A protein subcellular localization method based on graph convolutional neural network

Abstract

The present invention belongs to the field of biological information processing technology, in particular, a protein subcellular localization method based on a graph convolutional neural network, comprising the following steps: step 1, data collection and data preprocessing: 1-1, preparing a data set: the data set includes obtaining single-channel immunofluorescence images of the protein to be predicted under the action of four different staining substances and label data of the subcellular structure in the image; 1-2, graying the four single-channel immunofluorescence images of the protein described in step 1-1 and using them as the four channels of an image, obtaining a corresponding four-channel image of different substances staining the same area; step 2, data enhancement and training set and test set division. The present invention extracts features from protein immunofluorescence images. Although the difficulty is increased compared to single-label protein immunofluorescence images, the correlation characteristics between labels can be utilized to make the model more robust and the classification results more accurate.

Description

Technical Field

[0001] The present invention relates to the field of biological information processing technology, and in particular to a protein subcellular localization method based on graph convolutional neural network. Background Art

[0002] In the post-genomic era, with the advancement of biological science research and the sequencing of the human genome, most scholars have focused on functional genomics, encompassing both genomic and proteomic studies. Proteins are the primary functional factors that enable normal biological activities. Specific proteins must be located in specific cellular locations to function. To ensure proper protein function and understand its mechanisms of action, it is essential to precisely determine the organelle location of proteins, i.e., their subcellular localization.

[0003] Protein subcellular localization is a crucial area of research, with many years of accumulated experience. Early on, biologists relied primarily on traditional experimental techniques such as subcellular fractionation, green fluorescent protein fusion, and manual observation to annotate proteins. While these experimental results were highly reliable, they were also time-consuming and labor-intensive, making them difficult to process the massive amounts of data required. However, with the advancement of information technology, a vast amount of biological genetic information has been continuously discovered. Leyi Wei, Hang Zhou, and others have begun to leverage the characteristics of protein amino acid sequences to identify their subcellular localization. In recent years, advances in microscopic imaging have enabled researchers to combine digital image processing techniques with machine learning algorithms to pinpoint protein locations from protein images, making predictions of protein subcellular localization more accurate. Various protein labeling and staining techniques, such as chemical dyes labeled with specific protein antibodies, can reveal distinct shapes and sizes within different subcellular structures. Microscopic imaging, then, allows for accurate and efficient prediction of protein subcellular localization.

[0004] Currently, most research focuses on predicting protein subcellular locations using a single label. Under this assumption, each protein corresponds to only one subcellular location. However, in reality, nearly 20% of proteins in human cells are present in two or more subcellular locations, indicating that proteins have labels for multiple subcellular structures. Immunofluorescence cell images contain numerous subcellular structures, each with different scales and postures, and even overlapping and occluding features. In current multi-label protein subcellular classification studies, most use simple classification methods that do not consider the correlation between individual subcellular structures and between features, and prediction performance still needs to be improved.

[0005] In summary, in the existing technology, methods based on deep learning for protein subcellular localization still have problems such as inaccurate protein localization accuracy and neglect of label correlation in multi-label classification and positioning, and there is still a lack of effective solutions. Summary of the Invention

[0006] (1) Technical problems solved

[0007] In response to the shortcomings of the existing technology, the present invention provides a protein subcellular localization method based on graph convolutional neural network, which solves the problems of low accuracy of protein subcellular localization and neglect of label relevance in the existing technology.

[0008] (2) Technical solution

[0009] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0010] A protein subcellular localization method based on graph convolutional neural network, characterized by comprising the following steps:

[0011] Step 1, data collection and data preprocessing:

[0012] 1-1, preparing a data set: the data set includes obtaining single-channel immunofluorescence images of the protein to be predicted under the action of four different staining substances and label data of subcellular structures in the images;

[0013] 1-2, after graying, the four single-channel immunofluorescence images of the protein described in step 1-1 are used as the four channels of an image, thereby obtaining a corresponding four-channel image stained with different substances in the same area;

[0014] Step 2: Data enhancement and training set and test set division;

[0015] 2-1, perform mathematical statistics on the number of labels corresponding to the four-channel protein immunofluorescence image obtained in step 1-2, remove labels with too few labels, and perform data augmentation operations on labels with a small number of labels, flipping them upside down, left and right, and rotating them to obtain a supplemented data set;

[0016] 2-2, divide the data set obtained in step 2-1 into a training set and a test set;

[0017] Step 3: Build a feature extractor based on convolutional neural network;

