Molecular subtype method for identifying triple negative breast cancer by selecting optimal gene subset
By integrating multiple data and algorithms, combining gene importance scores and neuronal weight changes, and using genetic algorithms to optimize gene combinations, the accuracy and efficiency of TNBC subtype classification are solved, and high-precision subtype identification and treatment are achieved choose.
Patent Information
- Application Number
- CN202410927546.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing classification methods for triple-negative breast cancer (TNBC) molecular subtypes lack uniform standards, resulting in limited treatment choices. The existing feature selection methods have problems with computational complexity and inefficiency, making it difficult to achieve accurate subtype identification.
By integrating mRNA and lncRNA data, combining GeneRank scores with the gene importance score calculated by the automatic encoder, the topological overlap matrix and the automatic encoder are used to consider the interaction between genes and changes in neuron weights, reducing the number of genes and using genetic algorithms to simulate organisms In the evolution process, find the best gene combination to achieve more accurate TNBC molecular typing identification.
The accurate identification of TNBC subtypes was achieved, achieving an accuracy rate of 95.61%, improving the accuracy and efficiency of treatment choices, and making up for the shortcomings of the existing technology.
Smart Images

Figure CN118918942A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for identifying molecular subtypes of triple-negative breast cancer, and more particularly to a method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer. The present invention investigates feature selection methods that have been effective in TNBC subtype classification in the past two years, integrates mRNA and lncRNA data, combines GeneRank scores with gene importance scores calculated using an autoencoder, and uses a topological overlap matrix (TOM) and an autoencoder to consider interactions between genes and changes in neuronal weights, reduce the number of genes, and remove redundant genes. Finally, a genetic algorithm is used to simulate the process of biological evolution, change gene combinations to find the best gene combination, and thereby more accurately discriminate the molecular typing of TNBC. Background Art
[0002] 1. Introduction to triple-negative breast cancer
[0003] Triple-negative breast cancer (TNBC) is a subtype of breast cancer in which estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) are all negative, which means that this type of breast cancer will not respond significantly to conventional hormone therapy or targeted HER2 therapy, and treatment options are relatively limited. Based on current research results, TNBC is divided into six molecular subtypes, and each subtype requires targeted treatment. Therefore, accurate and stable subtype classification of TNBC is the key to treating TNBC.
[0004] Compared with other types of breast cancer, triple-negative breast cancer has the characteristics of high malignancy and poor prognosis. Since positive ER and PR expression are reference indicators for endocrine therapy, triple-negative breast cancer cannot be treated with endocrine therapy. In addition, since anti-HER-2 targeted therapy can only be performed when HER-2 expression is positive, patients with triple-negative breast cancer cannot undergo anti-HER-2 targeted therapy. Therefore, apart from surgical treatment, chemotherapy is still the main treatment for triple-negative breast cancer. In recent years, with the continuous progress of the concept of classification and treatment and research, the treatment methods and strategies for different types of triple-negative breast cancer have also been steadily improved and optimized. However, as far as the current research results are concerned, due to the progress of molecular biology, biomarkers have been continuously discovered and used as the basis for dividing TNBC subtypes. There is still no unified standard in the industry for the molecular typing of TNBC, so there are many different typing methods.
[0005] In different subtypes of cells, gene expression and clinical pathological characteristics vary, so multi-gene analysis is needed to determine specific targets to provide support for personalized clinical treatment. With the rapid development of genome sequencing technology, more and more researchers in the biomedical field have begun to pay attention to and apply this technology to the study of complex diseases such as cancer. For example, transcriptome sequencing data (RNA-Seq) can be used to analyze genes with significant expression differences in cancer samples. According to the characteristics of whether RNA encodes proteins, RNA in organisms can be divided into two categories: coding RNA and non-coding RNA. Coding RNA is mRNA, which guides protein synthesis and conveys genetic information. Although non-coding RNA does not guide protein synthesis, it plays an important role in biological processes such as cell differentiation and transcription. Initially, ncRNA was regarded as "dark matter" or "transcription noise", but more and more studies have shown that it plays a key role in tumorigenesis and development. ncRNA is involved in the regulation of various biological processes, such as chromatin structure, epigenetic memory, transcription, RNA splicing, translation, etc. The importance of long non-coding RNA (lncRNA) in tumorigenesis has been confirmed and has attracted the attention of many researchers.
[0006] 2. Current status of TNBC research
[0007] 2-1. Classification of TNBC subtypes
[0008] TNBC is divided into six molecular subtypes: basal-like 1 (BL1), basal-like 2 (BL2), mesenchymal (M), immunomodulatory (IM), lumial androgen receptor (LAR), and mesenchymal stem-like (MSL). Subsequently, the classification of TNBC subtypes has undergone a transition from gene expression profiling based on mRNA levels, DNA analysis to multi-omics integration. In recent years, long noncoding RNAs (lncRNAs) have been found to be involved in a variety of signaling pathways, such as chromatin organization and transcriptional regulation. More and more studies have shown that lncRNAs are not closely related to the diagnosis and treatment of cancer. Although lncRNAs are considered as potential therapeutic targets for the treatment of breast cancer due to their oncogenicity, the identification of lncRNA markers for breast cancer is still a difficult problem to overcome in this field, and its application in the field of comprehensive treatment still needs further exploration.
