consistency regularization loss. con .

[0091] The above single-cell protein subcellular localization model uses training loss. total During the optimization process, the hyperparameters of the negative samples are first fixed at their default values. =0.05 , =2 Then, through hyperparameter search, the positive sample hyperparameter in the asymmetric loss is set to... =0.01 , =0 .

[0092] The Adam optimizer was used for optimization, with an initial learning rate of 0.0002. Cosine annealing was employed for learning rate scheduling, combined with a linear warm-up over one cycle. The model was trained for a total of 15 epochs with a batch size of 20. A fixed random seed was used to ensure reproducible results.

[0093] III. Evaluate model performance.

[0094] To comprehensively evaluate the performance of the model described in this embodiment, the following evaluation metrics are used in this section, which aim to measure it from the two dimensions of prediction accuracy and generalization stability.

[0095] 1. Key performance indicators for each category.

[0096] In weakly supervised single-cell protein subcellular localization tasks, since real cell-level labels cannot be obtained during the training phase, the final evaluation of the model must rely on an independent test set with expert annotations. This embodiment uses the mean accuracy (mAP), the official evaluation metric of the Kaggle 2021 Single-Cell Protein Image Classification Challenge, as the core performance indicator.

[0097] mAP is calculated based on expert-annotated ground truth labels on the test set, comprehensively reflecting the model's precision and recall performance in multi-label classification tasks, especially imbalanced multi-label classification tasks. A higher mAP value indicates better overall prediction accuracy. Specifically, mAP quantifies the performance of the detection / classification model by calculating the mean of the average precision (AP) values ​​for all object classes. AP represents the area under the precision-recall curve, reflecting the model's performance at different thresholds. mAP@0.6 indicates evaluation at an intersection-union (IU) threshold of 0.6.

[0098] For the sample size is N The first in the dataset j There are several category labels, and the following basic quantities exist:

[0099]

[0100]

[0101]

[0102]

[0103] in, j , j , j and j They represent the first j The number of true positive, false positive, true negative, and false negative samples in a class. Based on this, the following indicators can be derived:

[0104]

[0105]

[0106]

[0107] Here, C represents the number of protein subcellular categories, P represents precision, and R represents recall.

[0108] 2. Generalization stability index.

[0109] To evaluate the model's robustness and generalization ability when facing different data distributions, this embodiment calculates the cross-dataset performance fluctuation score. Specifically, it calculates the absolute value of the difference in mAP obtained by the model on two independently constructed test sets (in this embodiment, the public test set and the private test set from the Kaggle 2021 challenge are used). The lower the fluctuation score, the less the model's performance is affected by changes in data distribution, and the stronger its generalization ability.

[0110] IV. Verification Results.

[0111] 1. Visualization of model prediction results.

[0112] To evaluate whether the model described in this embodiment can focus on biologically significant subcellular localization regions, Figure 2 The visualization results of spatial attention of the model on several typical multi-cell immunofluorescent protein images are presented using gradient-weighted class activation mapping.

[0113] Observations revealed that the model described in this embodiment can significantly focus attention on subcellular structural regions related to protein expression (such as the cell nucleus and cytoplasmic membrane) in multicellular images. This focusing ability indicates that the model does not process the entire image uniformly, but rather dynamically focuses on discriminative biological structural regions based on protein expression patterns, providing a foundation for accurate image-level and cell-level classification. For images containing multi-localized protein expression, the model's attention can be rationally allocated among regions corresponding to different localization categories, thereby supporting its analysis of complex localization patterns. Overall, the spatial attention pattern exhibited by the model in multicellular immunofluorescence protein images is consistent with the expected distribution of proteins in subcellular structures, reflecting its ability to capture key biological features.

[0114] 2. Validation of the class-aware adaptive pruning module.

[0115] To verify the key role of the class-aware adaptive pruning module in improving the quality of single-cell supervision signals, the performance difference between using and not using the module during model training was compared (Table 2). When the module is not used, the supervision of cell branches relies solely on the direct inheritance of image-level labels.

[0116] As shown in Table 2, after introducing the class-aware adaptive pruning module, the model's mAP on the test set was significantly improved, and the performance fluctuation score across datasets was significantly reduced. This demonstrates that the strategy of generating pseudo-labels by dynamically fusing and filtering multi-source weak supervision signals can effectively alleviate the deviation between image-level labels and the true localization of single cells, and is an effective method to improve model accuracy.

[0117] Table 2. Validation of the class-aware adaptive pruning module.

[0118]

[0119] 3. Validation of the effectiveness of the asymmetric loss function.