[0018] 3-1. Perform feature extraction on the preprocessed image data and build a multi-label dataset feature extractor based on a convolutional neural network: Build InceptionResnetV2 as a feature extractor for protein immunofluorescence images, put the training set and test set into the neural network for training and testing, resize the protein immunofluorescence images to the size of the convolutional neural network input, and then extract image features. Using the neural network feature extractor with InceptionResnetV2 as the backbone, we obtain the image feature matrix for each different label;

[0019] Step 4: Use the graph convolutional neural network model to construct a node feature extractor;

[0020] The label categories of protein immunofluorescence images were processed, and the number of subcellular structures in the training set and the number of co-occurrences with other subcellular structures were counted. The adjacency matrix of the training set labels was also calculated. The word vectors of the subcellular structures were extracted based on their names in the word2vec model. The word vectors of the subcellular structures were used as node features in the graph convolutional neural network, and the adjacency matrix was used as the weight of the graph convolutional neural network model.

[0021] Step 5: The image feature matrix and the node feature matrix are input into the feature fusion module to obtain the feature matrix of the fused label. The feature matrix is then passed through the image feature differentiation module to amplify the value of the label matrix of the fused label feature matrix, thereby making the difference between the feature matrices of the labels more obvious. Finally, the amplified feature matrix is input into the classifier to obtain the label corresponding to the protein immunofluorescence image. The entire model parameters are continuously updated through the back-propagation process, and the entire model is trained. When the preset stop condition is met, the entire model is saved, and the test set image is input into the entire model to obtain the final multi-label classification result.

[0022] Step 6: Based on the performance evaluation of the graph convolutional neural network classifier, a multi-label performance evaluation indicator is selected to evaluate the performance of the multi-label protein subcellular localization classifier.

[0023] Furthermore, the protein subcellular structure includes 19 subcellular locations, namely nucleoplasm, nuclear membrane, nucleolus, nucleolar fibril center, nuclear plaque, nuclear body, endoplasmic reticulum, Golgi apparatus, intermediate filaments, actin fibers, adhesion sites, microtubules, cell dynamics bridge, microtubule organizing center, centrosome, cytoplasmic membrane, cell junction, mitochondria, and cytoplasm.

[0024] Furthermore, the image feature extractor is used to extract the image features of each multi-label protein subcellular structure; the node feature extractor is used to extract the word vector of each protein subcellular structure and the association matrix between the subcellular structures as the node features and adjacency matrix of the graph convolution respectively, and the graph convolution neural network model updates the node features according to the node features and the adjacency matrix; the feature fusion device is used to perform the dot multiplication of the back-end features of the image feature matrix and the node feature matrix to obtain a feature fusion matrix; the classifier is used to input the feature fusion matrix into the activation function for judgment. When the predicted result is greater than the set threshold, it is judged as 1, otherwise it is judged as 0, and the prediction result is finally output.

[0025] Furthermore, the protein immunofluorescence image dataset in step 1 is collected online. The dataset is acquired in a highly standardized manner using a confocal microscope. Each image sample consists of four single-channel images (stored as a single file), representing the target protein (green) and three cell markers: cell nucleus (blue), microtubules (red), and endoplasmic reticulum (yellow). The four single-channel images of the sample are then processed to obtain a complete four-channel image of the corresponding sample. The association between the labels is statistically analyzed based on the label data of the subcellular locations in the collected protein immunofluorescence images to construct a label adjacency matrix.

[0026] Furthermore, in step 6, this paper evaluates network performance from the perspective of the overall sample, namely, overall precision (OP), overall recall (OR), and overall F1 score (OF1), and from the perspective of object labels, namely, per-class precision (CP), per-class recall (CR), and per-class F1 score (CF1). The experiment also calculates the mean average precision (mAP) across all classes to help accurately represent the accuracy of the entire predicted data.

[0027] Furthermore, the feature fusion in step 5 adopts point product fusion, which is divided into front-end fusion, mid-term fusion, and back-end fusion.

[0028] (3) Beneficial effects

[0029] Compared with the existing technology, the present invention provides a protein subcellular localization method based on graph convolutional neural network, which has the following beneficial effects:

[0030] The present invention extracts features from protein immunofluorescence images. Compared with single-label protein immunofluorescence images, although the difficulty is increased, the correlation characteristics between labels can be utilized to make the model more robust and the classification results more accurate.

[0031] The model in the present invention includes five parts: an image feature extractor, a node feature extractor, a feature fusion module, an image feature differentiation module, and a classifier. This makes the present invention an end-to-end method that integrates feature extraction, feature fusion, difference amplification, and subcellular structure classification.

[0032] The model construction in the present invention can be modified directly on the existing model, the code implementation is simple, and the pre-trained model parameters can be directly used. The model converges quickly, greatly reducing the model training time.