[0009] 2-2. Detection method
[0010] Feature selection is currently a very effective detection method. Its selection depends on the specific data set and task requirements. The appropriate feature selection method can improve the performance and generalization ability of the model. In biological research, it is necessary to use feature selection methods to reduce the dimensionality of genetic data. Feature selection methods can usually be divided into three types according to the basis and method of selection:
[0011] Filter Method: In the feature selection process, the features are first evaluated and ranked, and then the most relevant feature subset is selected. The filter method has a good computational complexity for high-dimensional small samples and has very good versatility. It can be applied to data sets in any field. However, it is precisely because the filter method does not consider the role of the classification model in data filtering that it is worse than other methods in terms of feature selection. Filter feature selection is independent of the specific learning algorithm, so it can be performed independently before feature selection.
[0012] Wrapper feature selection: In the feature selection process, a specific learning algorithm is used to evaluate different feature subsets to select the best feature combination. This method usually cross-validates different feature subsets, so when the amount of data is large or the data dimension is high, its computational cost will be much higher than other feature selection methods, but the selected feature subsets are more accurate.
[0013] Embedded feature selection: In the feature selection process, the feature selection process is combined with the model training process to select the best feature subset by learning the characteristics of the algorithm itself. Its disadvantages are similar to those of the packaging method, and it will be inefficient when the scale of the data set and the feature dimension are large.
[0014] Since the above three methods all have various defects, a suitable feature selection method can improve the performance and generalization ability of the model. Considering that using any feature selection algorithm alone has limitations and ignores the correlation between features, it is a very meaningful task to explore new feature selection methods and organically combine existing feature selection methods.
[0015] 3. Application of artificial intelligence in breast cancer treatment
[0016] With the widespread application of artificial intelligence technology in various fields, the medical field has also begun to use various efficient AI computing tools to assist diagnosis and treatment to reduce the morbidity and mortality of the disease. Machine Learning (ML) is an important branch of artificial intelligence. Its application in the medical field, especially in breast cancer diagnosis, can improve diagnostic efficiency. ML algorithms have the ability to enable computers to learn autonomously and adapt to new problems without writing new programs, thereby improving the efficiency and accuracy of classification and detection of different molecular subtypes of breast cancer. However, with the rapid development of computer science, artificial intelligence has entered the era of big data. Faced with large and high-dimensional data sets, reasonable training models has become a new challenge. Redundant data is a common feature of big data, which may lead to model accuracy and efficiency not meeting expectations, or even losing the original effect. Therefore, in gene expression profiling, feature dimensionality reduction becomes particularly important. At present, the field of statistical learning usually uses two methods to reduce dimensionality: feature selection and feature extraction. Feature extraction creates a new feature set with lower dimensionality through function mapping, which changes the nature of the original data; while feature selection selects important features from the original feature set to form a subset without changing the nature of the data. In biological research, feature selection is often used to retain data features of specific markers for biological interpretation. Summary of the invention
[0017] The purpose of the present invention is to make up for the defects of the existing technology and provide a method for selecting an optimal gene subset to identify the molecular subtypes of triple-negative breast cancer. The present invention changes the gene combination involved in classification detection by means of chromosome crossover, mutation, etc., and selects the best performing gene combination after multiple rounds of evolutionary iterations to accurately identify the molecular subtypes of triple-negative breast cancer.
[0018] The technical solution adopted to achieve the purpose of the present invention is:
[0019] The optimal gene subset selection method for identifying the molecular subtypes of triple-negative breast cancer includes the following steps:
[0020] Step 1, build the dataset:
[0021] Step 1-1, after obtaining the data from the TCGA official website, complete the screening of TNBC cases in the data and add labels to the data according to the six subtype classifications proposed by Lehmann et al.;
[0022] Step 1-2, screening samples with the same number value from breast cancer RNA-Seq expression data to reconstruct the expression matrix;
[0023] Steps 1-3: Use edgeR to perform differential gene analysis between cancer tissues and adjacent tissues, and select features with significant differences for gene annotation based on the GENCODE gene set;
[0024] Steps 1-4: After completing the annotation work, genes with the same alias are screened out according to gene symbols to construct gene expression data;
[0025] In steps 1-5, the SMOTE algorithm was used to amplify the data set, so that the number of samples of the six TNBC subtypes remained basically consistent.
[0026] Step 2: construct a weighted co-expression network of genes and screen out candidate gene sets:
[0027] Step 2-1, calculate the Pearson correlation coefficient h for each pair of genes, add power operations to enhance the connectivity between genes, and convert the Pearson matrix into an adjacency matrix Q = [q ij ], the formula is as follows:
[0028]
[0029] Among them, q ij represents the adjacency value between gene i and gene j, power(h,β) represents the βth power of distance h, which can be used to adjust the influence of distance on Pearson correlation coefficient, cor(i,j) represents the correlation coefficient between variable i and variable j, and β is a hyperparameter determined by the scale-free network;
[0030] Step 2-2: Simply using the adjacency matrix Q will result in low data utilization and a high error rate in module division. Therefore, a topological overlap matrix TOM is added on the basis of the adjacency matrix to more accurately describe the correlation between genes. The formula for calculating the TOM matrix elements is as follows:
[0031]
[0032] The larger the value of the element, the stronger the correlation between the two genes. ij represents the TOM value of variables i and j, q in and q nj represents the value of the variable i and variable j positions in the TOM matrix, and n represents the total number of genes in the TOM matrix.
[0033] Step 2-3, calculate the dissimilarity between genes based on the TOM matrix to construct a hierarchical clustering tree. Determine the number of genes in each module according to the set standard, and then calculate the correlation between the characteristic gene MEs of each module and the six subtypes to eliminate the features with low correlation.
