Disease classification model training method and apparatus, and disease classification system

Abstract

The application discloses a disease classification model training method and device and a disease classification system, and applies to the technical field of data processing. The method comprises the following steps: obtaining target disease gene data and a transcriptome data set associated with a target disease; performing classification processing on the transcriptome data set to obtain a discovery set, a training set and a test set; obtaining a plurality of target gene pair features associated with the target disease according to the discovery set and the target disease gene data; processing the training set and the test set according to the plurality of target gene pair features to obtain a processed training set and a processed test set; and performing model training and testing by using a machine learning method according to the processed training set and the processed test set to obtain a disease classification model. In the classification system, the transcriptome data of a to-be-tested sample associated with the target disease is classified according to the disease classification model, and the category of the to-be-tested sample is obtained. The application improves the training precision and the training efficiency of the disease classification model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a disease classification model training method and device and a disease classification system. BACKGROUND

[0002] Transcriptome data refers to the determination result of the collection of all ribonucleic acid molecules in a biological body and the expression amount thereof. In the related art, a classification model of a specific disease is obtained by using transcriptome data of the specific disease for model training, and then the classification model of the specific disease is used to implement classification processing of the specific disease. Before model training, the related art usually performs operations such as data uniform normalization and batch effect removal on the transcriptome data in the test set and the training set, aiming to reduce the differences between different batches and ensure the alignment of data in each batch. However, operations such as data uniform normalization and batch effect removal usually require the training set and the test set to be standardized to the same data distribution space, which means that when a new test set comes, the current training set and the current test set are used to integrate the new test set and re-standardize, or when a new training set comes, the current test set and the current training set are used to integrate the new training set and re-standardize, thereby causing the training efficiency of the disease classification model to decrease. In addition, the related art still uses a traditional data processing method to process different types of transcriptome data, which has poor effect in cross-dataset processing, thereby causing the precision of the disease classification model to decrease. SUMMARY

[0003] The embodiments of the present application provide a disease classification model training method, device and disease classification system, which are used to improve the training precision and training efficiency of the disease classification model, thereby improving the classification accuracy of the disease classification model.

[0004] In one aspect, the embodiments of the present application provide a disease classification model training method, including the following steps:

[0005] Obtaining target disease gene data and a transcriptome data set associated with a target disease; wherein the transcriptome data set includes transcriptome data of a plurality of initial samples and a category;

[0006] Performing classification processing on the transcriptome data set to obtain a discovery set, a training set and a test set;

[0007] According to the discovery set and the target disease gene data, a plurality of target gene pair features associated with the target disease are obtained;

[0008] According to a plurality of target gene pair features, the training set and the test set are processed to obtain a processed training set and a processed test set;

[0009] According to the processed training set and the processed test set, a disease classification model is obtained by using a machine learning method for model training and testing, and the disease classification model is used to identify the category of the sample associated with the target disease.

[0010] In another aspect, the embodiments of the present application provide a disease classification model training device, comprising:

[0011] The first obtaining module is configured to obtain target disease gene data and a transcriptome dataset associated with the target disease, wherein the transcriptome dataset comprises transcriptome data and categories of a plurality of initial samples.

[0012] The first processing module is configured to perform classification processing on the transcriptome dataset to obtain a discovery set, a training set, and a test set.

[0013] The second processing module is configured to obtain a plurality of target gene pair features associated with the target disease according to the discovery set and the target disease gene data.

[0014] The third processing module is configured to process the training set and the test set according to the plurality of target gene pair features to obtain a processed training set and a processed test set.

[0015] The fourth processing module is configured to obtain a disease classification model by using a machine learning method for model training and testing according to the processed training set and the processed test set, and the disease classification model is used to identify the category of the sample associated with the target disease.

[0016] In another aspect, the embodiments of the present application provide a disease classification system, comprising:

[0017] The second obtaining module is configured to obtain transcriptome data of a to-be-tested sample associated with the target disease.

[0018] The fifth processing module is configured to perform classification processing on the transcriptome data of the to-be-tested sample according to the disease classification model to obtain the category of the to-be-tested sample, and the disease classification model is obtained by using the disease classification model training method according to any one of claims 1-8.

[0019] The beneficial effects of the present application are: a disease classification model training method and device and a disease classification system are provided, in the training method and device, target disease gene data and a transcriptome data set associated with the target disease are obtained; the transcriptome data set is classified to obtain a discovery set, a training set and a test set; a plurality of target gene pair features associated with the target disease are obtained according to the discovery set and the target disease gene data; the training set and the test set are processed according to the plurality of target gene pair features to obtain a processed training set and a processed test set; and a disease classification model is obtained by model training and testing using a machine learning method according to the processed training set and the processed test set. In the classification system, the transcriptome data of a to-be-tested sample associated with the target disease is classified according to the disease classification model to obtain the category of the to-be-tested sample. The present application can improve the training accuracy and efficiency of the disease classification model, thereby improving the classification accuracy of the disease classification model. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a structural diagram of the disease classification model training method provided by the present application;

[0021] Figure 2 is a principle diagram of the disease classification model training method provided by the present application;

[0022] Figure 3 is an effect comparison diagram of the skin cancer classification task provided by the present application;

[0023] Figure 4 is an effect comparison diagram of the skin cancer immunotherapy response prediction task provided by the present application. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0025] Currently, transcriptome data can be obtained by DeoxyriboNucleic Acid (DNA) microarray, RNA sequencing (RNA-seq) technology or single-cell RNA sequencing (scRNA-seq) technology, but no matter which sequencing technology is used, the final result is often presented in the form of a gene expression matrix, which provides a basis for various downstream analyses, such as differential expression analysis, clustering analysis, principal component analysis, functional enrichment analysis and other tasks. In related technologies, a transcriptome data of a specific disease is used for model training to obtain a classification model of the specific disease, and then the classification model of the specific disease is used to implement classification processing of the specific disease. Due to the differences between different sequencing technologies, the differences between different data standardization methods and experimental errors caused by different experimental conditions, there is a large data difference between the test set and the training set. Therefore, before model training, related technologies usually perform operations such as data uniform standardization and batch effect removal on the transcriptome data in the test set and the training set, aiming to reduce the differences between different batches and ensure the alignment of data in each batch. However, operations such as data uniform standardization and batch effect removal often require the training set and the test set to be standardized to the same data distribution space, which means that when a new test set comes, the current training set and the current test set are used to integrate the new test set and re-standardize, or when a new training set comes, the current test set and the current training set are used to integrate the new training set and re-standardize, thereby causing the training efficiency of the disease classification model to decrease. In addition, related technologies still use traditional data processing methods to process different types of transcriptome data, which is not good at processing across data sets, thereby causing the accuracy of the disease classification model to decrease. It can be understood that, since the sample types of the test set and the training set are different, cross data set can refer to the test set and the training set. In addition, since the types of samples contained in the test set or the training set are also different, cross data set can also refer to data sets composed of different types of transcriptome data.

[0026] Therefore, the embodiments of the present application provide a disease classification model training method and device and a disease classification system, aiming to improve the training accuracy and efficiency of the disease classification model, thereby improving the classification accuracy of the disease classification model.

[0027] It should be emphasized that the input of the disease classification model in the embodiments of the present application is the data sample pre-stored in the database, not the data sampled from the living body, and the output of the disease classification model is the category of the sample associated with the target disease, which is not directly aimed at obtaining the disease diagnosis result or the health condition, and the output of the disease classification model is only used to assist doctors in diagnosis, treatment or medical learning, and the final diagnosis result or treatment result still needs to be determined by the doctor.

[0028] First, the implementation steps of the training method of the disease classification model provided by the present application will be described in detail below in combination with the drawings.

[0029] The training method of the disease classification model provided by the present application can be applied in a terminal, can be applied in a server, and can also be software running in a terminal or a server, etc. The terminal can be a tablet computer, a notebook computer, a desktop computer, etc., but is not limited thereto. The server can be a standalone physical server, can be a server cluster or a distributed system composed of multiple physical servers, can also be a cloud server providing basic cloud computing services such as cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content delivery network (CDN), and big data and artificial intelligence platform, etc. In addition, the server can also be a node server in a blockchain network, but is not limited thereto. The blockchain is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, and encryption algorithm.

[0030] Referring to Figure 1 , Figure 1 is a flowchart of the training method of the disease classification model provided by the present application, which can include the following steps S100-S500.

[0031] S100, obtaining target disease gene data and a transcriptome data set associated with the target disease.

[0032] It should be noted that the target disease refers to the classification object of the disease classification model in the embodiments of the present application, and its type can be set according to actual conditions, for example, the target disease can be skin cancer, liver cancer, etc., but is not limited thereto. In addition, the target disease gene data refers to the related gene type or high variable gene type of the target disease, which can be pre-stored in a specific database, wherein the high variable gene type refers to the type of genes with significant differences in expression amount in different cells or individuals.

[0033] It can be understood that the transcriptome dataset includes transcriptome data of a plurality of initial samples and categories. The transcriptome data of the initial samples can include at least one of microarray data of the initial samples, transcriptome sequencing data of the initial samples, or single-cell transcriptome sequencing data of the initial samples. The microarray data refers to preprocessed initial microarray data, the initial microarray data refers to data obtained by determining the initial samples by DNA microarray; the transcriptome sequencing data refers to preprocessed transcriptome sequencing expression matrix, the transcriptome sequencing expression matrix refers to data obtained by determining the initial samples by RNA-seq technology; and the single-cell transcriptome sequencing data refers to preprocessed single-cell transcriptome sequencing expression matrix, the single-cell transcriptome sequencing expression matrix refers to data obtained by determining the initial samples by scRNA-seq technology. In addition, the categories of the initial samples refer to target disease categories to which the initial samples belong. For example, in a skin cancer classification task, the categories of the initial samples can include melanoma, basal cell carcinoma, Merkel cell carcinoma, squamous cell carcinoma, and non-skin cancer, but are not limited thereto. For another example, in a skin cancer immunotherapy response prediction task, the categories of the initial samples can include two categories of skin cancer immunotherapy response abnormality and skin cancer immunotherapy response normality, but are not limited thereto.