[0120] To verify the crucial role of asymmetric loss functions in mitigating the interference of long-tailed data distribution and label noise, the performance differences between using asymmetric loss functions and standard cross-entropy loss functions during model training were compared (Table 3). When using standard cross-entropy loss, the model learns the same weights for head and tail categories, making it difficult to effectively distinguish hard negative samples.

[0121] As shown in Table 3, the introduction of the asymmetric loss function significantly improved the model's mAP on the test set, while also enhancing its stability across test datasets. This demonstrates that the strategy of differentially adjusting the weights of positive and negative samples and setting a hard probability threshold can effectively balance class contributions and suppress the negative impact of mislabeled samples, making it an effective way to improve the model's ability to discriminate complex samples.

[0122] Table 3 Validation of the Asymmetric Loss Function

[0123]

[0124] 4. Validation of the effectiveness of consistency regularization.

[0125] To verify the key role of the feature perturbation-based consistency regularization strategy in enhancing the model's generalization ability and robustness, the performance difference between introducing and not introducing this strategy during model training was compared (Table 4). When consistency regularization is not introduced, the model learns features only through standard data augmentation, resulting in limited adaptability to input variations.

[0126] As shown in Table 4, after introducing the consistency regularization strategy, the model's mAP on both test sets remained stable, while its performance fluctuation score across test datasets was significantly reduced. This demonstrates that by constructing and constraining the consistency of augmented views in the feature space, the robustness of feature representations can be effectively improved, making the model more stable in the face of data distribution differences and potential noise. This is an effective means to improve the model's generalization performance.

[0127] Table 4. Validation of Consistency Regularization Effectiveness

[0128]

[0129] 5. Ablation experiments of each component of the model.

[0130] To validate the class-aware adaptive pruning module, the asymmetric loss function and consistency regularization are equally important for improving model performance and stability across test datasets. The table below shows the ablation experiments of the three components. The results show that adding all three components to the model simultaneously results in a better mAP score and more stable cross-dataset performance.

[0131] Table 5 Ablation experiments of each component of the model.

[0132]

[0133] 6. Verification of reasoning strategies.

[0134] To achieve better performance, this embodiment employs a cell integrity detection strategy during the inference phase (Table 5). Compared to not using a cell integrity detection strategy, introducing this strategy to correct for incomplete cell predictions further improves the model's final performance, achieving the current best benchmark score. This verifies the effectiveness of the post-processing workflow in practical applications.

[0135] 7. Compare with other jobs.

[0136] Here, the model in this embodiment is compared with state-of-the-art methods reported on the same Kaggle 2021 benchmark dataset during the same period and previously (Table 6). The comparison includes two core metrics: mAP and performance fluctuation score across test datasets.

[0137] As shown in Table 6, the model described in this embodiment has achieved a new optimal level in mAP while keeping the performance volatility at a minimum, demonstrating significant advantages in both prediction accuracy and generalization robustness.

[0138] Table 6 compares the results with other existing single-cell level protein subcellular localization models.

[0139]

[0140] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A single-cell protein subcellular localization model based on weakly supervised multi-instance learning, characterized in that, Includes a cell feature extractor, image branching, cell branching, and a class-aware adaptive pruning module; The cell feature extractor is used to encode a single-cell image into a single-cell feature vector, which is input to the image branch, the class-aware adaptive pruning module, and the cell branch. The image branch includes: a feature aggregation unit and an image-level classifier; the feature aggregation unit is used to perform max pooling aggregation on the single-cell feature vectors to obtain an image-level feature representation, which is then input into the image-level classifier; the image-level classifier is used to output an image-level protein subcellular localization prediction result based on the image-level feature representation. The class-aware adaptive pruning module generates single-cell pseudo-labels through several label sources, including a first label source, a second label source, and a third label source. The first label source includes the protein subcellular localization prediction results output by the image-level classifier for each single cell. The second label source includes external single-cell pseudo-labels generated based on the single-cell feature vector and external clustering and heuristic algorithms. The third label source includes image-level protein subcellular localization annotations. The cell branch includes: a cell-level classifier; the cell-level classifier is used to output single-cell-level protein subcellular localization prediction results based on the single-cell feature vector; The method for constructing the single-cell protein subcellular localization model includes the following steps: acquiring multi-cell immunofluorescence images with image-level protein subcellular localization annotations, performing cell segmentation to obtain a dataset of single-cell images, inputting the single-cell images into the cell feature extractor to obtain single-cell feature vectors, and inputting the single-cell feature vectors into the image branch to obtain image-level protein subcellular localization prediction results. The single-cell feature vector is input into the image-level classifier to obtain the protein subcellular localization prediction results output by the image-level classifier for each single cell; The protein subcellular localization prediction results output by the image-level classifier of each single cell, the external single-cell pseudo-label, and the image-level protein subcellular localization annotation are input into the class-aware adaptive pruning module to obtain the single-cell pseudo-label; The training of the image branch is supervised by the image-level protein subcellular localization annotation, and the training of the cell branch is supervised by the single-cell pseudo-label. The optimization yields the single-cell protein subcellular localization model.

