A breast cancer prognosis prediction method and system based on deep neural network

Abstract

The present invention discloses a method and system for predicting the prognosis of breast cancer based on a deep neural network. By obtaining the gene expression data of triple-negative breast cancer samples, data preprocessing is performed to extract effective expression quantitative trait sites; then the data are grouped, and the core expression quantitative trait sites are screened using the t-test function, variance analysis function and correlation function in turn to obtain the core expression quantitative trait sites; a deep neural network is constructed using the core expression quantitative trait sites, and the deep neural network is trained using the gene expression data, and the parameters are optimized at the same time to obtain an optimized triple-negative breast cancer prognosis deep neural network for predicting the risk of recurrence and metastasis after neoadjuvant chemotherapy for breast cancer. The present invention can more accurately predict the risk of recurrence and metastasis after neoadjuvant chemotherapy for triple-negative breast cancer by effectively filtering and screening gene expression data and optimizing the deep neural network at the same time, thereby helping clinical diagnosis and treatment.

Description

Technical Field

[0001] The present invention relates to the field of medical diagnosis, and in particular to a method and system for predicting the prognosis of breast cancer by using gene expression data based on a deep neural network. Background Art

[0002] Breast cancer is one of the most common malignant tumors in women. According to data, there are about 1.2 million new female breast cancer patients and about 500,000 women die from breast cancer each year. According to research, breast cancer is a disease that easily causes metastasis. About 50% of cases can be cured after surgical treatment, while the remaining 50% of cases may relapse or metastasize.

[0003] Among them, triple-negative breast cancer (TNBC) is highly heterogeneous, with strong invasiveness, poor prognosis, and high risk of recurrence and metastasis, which has always been a difficult point in clinical treatment. Due to the lack of therapeutic targets for TNBC, the proportion of patients with locally advanced disease who achieve complete lesion removal (pathological complete response, pCR) after neoadjuvant chemotherapy (NAC) is much higher than that of other molecular subtypes, but the risk of recurrence and metastasis is also much higher than that of other molecular subtypes, which is 2.0~3.5 times that of other molecular subtypes. Therefore, accurately predicting the risk of recurrence and metastasis of patients with triple-negative breast cancer after neoadjuvant chemotherapy is of great clinical significance, which helps to assist in the formulation of individualized treatment plans and postoperative follow-up management.

[0004] In recent years, with the in-depth study of genetic testing in the probability of disease occurrence and the rise of bioinformatics, there have been many related studies. For example, the Shenzhen Institute of Advanced Technology proposed a breast cancer survival prediction method based on deep neural networks (publication number: CN 111161882 A). By constructing deep neural networks for clinical data, gene expression data and DNA methylation data respectively, breast cancer survival is predicted. However, there are problems such as simple data preprocessing and the accuracy is greatly affected by the data preprocessing results.

[0005] In order to solve the above problems, the present invention adds more effective steps to data preprocessing, focuses on the prognosis prediction of triple-negative breast cancer, and proposes a breast cancer prognosis prediction method and system based on deep neural network. Summary of the invention

[0006] The present invention provides a method for predicting breast cancer prognosis based on a deep neural network, which is characterized by comprising the following steps:

[0007] S1: Obtaining gene expression data of triple-negative breast cancer samples, wherein the triple-negative breast cancer samples include samples with completely removed lesions and samples with residual lesions;

[0008] S2: performing data preprocessing on the gene expression data to extract effective expression quantitative trait loci;

[0009] S3: grouping the expression quantitative trait loci, and then screening them using t-test function, variance analysis function and correlation function in turn to obtain core expression quantitative trait loci;

[0010] S4: constructing a deep neural network using the core expression quantitative trait loci, training the deep neural network using the gene expression data, and optimizing the parameters of the deep neural network to obtain an optimized triple-negative breast cancer prognosis deep neural network;

[0011] S5: Inputting the gene expression quantitative trait loci of the triple-negative breast cancer sample to be tested into the triple-negative breast cancer prognosis deep neural network to predict prognosis and recurrence.