[0034] Step 3, calculate the GeneRank (gene ranking model) ranking of the gene:
[0035] Step 3-1, construct a graph Q(W,E) representing the connectivity between genes, where W represents the gene node and E represents the connectivity between genes, and calculate the final gene importance ranking t;
[0036] (I - dWD -1 )t = (1 - d)fc (1-3)
[0037] Among them, I is a unit matrix; d is a constant between 0 and 1; D is a diagonal matrix, which represents the number of edges associated with genes, that is, deg i =∑ j w ik (deg i represents the degree of node i, w ij represents the connection weight between node i and node j); fc is the fold change of the gene list, indicating the degree of difference in expression of a gene between different classes;
[0038] Step 3-2, calculate the ranking of gene g in subtype i After η iterations, the GeneRank ranking of gene importance is calculated: the formula is as follows:
[0039]
[0040] in, represents the functional node of node i when the clustering coefficient is g, b jk represents the elements in the binary matrix B constructed on the basis of the TOM matrix, indicating the connectivity between genes, represents the value of node i when the clustering coefficient is g in the η-1th round of iteration. jk When b is greater than the mean of the matrix elements, jk With b kj Set it to 1, otherwise it is set to 0. After n iterations, the GeneRank ranking of gene importance is obtained by formula (1-5):
[0041]
[0042] Where M represents the total number of subtype categories in the training set; N x=i Represents the number of samples with label i in the sample set; N total Represents the total number of samples in the sample set.
[0043] Step 4, gene importance scores calculated by AE embedded in TOM matrix:
[0044] Step 4-1: Embed a TOM matrix that can express the correlation between genes
[0045] Generally speaking, in neural networks, feature importance is related to weight, which means that the higher the weight of a neuron after network training, the greater its role in classification tasks. The CW method proposed by Olden et al. believes that the importance of each feature can be obtained by calculating the sum of the absolute values of all weights related to it. Inspired by this concept, the present invention embeds a TOM matrix that can express the correlation between genes in the autoencoder; through the TOM matrix, the output Q of the first hidden layer is obtained. [1] , the calculation formula is as follows:
[0046]
[0047] Among them, ⊙ is the dot multiplication operation, representing the matrix W [1] The multiplication operation of the elements with the same subscript in the matrix TOM; ε1 is the neural network activation function, represents the bias term of the first layer, X T Represents the transpose of matrix X. The transpose operation swaps the rows and columns of matrix X to form a new matrix, namely the transposed matrix X. T ;
[0048] Step 4-2, feature importance score calculation
[0049] The importance score of gene g is calculated by the following formula:
[0050]
[0051] Among them, x represents the number of genes in the input layer, and y represents the number of neurons in the second hidden layer; and represents the connection weight between the gth hidden node in the autoencoder and the jth input node in the input layer, and Respectively and The standard deviation of , SD represents the standard deviation of all weights connected to neurons during AE training, which is calculated by formula (8):
[0052]
[0053] Among them, t represents the number of network iterations, indicating that the greater the influence of a unit node on the prediction performance, the greater the range of change of all weights connected to it, and l represents the weight w ij The index variable used when summing the square values of , l changes from 1 to t, which means traversing all t weight data values.
[0054] Since AE is embedded in the topological overlap matrix TOM, the relative importance of each gene is affected by both the first hidden layer and the second hidden layer. Therefore, it is necessary to take the second hidden layer into account during the calculation process and calculate the standard deviation (SD) of all weights connected to neurons during AE training;
[0055] Step 5, calculate the score ranking of gene feature importance:
[0056] In step 3 and step 4 of the above technical solution, this section will comprehensively consider the two calculation results to obtain the score ranking of feature importance.
[0057] Step 5-1, standardize the calculation results of step 3 and step 4:
[0058]
[0059]
[0060] Step 5-2, construct a new scoring criterion to evaluate the contribution of each gene to the classification result. The formula is as follows:
[0061]
[0062] Among them, e is a constant term, and The value range of is between 0 and 1, and based on the comprehensive score of each gene, the top k2 genes are selected for the second feature dimensionality reduction operation. At this time, the importance score of the gene is calculated and the feature dimension of the gene becomes k2.
[0063] Step 6: Use genetic algorithm to determine the final gene candidate set:
[0064] In the present invention, for the genes with high feature importance scores obtained in step 5, a genetic algorithm is used to search for a gene subset with the best classification performance from the candidate genes.
[0065] Step 6-1: Consider each sample in the sample set as a chromosome, and each feature as a gene on the chromosome. First, randomly initialize the chromosome to a k2-dimensional binary vector D = (D1, D2, ..., D n2 ), and then generate a uniformly distributed random variable between 0 and 1 by random encoding. when When the value is greater than 0.5, the gene at the corresponding position is selected, D i =0;
[0066] Step 6-2, select random forest as the classifier model, and use the classification accuracy as the fitness function of the genetic algorithm. When the classification accuracy is higher, the fitness of the genetic algorithm is higher, and it is easier to be selected to perform genetic operations in biological genetics, and the feature importance of the selected gene is higher;
[0067] Step 6-3, perform a genetic process to select a subset of genes with better performance;
[0068] Step 6-4, after completing t rounds of iterations, select the chromosome with the largest fitness from the population for decoding, and find the genes at the corresponding positions in the expression matrix based on the non-zero elements to form the optimal subset Gene.
[0069] In the above scheme, in step 1-1, the six subtypes proposed by Lehmann et al. are classified as BL1, BL2, IM, M, MSL, and LAR.
[0070] In the above scheme, in steps 1-3, edgeR is used to perform differential gene analysis between cancer tissues and adjacent tissues, and based on the GENCODE gene set, features with significant differential changes are selected for gene annotation.
