Tumor classification method, apparatus, terminal device and storage medium

Abstract

A tumor classification method, an apparatus, a terminal device and a storage medium. The method comprises: performing feature processing on specific tumor gene data to be classified, so as to obtain a target feature vector; and, by means of a trained biological neural network model, identifying the target feature vector to obtain a classification result of said specific tumor gene data. As the biological neural network model is constructed by introducing biological knowledge, classification results of specific tumor gene data can be predicted on the basis of specific cancer tumor gene attributes combined with hierarchical information in biology, so that the degree of credibility of the classification results is higher, thus improving the accuracy of tumor classification.

Description

Tumor classification method, device, terminal equipment and storage medium

[0001] Related applications

[0002] The present application claims priority to Chinese Patent Application No. 202410450985.1, filed on April 11, 2024, the contents of which are incorporated herein by reference in their entirety. TECHNICAL FIELD

[0003] The present application relates to the field of artificial intelligence, in particular to a tumor classification method, device, terminal equipment and storage medium. BACKGROUND

[0004] For tumor classification schemes, the current main method is to obtain medical image data combined with artificial intelligence models to classify tumors. However, existing artificial intelligence models (such as deep neural network models) often use black box models, and the internal decision-making process is difficult to explain, so it is difficult to meet the requirements of model explanation and credibility in the medical field, resulting in inaccurate tumor classification results.

[0005] SUMMARY

[0006] The main purpose of the present application is to provide a tumor classification method, device, terminal equipment and storage medium, which aims to improve the accuracy of tumor classification.

[0007] To achieve the above purpose, the present application provides a tumor classification method, which comprises:

[0008] performing feature processing on specific tumor gene data to be classified to obtain a target feature vector;

[0009] performing identification on the target feature vector through a trained biological neural network model to obtain a classification result of the specific tumor gene data to be classified; wherein the biological neural network model is a neural network model constructed based on genes, pathways and biological processes.

[0010] In an embodiment, the biological neural network model comprises at least a first network layer, a second network layer and a third network layer; the first network layer comprises a plurality of gene nodes, the second network layer comprises a plurality of pathway nodes; the third network layer comprises a plurality of biological process nodes; and the connection relationship between each gene node and each channel node is determined based on the inclusion relationship between genes and channels; the connection relationship between each channel node and each biological process node is determined based on the inclusion relationship between pathways and biological processes;

[0011] The step of identifying the target feature vector by using the trained biological neural network model to obtain the classification result of the specific tumor gene data to be classified comprises:

[0012] The target feature vector is input into each gene node in the first network layer for processing to obtain a first feature vector output by each gene node;

[0013] The first feature vector of each gene node is input into a channel node connected to each gene node for processing to obtain a second feature vector output by each channel node;

[0014] The second feature vector output by each channel node is output to a biological process node connected to each channel node for processing to obtain a third feature vector output by each biological process node;

[0015] The classification result is obtained based on at least one first feature vector, at least one second feature vector, and at least one third feature vector.

[0016] In an embodiment, the step of obtaining the classification result based on at least one first feature vector, at least one second feature vector, and at least one third feature vector comprises:

[0017] The at least one first feature vector, the at least one second feature vector, and the at least one third feature vector are aggregated by averaging to obtain the classification result.

[0018] In an embodiment, the step of performing feature processing on the specific tumor gene data to be classified to obtain a target feature vector comprises:

[0019] The target feature vector is obtained by performing feature extraction on the specific tumor gene data to be classified by using a target domain feature extraction model obtained in advance.

[0020] In an embodiment, the step of obtaining the target feature vector by performing feature extraction on the specific tumor gene data to be classified by using a target domain feature extraction model obtained in advance further comprises:

[0021] The source domain feature extraction model obtained by pre-training is fine-tuned by using a target domain dataset to obtain the target domain feature extraction model; wherein the target domain dataset comprises a plurality of first gene samples of specific tumor types.

[0022] In an embodiment, the step of obtaining the target domain feature extraction model by fine-tuning the source domain feature extraction model obtained by pre-training by using a target domain dataset further comprises:

[0023] obtaining the source domain dataset, wherein the source domain dataset contains second gene samples of multiple different tumor types;

[0024] training a pre-constructed feature extraction model through the source domain dataset to obtain the source domain feature extraction model.

[0025] In an embodiment, the step of training a pre-constructed feature extraction model through the source domain dataset to obtain the source domain feature extraction model comprises:

[0026] calculating correlation coefficients between each of the second gene samples;

[0027] determining gene pairs satisfying gene co-expression relationship based on the correlation coefficients between each of the second gene samples;

[0028] based on the gene pairs, performing word embedding processing on each gene using a pre-set language universal model to obtain gene representation vectors corresponding to each of the second gene samples;

[0029] vectorizing each of the second gene samples to obtain gene expression vectors corresponding to each of the second gene samples;

[0030] training a pre-constructed feature extraction model through the gene representation vectors and the gene expression vectors corresponding to each of the second gene samples to obtain the source domain feature extraction model.

[0031] The embodiments of the present application also propose a tumor classification device, which comprises:

[0032] a feature extraction module, configured to perform feature processing on specific tumor gene data to be classified to obtain a target feature vector

[0033] a classification module, configured to identify the target feature vector through a trained biological neural network model to obtain a classification result of the specific tumor gene data to be classified, wherein the biological neural network model is a neural network model constructed based on genes, pathways and biological processes.