[0012] The data preprocessing includes:

[0013] S21: Delete the data with no phenotype in the triple-negative breast cancer sample, and go to S22;

[0014] S22: extract gene expression data, delete duplicate or synonymous gene expression data, and go to S23;

[0015] S23: Calculate the expression ratio of the gene expression data, and exclude gene expression data with an expression ratio less than 30%.

[0016] The grouping of the expression quantitative trait loci is specifically:

[0017] For the effective expression quantitative trait loci, the expression quantitative trait loci of the sample with completely removed lesions are assigned as array variable X, and the expression quantitative trait loci of the sample with residual lesions are assigned as array variable Y.

[0018] Furthermore, the screening is performed using the t-test function, the variance analysis function and the correlation function in sequence, specifically:

[0019] S31: performing a t test on the array variable X and the array variable Y, and retaining the expression quantitative trait loci with a P value less than 0.01;

[0020] S32: performing variance analysis on the expression quantitative trait loci retained in S31, sorting and screening according to the variance values ​​between the array variable X and the array variable Y, and retaining the expression quantitative trait loci with a P value less than 0.001;

[0021] S33: Calculate the correlation value between each pair of the expression quantitative trait loci retained in S32 using the correlation function, and retain the P value less than 1.0e -16 expression quantitative trait loci.

[0022] Preferably, the t-test function and the correlation function both adopt a two-tailed test method.

[0023] The deep neural network is a deep neural network with 2 hidden layers.

[0024] The parameters of the deep neural network include:

[0025] The number of training samples each time is at least 80, 90, 100, 110, 120, 130 or 140;

[0026] The training step size is at least 10000, 12500, 15000, 17500 or 20000.

[0027] Furthermore, the hidden layer also includes:

[0028] The number of nodes in the first hidden layer is at least 20, 30, 40, 50 or 60;

[0029] The number of nodes in the second hidden layer is at least 8, 9, 10, 11 or 12.

[0030] Preferably, the present invention provides a breast cancer prognosis prediction system based on a deep neural network, characterized in that it comprises:

[0031] A data acquisition module, which acquires gene expression data of triple-negative breast cancer samples, wherein the triple-negative breast cancer samples include samples with completely removed lesions and samples with residual lesions;

[0032] A data filtering module performs data preprocessing on the gene expression data to obtain effective expression quantitative trait loci;

[0033] A site screening module groups the expression quantitative trait sites, and then screens them using a t-test function, a variance analysis function, and a correlation function in turn to obtain core expression quantitative trait sites;

[0034] A neural network module, using the core expression quantitative trait loci to construct a deep neural network, using the gene expression data to train the deep neural network, and optimizing the parameters of the deep neural network to obtain an optimized triple-negative breast cancer prognosis deep neural network;

[0035] The input-output module inputs the gene expression quantitative trait loci of the triple-negative breast cancer sample to be tested into the triple-negative breast cancer prognosis deep neural network for prognosis recurrence prediction, and converts the prediction results into paper reports or electronic reports.

[0036] The data preprocessing includes: deleting the triple-negative breast cancer samples without phenotype data; then extracting the gene expression data, deleting duplicate or synonymous gene expression data; then calculating the expression ratio of the gene expression data, and eliminating the gene expression data with an expression ratio less than 30%.

[0037] The grouping of the expression quantitative trait loci is specifically as follows: for the effective expression quantitative trait loci, the expression quantitative trait loci of the sample with completely removed lesions are assigned as array variable X; and the expression quantitative trait loci of the sample with residual lesions are assigned as array variable Y.

[0038] The screening is performed using the t-test function, variance analysis function and correlation function in sequence, specifically:

[0039] Performing a t test on the array variable X and the array variable Y, and retaining expression quantitative trait loci with a P value less than 0.01;

[0040] Then, variance analysis is performed, and the variance values ​​between the array variable X and the array variable Y are sorted and screened, and the expression quantitative trait loci with the P value less than 0.001 are retained;

[0041] The correlation function was then used to calculate the correlation values ​​between the expression quantitative trait loci, and the P value was less than 1.0e -16 expression quantitative trait loci.