[0071] edgeR is an analytical method for determining differentially expressed genes using empirical Bayesian estimation and an exact test based on a negative binomial model. The present invention uses a negative binomial generalized log-linear model to divide the genes of the sample into three types: up, down, and none according to the differential expression fold table output by the model, and completes the screening of genes with significant differential expression. Afterwards, the annotation data in the GENCODE database is used to annotate the screened data, and the ENSEMBLE gene numbers that are not easy to understand in the data set are converted into more easily understood gene names.
[0072] In the above scheme, in steps 1-5, the SMOTE algorithm refers to an algorithm for dealing with the class imbalance problem in cancer data. In cancer research, the number of samples in the minority class is usually small, while the number of samples in the majority class is large, resulting in an imbalanced data set.
[0073] To address this imbalance problem, the researchers used the SMOTE algorithm to generate synthetic minority class samples, thereby balancing the difference in the number of samples between different classes in the dataset.
[0074] In the above scheme, in step 2-1, the Pearson correlation coefficient h is introduced to measure the strength of the relationship between two variables:
[0075]
[0076] Among them, A and B represent two variables, a iand b i denote the i-th data point of variables A and B respectively, and denote the mean of A and B respectively. As can be seen from formula (1), h is a value between -1 and 1. When h is less than 0, A and B are negatively correlated, that is, the larger the value of A, the smaller the value of B; conversely, when the value of h is greater than 0, A and B are positively correlated. When h = 0, A and B are independent of each other and do not reflect correlation. Therefore, it can be considered that the absolute value of h can reflect the correlation between A and B, that is, the larger the h, the stronger the correlation between A and B, and vice versa.
[0077] In the above scheme, in step 4-2, the auto encoder (AE) is embedded in the TOM matrix to calculate the feature importance score of the gene. The auto encoder is an unsupervised learning model. It is based on the back propagation algorithm and the optimization algorithm, using the input data X itself as supervision, learning the implicit features of the input data, and using the learned new features to reconstruct the original input data.
[0078] In the above scheme, in step 6-3, when a round of genetic process is executed, the first few chromosomes with the highest fitness will be selected to directly enter the next round of iteration, and then instances will be randomly selected from the remaining chromosomes to simulate crossover and mutation operations, and the simulated chromosomes will be added to the previously selected chromosome group with the highest fitness to form a new generation population. Among them, mutation refers to changing the elements at a certain position of the chromosome with the mutation operator km as the probability, such as changing the element originally 0 to 1; crossover refers to randomly selecting a breakpoint from the chromosome with the crossover operator kc as the probability, and exchanging the elements after it.
[0079] In the above scheme, in step 6-4, for the genes with high feature importance scores obtained in step 5-2, the individuals with the best performance are gradually obtained by simulating the biological evolution process. In the present invention, the top 100 genes are extracted, and finally the best feature subset containing 54 genes is obtained. The TNBC subtype classification using the gene subset can achieve an accuracy of 95.61%.
[0080] Compared with the prior art, the present invention has the following beneficial effects:
[0081] The algorithm proposed in the present invention combines GeneRank with the autoencoder AE embedded in TOM. By utilizing the neuron connection weights in the network training process, the importance of genes is linked to the absolute value of the sum of neuron weights. Then, the genetic algorithm is applied to simulate the biological evolution process to effectively extract the gene subset with the greatest contribution to the prediction work, so as to more accurately identify TNBC subtypes. BRIEF DESCRIPTION OF THE DRAWINGS
[0082] Figure 1 is a flow chart of the present invention;
[0083] Figure 2 This is the structure diagram of the automatic encoder in step 4-2 of the present invention;
[0084] Figure 3 This is a schematic diagram of the random forest algorithm in step 6-2 of the present invention;
[0085] Figure 4 Outlier detection for the sample set in Example 1;
[0086] Figure 5 A schematic diagram of determining a soft threshold in Example 1;
[0087] Figure 6 Schematic diagram of gene modules divided by the hierarchical clustering tree in Example 1;
[0088] Figure 7 This is a result diagram of the genetic algorithm iteration process in Example 2;
[0089] Figure 8 The ranking performance results of the five classifiers in Example 2 under different number of gene combinations;
[0090] Fig. 9 This is a comparison diagram of the three gene importance scoring methods in the neural network in Example 2. DETAILED DESCRIPTION
[0091] The present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0092] Example 1
[0093] like Figure 1 As shown, the optimal gene subset selection method for identifying the molecular subtypes of triple-negative breast cancer includes the following steps:
[0094] Step 1: Build a dataset
[0095] Step 1-1 prepare a data set. The data set comes from the Cancer Genome Atlas
[0096] The TCGA Program is a research project jointly organized by the National Cancer Institute (NCI) and the National Human Genome Research Institute (NHGRI) in 2006. The project mainly focuses on large-scale genome sequencing.
[0097] The analysis technology was used to study more than 2,000 primary cancer cases of 33 cancer types at the molecular level.
[0098] Actively cooperate and communicate extensively with scientific research and health institutions around the world to improve people's scientific understanding of the molecular basis of cancer and their ability to diagnose, treat and prevent cancer, and ultimately form a set of cancer-related diagnosis and treatment maps. However, due to the strong heterogeneity of TNBC, its sample set and corresponding label information cannot be directly obtained from the TCGA official website, and it needs to be preprocessed before it can be applied to the classification task of the model.
[0099] Step 1-2 preprocesses the data obtained in step 1-1. First, the sample number and category label provided in the paper by Barche et al. are used as screening conditions to screen samples with the same number value from the breast cancer RNA-Seq expression data to reconstruct the expression matrix.
[0100] Step 1-3: After that, edgeR was used to perform differential gene analysis between cancer tissues and adjacent tissues, and based on the GENCODE gene set, features with significant differences were selected for gene annotation.
