Cancer typing method and device using gene characteristics

Abstract

According to the cancer typing method based on the gene characteristics, key gene characteristics in biological information data are efficiently extracted and utilized, and cancer typing challenges caused by cancer heterogeneity are overcome. Traditional methods often face the problem of insufficient classification accuracy when treating high heterogeneity of cancer samples. According to the method, a special gene feature loss function is designed, gene features are ensured to be fully utilized, and information loss is reduced. Meanwhile, the dependency relationship between the gene features is calculated by utilizing an attention mechanism, so that the model can automatically pay attention to important gene features and capture complex interaction of the important gene features, and the typing accuracy is improved. The method provides a new tool for biologists to research cancer gene characteristics, provides support for clinical diagnosis, and helps doctors to identify cancer types and formulate therapeutic schedules more accurately. The invention opens up a new path for cancer research and clinical practice, and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the field of AI+bioinformatics, and in particular to a cancer typing method using gene features. Background Art

[0002] In recent years, due to the development of deep learning technology, people have made significant progress in the fields of image and video classification, natural language processing, and bioinformatics processing. In the field of bioinformatics data analysis, deep learning models have demonstrated their powerful ability to achieve high-precision classification of complex data types. However, due to the heterogeneity of cancer data, cancer typing has become more difficult, so we designed a cancer typing model that uses genetic features to accurately and stably type cancer.

[0003] Traditional bioinformatics analysis methods often use some statistical methods to perform statistical analysis on bioinformatics data, thereby mining some statistical difference information in the data for differential analysis of various diseases. With the development of biological sequencing methods and the emergence of single-cell sequencing methods, its data characteristics show high dimensionality, sparsity, high noise, and large data volume. Therefore, traditional statistical methods cannot quickly and accurately mine biologically meaningful information in large and noisy data. More and more bioinformatics experts have begun to study the use of deep learning to achieve bioinformatics analysis. However, the designed network does not take into account the extraction and fusion of gene features, so such models are easily affected by cancer heterogeneity, resulting in low classification accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide biological scientists with a new cancer typing method. This method first pre-processes the single-cell sequencing data, and then uses clustering to classify genes with similar functions into different clusters; each cluster is placed in GFE to extract gene features from the genes in each cluster, and the extracted gene features are input into SWDF AM as tokens for mutual dependency calculation. SWGFAM calculates the dependencies between tokens, so genes in different clusters have different dependencies. We call this result a gene attention graph. Using the gene attention graph, we can calculate the probability of cancer typing, that is, the classification result. In order to better utilize gene features, we designed a gene loss function to improve the accuracy and stability of network classification. Use AI capabilities to help clinical diagnosis of cancer types and achieve more accurate diagnostic schemes. The technical solution is as follows:

[0005] A cancer classification method using genomic features, comprising the following steps:

[0006] (1) Data preprocessing: The bioinformatics data format is converted into a C×M×N bioinformatics matrix, where C is the dimension of the data matrix, M is the width of the matrix, and N is the length of the matrix. This bioinformatics matrix contains gene information and cell information;

[0007] (2) Gene clustering module: Genes are clustered using the biological information matrix to obtain n clusters representing genomes with different gene functions, denoted as Cg i ,i∈1,2,...,n;

[0008] (3) Feature extractor: Gene feature extraction is performed on n clusters of genomes to obtain gene features representing n different types. This gene feature contains the gene information of the cell and is used to implement cancer classification information in the network. The method is: convolution is performed on n gene clusters using a one-dimensional convolution kernel, and the genes of each gene cluster are converted into a gene feature matrix, recorded as The gene feature matrix is ​​used as the token of the gene attention module and is also used to calculate the gene loss function;

[0009] (4) Gene attention module: In order to capture the dependencies between the gene features of each cluster, As a token input gene attention module, the self-attention mechanism is used to calculate the dependencies between tokens, and the dependencies of each cluster are integrated to obtain the attention graph A. g Used for subsequent calculation of classification probability;

[0010] (5) Classification module: Attention map A g After passing through an MLP layer, it is sent to softmax for calculation to obtain the probability P of each sample being identified as each type of cancer. g , and use probability to get the classification result.

[0011] Furthermore, the calculation formula of the gene feature loss function is as follows:

[0012]

[0013] in is the gene feature matrix of each cluster in training, is the gene feature of each cluster of groundtruth, α i is the weight of each gene feature.

[0014] The beneficial effects of the present invention are:

[0015] The present invention proposes a cancer typing method using gene features, which can realize the typing of various cancer types in bioinformatics data. It better solves the problem of inaccurate classification due to heterogeneity of data of various cancer types. Through gene clustering, various gene features are extracted, and gene loss functions are constructed to achieve high-accuracy classification of various cancer types. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the overall algorithm framework diagram of the present invention;

[0017] Figure 2 It is a cancer typing network framework using gene features of the present invention;

[0018] Figure 3 This is the gene clustering result diagram DETAILED DESCRIPTION

[0019] In order to make the technical solution of the present invention clearer, the present invention is further described below in conjunction with the accompanying drawings.