[0034] In this step, target disease gene data is obtained, the target disease gene data being used to determine target gene pair features associated with a target disease, and transcriptome data and categories of a plurality of initial samples associated with the target disease are obtained and a transcriptome dataset is constructed, the transcriptome dataset being used to determine target gene pair features associated with the target disease and to perform model training and testing.

[0035] S200, performing classification processing on the transcriptome dataset to obtain a discovery set, a training set, and a test set.

[0036] It should be noted that the discovery set can include transcriptome data of a plurality of discovery samples and categories. The discovery samples refer to initial samples divided into the discovery set, the transcriptome data of the discovery samples can include any one of single-cell transcriptome sequencing data of the discovery samples or batch transcriptome data of the discovery samples, and the batch transcriptome data of the discovery samples can include microarray data of the discovery samples and transcriptome sequencing data of the discovery samples.

[0037] It can be understood that the training set can include transcriptome data and categories of a plurality of training samples. Among them, the training sample refers to the initial sample divided into the training set, and the transcriptome data of the training sample can include microarray data of the training sample and transcriptome sequencing data of the training sample. In addition, the test set can include transcriptome data and categories of a plurality of test samples. Among them, the test sample refers to the initial sample divided into the test set, and the transcriptome data of the test sample can include at least one of the microarray data of the test sample, the transcriptome sequencing data of the test sample or the single-cell transcriptome sequencing data of the test sample.

[0038] In this step, the transcriptome data set is divided into a discovery set, a training set and a test set according to a preset division ratio. Among them, the discovery set is used to determine the target gene pair features associated with the target disease, and the training set and the test set are used to participate in the training and testing of the model.

[0039] Optionally, the division ratio can be set according to actual conditions, and the embodiments of the present application do not make specific limitations. For example, a plurality of initial samples of transcriptome data are divided into a discovery set, a training set and a test set according to a division ratio of 2:5:3, but not limited to this.

[0040] Optionally, the discovery set and the training set can be the same type of data set, or can be different types of data set. For example, when the transcriptome data of each training sample in the training set is microarray data and transcriptome sequencing data, the transcriptome data of each discovery sample in the discovery set can be microarray data and transcriptome sequencing data, or single-cell transcriptome sequencing data.

[0041] S300, according to the discovery set and the target disease gene data, a plurality of target gene pair features associated with the target disease are obtained.

[0042] It should be noted that the target gene pair feature refers to a feature obtained by combining two gene data most related to the target disease, and the gene data is the gene type.

[0043] In this step, after obtaining the discovery set and the target disease gene data, the target disease gene data and the transcriptome data and categories of each discovery sample in the discovery set are processed, aiming to determine a plurality of gene data most related to the target disease, so as to construct a plurality of target gene pair features associated with the target disease.

[0044] Optionally, the total number of target gene pair features can be set according to actual conditions, and the embodiments of the present application do not make specific limitations. For example, the total number of target gene pair features can be 50 or 100, or an integer value between 50 and 100, but not limited to this.

[0045] S400, processing the training set and the test set according to the plurality of target gene pair features to obtain a processed training set and a processed test set.

[0046] It should be noted that the processed training set can include a plurality of gene pair features and categories corresponding to each training sample, and the plurality of gene pair features corresponding to the training sample are the features of the transcriptome data of each training sample, and the processed test set can include a plurality of gene pair features and categories corresponding to each test sample, and the plurality of gene pair features corresponding to the test sample are the features of the transcriptome data of the test sample.

[0047] It can be understood that the processed training set and the processed test set are located in the same data distribution space.

[0048] In this step, the training set and the test set are processed by using a plurality of target gene pair features to extract features of the training set and the test set, and to make the training set and the test set located in the same data distribution space, thereby obtaining a processed training set and a processed test set. It can be understood that in actual application, when a new test set arrives, the new test set, the current test set and the current training set can be standardized to the same data distribution space and feature extraction can be achieved by using a plurality of target gene pair features, or when a new training set arrives, the new training set, the current test set and the current training set can be standardized to the same data distribution space and feature extraction can be achieved by using a plurality of target gene pair features. In this way, using a plurality of target gene pair features as the processing reference of the test set and the training set can not only effectively ensure the data uniformity between the test set and the training set, but also ensure the data uniformity within the test set and the data uniformity within the training set, thereby improving the training accuracy and training efficiency of the disease classification model.

[0049] S500, model training and testing are performed according to the processed training set and the processed test set by using a machine learning method to obtain a disease classification model.

[0050] It should be noted that the disease classification model is used to identify the category of the sample associated with the target disease. For example, in a skin cancer classification task, the disease classification model can be used to identify the specific type of skin cancer, such as identifying melanoma, basal cell carcinoma, Merkel cell carcinoma, squamous cell carcinoma and non-skin cancer, etc. For another example, in a skin cancer immunotherapy response prediction task, the disease classification model can be used to identify whether the skin cancer immunotherapy response is normal, i.e., the disease classification model outputs a result of normal skin cancer immunotherapy response or a result of abnormal skin cancer immunotherapy response.

[0051] In this step, after obtaining the processed training set and the processed test set, a model training is performed on the processed training set using a machine learning method, and a model testing is performed on the processed test set, so as to obtain a disease classification model.

[0052] Optionally, the machine learning method can be set according to actual conditions, and embodiments of the present application do not make specific limitations thereon. For example, the machine learning method can include Logistic Regression (LR), Random Forest (RF), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), etc., but is not limited thereto.

[0053] As can be seen from the above, although the current gene pair analysis relies on the expression value of the gene, the expression value of the two genes is often consistent in different data sets of the same tissue, and therefore the characteristics of the gene pair can be applied to the research across data sets and the application of the disease classification task, especially the skin cancer classification task or the skin cancer immunotherapy prediction task, and of course the embodiments of the present application can also be applied to other disease classification tasks. Specifically, in the embodiments of the present application, firstly, the target disease gene data and the transcriptome data of a plurality of initial samples associated with the target disease are obtained, secondly, the transcriptome data of the plurality of initial samples is classified to obtain a discovery set, a training set and a test set, then the sample data of the discovery set is used to explore a plurality of target gene pair characteristics associated with the target disease, which are less affected by the differences between different sequencing technologies, the differences between different data standardization methods and experimental errors caused by different experimental conditions, have better cross-data set generalization ability, and can connect different modal transcriptome data, for example, can connect single-cell transcriptome sequencing data and transcriptome sequencing data, which is beneficial to cross-data set processing, then the training set and the test set are processed using the target gene pair characteristics, so as to not only accurately extract the gene pair characteristics of the data set as the sample characteristics of the data set, but also ensure that different types of transcriptome data are in the same data distribution space, effectively improving the processing efficiency and processing accuracy of cross-data set, finally, the model training and testing are performed using the processed training set and the processed test set, so as to effectively improve the training accuracy and training efficiency of the disease classification model, and further improve the explainability and accuracy of the disease classification model.

[0054] The above steps will be further described below.

[0055] In some embodiments, with reference to Figure 2 , Figure 2is a schematic diagram of the training method of the disease classification model provided in the present application. In the step S100, the implementation process of obtaining the target disease gene data and the transcriptome dataset associated with the target disease can include the following steps S110-S140.

[0056] S110, obtaining a gene type associated with the target disease as the target disease gene data.

[0057] It should be noted that the gene type associated with the target disease refers to the relevant gene type or the high-variation gene type of the target disease, which can be pre-stored in the corresponding database.

[0058] In this step, by performing a keyword search on the preset biological information database, the gene type associated with the target disease is retrieved from the biological information database as the target disease gene data and outputted.

[0059] Optionally, the biological information database can be set according to actual conditions, which is not specifically limited in the present embodiment. For example, the biological information database can be the database of the regional biological technology information center, but is not limited thereto.

[0060] Optionally, the retrieval keyword of the biological information database can be set according to actual conditions, which is not specifically limited in the present embodiment. For example, when facing the skin cancer classification task or the skin cancer immunotherapy response prediction task, the retrieval keyword can be skin cancer or Skin cancer, but is not limited thereto.

[0061] S120, obtaining initial transcriptome data and categories of a plurality of initial samples associated with the target disease.

[0062] It should be noted that the initial transcriptome data of the initial sample can include at least one of the transcriptome sequencing expression matrix of the initial sample, the single-cell transcriptome sequencing expression matrix of the initial sample, or the original microarray data of the initial sample. The initial microarray data refers to the data obtained by determining the initial sample through DNA microarray; the transcriptome sequencing expression matrix refers to the data obtained by determining the initial sample through RNA-seq technology, which is essentially a gene expression matrix; and the single-cell transcriptome sequencing expression matrix refers to the data obtained by determining the initial sample through scRNA-seq technology, which is essentially a gene expression matrix.

[0063] In this step, the initial transcriptome data of a plurality of initial samples associated with the target disease is obtained, and the categories of the plurality of initial samples associated with the target disease are also obtained, so as to construct the transcriptome dataset based on these data in the subsequent steps.