[0101] Step 1-4: After the annotation work was completed, a total of 469 lncRNAs and 1892 mRNAs were obtained. After merging these two types of non-coding RNAs, genes with the same aliases were screened out according to gene symbols to construct gene expression data D.
[0102] Step 1-5: Finally, the SMOTE algorithm was used to complete the data set expansion, so that the six TNBC subtypes were basically consistent in the number of samples. At this point, there were 301 samples and 2361 genes in the data set.
[0103] Step 2: construct a weighted co-expression network of genes and screen out candidate gene sets;
[0104] Before constructing a weighted co-expression network for genes, it is necessary to check the differences between samples and calculate the outliers of samples in the data set by constructing a hierarchical clustering tree based on the Euclidean distance between samples, so as to remove outliers in time. Figure 4It can be seen that most of the samples in the dataset are relatively similar, with only a few instances being quite different from the other samples, so they are manually removed.
[0105] After completing the construction of the expression matrix, it is necessary to determine the soft threshold before constructing the gene co-expression network. The soft threshold can set the features with absolute values below a certain threshold to 0, and adjust other features towards 0. In order to ensure that the connection between genes conforms to the scale-free distribution, the experiment conducted a network topology analysis on the soft threshold index with a value range of 1 to 30, as shown in the attached figure. Figure 5 shown.
[0106] Attached Figure 5 The horizontal axis in (a) represents the soft threshold, and the vertical axis represents the scale-free fitting index. In this experiment, the vertical axis selects the line at 0.85 as the screening mark, and the first horizontal axis that reaches the dividing line is selected as the power index to generate the adjacency matrix between genes. At this time, the soft threshold is 6. Figure 5 (b) represents the average connectivity of genes in the network. It can be seen that when the soft threshold is 6, the average connectivity corresponding to the ordinate is also relatively low, which is consistent with the distribution characteristics of real biological networks.
[0107] After determining the soft threshold, the Pearson correlation coefficient matrix generated by the gene data is added with a power operation to generate the correlation adjacency matrix Q. Then, according to formula (2), the adjacency matrix is converted into a topological overlap matrix TOM. Based on the TOM matrix and dynamic tree cutting, the following clustering tree is generated, and different colors are used to display different modules. Figure 6 shown.
[0108] After obtaining the module division results, the eigenvector MEs within the module were calculated, and the correlation between the module and TNBC subtype was analyzed in combination with the trait data. Finally, the sample of the constructed expression matrix contained a total of 1420 features.
[0109] Step 2-1 calculates the Pearson correlation coefficient h for each pair of genes, and adds a power operation to enhance the connectivity between genes, converting the Pearson matrix into an adjacency matrix Q = [q ij ]. The formula is as follows:
[0110]
[0111] Among them, q ij represents the adjacency value between gene i and gene j, and β is a hyperparameter determined by the scale-free network;
[0112] Step 2-2 Simply using the adjacency matrix Q will result in low data utilization and a high error rate in module division. Therefore, a topological overlap matrix TOM is added on the basis of the adjacency matrix to more accurately describe the correlation between genes. The formula for calculating the TOM matrix elements is as follows:
[0113]
[0114] It can be seen that the TOM matrix also takes into account other elements in the data set, and the larger the value of the element, the stronger the correlation between the two genes;
[0115] Step 2-3 calculates the dissimilarity between genes based on the TOM matrix to construct a hierarchical clustering tree. The number of genes in each module is determined according to the set standard, and then the correlation between the characteristic gene MEs of each module and the six subtypes is calculated to eliminate the features with low correlation.
[0116] In the embodiment of the present invention, after filtering out genes with low correlation, the dimension of the feature matrix obtained is k1;
[0117] Step 3, calculate the GeneRank ranking of the gene;
[0118] Step 3-1 constructs a graph Q(W,E) representing the connectivity between genes, where W represents the gene node and E represents the connectivity between genes. The final gene importance ranking t can be calculated by formula (3):
[0119] (I - dWD -1 )t = (1 - d)fc (1-3)
[0120] Among them, I is a unit matrix; d is a constant between 0 and 1 (this experiment defines it as 0.6); D is a diagonal matrix, which represents the number of edges associated with genes, that is, deg i =∑ j w ij ; fc is the fold change of the gene list, which can indicate the degree of difference in expression of a gene between different classes;
[0121] Step 3-2 Ordering of gene g in subtype i Formula (4) can be calculated based on the transformed form of formula (3):
[0122]
[0123] Among them, b jk represents the elements in the binary matrix B constructed on the basis of the TOM matrix, indicating the connectivity between genes. jk When b is greater than the mean of the matrix elements, jk With bkj Set to 1, otherwise set to 0. It is worth noting that because the diagonal element value of matrix B is 1, and the value of the non-diagonal element is usually much smaller than 1, it is not necessary to consider the element value on the diagonal when calculating the mean, otherwise it is likely to cause most of the element values on matrix B to be 0. Finally, after n iterations, the GeneRank ranking of gene importance is obtained by formula (5):
[0124]
[0125] Where M represents the total number of subtype categories in the training set; N x=i Represents the number of samples with label i in the sample set; N total Indicates the total number of samples in the sample set;
[0126] Step 4, calculate the importance score of the gene based on the embedded TOM matrix;
[0127] Step 4-1 Embed a TOM matrix that can express the correlation between genes
[0128] Generally speaking, in neural networks, feature importance is related to weight, which means that the higher the weight of a neuron after network training, the greater its role in classification tasks. The CW method proposed by Olden et al. believes that the importance of each feature can be obtained by calculating the sum of the absolute values of all weights related to it. Inspired by this concept, the present invention embeds a TOM matrix that can represent the correlation between genes in the autoencoder, as shown in the attached figure. Figure 2 shown.