[0034] The embodiments of the present application also propose a terminal device, which comprises a memory, a processor and a tumor classification program stored in the memory and executable on the processor, and the tumor classification program, when executed by the processor, implements the steps of the tumor classification method as described above.

[0035] The embodiment of the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores a tumor classification program, and the tumor classification program is executed by a processor to implement the steps of the tumor classification method.

[0036] The tumor classification method, device, terminal equipment and storage medium provided by the embodiment of the present application obtain a target feature vector by performing feature processing on specific tumor gene data to be classified, and obtain a classification result of the specific tumor gene data to be classified by identifying the target feature vector through a trained biological neural network model. Since the biological neural network model of the embodiment of the present application is constructed by introducing biological knowledge, the classification result of the specific tumor gene data can be predicted on the basis of specific cancer tumor gene attributes and combined with hierarchical information in biology, so that the credibility of the classification result is higher, and the accuracy of tumor classification is improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] FIG. 1 is a functional module schematic diagram of a terminal equipment to which the tumor classification device of the present application belongs;

[0038] FIG. 2 is a flow schematic diagram of a first embodiment of the tumor classification method of the present application;

[0039] FIG. 3 is a network structure schematic diagram of an embodiment of the biological neural network model involved in the present application;

[0040] FIG. 4 is a flow schematic diagram of a second embodiment of the tumor classification method of the present application;

[0041] FIG. 5 is a flow schematic diagram of a third embodiment of the tumor classification method of the present application;

[0042] FIG. 6 is a flow schematic diagram of a fourth embodiment of the tumor classification method of the present application;

[0043] FIG. 7 is a flow schematic diagram of a fifth embodiment of the tumor classification method of the present application.

[0044] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0045] It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0046] The main solution of the embodiment of the present application is: obtaining a target feature vector by performing feature processing on the specific tumor gene data to be classified; obtaining the classification result of the specific tumor gene data to be classified by recognizing the target feature vector through the trained biological neural network model. Since the biological neural network model of the embodiment of the present application introduces biological knowledge construction, it can predict the classification result of the specific tumor gene data based on the specific cancer tumor gene attribute combined with the hierarchical information in biology, thereby making the credibility of the classification result higher, and further improving the accuracy of tumor classification.

[0047] Technical terms involved in the present application:

[0048] Transfer learning: Transfer learning is a machine learning method that uses a model developed for task A as an initial point and reuses it in the process of developing a model for task B.

[0049] Source domain: refers to the data set or field that the model is pre-trained on. The source domain can be a task, a field or a data set, depending on the specific application scenario.

[0050] Target domain: refers to the new data set or field to which the model will be applied. The target domain can be a different task, a different field or a different data set. For example, when applying an image classification model to a medical image data set, the medical image data set is the target domain.

[0051] Genes: Genes are sequences that can encode RNA or proteins.

[0052] Pathways: Pathways refer to complex networks formed by the interaction of multiple biological molecules. In the present application, if a gene is involved in the construction of this pathway, it is said that the gene is included in the pathway.

[0053] Biology Processes (BP): Biology processes refer to a series of events resulting from the ordered combination of one or more molecular functions. Larger processes or "biological programs" completed through multiple molecular activities.

[0054] The embodiment of the present application considers that the related technical solutions perform tumor classification based on an artificial intelligence model, and the artificial intelligence model often adopts a black box model, whose internal decision-making process is difficult to explain, so it is difficult to meet the requirements of the medical field for the explainability and credibility of the model, resulting in inaccurate tumor classification results.

[0055] Based on this, the embodiment of the present application proposes a solution, by introducing biological knowledge into the construction of a biological neural network model, so as to predict the classification result of specific tumor gene data on the basis of specific tumor gene attributes in combination with hierarchical information in biology, thereby making the credibility of the classification result higher, and further improving the accuracy of tumor classification.

[0056] Specifically, referring to FIG. 1, FIG. 1 is a functional module schematic diagram of a terminal device to which the tumor classification device belongs. The tumor classification device can be a device independent of the terminal device and capable of data processing, which can be carried on the terminal device in the form of hardware or software. The terminal device can be a smart mobile terminal such as a mobile phone or a tablet computer with a touch screen, and can also be a fixed terminal with a touch screen, and can also be a server.

[0057] In the embodiment, the terminal device to which the tumor classification device belongs at least includes an output module 110, a processor 120, a memory 130, and a communication module 140.

[0058] The memory 130 stores an operating system and a tumor classification program, and the tumor classification device can store the obtained target feature vector, the classification result of the specific tumor gene data to be classified obtained by recognition, and other information in the memory 130; the output module 110 can be a display screen, a loudspeaker, etc. The communication module 140 can include a WIFI module, a mobile communication module, and a Bluetooth module, etc., and communicates with external devices or servers through the communication module 140.

[0059] When the tumor classification program in the memory 130 is executed by the processor, the following steps are implemented:

[0060] The target feature vector is input into each gene node in the first network layer for processing to obtain a first feature vector output by each gene node;

[0061] The first feature vector of each gene node is input into a channel node connected to each gene node for processing to obtain a second feature vector output by each channel node;

[0062] The second feature vector output by each channel node is output to a biological process node connected to each channel node for processing to obtain a third feature vector output by each biological process node;

[0063] The classification result is obtained based on at least one first feature vector, at least one second feature vector, and at least one third feature vector.

[0064] Further, the tumor classification program in the memory 130, when executed by the processor, further implements the following steps:

[0065] Based on at least one of the first feature vectors, at least one of the second feature vectors, and at least one of the third feature vectors, the classification result is obtained.