[0064] S130, normalizing the initial transcriptome data of each initial sample to obtain transcriptome data of each initial sample.

[0065] It should be noted that the transcriptome data of the initial sample can include at least one of the microarray data of the initial sample, the transcriptome sequencing data of the initial sample, or the single-cell transcriptome sequencing data of the initial sample.

[0066] In this step, for the initial transcriptome data of each initial sample, the initial transcriptome data of the current initial sample is preprocessed by a preset normalization method, so as to obtain the transcriptome data of the current initial sample. By traversing the initial transcriptome data of a plurality of initial samples, the transcriptome data of the plurality of initial samples can be obtained. Specifically, for the initial transcriptome data of each initial sample, the initial microarray data of the current initial sample is normalized by using a robust multi-array average (RMA) method to obtain the microarray data of the current initial sample, so as to reduce technical variation and improve data comparability; and the transcriptome sequencing expression matrix of the current initial sample is normalized and preprocessed by using a transcripts per million (TPM) method to obtain the transcriptome sequencing data of the current initial sample, and the single-cell transcriptome sequencing expression matrix of the current initial sample is normalized and preprocessed by using the transcripts per million method to obtain the single-cell transcriptome sequencing data of the current initial sample, so that the gene expression amount between different samples is comparable.

[0067] S140, obtaining a transcriptome data set associated with the target disease according to the transcriptome data of each initial sample and the category.

[0068] In this step, after the normalization preprocessing is completed, the transcriptome data set associated with the target disease is constructed by using the transcriptome data of each initial sample and the category, so that the construction processing of the transcriptome data set is realized, the accuracy of the transcriptome data set is effectively improved, and the accuracy of the target gene pair feature and the training accuracy of the disease classification model are ensured.

[0069] In some embodiments, referring to Figure 2 In the step S300, the implementation process of obtaining a plurality of target gene pair features associated with the target disease according to the discovery set and the target disease gene data can include steps S310-S340.

[0070] S310, obtaining a plurality of target gene data associated with the target disease according to the transcriptome data of each discovery sample in the discovery set, the transcriptome data of each training sample in the training set, and the target disease gene data.

[0071] It should be noted that the target gene data refers to the gene types existing in the discovery set, the training set and the target disease gene data.

[0072] In this step, the genes intersected from the discovery set, the training set and the target disease gene data are taken as the target gene data. Specifically, first, at least one discovery set gene data is obtained according to the transcriptome data of each discovery sample in the discovery set, the discovery set gene data refers to the gene data existing in the discovery set, and the gene data is the gene type, and at least one training set gene data is obtained according to the transcriptome data of each training sample in the training set, the training set gene data refers to the gene data existing in the training set, and then the intersection result of the at least one discovery set gene data, the at least one training set gene data and the target disease gene data is determined as a plurality of target genes associated with the target disease, that is, the gene types existing in the discovery set, the training set and the target disease gene data are determined as the target genes associated with the target disease and output.

[0073] S320, obtaining the gene expression matrix of the discovery set according to the plurality of target gene data and the transcriptome data of each discovery sample in the discovery set.

[0074] It should be noted that the gene expression matrix of the discovery set is a two-order matrix, in the gene expression matrix of the discovery set, the row represents the target gene data, the column represents the discovery sample, and the value of the element is the gene expression amount of the target gene data in the discovery sample. The gene expression amount is a specific, measured value, for example, the gene expression amount of gene A is 3 and the gene expression amount of gene B is 5 in a certain cell sample, but it is not limited thereto.

[0075] In this step, first, for each target gene data, the gene expression amount of the current target gene data in each discovery sample is determined according to the current target gene data and the transcriptome data of each discovery sample in the discovery set. It can be understood that the gene expression amount is determined by a preset sequencing technology and stored in the transcriptome data in the form of a specific numerical value. By traversing the plurality of target gene data, the gene expression amount of each target gene data in each discovery sample can be obtained. Then, the target gene data is taken as the row of the gene expression matrix of the discovery set, the discovery sample is taken as the column of the gene expression matrix of the discovery set, and the gene expression amount of each target gene data in each discovery sample is taken as the value of each element in the gene expression matrix of the discovery set, to construct the gene expression matrix of the discovery set, which contains each discovery sample and its corresponding plurality of target gene data, as shown in part (b) of Figure 2 .

[0076] Alternatively, the gene expression matrix is represented as as shown in the following formula (1):

[0077] (1);

[0078] In equation (1), This represents the total number of genes, i.e., the number of rows in the gene expression matrix; The number of samples indicates the number of columns in the gene expression matrix; elements It refers to the first gene expression matrix. Line 1 Column elements, elements Used to indicate the The gene in the first Gene expression levels in each sample , Understandably, the gene expression matrix representation is applicable to the discovery set, training set, and test set.

[0079] S330: Based on the gene expression matrix of the discovery set, obtain the gene pair matrix of the discovery set.

[0080] It should be noted that the gene pair matrix of the discovery set is a second-order matrix. In the gene pair matrix of the discovery set, the rows represent gene pairs of the discovery set, which are composed of two target gene data. The columns represent the discovery samples, and the values ​​of the elements are the numerical values ​​of the gene pairs of the discovery set in the discovery samples. The numerical values ​​of the gene pairs of the discovery set in the discovery samples include any one of the first preset value, the second preset value, or the third preset value.

[0081] In this step, to construct an adaptive, individualized gene pair matrix, the gene expression matrix of the discovery set is used to exhaustively compare the data of every two target genes in each discovery sample. The Welch test's fitness difference is introduced to capture the differential changes between different genes, and the differential changes between different genes are adaptively quantified, thus obtaining multiple discovery set gene pairs and the numerical values ​​of each discovery set gene pair in each discovery sample. Then, using the discovery set gene pairs as the rows of the discovery set gene pair matrix, the discovery samples as the columns, and the numerical values ​​of each discovery set gene pair in each discovery sample as the values ​​of each element in the discovery set gene pair matrix, a gene pair matrix of the discovery set is constructed. The gene pair matrix of the discovery set includes each discovery sample and its corresponding multiple discovery set gene pairs, such as... Figure 2 As shown in section (c), this enables the conversion of the gene expression matrix of the discovery set into a gene pair matrix of the discovery set.

[0082] Optionally, the gene pair matrix is ​​represented as As shown in the following formula (2):

[0083] (2);

[0084] In equation (2), This represents the total number of gene pairs, i.e., the number of rows in the gene pair matrix, where the total number of gene pairs is from... The number of combinations of extracting two target gene data from one gene; The number of samples is represented by the number of columns in the gene pair matrix; elements This refers to the first gene pair in the gene pair matrix. Line 1 Column elements, elements Used to indicate the The gene pair in the first... The values ​​of each sample, , Understandably, the gene pair matrix representation is applicable to the discovery set, training set, and test set.

[0085] S340 involves screening gene pairs in the gene pair matrix of the discovery set to obtain multiple target gene pair features associated with the target disease.

[0086] It should be noted that target gene pair features refer to features obtained by combining the data of two genes most relevant to the target disease. In other words, target gene pair features can include two gene data most relevant to the target disease, both of which are gene types that are significantly reversed in different categories of observed samples. Significant reversal means significantly improving the effect of immunotherapy response to the target disease or completely reversing the defects caused by the target disease.

[0087] In this step, after obtaining the gene pair matrix of the discovery set, cross-category analysis is performed. Specifically, Fisher's exact test or Pearson's chi-square test is applied to calculate the significance level value of each gene pair in the gene pair matrix of the discovery set. The significance level value of each gene pair is then used to screen the gene pairs, aiming to identify gene pairs that show significant reversal across different categories as target gene pair features associated with the target disease, thereby obtaining multiple target gene pair features associated with the target disease.

[0088] In some implementations, refer to Figure 2 In step S330 above, the process of obtaining the gene pair matrix of the discovery set based on the gene expression matrix of the discovery set may include the following steps S331-S335:

[0089] S331, the total number of gene pairs in the discovery set is determined by the number of combinations of two target gene data taken from multiple target gene data;

[0090] S332, Based on the total number of gene pairs in the discovery set and the gene expression matrix of the discovery set, determine multiple gene pairs in the discovery set;

[0091] S333, obtaining the variance of the gene expression amount of each target gene data in all discovery samples as the expression variance of each target gene data according to the gene expression matrix of the discovery set;

[0092] S334, obtaining the value of each discovery set gene pair in each discovery sample according to the gene expression matrix of the discovery set, the expression variance of each target gene data, and a preset adaptive coefficient;

[0093] S335, constructing a gene pair matrix of the discovery set by taking the discovery set gene pairs as the rows of the gene pair matrix of the discovery set, taking the discovery samples as the columns of the gene pair matrix of the discovery set, and taking the value of each discovery set gene pair in each discovery sample as the value of each element in the gene pair matrix of the discovery set.

[0094] In this embodiment, first, the number of combinations of two target gene data is determined as the total number of discovery set gene pairs, that is, the total number of discovery set gene pairs is and the combination operation is performed on the target gene data in the gene expression matrix of the discovery set according to the total number of discovery set gene pairs, to obtain discovery set gene pairs, each of which is composed of two target gene data.

[0095] Then, the variance of the gene expression amount of each target gene data in all discovery samples is calculated as the expression variance of each target gene data according to the value of each element in the gene expression matrix of the discovery set. The expression variance of the target gene data can describe the dispersion and volatility of the target gene data in all discovery samples. Since the calculation of variance belongs to the prior art, this embodiment will not be described again.