[0042] The present invention discloses a breast cancer prognosis prediction method and system based on a deep neural network. The method comprises the following steps: obtaining gene expression data of triple-negative breast cancer samples, performing data preprocessing to extract effective expression quantitative trait loci; grouping the samples, and screening them in turn using a t-test function, a variance analysis function, and a correlation function to obtain core expression quantitative trait loci; constructing a deep neural network using the core expression quantitative trait loci, training the deep neural network using the gene expression data, and optimizing parameters to obtain an optimized triple-negative breast cancer prognosis deep neural network for predicting the risk of recurrence and metastasis after neoadjuvant chemotherapy for breast cancer.

[0043] Compared with the prior art, the beneficial effects of the present invention are: by effectively filtering and screening the gene expression data of triple-negative breast cancer samples, the validity of the data is improved, and at the same time, the key parameters of the deep neural network are optimized, which can more accurately predict the risk of recurrence and metastasis of triple-negative breast cancer after neoadjuvant chemotherapy, and assist clinical diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0045] Figure 1 It is a schematic flow chart of a method for predicting breast cancer prognosis based on a deep neural network described in the present invention.

[0046] Figure 2 This is a schematic diagram of a breast cancer prognosis prediction system module based on a deep neural network described in the present invention. DETAILED DESCRIPTION

[0047] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0048] like Figure 1 As shown, the present invention proposes a breast cancer prognosis prediction method based on a deep neural network, which is characterized by comprising the following steps:

[0049] S1: Acquire triple-negative breast cancer samples, wherein the triple-negative breast cancer samples include samples with completely removed lesions and samples with residual lesions.

[0050] S2: Preprocessing the gene expression data of the triple-negative breast cancer sample to extract effective expression quantitative trait loci.

[0051] The data preprocessing includes:

[0052] S21: Delete the data with no phenotype in the triple-negative breast cancer sample, and go to S22;

[0053] S22: extract gene expression data, delete duplicate or synonymous gene expression data, and go to S23;

[0054] S23: Calculate the expression ratio of the gene expression data, and exclude gene expression data with an expression ratio less than 30%.

[0055] S3: Grouping the expression quantitative trait loci, specifically:

[0056] For the effective expression quantitative trait loci, the expression quantitative trait loci of the sample with completely removed lesions are assigned as array variable X, and the expression quantitative trait loci of the sample with residual lesions are assigned as array variable Y;

[0057] Then use the t-test function, variance analysis function and correlation function to screen in turn, specifically:

[0058] S31: performing a t test on the array variable X and the array variable Y, and retaining the expression quantitative trait loci with a P value less than 0.01;

[0059] S32: performing variance analysis on the expression quantitative trait loci retained in S31, sorting and screening according to the variance values ​​between the array variable X and the array variable Y, and retaining the expression quantitative trait loci with a P value less than 0.001;

[0060] S33: Calculate the correlation value between each pair of the expression quantitative trait loci retained in S32 using the correlation function, and retain the P value less than 1.0e -16 expression quantitative trait loci.

[0061] The t-test function and the correlation function both use a two-tailed test method to obtain core expression quantitative trait loci.

[0062] S4: constructing a deep neural network using the core expression quantitative trait loci, training the deep neural network using the gene expression data, and optimizing the parameters of the deep neural network to obtain an optimized triple-negative breast cancer prognosis deep neural network.

[0063] The deep neural network is a deep neural network with 2 hidden layers;

[0064] The number of nodes in the first hidden layer is at least 20, 30, 40, 50 or 60;

[0065] The number of nodes in the second hidden layer is at least 8, 9, 10, 11 or 12.

[0066] The parameters of the deep neural network include:

[0067] The number of training samples each time is at least 80, 90, 100, 110, 120, 130 or 140;

[0068] The training step size is at least 10000, 12500, 15000, 17500 or 20000.