[0033] The present invention uses deep learning methods to realize the identification of protein subcellular localization, reducing the workload of manual identification and solving the identification difficulties of non-professionals to a certain extent.

[0034] Compared to conventional multi-label classification models, this method not only considers the image features themselves but also incorporates subcellular associations. This makes the model more robust and the classification results more accurate, rather than relying solely on image features. This method can be modified directly on existing models, resulting in simple code implementation. It can directly utilize pre-trained model parameters, resulting in rapid model convergence and significantly reducing model training time. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall framework diagram of the present invention.

[0036] Figure 2 Flowchart of the present invention.

[0037] Figure 3 These are the protein immunofluorescence images of the four channels of the present invention and their processed samples. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] Example 1

[0040] like Figure 1-3 As shown, an embodiment of the present invention proposes a protein subcellular localization method based on graph convolutional neural network, comprising the following steps:

[0041] Step 1, data collection and data preprocessing:

[0042] 1-1, preparing a data set: the data set includes obtaining single-channel immunofluorescence images of the protein to be predicted under the action of four different staining substances and label data of subcellular structures in the images;

[0043] 1-2, after graying, the four single-channel immunofluorescence images of the protein described in step 1-1 are used as the four channels of an image, thereby obtaining a corresponding four-channel image stained with different substances in the same area;

[0044] Step 2: Data enhancement and training set and test set division;

[0045] 2-1, perform mathematical statistics on the number of labels corresponding to the four-channel protein immunofluorescence image obtained in step 1-2, remove labels with too few labels, and perform data augmentation operations on labels with a small number of labels, flipping them upside down, left and right, and rotating them to obtain a supplemented data set;

[0046] 2-2, divide the data set obtained in step 2-1 into a training set and a test set;

[0047] Step 3: Build a feature extractor based on convolutional neural network;

[0048] 3-1. Perform feature extraction on the preprocessed image data and build a multi-label dataset feature extractor based on a convolutional neural network: Build InceptionResnetV2 as a feature extractor for protein immunofluorescence images, put the training set and test set into the neural network for training and testing, resize the protein immunofluorescence images to the size of the convolutional neural network input, and then extract image features. Using the neural network feature extractor with InceptionResnetV2 as the backbone, we obtain the image feature matrix for each different label;

[0049] Step 4: Use the graph convolutional neural network model to construct a node feature extractor;

[0050] The label categories of protein immunofluorescence images were processed, and the number of subcellular structures in the training set and the number of co-occurrences with other subcellular structures were counted. The adjacency matrix of the training set labels was also calculated. The word vectors of the subcellular structures were extracted based on their names in the word2vec model. The word vectors of the subcellular structures were used as node features in the graph convolutional neural network, and the adjacency matrix was used as the weight of the graph convolutional neural network model.

[0051] Step 5: The image feature matrix and the node feature matrix are input into the feature fusion module to obtain the feature matrix of the fused label. The feature matrix is then passed through the image feature differentiation module to amplify the value of the label matrix of the fused label feature matrix, thereby making the difference between the feature matrices of the labels more obvious. Finally, the amplified feature matrix is input into the classifier to obtain the label corresponding to the protein immunofluorescence image. The entire model parameters are continuously updated through the back-propagation process, and the entire model is trained. When the preset stop condition is met, the entire model is saved, and the test set image is input into the entire model to obtain the final multi-label classification result.

[0052] Step 6: Based on the performance evaluation of the graph convolutional neural network classifier, a multi-label performance evaluation indicator is selected to evaluate the performance of the multi-label protein subcellular localization classifier.

[0053] Example 2

[0054] like Figure 2 As shown, a protein subcellular localization method based on graph convolutional neural network includes the following steps:

[0055] Step 1, data collection and data preprocessing:

[0056] 1-1, prepare a data set: the data set includes single-channel immunofluorescence images of the protein to be predicted under the action of four different staining substances and label data of subcellular structures in the images;

[0057] 1-2, after graying, the four single-channel immunofluorescence images of the protein described in step 1-1 are used as the four channels of an image, thereby obtaining a corresponding four-channel image stained with different substances in the same area;

[0058] Step 2: Data enhancement and training set and test set division;

[0059] 2-1, perform mathematical statistics on the corresponding labels (i.e., nucleoplasm, nuclear membrane, nucleolus, nucleolar fibril center, nuclear plaques, nuclear bodies, and other protein subcellular structures) in the four-channel protein immunofluorescence images obtained in step 1-2, remove labels with too few labels, and perform data enhancement operations on labels with a small number of labels (such as nuclear membrane, nucleolus, nuclear plaques, and other protein subcellular structures), flipping them upside down, left to right, and rotating them to obtain the supplemented data set;