[0066] Further, the tumor classification program in the memory 130, when executed by the processor, further implements the following steps:

[0067] The target feature vector is obtained by performing feature extraction on the specific tumor gene data to be classified by the pre-acquired target domain feature extraction model.

[0068] Further, the tumor classification program in the memory 130, when executed by the processor, further implements the following steps:

[0069] The source domain feature extraction model is obtained by fine-tuning the pre-trained source domain feature extraction model by using a target domain data set; wherein the target domain data set includes a plurality of first gene samples of specific tumor types.

[0070] Further, the tumor classification program in the memory 130, when executed by the processor, further implements the following steps:

[0071] The source domain data set is obtained, wherein the source domain data set contains a plurality of second gene samples of different tumor types;

[0072] The source domain feature extraction model is obtained by training the pre-constructed feature extraction model by using the source domain data set.

[0073] Further, the tumor classification program in the memory 130, when executed by the processor, further implements the following steps:

[0074] The correlation coefficients between each of the second gene samples are calculated;

[0075] Based on the correlation coefficients between each of the second gene samples, a gene pair satisfying a gene co-expression relationship is determined;

[0076] Based on the gene pair, each gene is subjected to word embedding processing by using a pre-set language universal model, to obtain a gene representation vector corresponding to each of the second gene samples;

[0077] Each of the second gene samples is subjected to vectorization processing to obtain a gene expression vector corresponding to each of the second gene samples;

[0078] The source domain feature extraction model is obtained by training the pre-constructed feature extraction model by using the gene representation vector and the gene expression vector corresponding to each of the second gene samples.

[0079] The embodiment obtains a target feature vector by performing feature processing on the specific tumor gene data to be classified. The biological neural network model trained is used to identify the target feature vector, so as to obtain the classification result of the specific tumor gene data to be classified. Since the biological neural network model of the embodiment is constructed by introducing biological knowledge, the classification result of the specific tumor gene data can be predicted based on the specific cancer tumor gene attributes and combined with the hierarchical information in biology, so that the credibility of the classification result is higher, and the accuracy of tumor classification is improved.

[0080] The method embodiment of the present application is based on, but not limited to, the terminal device architecture described above.

[0081] Referring to FIG. 2, FIG. 2 is a flowchart of a first embodiment of the tumor classification method of the present application. The tumor classification method of the embodiment includes the following steps.

[0082] In step S101, feature processing is performed on the specific tumor gene data to be classified, to obtain a target feature vector.

[0083] The execution subject of the method of the embodiment can be a tumor classification device or a terminal device. The terminal device is taken as an example in the embodiment.

[0084] To realize the classification of tumors, the embodiment extracts features from the specific tumor gene data to be classified after obtaining the specific tumor gene data, to obtain a target feature vector. Thus, the specific tumor classification task can be better performed according to the feature information (such as semantic feature information) of the specific tumor gene data.

[0085] The existing tumor classification is mainly performed by acquiring medical image data and combining an artificial intelligence model. However, the existing artificial intelligence model (such as a deep neural network model) often adopts a black box model, and the internal decision-making process is difficult to explain. Therefore, it is difficult to meet the requirements of explanation and credibility of the model in the medical field, resulting in inaccurate tumor classification results. To better explain the biological significance of the model in tumor classification prediction, the embodiment regards genes, pathways, and biological processes as nodes (i.e., neurons) in a biological neural network model. The connections between the nodes are established according to the containing and contained relationship between them, and finally the classification result of the specific tumor gene data is output. Thus, the state of different biological components can be better understood by giving constraints between nodes with biological significance, and the parameter quantity is less, and the computational complexity is reduced.

[0086] In the embodiment, the specific tumor gene data to be classified is a gene data belonging to a specific tumor type, and the gene data contains relevant information of the gene, such as RNA coding information. In the field of tumor classification, tumors include many types, such as lung cancer, pancreatic cancer, prostate cancer, breast cancer, colorectal cancer, and the like. The specific type is a type further divided from all tumor types, such as the specific type including breast cancer and colorectal cancer.

[0087] In step S101 of the embodiment, the specific tumor gene data to be classified is processed to obtain a target feature vector. The implementation scheme of the specific tumor gene data to be classified can include the following steps:

[0088] As an implementation manner, the terminal device can use a deep learning network model to process the specific tumor gene data to obtain a target feature vector. For example, a fast attention model (Performer) can be used for feature extraction.

[0089] As another implementation manner, the terminal device can use a bag-of-words model and a word embedding model (such as Word2Vec) to extract features to obtain a target feature vector.

[0090] In step S102, the target feature vector is identified by the trained biological neural network model to obtain a classification result of the specific tumor gene data to be classified. The biological neural network model is a neural network model constructed based on genes, pathways, and biological processes.

[0091] In order to enable the neural network model to predict the classification result (i.e., whether to be sick or not) of the tumor based on the specific tumor gene attribute and the hierarchical information in bioinformatics, the biological neural network model is designed in the embodiment. By using the inclusion relationship based on nodes and channels and the inclusion relationship between channels and biological processes, the node information in the biological neural network model is endowed with the hierarchical information in biology, and the connection between layers of the biological neural network model in the embodiment is not full connection but sparse connection less than full connection. Compared with the full connection network, the number of parameters is smaller, so that the training speed of the biological neural network model can be accelerated.