[0069] S5: Inputting the gene expression quantitative trait loci of the triple-negative breast cancer sample to be tested into the triple-negative breast cancer prognosis deep neural network to predict prognosis and recurrence.

[0070] In one embodiment of the present invention, the specific steps include:

[0071] The gene expression data of samples from triple-negative breast cancer patients treated with neoadjuvant chemotherapy were downloaded from the SRA (Sequence Read Archive) database of the National Center for Biotechnology Information (NCBI). A total of 222 samples were divided into two groups: PCR (complete removal of lesions) and RD (residual lesions), including 80 PCR samples and 142 RD samples.

[0072] 222 triple-negative breast cancer samples were filtered, of which BA00097 sample was deleted because it was data without phenotype; gene expression data were extracted from the remaining samples, totaling 60,279 gene expression data. The expression ratio of gene expression data was calculated, expression ratio = number of genes with expression data greater than 0 / total number of genes; gene expression data with expression ratio less than 30% were filtered, and 19,451 gene expression data were obtained.

[0073] For the above-mentioned effective expression quantitative trait loci, the expression quantitative trait loci of the sample with complete lesion removal are assigned to the array variable X, and the expression quantitative trait loci of the sample with residual lesions are assigned to the array variable Y.

[0074] The t.test function of the R language was used to calculate the difference between the array variable X and the array variable Y using a two-tailed test method. A p value less than 0.01 was used as a screening criterion, and expression quantitative trait loci with a p value less than 0.01 were retained, resulting in 527 expression quantitative trait loci.

[0075] The variance values ​​of these 527 quantitative trait loci were calculated using the aov function of the R language, and the quantitative trait loci with large variance values ​​and P values ​​less than 0.001 were screened, and 153 loci were obtained.

[0076] The cor.test correlation function in R language was used to calculate the correlation values ​​between 153 expression quantitative trait loci using a two-tailed test method, and the site correlation P value was retained to be less than 1.0e -16 sites; further, combined with the size sorting of the variance values ​​in the previous step, 46 sites were obtained.

[0077] The 221 groups of samples were divided into 161 training sets, 30 validation sets, and 30 test sets. The 161 training sets were used to train a deep neural network with 2 hidden layers.

[0078] The deep neural network parameter optimization is divided into two rounds:

[0079] In the first round, for the two parameters of the number of training samples each time (Batch size) and the step length of each training (step):

[0080] The number of training samples each time is at least 80, 90, 100, 110, 120, 130 or 140;

[0081] The training step size is at least 10000, 12500, 15000, 17500 or 20000.

[0082] The different values ​​of the two parameters, the number of training samples each time (Batch size) and the step size each time (step), are combined in pairs; under different parameter combinations, the deep neural network repeats the training set 10 times each time, and then verifies and tests it after the training is completed. Three indicators, verification accuracy, test accuracy and the combined accuracy of the two, are obtained in a single training.

[0083] Take the average accuracy and maximum accuracy of 10 trainings, and rank different parameter combinations according to six indicators: average combined accuracy, maximum combined accuracy, average test set accuracy, average test set maximum, validation set accuracy, and validation set maximum. Add up the indicator rankings, and the one with the lowest ranking of all indicators is the best parameter combination. After comprehensive comparison, the parameter combination with batch size of 130 and step of 20000 has the best effect. Some of the result data are detailed in the table below.

[0084]

[0085] In the second round, for the two parameters of the number of nodes in the first hidden layer and the number of nodes in the second hidden layer:

[0086] The number of nodes in the first hidden layer is at least 20, 30, 40, 50 or 60;

[0087] The number of nodes in the second hidden layer is at least 8, 9, 10, 11 or 12.

[0088] The different values ​​of the two parameters, the number of nodes in the first hidden layer and the number of nodes in the second hidden layer, are combined in pairs. Under different combinations of the number of nodes, the deep neural network repeats the training set 10 times each time, and then performs verification and testing after the training is completed. Three indicators, verification accuracy, test accuracy and the combined accuracy of the two, are obtained in a single training.