[0060] 2-2, randomly divide the data set obtained in step 2-1 into a training set and a test set in a ratio of 7:3;

[0061] Step 3: Build a feature extractor based on convolutional neural network;

[0062] 3-1. Perform feature extraction on the preprocessed image data and construct a multi-label dataset feature extractor based on a convolutional neural network: Build InceptionResnetV2 as the feature extractor for protein immunofluorescence images, temporarily discard the last downsampling layer of the original InceptionResNetv2, resize the protein immunofluorescence images to the size of the convolutional neural network input, then put the training set and test set into the neural network for training and testing, and obtain a spatial tensor feature set (Tensor T), thereby realizing the image feature matrix of each different label in the immunofluorescence image dataset;

[0063] Step 4: Use the graph convolutional neural network model to construct a node feature extractor;

[0064] The labels of various subcellular structures, including the nucleoplasm, nuclear membrane, nucleolus, nucleolar fibril center, nuclear plaques, and nuclear bodies, in a protein immunofluorescence cell image dataset were processed. The number of interactions between each subcellular structure and the number of co-occurrences with other subcellular structures in the training set were counted, and the adjacency matrix of the training set labels was calculated. Word vectors were extracted from the word2vec model based on the subcellular structure names. The word vectors were used as node features in a graph convolutional neural network model, and the adjacency matrix was used as edge weights. The parameters of the graph convolutional neural network were iteratively updated based on the original node features and the adjacency matrix, allowing the graph convolutional neural network to discover more optimal subcellular structure node features and improve the model's prediction accuracy.

[0065] Step 5: After extracting image features and node features, two feature matrices of different dimensions will be obtained.

[0066] To fuse image features with node features, feature fusion uses point multiplication fusion, which is divided into three methods: front-end fusion, mid-term fusion, and back-end fusion. These two feature matrices are input into the feature fusing unit for feature fusion to obtain the label feature matrix. The label feature matrix is then input into the classifier to obtain the label category corresponding to the protein immunofluorescence cell image. To better implement the backpropagation process and make the differences between the label feature matrices more obvious, BCE loss is used instead of the common cross-entropy loss. This makes the large value of the label feature matrix larger and the small value smaller, thereby improving its differentiation. The entire model is then trained to continuously update its various parameters. When the preset stopping condition is met, the entire model is saved, and the test set images are input into the entire model to obtain the final multi-label classification results.

[0067] Step 6: Based on the performance evaluation of the graph convolutional neural network classifier, a multi-label performance evaluation index is selected to evaluate the performance of the multi-label protein subcellular localization classifier;

[0068] In multi-label classification tasks, each sample is associated with a set of labels, which makes the performance evaluation of the model more complicated. Currently, there are more than 20 evaluation indicators in the field of multi-label learning. Among them, considering general evaluation indicators and evaluation indicators suitable for protein subcellular localization, this experiment is based on the performance evaluation of graph convolutional neural network classifiers, selects multi-label performance evaluation indicators, and evaluates the performance of multi-label protein subcellular localization classifiers. That is, the network performance is evaluated from the perspective of overall samples, and the network performance is evaluated from the perspective of object labels to make a comprehensive evaluation of the classifier performance.

[0069] Considering general evaluation indicators and evaluation indicators applicable to protein subcellular localization, network performance is evaluated from the perspective of overall samples, namely overall precision (OP), overall recall (OR), and overall F1 (OF1); and network performance is evaluated from the perspective of object labels, namely per-group precision (CP), per-class recall (CR), per-class F1 score (CF1), etc. to comprehensively evaluate classifier performance.

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077] Where: C is the number of labels; is the number of images correctly predicted for the i-th label; is the number of images predicted for the i-th label; is the number of true images predicted for the i-th label; AP is defined as the average of all accuracies at K when K = 1~K.