[0096] After that, for each discovery set gene pair, the value of the current discovery set gene pair in discovery samples is obtained according to the value of each element in the gene expression matrix of the discovery set, the expression variance of the target gene data, and a preset adaptive coefficient. By traversing the discovery set gene pairs, the value of each discovery set gene pair in discovery samples can be obtained.

[0097] ​​Specifically, for the two target gene data contained in a single discovery sample and a single discovery set gene pair, there are: first, obtaining the difference between the gene expression of one of the target gene data in the discovery sample and the gene expression of the other target gene data in the discovery sample as the expression difference of the discovery set gene pair; then, obtaining the sum of the variances of the expression of the two target gene data as the variance sum value, and taking the square root of the variance sum value to obtain the variance square root value; then, obtaining the product of the variance square root value and the adaptive coefficient as the quantitative change value of the discovery set gene pair, which can reflect the expression level difference between the two target gene data in the discovery set gene pair; finally, the size relationship between the gene expression of the two target gene data in the discovery sample is determined by comparing the expression difference of the discovery set gene pair and the quantitative change value of the discovery set gene pair, and the numerical value of the discovery set gene pair in the discovery sample is determined by using the size relationship.

[0098] More specifically, if the expression difference of the discovery set gene pair is greater than the quantitative change value of the discovery set gene pair, it indicates that the gene expression of one of the target gene data in the discovery sample is greater than the gene expression of the other target gene data in the discovery sample, and at this time the numerical value of the discovery set gene pair in the discovery sample is determined as a first preset value. If the expression difference of the discovery set gene pair is less than the quantitative change value of the discovery set gene pair, it indicates that the gene expression of one of the target gene data in the discovery sample is less than the gene expression of the other target gene data in the discovery sample, and at this time the numerical value of the discovery set gene pair in the discovery sample is determined as a second preset value. If the expression difference of the discovery set gene pair is equal to the quantitative change value of the discovery set gene pair, it indicates that the gene expression of the two target gene data in the discovery sample is the same, and at this time the numerical value of the discovery set gene pair in the discovery sample is determined as a third preset value.

[0099] Finally, taking the discovery set gene pair as the row of the gene pair matrix of the discovery set, taking the discovery sample as the column of the gene pair matrix of the discovery set, and taking the numerical value of each discovery set gene pair in each discovery sample as the value of each element in the gene pair matrix of the discovery set, an adaptive individualized gene pair matrix, i.e., the gene pair matrix of the discovery set, is constructed, as shown in part (c) of FIG. 1. Figure 2

[0100] It can be seen that in the process of constructing the gene pair matrix, the present embodiment first obtains a plurality of gene pairs from the gene expression matrix by a combination method, then determines the expression difference between the two genes in each gene pair based on the rank method, and uses these expression differences to assign values to each element of the gene pair matrix, and finally constructs the gene pair matrix by taking the gene pair as the row of the gene pair matrix and taking the sample as the column of the gene pair matrix.​

[0101] Current rank-based methods often measure the relative expression levels of different genes by the difference between their expression levels. This approach often ignores the true differences between gene expression levels, and its accuracy needs improvement. Therefore, this implementation introduces adaptive quantification of gene pair difference changes into the rank-based method. Specifically, it uses the Welch test to capture the differences in gene expression levels between the two genes in a gene pair and quantifies these differences to better reflect the expression differences between the two genes in the gene pair and assign values ​​to each element of the gene pair matrix. This preserves the quantitative information between the gene expression levels of the two genes in the gene pair, thus retaining more quantitative information for the rank-based method and reducing the impact of differences in gene pair relative expression levels caused by different sequencing technologies, different data standardization methods, and experimental errors caused by different experimental conditions. Since the relative expression levels of gene pairs are less affected by technological variations, gene pairs can concatenate transcriptome data from different modalities, thereby effectively improving the generalization ability of gene pairs in cross-dataset processing. Furthermore, this implementation method can accurately determine the values ​​of each element in the gene pair matrix, improve the accuracy of the gene pair matrix, ensure the accuracy of subsequent target gene pair features, and thus improve the processing accuracy and efficiency of cross-dataset processing.

[0102] Optionally, the adaptive coefficient can be set according to the actual situation, and this embodiment does not specifically limit it. For example, the adaptive coefficient can be 0, 1, or 2, but is not limited to this. According to the actual experimental results of the embodiments of this application, when the transcriptome data of each discovery sample in the discovery set is single-cell transcriptome sequencing data, the adaptive coefficient can be 2, and when the transcriptome data of each discovery sample in the discovery set is batch transcriptome data, the adaptive coefficient can be 0, which can ensure the best overall performance.

[0103] Optionally, for the first The first discovery set of gene pairs and the first The first discovered sample, the first The first set of gene pairs was discovered by the first The target gene data and the first The data consists of several target genes. If the first preset value is 1, the second preset value is -1, and the third preset value is 0, then the first... The first set of gene pairs was discovered in the first The numerical value of each found sample can be expressed as the following formula (3):

[0104] , (3);

[0105] In equation (3), element refers to the element in the matrix at row and column , , , element is used to indicate the value of the th discovery set gene pair at the th discovery sample; represents the variance of the gene expression of the th target gene data across all discovery samples; represents the variance of the gene expression of the th target gene data across all discovery samples; represents the adaptive coefficient; can describe the quantitative change in the gene expression level between the th target gene data and the th target gene data. In this formula, when , it is considered that the gene expression of the th target gene data at the th sample is greater than the gene expression of the th target gene data at the th sample, and the value of the th discovery set gene pair at the th discovery sample is 1. When , it is considered that the gene expression of the th target gene data at the th sample is less than the gene expression of the th target gene data at the th sample, and the value of the th discovery set gene pair at the th discovery sample is -1. When , it is considered that the gene expression of the th target gene data at the th sample is equal to the gene expression of the th target gene data at the th sample, and the value of the th discovery set gene pair at the th discovery sample is 0. Exemplarily, for a gene pair (A, B) composed of gene A and gene B, the gene expression of gene pair A at a certain sample is 3, and the gene expression of gene B at a certain sample is 5, then the difference between the gene expression of gene pair A and the gene expression of gene pair B is calculated as the expression difference of gene pair (A, B), i.e. 3-5=-2; represents the variance of gene A across all samples, represents the variance of gene B in all samples, the adaptive coefficient If the value is 1, the quantitative change value of the gene pair (A, B) is 1; since the expression difference of the gene pair (A, B) is less than the quantitative change value of the gene pair (A, B), the actual expression of gene B is greater than the actual expression of gene A, and therefore the value of the gene pair (A, B) in the sample is -1.

[0106] In some embodiments, the reference Figure 2 The implementation process of obtaining the plurality of target gene pair features associated with the target disease by screening each discovery set gene pair in the gene pair matrix of the discovery set in step S340 can include steps S341-S343.

[0107] S341, for each discovery set gene pair, according to the values of the discovery set gene pair in each discovery sample and the categories of all discovery samples, obtaining the first target number and the second target number of each category relative to the discovery set gene pair as the category number features of the discovery set gene.

[0108] It should be noted that the category refers to the category of the discovery sample, such as melanoma, basal cell carcinoma, Merkel cell carcinoma, squamous cell carcinoma, and non-skin cancer, etc. Among them, for each category, the first target number of the current category relative to the discovery set gene pair refers to the total number of discovery set gene pairs with a value of the first preset value in the discovery sample belonging to the current category, and the second target number of the current category relative to the discovery set gene pair refers to the total number of discovery set gene pairs with a value of the second preset value in the discovery sample belonging to the current category.

[0109] In this step, first, for each discovery set gene pair, there is: for each category, according to the values of each element in the gene pair matrix of the discovery set, obtaining the total number of discovery set gene pairs with a value of the first preset value in the discovery sample belonging to the current category as the first target number of the current category relative to the discovery set gene pair, and obtaining the total number of discovery set gene pairs with a value of the second preset value in the discovery sample belonging to the current category as the second target number of the current category relative to the discovery set gene pair; by traversing the categories of all discovery samples, the first target number and the second target number of all categories relative to the discovery set gene pair can be obtained as the category number features of the discovery set gene pair. By traversing all discovery set gene pairs, the category number features of all discovery set gene pairs can be obtained, and the category number features of each discovery set gene pair are used to calculate the significance level value of the discovery set gene pair.

[0110] Alternatively, the category number features of a single discovery set gene pair can be represented by a matrix , such as Figure 2The first row of the matrix represents the first preset value, the second row represents the second preset value, and the columns represent the categories of the discovery samples, wherein the value of an element in the first row and all columns of the matrix is the first target number of each category with respect to the discovery set gene pair, and the value of an element in the second row and all columns of the matrix is the second target number of each category with respect to the discovery set gene pair. The category number feature of a single discovery set gene pair satisfies the following formula (4):

[0111] (4);

[0112] In formula (4), represents the category with respect to the first target number of the discovery set gene pair, represents the category with respect to the second target number of the discovery set gene pair. represents the total number of categories of the discovery samples. It can be understood that if the first preset value is 1 and the second preset value is -1, then represents the first target number of the first category with respect to the discovery set gene pair, that is, the total number of discovery set gene pairs with a value of 1 in the discovery samples belonging to the first category, represents the number of the category, represents the category with respect to the first target number of the discovery set gene pair, that is, the total number of discovery set gene pairs with a value of 1 in the discovery samples belonging to the category , represents the second target number of the first category with respect to the discovery set gene pair, that is, the total number of discovery set gene pairs with a value of -1 in the discovery samples belonging to the first category, represents the category with respect to the second target number of the discovery set gene pair, that is, the total number of discovery set gene pairs with a value of -1 in the discovery samples belonging to the category .