[0092] Specifically, the biological neural network model can include at least a first network layer, a second network layer, and a third network layer. The first network layer includes a plurality of gene nodes, the second network layer includes a plurality of pathway nodes, and the third network layer includes a plurality of biological process nodes. The connection relationship between each gene node and each channel node is determined based on the inclusion relationship between genes and channels, and the connection relationship between each channel node and each biological process node is determined based on the inclusion relationship between pathways and biological processes.

[0093] The number of gene nodes contained in the first network layer can be determined from the gene expression data of interest obtained from the National Center for Biotechnology Information (NCBI). The gene expression data is usually represented in the form of a matrix, referred to as a gene expression matrix. The rows of the gene expression matrix represent the expression of a gene under different environmental conditions or at different time points, and the columns represent the expression of all genes under different conditions or samples (such as tissues, experimental conditions, treatment factors, etc.). The data in each cell represents the expression level of a specific gene in a specific sample. The gene expression matrix describes the relevant information of multiple genes, from which the number of gene nodes can be determined, i.e., the number of gene nodes is set equal to the number of genes contained in the gene expression matrix, and each gene node can represent a different gene.

[0094] The inclusion relationship between genes and channels means that if a gene participates in a channel, it can be determined that the gene is included in the channel; if the gene does not participate in the channel, it can be determined that the gene is not included in the channel. The inclusion relationship between channels and biological processes means that if a channel belongs to a biological process, it can be determined that the channel is included in the biological process; if the channel does not belong to the biological process, it can be determined that the channel is not included in the biological process. The inclusion relationship between genes and channels, and the inclusion relationship between channels and biological processes are known biological knowledge, which can be obtained by a terminal device through a related platform in specific implementation, for example, from the KEGG platform or the NCBI platform.

[0095] It can be understood that the neural network can generally include an input layer, a hidden layer and an output layer. In the construction of the biological neural network model, the number of layers of the hidden layer and the number of nodes in the hidden layer are important. The existing scheme is generally based on artificial experience judgment or borrowing classical models in other fields for design. This way fails to apply knowledge in a specific field to the construction of the neural network, resulting in poor learning effect of the neural network. In order to enable the biological neural network model to predict the classification result of the specific tumor gene data on the basis of the specific cancer tumor gene attribute combined with the hierarchical information in biology, in the embodiment scheme, the first network layer, the second network layer and the third network layer integrated with biological knowledge can be used as the hidden layer of the biological neural network model. Specifically, the number of nodes of the input layer can be designed according to the dimension of the target feature vector. For example, if the target feature vector is an n-dimensional vector [x1, x2,..., xn], the number of nodes of the input layer includes n, and the input variable and the output variable of the first node of the input layer are x1. The first network layer, the second network layer and the third network layer are used as the hidden layer, each node in the input layer and each gene node in the first network layer are connected one by one, then the connection relationship between each gene node in the first network layer and each channel node in the second network layer is designed according to the inclusion relationship between the gene and the channel, and the connection relationship between each channel node in the second network layer and each biological process node in the third network layer is designed according to the inclusion relationship between the channel and the biological process, to obtain the network structure diagram of the biological neural network model as shown in FIG. 3. Further, the third network layer can be directly connected with the output layer, and the number of nodes of the output layer can be set according to actual needs. For example, in the embodiment scheme, the tumor classification is a binary classification result, and the number of nodes of the output layer can be set to two.

[0096] Of course, in other embodiments, more hidden layers can be set between the third network layer and the output layer according to actual needs. Specifically, the fourth network layer can be designed as another hidden layer based on biological knowledge construction. For example, if other types of biological attributes are included, and the biological process has an inclusion relationship with other types of biological attributes, a fourth network layer can be designed as another hidden layer in the next level of the third network layer.

[0097] Specifically, the biological neural network model can be trained by obtaining a specific tumor gene sample data set, and performing feature processing on each specific tumor gene sample to obtain a sample feature vector of each sample. Then, each sample feature vector is divided into a training set and a validation set, and then the biological neural network model is trained through the training set, and the training result is evaluated through the validation set. Specifically, the model evaluation indicators AUC (area under the curve), AUC-PR (Area Under the Precision versus Recall Curve) and F1 score can be used to measure the performance of the model classification prediction. Perform 5-fold cross-validation on the validation set and the training set together, and take the average of the predicted values to alleviate data bias.

[0098] Further, in order to evaluate the relative importance of the specific genes that contribute to the model prediction, the total importance score of the genes is obtained using the DeepLIFT attribution method in the DeepExplain library, and the higher the score, the more relevant the specific tumor disease is. DeepLIFT is a backpropagation-based attribution method for assigning sample-level importance scores to each feature. Therefore, the importance score of each node in each layer can be obtained.

[0099] When the biological neural network model training reaches convergence, the biological neural network model reaching convergence is obtained, that is, the trained biological neural network model is obtained, and the target feature vector is identified through the trained biological neural network model to obtain the classification result of the specific tumor gene data to be classified.

[0100] The above scheme is used to obtain the target feature vector by performing feature processing on the specific tumor gene data to be classified. The trained biological neural network model is used to identify the target feature vector to obtain the classification result of the specific tumor gene data to be classified. Since the biological neural network model of the embodiment introduces biological knowledge construction, it can predict the classification result of the specific tumor gene data based on the specific cancer tumor gene attributes and combine the hierarchical information in biology, thereby improving the accuracy of tumor classification.