[0078] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A protein subcellular localization method based on graph convolutional neural network, characterized by: The following steps are included: Step 1, data collection and data preprocessing: 1-1, preparing a data set: the data set includes obtaining single-channel immunofluorescence images of the protein to be predicted under the action of four different staining substances and label data of subcellular structures in the images; 1-2, after graying, the four single-channel immunofluorescence images of the protein described in step 1-1 are used as the four channels of an image, thereby obtaining a corresponding four-channel image stained with different substances in the same area; Step 2: Data enhancement and training set and test set division; 2-1, perform mathematical statistics on the number of labels corresponding to the four-channel protein immunofluorescence image obtained in step 1-2, remove labels with too few labels, and perform data augmentation operations on labels with a small number of labels, flipping them upside down, left and right, and rotating them to obtain a supplemented data set; 2-2, divide the data set obtained in step 2-1 into a training set and a test set; Step 3: Build a feature extractor based on convolutional neural network; 3-1. Perform feature extraction on the preprocessed image data and build a multi-label dataset feature extractor based on a convolutional neural network: Build InceptionResnetV2 as a feature extractor for protein immunofluorescence images, put the training set and test set into the neural network for training and testing, resize the protein immunofluorescence images to the size of the convolutional neural network input, and then extract image features. Using the neural network feature extractor with InceptionResnetV2 as the backbone, we obtain the image feature matrix for each different label; Step 4: Use the graph convolutional neural network model to construct a node feature extractor; The label categories of protein immunofluorescence images were processed, the number of subcellular structures in the training set and the number of co-occurrences with other subcellular structures were counted, and the adjacency matrix of the training set labels was calculated; In the word2vec model, word vectors of the subcellular structures are extracted based on their names. The word vectors of the subcellular structures are used as node features of the graph convolutional neural network, and the adjacency matrix is used as the weight of the graph convolutional neural network model. Step 5: The image feature matrix and the node feature matrix are input into the feature fusion module to obtain the feature matrix of the fused label. The feature matrix is then passed through the image feature differentiation module to amplify the value of the label matrix of the fused label feature matrix, thereby making the difference between the feature matrices of the labels more obvious. Finally, the amplified feature matrix is input into the classifier to obtain the label corresponding to the protein immunofluorescence image. The entire model parameters are continuously updated through the back-propagation process, and the entire model is trained. When the preset stop condition is met, the entire model is saved, and the test set image is input into the entire model to obtain the final multi-label classification result. Step 6: Based on the performance evaluation of the graph convolutional neural network classifier, a multi-label performance evaluation indicator is selected to evaluate the performance of the multi-label protein subcellular localization classifier.

2. The protein subcellular localization method based on graph convolutional neural network according to claim 1, characterized in that: The protein subcellular structure includes 19 subcellular locations, namely nucleoplasm, nuclear membrane, nucleolus, nucleolar fibril center, nuclear plaque, nuclear body, endoplasmic reticulum, Golgi apparatus, intermediate filaments, actin fibers, adhesion sites, microtubules, cell dynamics bridge, microtubule organizing center, centrosome, cytoplasmic membrane, cell junction, mitochondria, and cytoplasm.

3. The protein subcellular localization method based on graph convolutional neural network according to claim 1, characterized in that: The image feature extractor is used to extract the image features of each multi-label protein subcellular structure; the node feature extractor is used to extract the word vector of each protein subcellular structure and the association matrix between subcellular structures as the node features and adjacency matrix of the graph convolution respectively, and the graph convolution neural network model updates the node features according to the node features and the adjacency matrix; The feature fusion device is used to perform back-end feature dot multiplication of the image feature matrix and the node feature matrix to obtain a feature fusion matrix; the classifier is used to input the feature fusion matrix into the activation function for judgment. When the predicted result is greater than the set threshold, it is judged as 1, otherwise it is judged as 0, and the prediction result is finally output.

4. The protein subcellular localization method based on graph convolutional neural network according to claim 1, characterized in that: The protein immunofluorescence image dataset in step 1 is collected online. The dataset is acquired in a highly standardized manner using a confocal microscope. Each image sample consists of four single-channel images (stored as a single file), representing the target protein (green) and three cell markers: the nucleus (blue), microtubules (red), and the endoplasmic reticulum (yellow). The four single-channel images of the sample are then processed to obtain a complete four-channel image of the corresponding sample. The association between the labels is statistically analyzed based on the label data of the subcellular locations in the collected protein immunofluorescence images to construct a label adjacency matrix.

5. The protein subcellular localization method based on graph convolutional neural network according to claim 1, characterized in that: In step 6, this paper evaluates the network performance from the perspective of the overall samples, namely the overall precision (OP), the overall recall rate (OR), and the overall F1 score (OF1), and evaluates the network performance from the perspective of the object label, namely the precision of each class (CP), the recall rate of each class (CR), and the F1 score of each class (CF1); the experiment also calculates the mean average precision (mAP) of all categories, which helps to accurately represent the accuracy of the entire prediction data.

6. The protein subcellular localization method based on graph convolutional neural network according to claim 1, characterized in that: The feature fusion in step 5 adopts point multiplication fusion, which is divided into front-end fusion, mid-term fusion, and back-end fusion.

Images