[0113] S342, according to the category number feature of each discovery set gene pair and the total number of categories of the discovery samples, the significance level value of each discovery set gene pair is obtained.

[0114] In this step, different significance level value calculation methods are used according to the total number of categories of the discovery samples. Specifically, for each discovery set gene pair, when the total number of categories is 2, that is, , the significance level value of the discovery set gene pair is calculated using Fisher's exact test, as shown in the following formula (5):

[0115]

[0116] (5);

[0117] In formula (5), represents a significance level value of the discovery set gene pair; represents a first target number of the discovery set gene pairs in the second category, that is, a total number of the discovery set gene pairs with a value of 1 in the discovery samples belonging to the second category; represents a second target number of the discovery set gene pairs in the second category, that is, a total number of the discovery set gene pairs with a value of -1 in the discovery samples belonging to the second category.

[0118] When the total number of categories is greater than 2, that is, the significance level value of the discovery set gene pair is calculated using the Pearson chi-square test, and the significance level value of the discovery set gene pair is a right tail probability of a test statistic of the Pearson chi-square test, as shown in the following formula (6):

[0119] (6);

[0120] In formula (3), represents a test statistic of the discovery set gene pair; represents a sum of the number of all discovery samples satisfying or ; represents a chi-square distribution with a degree of freedom of .

[0121] S343, according to the significance level value of each discovery set gene pair, screening each discovery set gene pair in the gene pair matrix of the discovery set to obtain a plurality of target gene pair features associated with the target disease.

[0122] It should be noted that the significance level value of the discovery set gene pair is negatively correlated with the significant reversal degree of the discovery set gene pair, that is, the smaller the significance level value of the discovery set gene pair, the higher the significant reversal degree of the discovery set gene pair, and the more likely the discovery set gene pair is significantly reversed; the greater the significance level value of the discovery set gene pair, the lower the significant reversal degree of the discovery set gene pair, and the less likely the discovery set gene pair is significantly reversed.

[0123] It can be understood that the target gene pair feature refers to the discovery set gene pair with a significance level value greater than a preset level threshold, or the target gene pair feature refers to the discovery set gene pair with an arrangement number less than or equal to a preset number threshold when all the discovery set gene pairs are arranged from small to large according to the size of the significance level value.

[0124] In this step, since the smaller the significance level value of a discovery set gene pair is, the more likely the discovery set gene pair is significantly reversed, each discovery set gene pair in the gene pair matrix of the discovery set is screened according to the significance level value of each discovery set gene pair, aiming to screen discovery set gene pairs with lower significance level values as target gene pair features. In this way, representative target gene pair features that are significantly reversed in each category can be accurately screened, the accuracy of the target gene pair features is improved, and the processing accuracy and efficiency across datasets are improved.

[0125] For example, in some embodiments, the screening is performed in a ranking manner. Specifically, first, all the discovery set gene pairs are ranked according to the significance level values of the discovery set gene pairs, to obtain each ranked discovery set gene pair. The ranking number of a discovery set gene pair is positively correlated with the significance level value of the discovery set gene pair, that is, the smaller the significance level value of a discovery set gene pair is, the smaller the ranking number of the discovery set gene pair is, and the larger the significance level value of a discovery set gene pair is, the larger the ranking number of the discovery set gene pair is. Then, the first N discovery set gene pairs (i.e., the discovery set gene pairs with ranking numbers less than or equal to N, and N is a preset number threshold) are selected from all the ranked discovery set gene pairs as the target gene pair features associated with the target disease. The number threshold can be flexibly set according to actual conditions, and the present embodiment does not make specific limitations thereon. For example, the number threshold can be 50 or 100, but is not limited thereto.

[0126] For example, in some embodiments, the screening is performed in a threshold comparison manner. Specifically, the discovery set gene pairs with significance level values less than a preset level threshold are selected as the target gene pair features. The level threshold can be flexibly set according to actual conditions, and the present embodiment does not make specific limitations thereon.

[0127] In some embodiments, the implementation process of processing the training set and the test set according to the target gene pair features in step S400 can include the following steps S01-S03.

[0128] S01, obtaining a gene pair matrix of the training set according to the transcriptome data of each training sample in the training set.

[0129] It should be noted that the gene pair matrix of the training set is a two-order matrix, in the gene pair matrix of the training set, the row represents the training set gene pair, the column represents the training sample, and the value of the element is the numerical value of the training set gene pair in the training sample. The numerical value of the training set gene pair in the training sample includes any one of the first preset value, the second preset value, or the third preset value.

[0130] In this step, first, at least one training set gene data is obtained according to the transcriptome data of each training sample in the training set, and the training set gene data refers to the gene data existing in the training set, and the gene data is the gene type. Then, the gene expression matrix of the training set is obtained according to the at least one training set gene data and the transcriptome data of each training sample in the training set, and the gene expression matrix of the training set contains each training sample and its corresponding multiple training set gene data. Wherein, the gene expression matrix of the training set is a two-order matrix, in the gene expression matrix of the training set, the row represents the training set gene data, the column represents the training sample, and the value of the element is the gene expression amount of the training set gene data in the training sample. The representation form of the gene expression matrix of the training set is the same as the gene expression matrix shown in the above formula (1), and the specific implementation of constructing the gene expression matrix of the training set is the same as the implementation mode of the foregoing step S320, and this embodiment will not be repeated. Then, the gene pair matrix of the training set is obtained according to the gene expression matrix of the training set, and the gene pair matrix of the training set contains each training sample and its corresponding multiple training set gene pairs. Wherein, the implementation mode of converting the gene expression matrix of the training set into the corresponding gene pair matrix is the same as the implementation mode of the foregoing step S330, and the representation form of the gene pair matrix of the training set is the same as the gene pair matrix shown in the above formula (2), and this embodiment will not be repeated.

[0131] S02, obtaining the multiple gene pair features corresponding to each training sample according to the multiple target gene pair features and the gene pair matrix of the training set.

[0132] It should be noted that the multiple gene pair features corresponding to the training sample are the features of the transcriptome data of the training sample.

[0133] In this step, first, for each training set gene pair, if the current training set gene pair exists in the set composed of the multiple target gene pair features, it means that the current training set gene is the same as a certain target gene pair feature, and the current training set gene has model training significance, at this time, the current training set gene pair is retained, otherwise, it means that the current training set gene is different from any one of the target gene pair features, and the current training set gene does not have model training significance, at this time, the current training set gene pair is excluded. By traversing the multiple target gene pair features and the multiple training set gene pairs in the gene pair matrix of the training set, the screened gene pair matrix of the training set can be obtained. Then, the multiple training set gene pairs corresponding to each training sample are extracted from the screened gene pair matrix of the training set as the multiple gene pair features corresponding to each training sample.

[0134] S03, obtaining the processed training set according to the multiple gene pair features corresponding to each training sample and the category.

[0135] In this step, for each training sample, the plurality of gene pair features of the training sample are taken as the sample features, and the category of the training sample is taken as the sample label. Through the plurality of gene pair features and the category corresponding to each training sample, a processed training set can be constructed, which will be used for training the model.

[0136] It can be seen that the embodiment can accurately extract the gene pair features of the training set as the sample features of the training set, and ensure the data uniformity of different types of transcriptome data in the training set. In addition, the plurality of target gene pair features are taken as the processing reference of the training set and the test set. Since the processing reference of the training set and the test set is the same, the training set and the test set can be effectively ensured to be in the same data distribution space during the processing of the training set and the test set, thereby effectively improving the processing efficiency across the data sets and improving the sample accuracy across the data sets.

[0137] In some embodiments, the specific implementation of processing the test set and the test set according to the plurality of target gene pair features in the above step S400 to obtain the processed test set and the processed test set can further include steps S11-S13.

[0138] S11, obtaining a gene pair matrix of the test set according to the transcriptome data of each test sample in the test set.

[0139] It should be noted that the gene pair matrix of the test set is a second-order matrix, and in the gene pair matrix of the test set, the row represents the test set gene pair, the column represents the test sample, and the value of the element is the numerical value of the test set gene pair in the test sample. The numerical value of the test set gene pair in the test sample includes any one of the first preset value, the second preset value or the third preset value.

[0140] In this step, first, at least one test set gene data is obtained according to the transcriptome data of each test sample in the test set, and the test set gene data refers to the gene data existing in the test set, and the gene data is of a gene type. Then, a gene expression matrix of the test set is obtained according to the at least one test set gene data and the transcriptome data of each test sample in the test set, and the gene expression matrix of the test set contains each test sample and a plurality of test set gene data corresponding thereto. The gene expression matrix of the test set is a second-order matrix, in which the row represents the test set gene data, the column represents the test sample, and the value of the element is the gene expression amount of the test set gene data in the test sample. The representation form of the gene expression matrix of the test set is the same as that of the gene expression matrix shown in the above formula (1), and the specific implementation of constructing the gene expression matrix of the test set is the same as that of the implementation mode of the foregoing step S320, which will not be described herein again. Subsequently, a gene pair matrix of the test set is obtained according to the gene expression matrix of the test set, and the gene pair matrix of the test set contains each test sample and a plurality of test set gene pairs corresponding thereto. The implementation mode of converting the gene expression matrix of the test set into the corresponding gene pair matrix is the same as that of the implementation mode of the foregoing step S330, and the representation form of the gene pair matrix of the test set is the same as that of the gene pair matrix shown in the above formula (2), which will not be described herein again.

