Gastric cancer diagnosis method based on gastric cancer gene target and neural network

Abstract

The present application relates to a gastric cancer diagnosis method based on gastric cancer gene targets and neural networks, comprising: obtaining gastric cancer samples and general sample data, preprocessing; and establishing a neural network model, obtaining patient gene expression data for diagnosis. The beneficial effects of the present application are: the present application establishes a neural network model based on gene targets INHBA, LYVE1, CD36 and COL10A1, and proposes a low-cost, convenient and minimally invasive method for preliminary screening of gastric cancer, so as to reduce the situation that the detection is abandoned by the detected person due to high cost, strong trauma and inconvenience of methods such as gastroscopy, resulting in delay of treatment opportunity.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biological information, more particularly, it relates to a gastric cancer diagnosis method based on gastric cancer gene targets and neural networks. BACKGROUND

[0002] Gastric cancer is one of the most important cancers in the world, with high morbidity and mortality. Gastric cancer usually has no obvious symptoms in the early stage, so it is often detected in the late stage. However, the median survival of advanced gastric cancer is generally 8 to 12 months, so it is very important to diagnose gastric cancer as soon as possible.

[0003] Gastroscopy is the standard diagnostic method for gastric cancer, but it is costly, invasive and inconvenient for patients. Many asymptomatic patients do not undergo gastroscopy due to the inconvenience of gastroscopy, delaying the opportunity for treatment. Gene-based detection methods are less invasive and less traumatic, and the fewer types of genes they rely on, the lower the cost. Gene detection based on specific targets is an excellent preliminary auxiliary diagnostic method for gastric cancer. SUMMARY

[0004] The purpose of the present application is to address the high cost and inconvenience of gastric cancer detection, and to provide a gastric cancer diagnosis method based on gastric cancer gene targets and neural networks.

[0005] In a first aspect, a gastric cancer diagnosis method based on gastric cancer gene targets and neural networks is provided, comprising:

[0006] Step 1, data acquisition: a certain number of gastric cancer patient samples and healthy person samples are obtained, including their health status labels and INHBA, LYVE1, CD36, and COL10A1 4 gene expression data;

[0007] Step 2, data preprocessing: for the data in step 1, obtain grouping information, standardize the data, obtain data annotation information based on the platform, complete the conversion of probes and corresponding gene names and the deduplication of identical genes, take the maximum value of repeated genes, and normalize the data;

[0008] Step 3, model training: based on the preprocessed data, a neural network model is established and trained;

[0009] Step 4, obtaining patient gene expression data: for the patient to be detected, obtain the INHBA, LYVE1, CD36, and COL10A1 4 gene expression data;

[0010] Step 5, diagnosis: for the gene expression data in step 4, input the normalized data into the neural network model to obtain the prediction result;

[0011] Step 6, reuse: for new patients, return to step 4.

[0012] As preferred, the normalization method in step 2 and step 5 is max-min normalization between each sample data dimension.

[0013] As preferred, in step 3, the neural network model is a five-layer fully connected layer neural network model.

[0014] As preferred, in step 3, the neural network model uses cross-entropy as the loss function, and the weight ratio is the class ratio in the training data.

[0015] As preferred, in step 3, the neural network model uses ReLU function as the activation function, and its formula is as follows:

[0016] ReLU(x) = max(0, x)

[0017] Where x represents the input of the activation function, max(0, x) represents taking the maximum value, and the activation function is placed after the fully connected layer.

[0018] As preferred, in step 3, the neural network model adopts the forward propagation algorithm, and in the fully connected layer, the value of each neuron in the current layer is obtained by multiplying the value of each neuron in the previous layer by the corresponding coefficient matrix and bias and then passing through the activation function, which is represented as:

[0019]

[0020] Where X represents the input of the current neural network layer, W represents the coefficient matrix, b represents the bias of the neural network, and a represents the calculated neural value. represents the activation function.

[0021] As preferred, in step 3, the neural network model is optimized by the backpropagation algorithm, which is represented as:

[0022] argminloss(w, b)

[0023] Where w represents the coefficient of the neural network, b represents the bias of the neural network, and loss represents the loss function.

[0024] Therefore, the following needs to be calculated: and We can get:

[0025]

[0026]

[0027]

[0028]

[0029] wherein lr represents a learning rate; z represents the output of the last layer; w * and b * represent the updated w and b.

[0030] As a preference, when the loss function is cross-entropy, The calculation formula is as follows:

[0031] loss (a) = CE (softMax (a))

[0032]

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] wherein CE represents cross-entropy; y represents the output of each bit of the softMax function; j and i and k represent the number of each bit; and n represents the number of bits.

[0040] In a second aspect, a gastric cancer diagnosis device based on gene targets and neural networks is provided, which is used to execute the gastric cancer diagnosis method based on gene targets and neural networks of any one of the first aspect, and comprises:

[0041] A first acquisition module is configured to acquire a certain number of gastric cancer patient samples and healthy person samples, including health status labels and INHBA, LYVE1, CD36, and COL10A1 four gene expression data;

[0042] A preprocessing module is configured to acquire grouping information for the data in the acquisition module, perform standardization processing on the data, acquire data annotation information based on a platform, complete conversion of probes and corresponding gene names and deduplication of the same genes, take the maximum value of repeated genes, and perform normalization processing on the data.