[0089] Take the average accuracy and maximum accuracy of 10 trainings, and sort different parameter combinations according to six indicators: average combined accuracy, maximum combined accuracy, average test set accuracy, average test set maximum, validation set accuracy, and validation set maximum. Add up the indicator rankings, and the one with the lowest ranking of all indicators is the best parameter combination. After comprehensive comparison, the combination with 50 nodes in the first hidden layer and 10 nodes in the second hidden layer has the best effect.

[0090] The optimization results of the above four parameters are: batch size is 130, step is 20000, the number of nodes in the first hidden layer is 50, and the number of nodes in the second hidden layer is 10; the optimized triple-negative breast cancer prognosis deep neural network was tested multiple times, and the accuracy of the test set was 90%, the accuracy of the validation set was 82.5%, and the combined accuracy of the two was 85%.

[0091] Preferably, Figure 2 As shown, the present invention also provides a breast cancer prognosis prediction system based on a deep neural network, characterized in that it includes:

[0092] A data acquisition module, which acquires gene expression data of triple-negative breast cancer samples, wherein the triple-negative breast cancer samples include samples with completely removed lesions and samples with residual lesions;

[0093] A data filtering module performs data preprocessing on the gene expression data to obtain effective expression quantitative trait loci;

[0094] A site screening module groups the expression quantitative trait sites, and then sequentially screens them using a t-test function, a variance analysis function, and a correlation function to obtain core expression quantitative trait sites;

[0095] A neural network module, using the core expression quantitative trait loci to construct a deep neural network, using the gene expression data to train the deep neural network, and optimizing the parameters of the deep neural network to obtain an optimized triple-negative breast cancer prognosis deep neural network;

[0096] The input-output module inputs the gene expression quantitative trait loci of the triple-negative breast cancer sample to be tested into the triple-negative breast cancer prognosis deep neural network for prognosis recurrence prediction, and converts the prediction results into paper reports or electronic reports.

[0097] The data preprocessing includes: deleting the triple-negative breast cancer samples without phenotypic data; then extracting the gene expression data, deleting duplicate or synonymous gene expression data; then calculating the expression ratio of the gene expression data, and eliminating the gene expression data with an expression ratio less than 30%.

[0098] The grouping of the expression quantitative trait loci is specifically as follows: for the effective expression quantitative trait loci, the expression quantitative trait loci of the sample with completely removed lesions are assigned as array variable X; and the expression quantitative trait loci of the sample with residual lesions are assigned as array variable Y.

[0099] The screening is performed using the t-test function, variance analysis function and correlation function in sequence, specifically:

[0100] Performing a t test on the array variable X and the array variable Y, and retaining expression quantitative trait loci with a P value less than 0.01;

[0101] Then, variance analysis is performed, and the variance values ​​between the array variable X and the array variable Y are sorted and screened, and the expression quantitative trait loci with the P value less than 0.001 are retained;

[0102] The correlation function was then used to calculate the correlation values ​​between the expression quantitative trait loci, and the P value was less than 1.0e -16 expression quantitative trait loci.