[0101] Based on the first embodiment, the second embodiment of the present application is proposed. Referring to FIG. 4, FIG. 4 is a flowchart of the third embodiment of the tumor classification method of the present application. Compared with the first embodiment, step S102 in the embodiment includes:

[0102] Step S201, input the target feature vector into each gene node in the first network layer for processing to obtain the first feature vector output by each gene node.

[0103] Specifically, the first network layer is the first layer hidden layer of the biological neural network model, and each gene node in the first network layer is regarded as a neural unit. After the target feature vector is input into each gene node for processing, the target feature vector is processed by each gene node through the following formula to obtain the first feature vector output by each gene node: y = f[(M * W) T x + b] (1),

[0104] Wherein, the parameter f is an activation function, M is a mask matrix, which is set according to actual conditions. W is a weight matrix, x is an input matrix, i.e. the target feature vector, and b is a bias vector. * represents Hadamard product. The activation function f can adopt the activation sigmoid function, and higher loss weights are used for the results of the later layers. The sigmoid function f = tanh = (e 2x -1) / (e 2x +1).

[0105] Step S202, output the first feature vector of each gene node to the channel node to which each gene node is respectively connected for processing to obtain the second feature vector output by each channel node.

[0106] Specifically, after obtaining the first feature vector of each gene node, the first feature vector output by each gene node is taken as the input value of the corresponding node of the next layer, i.e. the first feature vector of each gene node is input into the channel node in the second network layer to which each gene node is respectively connected. Each channel node also corresponds to a neural unit, and each first feature vector can be processed through the above formula (1) to obtain the second feature vector output by each channel node.

[0107] Step S203, output the second feature vector output by each channel node to the biological process node to which each channel node is respectively connected for processing to obtain the third feature vector output by each biological process node.

[0108] Specifically, after obtaining the second feature vectors output by the respective channel nodes, the second feature vectors output by the respective channel nodes are taken as input values of corresponding nodes of the next layer, i.e., the second feature vectors of the respective channel nodes are input to biological process nodes in the second network layer to which the respective channel nodes are respectively connected. Each biological process node also corresponds to a neural unit, and each second feature vector can be processed by the above formula (1) to obtain a third feature vector output by each biological process node.

[0109] In step S204, a classification result is obtained based on the at least one first feature vector, the at least one second feature vector, and the at least one third feature vector.

[0110] In the embodiment, the terminal device can aggregate the at least one first feature vector, the at least one second feature vector, and the at least one third feature vector by averaging to obtain the classification result. The embodiment can fuse the feature information extracted by the first network layer, the second network layer, and the third network layer with biological knowledge to aggregate, thereby better explaining the tumor classification result and improving the reliability of tumor classification.

[0111] In the embodiment, the terminal device can aggregate the at least one first feature vector, the at least one second feature vector, and the at least one third feature vector by averaging to obtain the classification result. The embodiment can fuse the feature information extracted by the first network layer, the second network layer, and the third network layer with biological knowledge to aggregate, thereby better explaining the tumor classification result and improving the reliability of tumor classification.

[0112] The third embodiment of the present application is based on the first embodiment. Please refer to FIG. 5, which is a flowchart of the third embodiment of the present application. Compared with the first embodiment, in the third embodiment, the step S102 of performing feature processing on the specific tumor gene data to be classified to obtain a target feature vector includes:

[0113] In step S301, a target feature vector is obtained by performing feature extraction on the specific tumor gene data to be classified by using a target domain feature extraction model obtained in advance.

[0114] The target domain feature extraction model can be a deep neural network-based model, for example, a fast attention mechanism model.

[0115] In the embodiment, after obtaining the target domain feature extraction model, the target domain feature extraction model is used to extract features of the specific gene data to be classified to obtain a target feature vector containing multiple dimension information.

[0116] In the embodiment, the target domain feature extraction model is used to extract features of the specific tumor gene data to be classified to obtain a target feature vector. The target domain feature extraction model can quickly identify the feature information of the specific tumor gene data.

[0117] The fourth embodiment of the present application is based on the third embodiment. Please refer to FIG. 6, which is a flowchart of the fourth embodiment of the present application. Compared with the third embodiment, the fourth embodiment further includes the following steps before step S301 of obtaining a target feature vector by using the pre-obtained target domain feature extraction model to extract features of the specific tumor gene data to be classified.

[0118] In step S401, the source domain feature extraction model is fine-tuned by using the target domain data set to obtain a target domain feature extraction model.

[0119] The target domain data set can include multiple first gene samples of specific tumor types.

[0120] The present embodiment considers that the existing computing power is insufficient when dealing with large-scale high-dimensional gene expression data, and it is difficult to mine important genes or features while discovering the complex nonlinear relationships between them. In some cases, the prediction effect is not good due to the lack of a large amount of labeled data. Therefore, the present embodiment uses the idea of cross-domain transfer learning, first uses large-scale gene expression data to obtain general knowledge about tumor diseases, then introduces a model migration method to fine-tune the parameters of the source domain feature extraction model to obtain a target domain feature extraction model, and applies the target domain feature extraction model to the feature processing of specific tumor gene data, so that the extracted target feature vector can better represent the feature information of the specific tumor gene data, and then more accurately predict the classification result of the specific tumor gene data.