[0129] After the TOM matrix, the output Q of the first hidden layer can be obtained [1] The calculation formula is as follows
[0130]
[0131] Among them, ⊙ is the dot multiplication operation, representing the matrix W [1] The multiplication operation is performed with the elements at the same subscript in the matrix TOM; ε1 is the neural network activation function, and the present invention selects ReLU as the activation function to prevent overfitting; represents the deviation term of the first layer;
[0132] Step 4-2: Calculation of feature importance score;
[0133] The importance score of gene g can be calculated by the following formula:
[0134]
[0135] Where x represents the number of genes in the input layer, and y represents the number of neurons in the second hidden layer. Since AE embeds the topological overlap matrix TOM, the relative importance of each gene is affected by both the first hidden layer and the second hidden layer. Therefore, the second hidden layer needs to be taken into account during the calculation process. SD represents the standard deviation of all weights connected to neurons during AE training, which is calculated by formula (1-8):
[0136]
[0137] Where t represents the number of network iterations, which means that the greater the influence of a unit node on the prediction performance, the greater the range of changes in all weights connected to it;
[0138] Step 5: Calculate the importance score of the gene based on the embedded TOM matrix
[0139] In step 3 and step 4 of the above technical solution, this section will comprehensively consider the two calculation results to obtain the score ranking of feature importance.
[0140] Step 5-1: First, standardize the two calculation results:
[0141]
[0142]
[0143] Step 5-2: After completing the standardization work, a new scoring standard is constructed to evaluate the contribution of each gene to the classification result. The formula is as follows:
[0144]
[0145] in, is a constant term, which is manually set to 6 in the present invention. and The value range of is between 0 and 1, and based on the comprehensive score of each gene, the top k2 genes are selected for the second feature dimension reduction operation. At this point, the importance score of the gene is calculated, and the feature dimension of the gene becomes k2;
[0146] Step 6, optimal gene subset selection based on genetic algorithm;
[0147] In the present invention, for the genes with high feature importance scores obtained in step 5, a genetic algorithm is used to search for a gene subset with the best classification performance from the candidate genes.
[0148] Step 6-1: Consider each sample in the sample set as a chromosome, and each feature as a gene on the chromosome. First, randomly initialize the chromosome to a k2-dimensional binary vector D = (D1, D2, ..., D n2 ), and then generate a uniformly distributed random variable between 0 and 1 by random encoding. when When the value is greater than 0.5, the gene at the corresponding position is selected, D i =0;
[0149] Step 6-2, select random forest as the classifier model, and use the classification accuracy as the fitness function of the genetic algorithm. When the classification accuracy is higher, the fitness of the genetic algorithm is higher, and it is easier to be selected to perform genetic operations in biological genetics, and the feature importance of the selected gene is higher; Figure 3 As shown;
[0150] Step 6-3, perform a genetic process to select a subset of genes with better performance;
[0151] Step 6-4, after completing t rounds of iterations, select the chromosome with the largest fitness from the population for decoding, and find the genes at the corresponding positions in the expression matrix based on the non-zero elements to form the optimal subset Gene.
[0152] Genetic algorithms are mainly used on genes ranked in the top 100, and gradually obtain the best performing individuals by simulating the biological evolution process. In this experiment, the mutation probability km is 0.2, the crossover probability kc is 0.9, and the fitness function is the prediction accuracy of random forest under a specific genome. Figure 7 The fitness value of the best individual in each iteration is shown. When the number of evolutions reaches 100, the iteration ends. By decoding the chromosome with the highest fitness in the last generation, the best feature subset containing 54 genes is finally obtained. The classification accuracy of this gene subset can reach 95.61%.
[0153] Example 2
[0154] The experimental results were verified by three cross-validation methods: the effectiveness of the autoencoder in calculating the importance of features, the role of the autoencoder embedded in the TOM layer, and the comparison of the importance calculation methods of the autoencoder.
[0155] Evaluation criteria:
[0156] In most machine learning related research, accuracy (ACC), sensitivity (SEN), specificity (SPE), F1 Score (F1) and area under the ROC curve AUC are usually used as evaluation indicators to measure the quality of classification results. The present invention also uses these indicators as the criteria for evaluating the quality of the model.
[0157] ACC is the most basic and intuitive way to measure the accuracy of model classification (or regression) prediction results. It represents the proportion of correctly classified samples in the total number of samples. The specific calculation formula is as follows:
[0158]
[0159] Among them, TP, TN, FP, and FN are four-quadrant confusion matrices composed of two variables: the true or false classification results and the true or false actual labels.
[0160] Simply using ACC as an evaluation metric has great limitations, especially when the distribution of sample data is uneven, the accuracy result is easily affected by the maximum sample label ratio. In other words, as long as the evaluation function is simply set to be equal to the identity of the maximum class label, a very high accuracy rate can be obtained. Therefore, it is necessary to introduce other scientific indicators to more comprehensively evaluate the performance of the model.
[0161] SEN is also known as recall, which indicates the ratio of the number of samples predicted by the model to the number of samples that are actually positive. The higher the sensitivity, the stronger the model's ability to identify positive examples in samples. The specific definition is as follows:
[0162]
[0163] SPE is the opposite of SEN. It represents the ratio of the number of samples that are both predicted by the model and are actually counterexamples to the number of samples that are actually counterexamples. The higher the specificity, the stronger the model's ability to identify counterexamples in samples. The specific definition is as follows:
[0164]
[0165] The F1 metric is a special case of the F-ε metric where ε=1. It is a comprehensive metric of sensitivity and accuracy. The larger the F1 value, the more stable the model's performance. The specific calculation method is as follows:
[0166]
[0167] AUC is defined as the area under the ROC curve and the coordinate axis. Obviously, this area is between 0 and 1. Since the ROC curve is generally above the straight line y=x, the value of AUC is generally between 0.5 and 1, and the more stable the classification results of a classifier, the larger its AUC value.