[0103] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A breast cancer prognosis prediction method based on deep neural network, characterized in that: The following steps are involved: S1: Obtaining gene expression data of triple-negative breast cancer samples, wherein the triple-negative breast cancer samples include samples with completely removed lesions and samples with residual lesions; S2: performing data preprocessing on the gene expression data to extract effective expression quantitative trait loci; S3: grouping the expression quantitative trait loci, and then screening them using t-test function, variance analysis function and correlation function in turn to obtain core expression quantitative trait loci; S4: constructing a deep neural network using the core expression quantitative trait loci, training the deep neural network using the gene expression data, and optimizing the parameters of the deep neural network to obtain an optimized triple-negative breast cancer prognosis deep neural network; S5: inputting the gene expression quantitative trait loci of the triple-negative breast cancer sample to be tested into the triple-negative breast cancer prognosis deep neural network to perform prognosis recurrence prediction; Wherein, the data preprocessing includes: S21: Delete the data with no phenotype in the triple-negative breast cancer sample, and go to S22; S22: extract gene expression data, delete duplicate or synonymous gene expression data, and go to S23; S23: calculating the expression ratio of the gene expression data, and eliminating gene expression data with an expression ratio less than 30%; The grouping of the expression quantitative trait loci is specifically: For the effective expression quantitative trait loci, the expression quantitative trait loci of the sample with completely removed lesions are assigned as array variable X, and the expression quantitative trait loci of the sample with residual lesions are assigned as array variable Y; The screening is performed using the t-test function, variance analysis function and correlation function in sequence, specifically: S31: performing a t test on the array variable X and the array variable Y, and retaining the expression quantitative trait loci with a P value less than 0.01; S32: performing variance analysis on the expression quantitative trait loci retained in S31, sorting and screening according to the variance values ​​between the array variable X and the array variable Y, and retaining the expression quantitative trait loci with a P value less than 0.001; S33: Calculate the correlation value between each pair of the expression quantitative trait loci retained in S32 using the correlation function, and retain the P value less than 1.0e -16 expression quantitative trait loci.

2. The method for predicting breast cancer prognosis based on a deep neural network according to claim 1, characterized in that: The S3 further includes: The t-test function and the correlation function both adopt a two-tailed test method.

3. The method for predicting breast cancer prognosis based on a deep neural network according to claim 2, characterized in that: The deep neural network is a deep neural network with 2 hidden layers.

4. The method for predicting breast cancer prognosis based on a deep neural network according to claim 3, characterized in that: The parameters include: The number of training samples each time is at least 80, 90, 100, 110, 120, 130 or 140; The training step size is at least 10000, 12500, 15000, 17500 or 20000.

5. The method for predicting breast cancer prognosis based on a deep neural network according to claim 4, characterized in that: The hidden layer also includes: The number of nodes in the first hidden layer is at least 20, 30, 40, 50 or 60; The number of nodes in the second hidden layer is at least 8, 9, 10, 11 or 12.

6. A breast cancer prognosis prediction system based on deep neural network, characterized in that: include: A data acquisition module, which acquires gene expression data of triple-negative breast cancer samples, wherein the triple-negative breast cancer samples include samples with completely removed lesions and samples with residual lesions; A data filtering module performs data preprocessing on the gene expression data to obtain effective expression quantitative trait loci; including deleting data with a data type of no phenotype in the triple-negative breast cancer sample, extracting gene expression data, deleting duplicate or synonymous gene expression data, calculating the expression ratio of the gene expression data, and eliminating gene expression data with an expression ratio of less than 30%; A site screening module is provided for grouping the expression quantitative trait sites, and then screening them in turn with a t-test function, a variance analysis function and a correlation function to obtain core expression quantitative trait sites; wherein, the grouping of the expression quantitative trait sites is specifically as follows: for the effective expression quantitative trait sites, the expression quantitative trait sites of the sample with complete lesion removal are assigned as array variable X, and the expression quantitative trait sites of the sample with residual lesions are assigned as array variable Y; wherein, the screening in turn with a t-test function, a variance analysis function and a correlation function is specifically as follows: performing a t test on the array variable X and the array variable Y, and retaining the expression quantitative trait sites with a P value less than 0.01; performing variance analysis on the retained expression quantitative trait sites, sorting and screening according to the variance value between the array variable X and the array variable Y, and retaining the expression quantitative trait sites with a P value less than 0.001; using the correlation function to calculate the correlation value between the retained expression quantitative trait sites, and retaining the expression quantitative trait sites with a P value less than 1.0e -16 expression quantitative trait loci; A neural network module, using the core expression quantitative trait loci to construct a deep neural network, using the gene expression data to train the deep neural network, and optimizing the parameters of the deep neural network to obtain an optimized triple-negative breast cancer prognosis deep neural network; The input-output module inputs the gene expression quantitative trait loci of the triple-negative breast cancer sample to be tested into the triple-negative breast cancer prognosis deep neural network for prognosis recurrence prediction, and converts the prediction results into paper reports or electronic reports.

Images