[0043] A training module is configured to establish and train a neural network model based on the preprocessed data.

[0044] A second acquisition module is configured to acquire INHBA, LYVE1, CD36, and COL10A1 four gene expression data of a patient to be detected.

[0045] a diagnosis module configured to normalize the gene expression data in the second acquisition module and input the normalized gene expression data into a neural network model to obtain a prediction result.

[0046] a repetition module configured to return to the second acquisition module for a new patient.

[0047] In a third aspect, a computer storage medium is provided, and the computer storage medium stores a computer program.

[0048] The present application has the advantages that the present application establishes a neural network model based on the gene targets INHBA, LYVE1, CD36, and COL10A1, and proposes a low-cost, convenient, and low-invasive method for preliminary screening of gastric cancer, so as to reduce the situation that the detection is abandoned due to high cost and strong invasiveness of gastroscopy and other methods, and the treatment opportunity is delayed. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 a flowchart of the process of establishing the neural network diagnosis model and testing in the present application;

[0050] Figure 2 a testing result of the diagnosis model on GSE29272.

[0051] Figure 3 a testing result of the diagnosis model on GSE29272. DETAILED DESCRIPTION

[0052] The present application will be further described below in conjunction with the embodiments. The following description of the embodiments is only for the purpose of understanding the present application. It should be noted that those skilled in the art can make several modifications to the present application without departing from the principles of the present application, and these modifications and improvements also fall within the scope of protection of the claims of the present application.

[0053] As an embodiment, the present application provides a gastric cancer diagnosis method based on gastric cancer gene targets and a neural network, Figure 1 which shows the overall implementation process of the present method. Figure 1 As shown in the above table, the present method includes the steps of collecting data, modeling training, five-fold cross-validation, and independent dataset verification, including:

[0054] Step 1, data acquisition: a certain number of gastric cancer patient samples and healthy person samples are obtained, including their health status labels and INHBA, LYVE1, CD36, and COL10A1 four gene expression data.

[0055] In step 1, samples can be obtained through the open-source dataset GSE66229. The dataset is publicly available at https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=gse66229 and contains 300 gastric cancer samples and 100 general samples.

[0056] Step 2, Data Preprocessing: For the data in Step 1, obtain the grouping information, standardize the data, obtain the data annotation information based on the GPL570 platform, complete the conversion of probe and corresponding gene names and remove duplicates of the same genes, take the maximum value of duplicate genes, and normalize the data.

[0057] Step 3: Train the model: Based on the preprocessed data, build and train the neural network model.

[0058] In step 3, the neural network model is a five-layer fully connected neural network model. The neural network model uses cross-entropy as the loss function and the weight ratio is the proportion of classes in the training data. Furthermore, the neural network model uses the ReLU function as the activation function, and its formula is as follows:

[0059] ReLU(x) = max(0,x)

[0060] Here, x represents the input to the activation function, and max(0,x) represents taking the maximum value. The activation function is generally placed after a fully connected layer or other linear structure, making the neural network a non-linear structure that can well fit non-linear relationships.

[0061] The neural network model uses the forward propagation algorithm. In the fully connected layer, the value of each neuron in the current layer is obtained by multiplying the values ​​of each neuron in the previous layer as input by the corresponding coefficient matrix and the bias, and then passing the result through the activation function. This can be represented as:

[0062]

[0063] Where X represents the input to the current neural network layer, W represents the coefficient matrix, b represents the bias of the neural network, and α represents the calculated value of the neuron. This represents the activation function.

[0064] The forward propagation process involves inputting the data to be input into a neural network, which then processes the data to produce an output, such as a classification result or a regression result. In a well-trained and effective neural network, an appropriate structure and appropriately trained parameters such as W and b ensure that the input, in this invention, represents gene expression values, and in most cases, the output represents disease conditions.

[0065] Neural network models can be optimized using the backpropagation algorithm, as follows:

[0066] argmin loss(w,b)

[0067] where w represents the coefficients of the neural network, b represents the bias of the neural network, and loss represents the loss function;

[0068] Therefore, the following needs to be calculated and It can be obtained that

[0069]

[0070]

[0071]

[0072]

[0073] where lr represents the learning rate, z represents the output of the last layer, and w * and b * represent the updated w and b.

[0074] When the loss function is cross-entropy, the following formula is used The calculation formula is as follows:

[0075] loss(a) = CE(softMax(a))

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] where CE represents cross-entropy, y represents the output of each bit of the softMax function, j and i and k represent the number of each bit, and n represents the number of bits. The back propagation process is a process in which the neural network adjusts the parameters such as W and b in order to obtain correct results in the forward propagation and to minimize the loss function. For this purpose, the derivatives of the parameters with respect to the loss function are calculated and used as the basis for updating the parameter values.

[0084] Step 4, obtaining patient gene expression data: For the patient to be detected, obtain the INHBA, LYVE1, CD36, and COL10A1 4 gene expression data of the patient.