[0121] In specific implementation, the terminal device can download the gene expression data of the specified number from the public database of the National Center for Biotechnology Information (NCBI) platform. The gene expression data is usually represented in the form of a matrix, which is called a gene expression matrix. The rows of the gene expression matrix represent the expression of a gene under different environmental conditions or at different time points, and the columns represent the expression of all genes under different conditions or samples (such as tissues, experimental conditions, treatment factors, etc.). The data in each cell represents the expression level of a specific gene in a specific sample. After obtaining the gene expression data of the specified number, the terminal device can use the affy package and the inSilicoMerging package to arrange and merge the data set, and then parse the gene-related information contained in the gene expression data of the specified number, that is, obtain the gene sample data set of interest. The obtained gene sample data set can include gene samples of multiple tumor types, such as lung cancer, pancreatic cancer, prostate cancer, etc. After obtaining the gene sample data set, the terminal device can extract a plurality of first gene samples of specific tumor types and a second gene sample containing different tumor types from the gene sample data set, and use the extracted plurality of first gene sample data as the target domain data set, and use the extracted plurality of second gene sample data as the source domain target domain data set. Then, the terminal device can pre-train the feature extraction model through the source domain target domain data set to obtain the source domain feature extraction model to learn the general implicit knowledge of tumor diseases, and then the terminal device can adjust the parameters in the source domain feature extraction model obtained in the pre-training stage using the target domain data set to adjust the distribution difference (i.e., gene difference) between the source domain and the target domain, so that the learned general implicit knowledge of tumor diseases is more suitable for the target domain, and a target domain feature extraction model suitable for specific tumor gene data is obtained.

[0122] Further, considering that the number of first gene samples and the number of second gene samples are different, the embodiment of the present application can modify the last layer of the feature extraction model by replacing the last layer with a one-layer convolutional network to extract abstract information from the gene features represented by each node. At the same time, considering that the gene feature information expressed by the parameters in the source domain feature extraction model and the gene feature information in the target domain data set are not necessarily completely consistent, appropriate pruning is needed. Then, the modified source domain feature extraction model is trained using a smaller learning rate and fewer iterations.

[0123] Further, in the embodiment scheme, before the source domain feature extraction model pre-trained through the target domain data set is fine-tuned to obtain the target domain feature extraction model, the following steps are further included:

[0124] In step S4011, a source domain data set is obtained.

[0125] wherein the source domain dataset can include a plurality of second gene samples of different tumor types.

[0126] At step S4012, the pre-constructed feature extraction model is trained by using the source domain dataset to obtain a source domain feature extraction model.

[0127] Considering that the second gene samples can include many samples lacking label information, manual marking of data is both time-consuming and expensive. In this embodiment, some attributes in the second gene sample data are used to set a pseudo-supervised task to replace human annotation, i.e., the source domain feature extraction model adopts a self-supervised training manner. For example, a mask model can be used in the source domain feature extraction model to randomly mask the positions of non-zero values in the first gene sample data, and then the real data of the masked gene is predicted by using other gene data, so that general deep representations of gene expression patterns can be learned on a large amount of unlabeled data, which helps to reduce the workload of the downstream fine-tuning process.

[0128] In this embodiment, the feature extraction model can adopt an attention model Transformer or a fast attention calculation model Performer based on orthogonal random features, or other machine learning models. This application does not limit this.

[0129] After the terminal device obtains the source domain feature extraction model, the target domain dataset can be used to adjust the parameters in the source domain feature extraction model in a supervised training manner, and after the parameters are adjusted to meet the preset convergence adjustment, the target domain feature extraction model is output. The output target domain extraction model can be used for feature extraction of specific tumor gene data.

[0130] Through the above scheme, the source domain feature extraction model pre-trained is fine-tuned by using the target domain dataset to obtain the target domain feature extraction model. By using the method of cross-domain transfer learning based on tumor general knowledge, important genes or features can be mined while discovering the complex nonlinear relationships between them, so as to improve the accuracy of feature extraction of specific tumor gene data.

[0131] The fifth embodiment of the present application is based on the fourth embodiment. Please refer to FIG. 7, which is a flowchart provided by the fifth embodiment of the present application. Compared with the fourth embodiment, at step S4012, the pre-constructed feature extraction model is trained by using the source domain dataset to obtain a source domain feature extraction model, which can include:

[0132] At step S501, the correlation coefficients between each second gene sample are calculated.

[0133] The terminal device can obtain a gene expression matrix from the NCBI platform, the gene expression matrix including information of a plurality of genes, and then can obtain a correlation coefficient between any two genes in the gene expression matrix by using a co-expression network construction method. The co-expression network can define an individual relationship between any two genes based on the correlation or mutual information between each pair of genes. Thus, the similarity between the expression patterns of gene pairs in all samples can be described by these relationships.

[0134] Further, the correlation coefficient between two genes can be calculated by Pearson or Spearman.

[0135] In step S502, a gene pair satisfying a gene co-expression relationship is determined based on the correlation coefficient between each second gene sample.

[0136] Specifically, the terminal device can take two second gene samples with a correlation coefficient greater than a preset correlation coefficient threshold as a gene pair satisfying a gene co-expression relationship. For example, if the correlation coefficient of gene A and gene B is greater than the correlation coefficient threshold, then gene A and gene B are a gene pair. The two genes in each gene pair are semantically similar.

[0137] After obtaining the gene pair, the relationship of each gene pair can be determined and the gene pair can be saved to a text file.

[0138] In step S503, based on the gene pair, a preset language universal model is used to perform word embedding processing on each gene to obtain a gene representation vector corresponding to each second gene sample.

[0139] In order to enable each obtained gene sample to include more semantic information, the embodiment scheme uses a language universal model to perform word embedding processing on each gene based on gene pairs with similar semantic information, i.e., converts each second gene sample data into a word vector to obtain a gene representation vector corresponding to each second gene sample.