[0141] S12, obtaining a plurality of gene pair features corresponding to each test sample according to the plurality of target gene pair features and the gene pair matrix of the test set.

[0142] It should be noted that the plurality of gene pair features corresponding to the test sample are features of the transcriptome data of the test sample.

[0143] In this step, first, for each test set gene pair, if the current test set gene pair exists in the set composed of the plurality of target gene pair features, it indicates that the current test set gene is the same as a certain target gene pair feature, and the current test set gene has model test significance, at which time the current test set gene pair is retained, otherwise it indicates that the current test set gene is different from any one of the target gene pair features, and the current test set gene does not have model test significance, at which time the current test set gene pair is excluded. By traversing the plurality of target gene pair features and the plurality of test set gene pairs in the gene pair matrix of the test set, a screened gene pair matrix of the test set can be obtained. Then, a plurality of test set gene pairs corresponding to each test sample are extracted from the screened gene pair matrix of the test set as a plurality of gene pair features corresponding to each test sample.

[0144] S13, obtaining a processed test set according to the plurality of gene pair features corresponding to each test sample and the category.

[0145] In this step, for each test sample, the gene pair features of the test sample are taken as the sample features, and the category of the test sample is taken as the sample label. Through the gene pair features and the category corresponding to each test sample, a processed test set can be constructed, which will be used for testing the model.

[0146] It can be seen that the embodiment can accurately extract the gene pair features of the test set as the sample features of the test set, ensuring the data uniformity of different types of transcriptome data in the test set. In addition, the plurality of target gene pair features are taken as the processing reference common to the training set and the test set. Since the processing reference of the training set and the test set is the same, the training set and the test set can be effectively ensured to be in the same data distribution space during the processing of the training set and the test set, thereby effectively improving the processing efficiency across the data sets and improving the sample accuracy across the data sets.

[0147] In some embodiments, with reference to Figure 2 The implementation process of the disease classification model obtained by using the machine learning method on the processed training set and the processed test set in step S500 can include the following steps S510-S530:

[0148] S510, using a machine learning method to train the processed training set to obtain each machine learning model;

[0149] S520, using the processed test set to test each machine learning model to obtain the test result of each machine learning model;

[0150] S530, obtaining the disease classification model according to the test result of each machine learning model.

[0151] In the embodiment, first, an initial machine learning model corresponding to each machine learning method is constructed. Then, for each initial machine learning model, the training set is input into the current initial machine learning model, the gene pair features of the training sample are taken as the features, and the category of the training sample is taken as the target. The current initial machine learning model is trained to enable the current initial machine learning model to learn the mapping relationship between the gene pair features and the category, and a machine learning model is obtained. Then, the test set is used to test each machine learning model, so that the test result of each machine learning model can reflect the performance of the machine learning model. Finally, according to the test result of each machine learning model, the machine learning model with the best test result is selected as the disease classification model output. According to the actual experimental results of the embodiments of the present application, the performance of the random forest model is the best, and therefore the disease classification model is the trained random forest model.

[0152] Optionally, in the process of model training, in order to improve the accuracy of the model training set, after the processed training set is trained by the machine learning method to obtain each machine learning model, it is judged whether the type of the discovery set is the same as that of the training set; if yes, it means that there is no sample feature with zero feature importance in the processed training set, and at this time, step S520 is directly entered to perform model testing; if not, it means that there is a sample feature with zero feature importance in the processed training set, and at this time, for each machine learning model, the feature importance of each sample feature in the processed training set is obtained by the machine learning model, the sample feature with zero feature importance in the processed training set is screened out, a new training set is obtained, and model training is re-performed on the new training set.

[0153] Optionally, in the process of model testing, the indicators for evaluating the performance of each machine learning model can be flexibly set according to actual conditions, for example, the indicators can be F1 score, accuracy, etc., but are not limited thereto. According to the actual experimental results of the embodiments of the present application, in the skin cancer classification task, the indicator is F1 score; in the skin cancer immunotherapy response prediction task, the indicator is accuracy.

[0154] Optionally, in the overall process of model training and testing, cross-validation methods such as three-fold cross-validation method can be introduced to improve the training efficiency and training accuracy of the machine learning model. For example, according to the three-fold cross-validation method, the training set is divided into a training subset and a test subset of each training round, and the test subset of each training round is combined with the processed test set to perform model testing, but is not limited thereto.

[0155] Secondly, the embodiments of the present application also provide a training device of a disease classification model, which can include a first acquisition module, a first processing module, a second processing module, a third processing module and a fourth processing module. The first acquisition module is used to acquire target disease gene data and a transcriptome data set associated with the target disease; the transcriptome data set includes transcriptome data and categories of a plurality of initial samples; the first processing module is used to perform classification processing on the transcriptome data set to obtain a discovery set, a training set and a test set; the second processing module is used to obtain a plurality of target gene pair features associated with the target disease according to the discovery set and the target disease gene data; the third processing module is used to process the training set and the test set according to the plurality of target gene pair features to obtain a processed training set and a processed test set; and the fourth processing module is used to train and test a model by a machine learning method according to the processed training set and the processed test set to obtain a disease classification model, and the disease classification model is used to identify the categories of samples associated with the target disease.

[0156] The contents in the method embodiments are applicable to the device embodiments, the device embodiments specifically implement the functions same as the method embodiments, and achieve the same beneficial effects as the method embodiments.

[0157] Finally, the embodiment of the present application provides a disease prediction system, which comprises a second acquisition module and a fifth processing module. The second acquisition module is configured to acquire transcriptome data of a target sample associated with a target disease. It should be noted that the target sample is a sample pre-stored in a database, rather than a sample directly obtained from a living body. The fifth processing module is configured to perform classification processing on the transcriptome data of the target sample according to a disease classification model to obtain a category of the target sample. The disease classification model is trained by the training method of the disease classification model.

[0158] In some embodiments, the fifth processing module can comprise a preprocessing unit and a prediction unit. The prediction unit is configured with the disease classification model. The preprocessing unit is configured to obtain at least one gene pair feature of the target sample according to the transcriptome data of the target sample. The prediction unit is configured to perform classification processing on the at least one gene pair feature of the target sample according to the disease classification model to obtain the category of the target sample. When the preprocessing unit is used to obtain the at least one gene pair feature of the target sample according to the transcriptome data of the target sample, it is specifically configured to perform the following operations: first, obtain at least one gene data of the target sample according to the transcriptome data of the target sample, the gene data being a gene type, and determine the gene expression amount of each gene data in the target sample according to the at least one gene data and the transcriptome data of the target sample, and then construct a gene expression matrix of the target sample by taking the gene data as the row, taking the target sample as the column, and taking the gene expression amount of each gene data in the target sample as the value of each element. Then, the gene expression matrix of the target sample is obtained according to the gene expression matrix of the target sample. The specific implementation of converting the gene expression matrix into the gene pair matrix is the same as the step S330 described above, and will not be described here. In the gene pair matrix of the target sample, the row represents the gene pair feature of the target sample, the column represents the target sample, and the value of each element is the numerical value of each gene pair feature in the target sample. The representation form of the gene pair matrix of the target sample is the same as the formula (2) described above, and will not be described here. Finally, at least one gene pair feature of the target sample can be extracted from the gene pair matrix of the target sample.

[0159] The contents in the method embodiments are applicable to the system embodiments, the system embodiments specifically implement the functions same as the method embodiments, and achieve the same beneficial effects as the method embodiments.

[0160] The training method and device of the disease classification model and the disease classification system in the embodiments of the present application are described below in terms of a specific application scenario. In this application scenario, the disease classification model is used to identify the category of a skin cancer sample to be tested.

[0161] T01, obtain high variable genes or related genes of skin cancer as skin cancer gene data through a preset database, and construct a transcriptome dataset of skin cancer. Specifically, first, obtain initial transcriptome data and categories of a plurality of skin cancer samples. The initial transcriptome data of the skin cancer samples includes transcriptome sequencing expression matrices, single-cell transcriptome sequencing expression matrices, and original microarray data of the skin cancer samples, and the categories of the skin cancer samples include melanoma, basal cell carcinoma, Merkel cell carcinoma, squamous cell carcinoma, and non-skin cancer (i.e., healthy). Then, for the initial transcriptome data of each skin cancer sample, standardize the initial microarray data using a robust multi-array average method to obtain microarray data, standardize the transcriptome sequencing expression matrices using a million transcript method to obtain transcriptome sequencing data, and standardize the single-cell transcriptome sequencing expression matrices using the million transcript method to obtain single-cell transcriptome sequencing data. Finally, construct a transcriptome dataset of skin cancer by using the transcriptome data and categories of each skin cancer sample. The transcriptome data of the skin cancer samples includes microarray data, transcriptome sequencing data, and single-cell transcriptome sequencing data of the skin cancer samples.

[0162] T02, divide the transcriptome dataset of skin cancer into a discovery set, a training set, and a test set according to a preset division ratio. Specifically, the discovery set can include transcriptome data and categories of a plurality of discovery samples. The discovery samples refer to initial samples divided into the discovery set, and the transcriptome data of the discovery samples can include any one of single-cell transcriptome sequencing data of the discovery samples or batch transcriptome data of the discovery samples, and the batch transcriptome data of the discovery samples can include microarray data of the discovery samples and transcriptome sequencing data of the discovery samples. The training set can include transcriptome data and categories of a plurality of training samples. The training samples refer to initial samples divided into the training set, and the transcriptome data of the training samples can include microarray data of the training samples and transcriptome sequencing data of the training samples. In addition, the test set can include transcriptome data and categories of a plurality of test samples. The test samples refer to initial samples divided into the test set, and the transcriptome data of the test samples can include at least one of microarray data of the test samples, transcriptome sequencing data of the test samples, or single-cell transcriptome sequencing data of the test samples.