2. The single-cell protein subcellular localization model according to claim 1, characterized in that, The single-cell image is obtained by segmentation of a multi-cell immunofluorescence image, and the image-level protein subcellular localization annotation comes from the multi-cell immunofluorescence image. The class-aware adaptive pruning module generates single-cell pseudo-labels through several label sources, including the following steps: performing dynamic threshold filtering on the confidence scores of the first label source and the second label source, weightedly fusing the filtered second label source with the third label source, and then weightedly summing it with the filtered first label source to obtain single-cell pseudo-labels.

3. The single-cell protein subcellular localization model according to any one of claims 1-2, characterized in that, Training loss of the single-cell protein subcellular localization model total Includes: Image branch classification loss img Loss of cell branching cell Consistency regularization loss con ; The image branch classification loss img Loss of cell branching cell Consistency regularization loss con The weight allocation is as follows: in, and Let the loss weight hyperparameter be set to . =0.1, =0.

05.

4. The single-cell protein subcellular localization model according to claim 3, characterized in that, The image branch classification loss and the cell branch classification loss are calculated using an asymmetric loss function, which is shown below: in, For positive sample loss, For negative sample loss, The focusing parameters are set for positive samples. The focusing parameters set for negative samples. A hard threshold set for positive samples. A hard threshold is set for negative samples. For positive sample prediction, For negative sample prediction.

5. The single-cell protein subcellular localization model according to claim 4, characterized in that, Positive sample prediction The negative sample prediction As shown below: in, p This represents the prediction of a single-cell protein subcellular localization model for a sample, when... or If the prediction fails, then the sample prediction will fail.

6. The single-cell protein subcellular localization model according to claim 3, characterized in that, The consistency regularization loss con This is the sum of image branch differences, cell branch differences, and single-cell differences. The consistency regularization loss con The calculation method includes: applying weak data augmentation and strong data augmentation to the single-cell image to generate weak data augmentation view and strong data augmentation view respectively; and processing the weak data augmentation view and strong data augmentation view through the cell feature extractor and inputting them into the image branch, the cell branch and the image-level classifier respectively. The image branch outputs image-level protein subcellular localization prediction results for weak data and image-level protein subcellular localization prediction results for strong data, and the difference between the two is the image branch difference. The difference between the single-cell level protein subcellular localization prediction results with weak data and the single-cell level protein subcellular localization prediction results with strong data output by the cell branch is the cell branch difference. The difference between the image-level classifier outputting the protein subcellular localization prediction results for each single cell with weak data and the image-level classifier outputting the protein subcellular localization prediction results for each single cell with strong data is the single-cell difference.

7. The method for constructing a single-cell protein subcellular localization model according to any one of claims 1-6, characterized in that, The process includes the following steps: acquiring multi-cell immunofluorescence images with image-level protein subcellular localization annotations, performing cell segmentation to obtain a dataset of single-cell images, inputting the single-cell images into the cell feature extractor to obtain single-cell feature vectors, and inputting the single-cell feature vectors into the image branch to obtain image-level protein subcellular localization prediction results. The single-cell feature vector is input into the image-level classifier to obtain the protein subcellular localization prediction results output by the image-level classifier for each single cell; The protein subcellular localization prediction results output by the image-level classifier of each single cell, the external single-cell pseudo-label, and the image-level protein subcellular localization annotation are input into the class-aware adaptive pruning module to obtain the single-cell pseudo-label; The training of the image branch is supervised by the image-level protein subcellular localization annotation, and the training of the cell branch is supervised by the single-cell pseudo-label. The optimization yields the single-cell protein subcellular localization model.

8. The construction method according to claim 7, characterized in that, The optimization includes: employing the training loss. total The single-cell protein subcellular localization model was optimized.

9. A method for predicting subcellular localization of proteins in a single cell, characterized in that, The method includes the following steps: segmenting the multi-cell immunofluorescence image to be predicted into single-cell images; inputting the single-cell images into the single-cell protein subcellular localization model according to any one of claims 1-6; the image branch outputs the image-level protein subcellular localization prediction result, while the cell branch outputs the single-cell-level protein subcellular localization prediction result; and multiplying the image-level protein subcellular localization prediction result with the single-cell-level protein subcellular localization prediction result to obtain the single-cell protein subcellular localization prediction result.

10. The application of the single-cell protein subcellular localization model according to any one of claims 1-6 in the analysis of single-cell protein subcellular localization heterogeneity and / or protein function analysis of single-cell immunofluorescence images.

Images