[0085] Step 5, diagnosis: For the gene expression data in step 4, input the normalized neural network model to obtain the prediction result. The diagnosis effect in step 5 is verified by cross-validation and independent data set, as shown in Figures 2-3

[0086] In addition, the normalization method in step 2 and step 5 is: for each sample data dimension, maximum and minimum normalization is performed.

[0087] Step 6, reuse: For new patients, return to step 4.

[0088] Step 4 to step 5 uses open source data set GSE26899 and GSE29272 data as patient data for testing. Figure 2 And Figure 3 The ROC curve of the model on the independent data set GSE26899 and the independent data set GSE29272 is shown, and the AUC value and the accuracy are marked. In the ROC curve, the horizontal axis is the probability of false samples being judged as correct samples, that is, 1-precision, and the vertical axis is the recall rate. The closer the curve is to the upper left corner, the higher the precision and accuracy can be maintained at the same time, and the better the model effect. The AUC value is the probability that a positive sample is more likely to be a positive sample than a negative sample from the test sample. Figures 2-3 The test results in steps 4 to 5 show that the model has high accuracy.​

Claims

1. A gastric cancer diagnostic method based on gastric cancer gene targets and neural networks, characterized in that, include: Step 1: Data Acquisition: Obtain a certain number of gastric cancer patient samples and healthy person samples, including their health status labels and expression data of four genes: INHBA, LYVE1, CD36, and COL10A1. Step 2, Data Preprocessing: For the data in Step 1, obtain the grouping information, standardize the data, obtain the data annotation information based on the platform, complete the conversion of probe and corresponding gene names and remove duplicates of the same gene, take the maximum value of duplicate genes, and normalize the data. Step 3: Train the model: Based on the preprocessed data, build and train the neural network model; In step 3, the neural network model uses the ReLU function as the activation function, and its formula is as follows: Where x represents the input to the activation function, max(0,x) represents taking the maximum value, and the activation function is placed after the fully connected layer; In step 3, the neural network model employs a forward propagation algorithm. In the fully connected layer, the value of each neuron in the current layer is obtained by multiplying the values ​​of each neuron in the previous layer as input by the corresponding coefficient matrix and bias, and then passing the result through an activation function, as shown below: Where X represents the input of the current neural network layer, Represents the coefficient matrix. Indicates the bias of the neural network; This represents the numerical value of the calculated neuron. Indicates the activation function; In step 3, the neural network model is optimized using the backpropagation algorithm, as follows: in, The coefficients of the neural network, This represents the bias of the neural network. Represents the loss function; Therefore, calculation is required. and We can obtain: in, Indicates the learning rate; This indicates the output of the last layer; and Indicates the updated and ; Step 4: Obtain patient gene expression data: For the patient to be tested, obtain the expression data of four genes: INHBA, LYVE1, CD36, and COL10A1. Step 5, Diagnosis: The gene expression data from Step 4 is normalized and then input into the neural network model to obtain the prediction results; Step 6, Reuse: For new patients, return to step 4.

2. The gastric cancer diagnosis method based on gastric cancer gene targets and neural networks according to claim 1, characterized in that, The normalization method in steps 2 and 5 is to perform maximum and minimum value normalization for each dimension of the sample data.

3. The gastric cancer diagnosis method based on gastric cancer gene targets and neural networks according to claim 2, characterized in that, In step 3, the neural network model is a five-layer fully connected neural network model.

4. The gastric cancer diagnosis method based on gastric cancer gene targets and neural networks according to claim 3, characterized in that, In step 3, the neural network model uses cross-entropy as the loss function and the weight ratio is the proportion of classes in the training data.

5. The gastric cancer diagnosis method based on gastric cancer gene targets and neural networks according to claim 4, characterized in that, when When the function is cross-entropy, The calculation formula is as follows: in, Represents cross-entropy; express The function outputs each bit. and and Indicates the number of each digit; Indicates the number of digits.

6. A gastric cancer diagnostic device based on gastric cancer gene targets and neural networks, characterized in that, The method for performing the gastric cancer diagnosis method based on gastric cancer gene targets and neural networks according to any one of claims 1 to 5 includes: The first acquisition module is used to acquire a certain number of gastric cancer patient samples and healthy person samples, including their health status labels and expression data of four genes: INHBA, LYVE1, CD36, and COL10A1. The preprocessing module is used to obtain grouping information from the data in the acquisition module, perform data standardization, obtain data annotation information based on the platform, complete the conversion between probe and corresponding gene name, remove duplicates of the same gene, take the maximum value of duplicate genes, and perform data normalization. The training module is used to build and train neural network models based on preprocessed data; The second acquisition module is used to acquire data on the expression of four genes, INHBA, LYVE1, CD36, and COL10A1, for the patient to be tested. The diagnostic module is used to normalize the gene expression data from the second acquisition module and input it into the neural network model to obtain prediction results. The repeat module is used to return to the second acquisition module for new patients.

7. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to execute the gastric cancer diagnosis method based on gastric cancer gene targets and neural networks as described in any one of claims 1 to 5.

Images