[0020] The present invention proposes a cancer typing method using gene feature information, and designs a gene feature extraction module and a gene loss function. First, the single-cell sequencing data is preprocessed, and the single-cell sequencing data is converted into a behavior gene, which is listed as a cell matrix of cells. Each cell matrix contains a cell group of a cancer type and thousands of genes. This method first proposes a gene clustering module. In order to form a genome with genes with the same function, we first cluster the genes. Then, according to the clustering results, the genes of each cluster are respectively placed in a gene feature extractor (gene feature extractor, GFE), and the GFE is used to extract gene features from each genome, and each gene feature is input as a token into the subsequent attention module. At the same time, a sliding window gene feature attention module (sliding window gene feature attention module, SWGFAM) is designed, and the relationship between each gene feature is calculated using SWGFAM, so as to calculate the correlation of each genome gene feature. The gene attention map is calculated, and the attention map is sent to the softmax module to calculate the probability value of the cancer type, which is used as the final cancer typing result. This method can classify data of different cancer types in single-cell sequencing data, providing a new perspective for clinical diagnosis of cancer types in patients. The present invention is specifically implemented in the following steps:

[0021] Step 1: Classification module;

[0022] (1) Data preprocessing:

[0023] In order to adapt the classification model to the bioinformatics data of cancer classification, we need to convert the data format into a C×M×N 3D matrix I C×M×N As the input of the classification module, where C is the dimension of the data matrix (usually 1 dimension), M is the width of the data matrix, and N is the length of the data matrix.

[0024] (2) Gene clustering:

[0025] Since different genes are responsible for different functions, we cluster the genes to obtain some gene clusters with the same function. These gene clusters are used for subsequent gene feature extraction. Here we use Kmeans clustering for feature clustering. The calculation formula is as follows:

[0026] Cg=Kmeans(I C×M×N ) (1)

[0027] Where Cg is the cluster information, which can be written as Cg i ,i∈1,2,...,n.

[0028] (3) Gene feature extractor:

[0029] In order to convert the data of each cluster into a higher-level genetic feature, we use the feature extraction capability of the convolution kernel to convert each cluster of data into a type of genetic feature. The result obtained after convolution is used as the input of the subsequent gene attention module, namely token, and the calculation formula is as follows:

[0030]

[0031] (4) Attention module:

[0032] In order to calculate the dependencies between gene features, we As the input of this module, gene features Composed of multiple tokens, the dependency structure between tokens is calculated using the window self-attention mechanism, as shown in the following formula:

[0033]

[0034] WMSA is a window attention module used to calculate the local relationship between tokens. We use gene features First, it is processed by the linear layer LN, then put into the window attention module to calculate the local dependency between tokens, and finally the window feature A is obtained by the linear layer LN and multi-layer perception processing. w .

[0035]

[0036] Among them, SWMSA is a moving window attention module used to calculate the global relationship between tokens. We first process the feature F through the linear layer LN, then put it into the moving window attention module to calculate the global dependency between tokens, and finally obtain the window feature A through the linear layer LN and multi-layer perception processing. g .

[0037] (5) Classification module:

[0038] The output of the attention module A g Send it to softmax for calculation to get the probability P of each sample belonging to each category g The calculation formula for the classification result is as follows:

[0039] P g =softmax(A g )(7)

[0040] (6) Gene feature loss function

[0041] In order to better utilize the gene features extracted in this paper, we designed a gene feature loss function, the calculation formula is as follows:

[0042]

[0043] in is the gene feature of each cluster in training, is the gene feature of each cluster of groundtruth, α i is the weight of each gene feature, which can be dynamically updated to achieve the best training effect.

Claims

1. A cancer typing method using genomic features, comprising the following steps: (1) Data preprocessing: The bioinformatics data format is converted into a C×M×N bioinformatics matrix, where C is the dimension of the data matrix, M is the width of the matrix, and N is the length of the matrix. This bioinformatics matrix contains gene information and cell information; (2) Gene clustering module: Genes are clustered using the biological information matrix to obtain n clusters representing genomes with different gene functions, denoted as Cg i ,i∈1,2,...,n; (3) Feature extractor: Gene feature extraction is performed on n clusters of genomes to obtain gene features representing n different types. This gene feature contains the gene information of the cell and is used in the network to implement cancer classification information. The method is as follows: The n gene clusters are convolved using a one-dimensional convolution kernel, and the genes of each gene cluster are converted into a gene feature matrix, recorded as The gene feature matrix is ​​used as the token of the gene attention module and is also used to calculate the gene loss function; (4) Gene attention module: In order to capture the dependencies between the gene features of each cluster, As a token input gene attention module, the self-attention mechanism is used to calculate the dependencies between tokens, and the dependencies of each cluster are integrated to obtain the attention graph A. g Used for subsequent calculation of classification probability; (5) Classification module: Attention map A g After passing through an MLP layer, it is sent to softmax for calculation to obtain the probability P of each sample being identified as each type of cancer. g , and use probability to get the classification result.

2. The method for cancer typing using genomic features according to claim 1, characterized in that: The calculation formula of the gene feature loss function is as follows: in is the gene feature matrix of each cluster in training, is the gene feature of each cluster of ground truth, α i is the weight of each gene feature.

3. A cancer typing device, characterized in that: The cancer typing device comprises: a processor and a memory, wherein program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method according to any one of claims 1 and 2.

Images