[0163] T03, constructing a gene pair matrix of the discovery set. Specifically, first, genes intersected from the discovery set, the training set and the skin cancer gene data are taken as target gene data. Then, according to the multiple target gene data and the transcriptome data of each discovery sample in the discovery set, a gene expression matrix of the discovery set is obtained, which contains each discovery sample and its corresponding multiple target gene data, as shown in part (b) of FIG. 6. Figure 2 Wherein, the gene expression matrix of the discovery set is a two-order matrix, in which the row represents the target gene data, the column represents the discovery sample, and the value of the element is the gene expression amount of the target gene data in the discovery sample. Then, the gene expression matrix of the discovery set is converted into a gene pair matrix of the discovery set by the above formula (3), which includes each discovery sample and its corresponding multiple discovery set gene pairs, as shown in part (c) of FIG. 6. Figure 2 Wherein, the gene pair matrix of the discovery set is a two-order matrix, in which the row represents the discovery set gene pair, the discovery set gene pair is composed of two target gene data, the column represents the discovery sample, and the value of the element is the value of the discovery set gene pair in the discovery sample. The value of the discovery set gene pair in the discovery sample is 1, -1 or 0.

[0164] T04, extracting multiple skin cancer target gene pair features. Specifically, first, for each discovery set gene pair, there is: for each category, according to the value of each element in the gene pair matrix of the discovery set, the total number of discovery set gene pairs with a value of 1 in the discovery samples belonging to the current category is obtained as the first target number of the current category relative to the discovery set gene pair, and the total number of discovery set gene pairs with a value of -1 in the discovery samples belonging to the current category is obtained as the second target number of the current category relative to the discovery set gene pair; by traversing the categories of all discovery samples, the first target number and the second target number of all categories relative to the discovery set gene pair can be obtained as the category number feature of the discovery set gene pair. And by traversing all discovery set gene pairs, the category number features of all discovery set gene pairs can be obtained. Then, the significance level value of each discovery set gene pair is calculated according to the category number features of each discovery set gene pair combined with the above formula (5) or the above formula (6). After that, all discovery set gene pairs are sorted from small to large according to the size of the significance level value, and the top 50 or 100 discovery set gene pairs are selected as the target gene pair feature.

[0165] T05, updating the training set and the test set. Specifically, for the training set, first, a gene expression matrix of the training set is constructed according to the transcriptome data of each training sample in the training set, and then the gene expression matrix of the training set is converted into a gene pair matrix of the training set by the above formula (3). The gene pair matrix of the training set is a second-order matrix, in which the row represents a training set gene pair, the column represents a training sample, and the value of the element is the value of the training set gene pair in the training sample, which is 1, -1 or 0. Then, for each training set gene pair in the gene pair matrix of the training set, if the current training set gene pair exists in the set composed of the plurality of target gene pair features, the current training set gene pair is retained, otherwise the current training set gene pair is excluded. By traversing the plurality of target gene pair features and the plurality of training set gene pairs in the gene pair matrix of the training set, the screened gene pair matrix of the training set can be obtained. Finally, the plurality of training set gene pairs corresponding to each training sample are extracted from the screened gene pair matrix of the training set as the plurality of gene pair features corresponding to each training sample. For the test set, the processing steps are the same as the above processing steps for the training set, and will not be described here.

[0166] T06, training and testing of the disease classification model. Specifically, first, an initial machine learning model corresponding to each machine learning method is constructed. Then, for each initial machine learning model, the training set is input into the current initial machine learning model to train the current initial machine learning model and obtain a machine learning model. Then, it is determined whether the type of the discovery set is the same as that of the training set; if yes, the test set is used to test each machine learning model to obtain the test results of each machine learning model, and the machine learning model with the best test result is selected as the disease classification model output; if not, for each machine learning model, the feature importance of each sample feature in the processed training set is obtained by the machine learning model, the sample features with zero feature importance in the processed training set are excluded, a new training set is obtained, and model training is performed again on the new training set to improve the training accuracy. In addition, a three-fold cross-validation method can be introduced to realize model training and testing.

[0167] T07, real-time classification of skin cancer. Specifically, first, the transcriptome data of the to-be-tested sample of skin cancer is obtained through a preset database. Then, the to-be-tested gene pair features of the to-be-tested sample are obtained according to the transcriptome data of the to-be-tested sample. Finally, the to-be-tested gene pair features are input into the disease classification model, and the category of the to-be-tested sample is obtained by the disease classification model.

[0168] The technical effects provided by the embodiments of the present application are verified by two examples below.

[0169] Example 1: Skin cancer classification task. In the skin cancer classification task, the classes of skin cancer are melanoma (MEL), basal cell carcinoma (BCC), Merkel cell carcinoma (MCC), squamous cell carcinoma (SCC) and healthy (Healthy), the transcriptome profiles of 330 samples of melanoma class, 44 samples of basal cell carcinoma class, 68 samples of Merkel cell carcinoma class, 88 samples of squamous cell carcinoma class and 320 samples of healthy class are collected by DNA microarray technology, a total of 850 samples and their transcriptome profiles, to construct a skin cancer dataset, as shown in Table 1 below. Among them, the samples of the healthy class include samples of the healthy skin class and samples of the healthy nevus class. The samples of the melanoma class include samples of the primary melanoma class and samples of the metastatic melanoma class. The gene chips applied to the DNA microarray include Affymetrix and Illumina.

[0170] Table 1: Skin cancer dataset

[0171]

[0172]

[0173] Three classification tasks are defined, including a binary classification task (i.e. 2class, identifying whether a sample is a sample of the healthy class or a sample of the non-healthy class), a ternary classification task (i.e. 3class, identifying whether a sample is a sample of the healthy class, a sample of the melanoma class (MEL) and a sample of the non-healthy class and the non-melanoma class (Non-MEL)) and a quinary classification task (i.e. 5class, identifying which one of the healthy, BCC, SCC, MCC and MEL a sample belongs to). For each classification task, the present embodiment finds 30 pairs of target gene pair features (AIGPS-30) from the skin cancer dataset and trains and tests the model to obtain skin cancer classification models for each classification task. The skin classification models corresponding to the three classification tasks are compared with the existing benchmark method (galvez degs), respectively, and the evaluation indicators of the model include F1-weighted, F1-macro, precision-weighted, precision-macro, Recall-weighted, Recall-macro and acc. The results of the comparison are as follows Figure 3As shown, in the binary classification task and the three-class classification task, all the indexes of the skin cancer classification model of the present application are higher than those of the existing benchmark method, and in the five-class classification task, the macro precision index of the skin cancer classification model of the present application is slightly lower than that of the existing benchmark method, and the indexes other than the macro precision are higher than those of the existing benchmark method. Thus, it can be seen that, compared with the existing benchmark method, the embodiment of the present application can significantly improve the performance of the skin cancer classification model.

[0174] Example 2: Melanoma immunotherapy response prediction task

[0175] In the melanoma immunotherapy response prediction task, a total of 252 skin cancer samples were extracted from the melanoma single-cell transcriptome dataset GSE12057552 (which includes 5928 pre-processed cells) and 6 RNA-seq datasets of melanoma Riaz, Van, Hugo, Gide, Lee and MGH, to construct a melanoma dataset, as shown in Table 2 below. The melanoma single-cell cohort can be found in the Gene Expression Omnibus (GEO) database, with sequence ID GSE120575. The Riaz dataset can be found in the GEO database, with sequence ID GSE91061. The Hugo dataset can be found in the GEO database, with sequence ID GSE78220. The MGH dataset can be found in the GEO database, with sequence IDs GSE115821 and GSE168204. The Gide cohort can be found in the BioProject database, with accession number PRJEB23709. The Lee dataset can be found in the European Genome-phenome Archive (EGA), with dataset accession number EGAD00001005738. The Van dataset can be found in The database of Genotypes and Phenotypes (dbGaP), with accession number phs000452.v2.p1.

[0176] Table 2: Melanoma dataset

[0177]

[0178] The melanoma immunotherapy response prediction task is mainly a binary classification task, i.e., identifying whether a melanoma sample responds to immunotherapy. For this prediction task, the embodiment of the present application finds 27 pairs of target gene pair features (AIGPS-27), compares AIGPS-27 with existing transcriptome-based prediction markers, and the existing prediction markers include PASS_PRE, IMPRES, , T cell inflamed, MHC-I, MHC-II, Immune Score, NRS, Chemokine and CYT, the results of comparison are shown in Hugo, Lee, Gide, MGH and all data sets (Overall), the average AUC values of embodiments of the present application are 0.59, 0.65, 0.70, 0.75 and 0.68 respectively, which indicates that the overall performance of AIGPS-27 is better than other markers, indicating the universality of AIGPS-27 across data sets. Among them, the AUC value refers to the area under the receiver operating characteristic curve (ROC). Figure 4

[0179] From the above, it can be seen that the embodiments of the present application can improve the training accuracy and training efficiency of the disease classification model, thereby improving the explainability and classification accuracy of the disease classification model.