[0140] Specifically, the language universal model in the embodiment scheme can use a vector space model (VSM). The core idea of the vector space model is that words appearing in a context have similar semantics. Research methods using this assumption are roughly divided into two categories: count-based methods (such as Glove), and prediction methods (such as the neural probabilistic language model Word2Vec). Taking the use of Word2Vec as an example, this method can directly predict the neighboring gene words (i.e., genes having a gene pair relationship) of a certain gene word based on the relationship of the gene pair, and in this process, a small and dense nested vector that has been learned is used. Since it is prior art, it will not be described here.

[0141] In step S504, the second gene samples are vectorized to obtain gene expression vectors corresponding to the second gene samples.

[0142] Specifically, the terminal device can collect all different gene names in the tumor gene sample data set, construct a gene vocabulary, and then use the bag-of-words model to vectorize the second gene samples to obtain gene expression vectors corresponding to the second gene samples. The converted gene expression vectors are n-dimensional vectors. For example, the gene expression vector converted from gene 1 can be [0, 1, 0, 0, 0].

[0143] In step S505, the pre-constructed feature extraction model is trained by using the gene representation vectors and the gene expression vectors corresponding to the second gene samples to obtain a source domain feature extraction model.

[0144] Considering that the second gene samples can include many samples lacking label information, manual marking of data is time-consuming and expensive. Specifically, the gene representation vectors and the gene expression vectors corresponding to the second gene samples can be input into the feature extraction model for self-supervised training.

[0145] Specifically, the feature extraction model can use a fast attention calculation model Performer based on orthogonal random features. In the pre-training process of the second gene sample data using the fast attention model Performer, there are two embedded vectors for input data, which are the gene representation vectors and the gene expression vectors. Then, according to the self-supervised learning strategy, the gene expression vectors are used as a Masked Language Model (MLM), that is, the tokenization in the gene expression vectors is randomly masked, such as the gene expression vector of gene 1 [0, 1, 0, 0, 0] is converted to [0, 1, MASK, 0, 0] by the Masked Language Model. Then, the true value and the predicted value at the corresponding masked position in the output result of the fast attention model Performer are taken to calculate the cross-entropy loss function, so as to evaluate the model training effect. Through this self-supervised training strategy, the model can learn the general deep representation of gene expression patterns on a large amount of unlabeled data, which helps to reduce the workload of the downstream fine-tuning process.

[0146] Specifically, the cross-entropy loss function can use the following formula:

[0147] where M is the number of second gene samples, N is the number of masked gene expression values in the gene expression vector, y i,j and p i,j represent the true value and the predicted value of the jth gene expression value in the ith second gene sample, respectively, and L Rec represents the cross-entropy loss.

[0148] Further, in order to facilitate the hyperparameter optimization in the model pre-training process, the hyperparameters in the feature extraction model are searched by using the grid optimization. For example, in the pre-training stage of the source domain feature extraction model, four parameters are mainly involved, which are the number of tokens (num_tokens) in the gene expression vector, the dimension (dim) of the gene expression vector, the number of heads (heads) of each layer of the Performer encoder, and the number of layers (depth) of the Performer encoder. Through the grid optimization, it can be obtained that the range of num_tokens is [5, 8, 12, 15, 20], the range of dim is [100, 150, 200], the range of heads is [8, 10, 16, 20], and the range of depth is [4, 6, 8].

[0149] According to the above scheme, the correlation coefficients between the second gene samples are calculated, the gene pairs satisfying the gene co-expression relationship are determined based on the correlation coefficients between the second gene samples, the genes are subjected to word embedding processing based on the gene pairs and the preset language universal model to obtain the gene expression vectors corresponding to the second gene samples, the second gene samples are subjected to vectorization processing to obtain the gene expression vectors corresponding to the second gene samples, and the feature extraction model is trained based on the gene expression vectors and the gene expression vectors corresponding to the second gene samples to obtain the source domain feature extraction model. The multi-type tumor data is used for pre-training, and the gene embedding and expression embedding are fused in the pre-training process, so that the general knowledge of tumor disease can be learned more accurately.

[0150] In addition, the present application also provides a tumor classification device, which comprises:

[0151] a feature extraction module, configured to perform feature processing on specific tumor gene data to be classified to obtain a target feature vector

[0152] a classification module, configured to identify the target feature vector by using the trained biological neural network model to obtain a classification result of the specific tumor gene data to be classified, wherein the biological neural network model is a neural network model constructed based on genes, pathways and biological processes.

[0153] The principle and implementation process of the tumor classification device are realized, and please refer to the above embodiments, which will not be described here.

[0154] Furthermore, the application also provides a terminal device, which comprises a memory, a processor, and a tumor classification program stored in the memory and capable of running on the processor. The tumor classification program, when executed by the processor, implements the steps of the tumor classification method.

[0155] Since the tumor classification program is executed by the processor, all the technical solutions of all the foregoing embodiments are adopted, and thus all the beneficial effects brought by all the technical solutions of all the foregoing embodiments are achieved, which will not be repeated here.

[0156] Furthermore, the application also provides a computer readable storage medium, which stores a tumor classification program. The tumor classification program, when executed by the processor, implements the steps of the tumor classification method.

[0157] Since the tumor classification program is executed by the processor, all the technical solutions of all the foregoing embodiments are adopted, and thus all the beneficial effects brought by all the technical solutions of all the foregoing embodiments are achieved, which will not be repeated here.