[0168] 1. Verification of the effectiveness of automatic encoder calculation of feature importance
[0169] In order to determine whether AE can achieve better performance results, the present invention selected prediction models with higher cumulative accuracy for comparative experiments, and used random forest as a unified evaluator to compare the ranking performance of different models. In order to reserve enough genes for the genetic algorithm to simulate the fireworks experiment, the present invention selected the number of gene combinations to increase from 40, and increased to 90 in steps of 10, and constructed 20 sets of gene subsets in total. The performance results are shown in the attached figure. Figure 7 shown.
[0170] From the attached Figure 7 It can be seen that the performance of AE is significantly better than other predictive values, indicating that the importance score generated by AE is more meaningful in identifying TNBC subtypes.
[0171] 2. Verification of the role of autoencoder embedded in TOM layer
[0172] The topological overlap matrix TOM constructed by the Pearson coefficient shows the connectivity between genes, which means that the higher the connectivity between two genes, the larger the value of TOM at the corresponding position. Therefore, if the TOM matrix is embedded in the neural network, the connectivity and importance between neurons can be amplified. To verify this idea, this paper conducted a cross-validation experiment, comparing the traditional AE with the AE embedded in the TOM layer on the cross-validation indicators to show that the AE embedded in the TOM layer performs better. The performance results are shown in Table 1.
[0173] Table 1 Comparison of AE and AE embedded in TOM
[0174]
[0175] 3. Comparative Verification of Importance Calculation Methods of Autoencoders
[0176] The importance ranking of genes can be calculated by the connection weights of the neural network without changing the gene features themselves. In the neural network model embedded with TOM, the relative contribution of features to the prediction results depends on the size and direction of CWs. In this experiment, since the weight values of each layer of the neural network will continue to change during the training process, it is necessary to conduct comparative tests on three different weight calculation methods: LAST, SD, and embedded TOM. Among them, the LAST metric refers to the consideration of only the sum of the absolute values of all weights related to each gene after the network training is completed in the importance calculation; LAST+SD means that on the basis of the LAST metric, the standard deviation of the weights is additionally considered in the back propagation to evaluate the magnitude of the weight change; TOM+LAST+SD means that on the basis of LAST+SD, between the first hidden layer and the input layer
[0177] Insert the TOM layer to strengthen the correlation and connectivity between genes. The experiment sorted the screening results of these three cases respectively, and selected the first [5,6,...,20,30,...,100] genes as input to verify their performance on the random forest classifier. The results are shown in the attached Figure 8 shown.
[0178] By attaching Fig. 9 It can be seen that the performance of the embedded TOM+LAST+SD is better than the other two. Therefore, the performance of the autoencoder embedded with TOM has higher accuracy, which also proves the effectiveness of the experimental method used in the present invention.
Claims
1. A method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer, characterized in that: The following steps are involved: Step 1, build a data set; Step 2: construct a weighted co-expression network of genes and screen out candidate gene sets; Step 3, calculate the GeneRank ranking of the gene; Step 4, gene importance scores calculated by AE embedded in the TOM matrix; Step 5, calculate the score ranking of gene feature importance; Step 6: Use genetic algorithm to determine the final gene candidate set.
2. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 1, characterized in that: The specific steps of constructing the data set described in step 1 are as follows: Step 1-1, obtain official website data, complete the screening of TNBC cases in the data and add labels to the data according to the six subtypes; Step 1-2, screening samples with the same number value from breast cancer RNA-Seq expression data to reconstruct the expression matrix; Steps 1-3: Use edgeR to perform differential gene analysis between cancer tissues and adjacent tissues, and select features with significant differences for gene annotation based on the GENCODE gene set; Steps 1-4: After completing the annotation work, select genes with the same alias according to gene symbols and construct gene expression data; In steps 1-5, the SMOTE algorithm is used to expand the data set so that the number of samples for the six subtype classifications remains consistent.
3. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 2, characterized in that: The weighted co-expression network of genes is constructed as described in step 2 to screen out candidate gene sets. The specific steps are as follows: Step 2-1, calculate the Pearson correlation coefficient h for each pair of genes, add power operations to enhance the connectivity between genes, and convert the Pearson matrix into an adjacency matrix Q = [q ij ], the formula is as follows: Among them, q ij represents the adjacency value between gene i and gene j, power(h,β(represents the βth power of distance h, which is used to adjust the influence of distance on Pearson correlation coefficient, cor(i,j) represents the correlation coefficient between variable i and variable j, and β is a hyperparameter determined by the scale-free network; Step 2-2, add a topological overlap matrix TOM based on the adjacency matrix. The formula for calculating the TOM matrix elements is as follows: The larger the value of the element, the stronger the correlation between the two genes; ij represents the TOM value of variables i and j, q in and q nj represents the value of the variable i and variable j positions in the TOM matrix, and n represents the total number of genes in the TOM matrix; Step 2-3, based on the TOM matrix, calculate the dissimilarity between genes to construct a hierarchical clustering tree, determine the number of genes in each module according to the set standard, and then calculate the correlation between the characteristic gene MEs of each module and the six subtypes to eliminate the features with low correlation.
4. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 3, characterized in that: The GeneRank ranking of the calculated gene described in step 3 is specifically performed as follows: Step 3-1, construct a graph Q(W,E) representing the connectivity between genes, where W represents the gene node and E represents the connectivity between genes, and calculate the final gene importance ranking t: (I - dWD -1 )t = (1 - d)fc (1-3) Among them, I is a unit matrix; d is a constant between 0 and 1; D is a diagonal matrix, which represents the number of edges associated with genes, that is, deg i =∑ j w ij , deg i represents the degree of node i, w ij represents the connection weight between node i and node j; fc is the fold change of the gene list, indicating the degree of difference in expression of a gene between different classes; Step 3-2, calculate the ranking of gene g in subtype i The formula is as follows: in, represents the functional node of node i when the clustering coefficient is g, b jk represents the elements in the binary matrix B constructed on the basis of the TOM matrix, indicating the connectivity between genes, represents the value of node i in the η-1th iteration when the clustering coefficient is g. jk When b is greater than the mean of the matrix elements, jk With b kj Set it to 1, otherwise it is set to 0. After n iterations, the GeneRank ranking of gene importance is obtained by formula (1-5): Where M represents the total number of subtype categories in the training set; N x=i Represents the number of samples with label i in the sample set; N total Represents the total number of samples in the sample set.
5. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 4, characterized in that: The gene importance score calculated by AE embedded in the TOM matrix described in step 4 is as follows: Step 4-1, embed a TOM matrix that can express the correlation between genes; After the TOM matrix, we get the output Q of the first hidden layer. [1] , the calculation formula is as follows: Among them, ⊙ is the dot multiplication operation, representing the matrix W [1] The multiplication operation of the elements with the same subscript in the matrix TOM; ε1 is the neural network activation function, represents the bias term of the first layer, X T Represents the transpose of matrix X. The transpose operation swaps the rows and columns of matrix X to form a new matrix, namely the transposed matrix X. T ; Step 4-2, feature importance score calculation; The importance score of gene g is calculated by the following formula: Among them, x represents the number of genes in the input layer, and y represents the number of neurons in the second hidden layer; and represents the connection weight between the gth hidden node in the autoencoder and the jth input node in the input layer, and Respectively and , SD represents the standard deviation of all weights connected to neurons during AE training, which is calculated by formula (8): Among them, t represents the number of network iterations, indicating that the greater the influence of a unit node on the prediction performance, the greater the range of change of all weights connected to it, and l represents the weight w ij The index variable used when summing the square values of , l changes from 1 to t, which means traversing all t weight data values.
6. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 5, characterized in that: The specific steps of calculating the score ranking of gene feature importance described in step 5 are as follows: Step 5-1, standardize the calculation results of step 3 and step 4: Step 5-2, construct a new scoring criterion to evaluate the contribution of each gene to the classification result. The formula is as follows: in, is a constant term, and The value range of is between 0 and 1, and based on the comprehensive score of each gene, the top k2 genes are selected for the second feature dimensionality reduction operation. At this time, the importance score of the gene is calculated and the feature dimension of the gene becomes k2.
7. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 6, characterized in that: The genetic algorithm described in step 6 is used to determine the final gene candidate set, as follows: For the genes with high feature importance scores obtained in step 5, a genetic algorithm is used to search for the gene subset with the best classification performance from the candidate genes: Step 6-1, each sample in the sample set is regarded as a chromosome, and each feature is a gene on the chromosome; first, the chromosome is randomly initialized as a k2-dimensional binary vector D = (D1, D2, ..., D n2 ), and then generate a uniformly distributed random variable between 0 and 1 by random encoding. when When the value is greater than 0.5, the gene at the corresponding position is selected, D i =0; Step 6-2, select random forest as the classifier model, and use the classification accuracy as the fitness function of the genetic algorithm. When the classification accuracy is higher, the fitness of the genetic algorithm is higher, and it is easier to be selected to perform genetic operations in biological genetics, and the feature importance of the selected gene is higher; Step 6-3, perform a genetic process to select a subset of genes with better performance; Step 6-4, after completing t rounds of iterations, select the chromosome with the largest fitness from the population for decoding, and find the genes at the corresponding positions in the expression matrix based on the non-zero elements to form the optimal subset Gene.
8. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 2, characterized in that: The six subtypes described in step 1-1 are classified as BL1, BL2, IM, M, MSL, and LAR.
9. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 3, characterized in that: In step 2-1, the Pearson correlation coefficient h is introduced to measure the strength of the relationship between two variables: Among them, A and B represent two variables, a i and b i denote the i-th data point of variables A and B respectively, and and They represent the means of A and B respectively, h is a value between -1 and 1. When h is less than 0, A and B are negatively correlated, that is, the larger the value of A, the smaller the value of B. Conversely, when the value of h is greater than 0, A and B are positively correlated. When h = 0, A and B are independent of each other and do not reflect correlation. Therefore, it is believed that the absolute value of h can reflect the correlation between A and B, that is, the larger h is, the stronger the correlation between A and B, and vice versa.
10. The method for selecting an optimal gene subset to identify molecular subtypes of triple-negative breast cancer according to claim 7, characterized in that: In step 6-3, when a round of genetic process is executed, the first few chromosomes with the highest fitness are selected to directly enter the next round of iteration, and then instances are randomly selected from the remaining chromosomes to simulate crossover and mutation operations, and the simulated chromosomes are added to the chromosome group with the highest fitness previously selected to form a new generation population; wherein, mutation refers to changing an element at a certain position of the chromosome with a probability of mutation operator km, and crossover refers to randomly selecting a breakpoint from the chromosome with a probability of crossover operator kc and exchanging the elements after it.
Citation Information
Patent Citations
New method for lung adenocarcinoma biomarker screening, prognosis model construction and biological verification
CN113130002A