[0180] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

[0181] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the described embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application, and these equivalent modifications or replacements are all included in the scope defined by the claims of the present application.​

Claims

1. A training method for a disease classification model, characterized in that, The method comprises the following steps: obtaining target disease gene data and a transcriptome dataset associated with a target disease; wherein the transcriptome dataset comprises transcriptome data and a category of a plurality of initial samples; performing classification processing on the transcriptome dataset to obtain a discovery set, a training set and a test set; obtaining a plurality of target gene pair features associated with the target disease according to the discovery set and the target disease gene data; processing the training set and the test set according to a plurality of target gene pair features to obtain a processed training set and a processed test set; performing model training and testing by using a machine learning method according to the processed training set and the processed test set to obtain a disease classification model, wherein the disease classification model is used to identify the category of a sample associated with the target disease; wherein the obtaining a plurality of target gene pair features associated with the target disease according to the discovery set and the target disease gene data comprises: obtaining a plurality of target gene data associated with the target disease according to the transcriptome data of each discovery sample in the discovery set, the transcriptome data of each training sample in the training set and the target disease gene data; obtaining a gene expression matrix of the discovery set according to a plurality of target gene data and the transcriptome data of each discovery sample in the discovery set; wherein the gene expression matrix of the discovery set is a second-order matrix, in the gene expression matrix of the discovery set, the rows represent the target gene data, the columns represent the discovery samples, and the values of the elements are the gene expression amounts of the target gene data in the discovery samples; obtaining a gene pair matrix of the discovery set according to the gene expression matrix of the discovery set; wherein the gene pair matrix of the discovery set is a second-order matrix, in the gene pair matrix of the discovery set, the rows represent the discovery set gene pairs, the columns represent the discovery samples, and the values of the elements are the numerical values of the discovery set gene pairs in the discovery samples; the discovery set gene pair comprises two target gene data, and the numerical value of the discovery set gene pair in the discovery sample comprises any one of a first preset value, a second preset value or a third preset value; performing screening processing on each discovery set gene pair in the gene pair matrix of the discovery set to obtain a plurality of target gene pair features associated with the target disease; wherein the obtaining the gene pair matrix of the discovery set according to the gene expression matrix of the discovery set comprises: comparing each two target gene data in each discovery sample exhaustively according to the gene expression matrix of the discovery set, introducing an adaptive difference of Welch's test to capture the difference changes between different genes, and adaptively quantifying the difference changes between different genes, thereby obtaining a plurality of discovery set gene pairs and the numerical values of each discovery set gene pair in each discovery sample. 2.The method of claim 1, wherein, The obtaining target disease gene data and a transcriptome dataset associated with a target disease comprises: obtaining a gene type associated with the target disease as the target disease gene data; obtaining initial transcriptome data and categories of a plurality of initial samples associated with the target disease, the initial transcriptome data of the initial samples comprising at least one of a transcriptome sequencing expression matrix of the initial samples, a single-cell transcriptome sequencing expression matrix of the initial samples, or initial microarray data of the initial samples; standardizing the initial transcriptome data of each of the initial samples to obtain transcriptome data of each of the initial samples, wherein the transcriptome data of the initial samples comprises at least one of microarray data of the initial samples, transcriptome sequencing data of the initial samples, or single-cell transcriptome sequencing data of the initial samples; obtaining a transcriptome dataset associated with the target disease according to the transcriptome data and categories of each of the initial samples. 3.The method of claim 1, wherein, The gene pair matrix of the discovery set is obtained according to the gene expression matrix of the discovery set, comprising: determining the number of combinations of two target gene data taken from a plurality of target gene data as the total number of discovery set gene pairs; determining a plurality of discovery set gene pairs according to the total number of discovery set gene pairs and the gene expression matrix of the discovery set; obtaining the variance of the gene expression amount of each target gene data in all discovery samples as the expression amount variance of each target gene data according to the gene expression matrix of the discovery set; obtaining the numerical value of each discovery set gene pair in each discovery sample according to the gene expression matrix of the discovery set, the expression amount variance of each target gene data, and a preset adaptive coefficient; constructing the gene pair matrix of the discovery set by taking the discovery set gene pairs as the rows of the gene pair matrix of the discovery set, taking the discovery samples as the columns of the gene pair matrix of the discovery set, and taking the numerical value of each discovery set gene pair in each discovery sample as the value of each element in the gene pair matrix of the discovery set. 4.The method of claim 1, wherein, The discovery set gene pairs in the gene pair matrix of the discovery set are screened to obtain a plurality of target gene pair features associated with the target disease, comprising: for each discovery set gene pair, obtaining the first target number and the second target number of each category relative to the discovery set gene pair as the category number feature of the discovery set gene pair according to the numerical value of the discovery set gene pair in each discovery sample and the categories of all discovery samples; wherein the first target number of the category relative to the discovery set gene pair is the total number of the discovery set gene pairs with the first preset value in the discovery samples belonging to the category, and the second target number of the category relative to the discovery set gene pair is the total number of the discovery set gene pairs with the second preset value in the discovery samples belonging to the category; obtaining the significance level value of each discovery set gene pair according to the category number feature of each discovery set gene pair and the total number of categories of the discovery samples; Screening each gene pair in the gene pair matrix of the discovery set according to the significance level value of each gene pair in the gene pair matrix of the discovery set, to obtain a plurality of target gene pair features associated with the target disease. 5.The method of claim 1, wherein, The processing of the training set and the test set according to the plurality of target gene pair features comprises: According to the transcriptome data of each training sample in the training set, obtain the gene pair matrix of the training set; wherein the gene pair matrix of the training set is a two-order matrix, in the gene pair matrix of the training set, the row represents a training set gene pair, the column represents the training sample, and the value of the element is the numerical value of the training set gene pair in the training sample, the numerical value of the training set gene pair in the training sample includes any one of the first preset value, the second preset value or the third preset value; According to the plurality of target gene pair features and the gene pair matrix of the training set, obtain a plurality of gene pair features corresponding to each training sample; According to the plurality of gene pair features corresponding to each training sample and the category, obtain the processed training set. 6.The method of claim 1, wherein, The processing of the training set and the test set according to the plurality of target gene pair features comprises: According to the transcriptome data of each test sample in the test set, obtain the gene pair matrix of the test set; wherein the gene pair matrix of the test set is a two-order matrix, in the gene pair matrix of the test set, the row represents a test set gene pair, the column represents the test sample, and the value of the element is the numerical value of the test set gene pair in the test sample, the numerical value of the test set gene pair in the test sample includes any one of the first preset value, the second preset value or the third preset value; According to the plurality of target gene pair features and the gene pair matrix of the test set, obtain a plurality of gene pair features corresponding to each test sample; According to the plurality of gene pair features corresponding to each test sample and the category, obtain the processed test set. 7.The method of claim 1, wherein, The model training and testing according to the processed training set and the processed test set by using the method of machine learning, to obtain a disease classification model, comprises: Model training on the processed training set by using the method of machine learning, to obtain each machine learning model; Testing each machine learning model by using the processed test set, to obtain the test result of each machine learning model; According to the test result of each machine learning model, obtaining the disease classification model. 8.A device for training a disease classification model, comprising: Comprise: The first acquisition module is used for acquiring target disease gene data and a transcriptome data set associated with a target disease; wherein the transcriptome data set comprises transcriptome data and categories of a plurality of initial samples; The first processing module is used for classifying the transcriptome data set to obtain a discovery set, a training set and a test set; The second processing module is used for obtaining a plurality of target gene pair features associated with the target disease according to the discovery set and the target disease gene data. a third processing module, configured to process the training set and the test set according to a plurality of target gene pair features, to obtain a processed training set and a processed test set; a fourth processing module, configured to train and test a model according to the processed training set and the processed test set by using a machine learning method, to obtain a disease classification model, the disease classification model being used to identify a category of a sample associated with the target disease; wherein the target gene pair features associated with the target disease are obtained according to the discovery set and the target disease gene data, including: target gene data associated with the target disease is obtained according to the transcriptome data of each discovery sample in the discovery set, the transcriptome data of each training sample in the training set, and the target disease gene data; a gene expression matrix of the discovery set is obtained according to the target gene data and the transcriptome data of each discovery sample in the discovery set, wherein the gene expression matrix of the discovery set is a second-order matrix, in the gene expression matrix of the discovery set, rows represent the target gene data, columns represent the discovery sample, and the value of an element is the gene expression amount of the target gene data in the discovery sample; a gene pair matrix of the discovery set is obtained according to the gene expression matrix of the discovery set, wherein the gene pair matrix of the discovery set is a second-order matrix, in the gene pair matrix of the discovery set, rows represent a discovery set gene pair, columns represent the discovery sample, and the value of an element is a numerical value of the discovery set gene pair in the discovery sample; the discovery set gene pair includes two target gene data, and the numerical value of the discovery set gene pair in the discovery sample includes any one of a first preset value, a second preset value, or a third preset value; each discovery set gene pair in the gene pair matrix of the discovery set is filtered to obtain a plurality of target gene pair features associated with the target disease; wherein the gene pair matrix of the discovery set is obtained according to the gene expression matrix of the discovery set, including: each two target gene data in each discovery sample is exhaustively compared according to the gene expression matrix of the discovery set, an adaptive difference of Welch's test is introduced to capture the difference between different genes, and the difference between different genes is adaptively quantified, so as to obtain a plurality of discovery set gene pairs and the numerical value of each discovery set gene pair in each discovery sample.

9. A disease classification system, characterized by, including: a second acquisition module, configured to acquire transcriptome data of a to-be-tested sample associated with a target disease; a fifth processing module, configured to perform classification processing on the transcriptome data of the to-be-tested sample according to the disease classification model, to obtain a category of the to-be-tested sample; wherein the disease classification model is trained by using the disease classification model training method in any one of claims 1-7.

Images