[0158] Compared with the prior art, the tumor classification method, device, terminal device and storage medium provided by the application specifically perform feature processing on specific tumor gene data to be classified to obtain a target feature vector, and perform identification on the target feature vector by using a trained biological neural network model to obtain a classification result of the specific tumor gene data to be classified. Since the biological neural network model of the embodiment introduces biological knowledge for construction, the classification result of the specific tumor gene data can be predicted on the basis of specific cancer tumor gene attributes in combination with hierarchical information in biology, so that the credibility of the classification result is higher, and thus the accuracy of tumor classification is improved.

[0159] It should be noted that, in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article or system including a series of elements does not only include those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article or system. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of another identical element in the process, method, article or system including the element.

[0160] The serial numbers of the above application embodiments are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0161] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, can also be through hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the prior art contribution can be embodied in the form of software products, the computer software product is stored in the above-mentioned storage medium (such as ROM / RAM, magnetic disc, optical disc), including a number of instructions to make a terminal device (may be a mobile phone, computer, server, controlled terminal, or network equipment, etc.) executes the method of each embodiment of the present application.

[0162] The above is only some embodiments of the present application, not therefore limit the patent scope of the present application, any equivalent structure or equivalent flow transformation made by using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for classifying tumors, wherein: The tumor classification method comprises: Perform feature processing on the specific tumor gene data to be classified to obtain the target feature vector; The target feature vector is identified by a trained biological neural network model to obtain a classification result of the specific tumor gene data to be classified; wherein the biological neural network model is a neural network model constructed based on genes, pathways and biological processes.

2. The tumor classification method according to claim 1, wherein: The biological neural network model includes at least a first network layer, a second network layer, and a third network layer; the first network layer includes a plurality of gene nodes, the second network layer includes a plurality of pathway nodes; the third network layer includes a plurality of biological process nodes; and the connection relationship between each of the gene nodes and each of the pathway nodes is determined based on the inclusion relationship between genes and pathways; the connection relationship between each of the pathway nodes and each of the biological process nodes is determined based on the inclusion relationship between pathways and biological processes; The step of identifying the target feature vector by using the trained biological neural network model to obtain the classification result of the specific tumor gene data to be classified includes: Inputting the target feature vector into each of the gene nodes in the first network layer for processing to obtain a first feature vector output by each of the gene nodes; Inputting the first eigenvector of each gene node into the channel node corresponding to each gene node for processing, to obtain the second eigenvector output by each channel node; Outputting the second eigenvector output by each of the channel nodes to the biological process nodes correspondingly connected to each of the channel nodes for processing, to obtain a third eigenvector output by each of the biological process nodes; The classification result is obtained based on at least one first feature vector, at least one second feature vector, and at least one third feature vector.

3. The tumor classification method according to claim 2, wherein: The step of obtaining the classification result based on at least one first feature vector, at least one second feature vector, and at least one third feature vector comprises: At least one of the first feature vectors, at least one of the second feature vectors, and at least one of the third feature vectors are aggregated by averaging to obtain the classification result.

4. The tumor classification method according to claim 1, wherein: The step of performing feature processing on the specific tumor gene data to be classified to obtain a target feature vector includes: The target feature vector is obtained by performing feature extraction on the specific tumor gene data to be classified using a pre-acquired target domain feature extraction model.

5. The tumor classification method according to claim 4, wherein: Before the step of extracting features from the specific tumor gene data to be classified using the pre-acquired target domain feature extraction model to obtain the target feature vector, the following step is also included: The pre-trained source domain feature extraction model is fine-tuned using a target domain dataset to obtain the target domain feature extraction model; wherein the target domain dataset includes a plurality of first gene samples of a specific tumor type.

6. The tumor classification method according to claim 5, wherein: Before the step of fine-tuning the pre-trained source domain feature extraction model using the target domain dataset to obtain the target domain feature extraction model, the following steps are further included: Acquire a source domain dataset, wherein the source domain dataset includes a plurality of second gene samples of different tumor types; The pre-built feature extraction model is trained using the source domain dataset to obtain the source domain feature extraction model.

7. The tumor classification method according to claim 6, wherein: The step of training a pre-built feature extraction model using the source domain dataset to obtain the source domain feature extraction model includes: Calculating the correlation coefficient between each of the second gene samples; Determining gene pairs that satisfy a gene co-expression relationship based on the correlation coefficients between the respective second gene samples; Based on the gene pairs, a preset universal language model is used to perform word embedding processing on each gene to obtain a gene representation vector corresponding to each second gene sample; performing vectorization processing on each of the second gene samples to obtain a gene expression vector corresponding to each of the second gene samples; The pre-constructed feature extraction model is trained using the gene representation vectors and gene expression vectors corresponding to each of the second gene samples to obtain the source domain feature extraction model.

8. A tumor classification device, wherein: The tumor classification device comprises: Feature extraction module, used to process the features of the specific tumor gene data to be classified and obtain the target feature vector A classification module is used to identify the target feature vector through a trained biological neural network model to obtain a classification result of the specific tumor gene data to be classified; wherein the biological neural network model is a neural network model constructed based on genes, pathways and biological processes.

9. A terminal device, wherein: The terminal device includes a memory, a processor, and a tumor classification program stored in the memory and executable on the processor. When the tumor classification program is executed by the processor, the steps of the tumor classification method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, wherein: The computer-readable storage medium stores a tumor classification program, which, when executed by a processor, implements the steps of the tumor classification method according to any one of claims 1 to 7.

Images