A gastric cancer proteomic classification framework identification method based on deep learning feature extraction

Abstract

The present invention belongs to the field of artificial intelligence and proteomics, and specifically relates to a method for identifying a gastric cancer proteomic typing framework based on deep learning feature extraction, the method comprising: obtaining a protein expression spectrum to be identified, and preprocessing the protein expression spectrum; inputting the preprocessed protein expression spectrum into a trained autoencoder, extracting nodes of the bottleneck layer of the autoencoder, and using the nodes to perform consistency clustering to obtain clustering labels of the protein expression spectrum to be identified; training a classifier according to the clustering labels, predicting unknown label samples, and completing gastric cancer proteomic typing framework identification; the present invention uses a deep learning autoencoder to perform consistency clustering on feature extraction of stage II and stage III gastric cancer patients from multiple centers, and obtains molecular subtypes with significant survival differences, and the subtypes are characterized by good prognosis, chemotherapy benefit, and poor prognosis, chemotherapy ineffectiveness.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and proteomics, and specifically relates to a gastric cancer proteomics typing framework recognition method based on deep learning feature extraction. Background Art

[0002] The high incidence and mortality of gastric cancer reflect the inadequacy of its diagnosis and treatment methods. At present, trastuzumab is still the only targeted drug approved as a first-line treatment for patients with HER2-positive advanced gastric cancer, but globally, the HER2-positivity rate of gastric cancer is only between 10.4% and 20.2%. For patients with HER2-negative advanced gastric cancer, chemotherapy is still the main treatment. However, the overall benefit of chemotherapy is limited. Some patients can benefit from chemotherapy and obtain a better prognosis, but some patients may even suffer from the damage of chemotherapy and lead to a worse prognosis. Obviously, not all gastric cancer patients are suitable for chemotherapy, so it is particularly important to distinguish the population that benefits from chemotherapy for the treatment of HER2-negative gastric cancer patients.

[0003] In order to better predict the prognosis of patients and guide individual treatment plans, gastric cancer classification has always been a hot topic for exploration. Although the Lauren classification based on pathological morphology plays a certain guiding role in the treatment of gastric cancer, the efficacy of different Lauren classifications on different chemotherapy drugs varies greatly, and the efficacy of the same drug on different classifications also varies. With the development of high-throughput omics technology, the focus of gastric cancer classification research has shifted from pathological classification to molecular classification. For example, the Cancer Genome Consortium (TCGA) and the Asian Cancer Research Group (ACRG) defined four molecular subtypes based on the genomic and transcriptomic data of gastric cancer, and found that they have their own unique molecular characteristics. For example, the four subtypes defined by TCGA are chromosomal instability (CIN), microsatellite instability (MSI), genomic stability (GS), and Epstein–Barr virus (EBV) positivity.

[0004] However, there are still large differences between genes and proteins, the ultimate executors of life activities, and it is difficult to directly reflect life changes. The gene level reflects more the probability (i.e. possibility) of disease occurrence, while proteins, as effector molecules of genes, directly reflect the phenotype of individuals, and changes in their status directly reflect the occurrence and development of diseases. Summary of the invention

[0005] In order to solve the problems existing in the above prior art, the present invention proposes a method for identifying a gastric cancer proteomic typing framework based on deep learning feature extraction, the method comprising: obtaining a protein expression spectrum to be identified, and preprocessing the protein expression spectrum; inputting the preprocessed protein expression spectrum into a trained classifier to obtain a classification label of the protein expression spectrum to be identified; classifying the protein expression spectrum according to the classification label to complete the identification of the gastric cancer proteomic typing framework;

[0006] The process of training a classifier includes:

[0007] S1: Obtain the patient's protein expression profile dataset and preprocess the protein expression profile in the dataset;

[0008] S2: Use an autoencoder to extract features from the preprocessed protein expression profile to obtain subtype labels;

[0009] S3: Use a consistent clustering algorithm to consistently cluster the features extracted by the autoencoder to obtain subtype labels; divide all subtype labels into a training set and a test set, where the training set is used to train the classifier and the test set is used to test the trained classifier;

[0010] S4: The univariate Cox proportional hazards model was used to select proteins in the training set, and the selected proteins were standardized;

[0011] S5: Input the standardized protein into the classifier for training to obtain a trained classifier.

[0012] Preferably, preprocessing the protein expression profile includes: screening out high-confidence proteins in the protein expression profile data set; removing high-abundance proteins and redundant proteins from the high-confidence proteins; performing comprehensive quantitative data standardization on the high-confidence proteins after deleting the high-abundance proteins and redundant proteins, and removing proteins with an identification frequency lower than 10% of the total number of samples to obtain a standard protein expression profile.

[0013] Preferably, the process of extracting bottleneck layer features from a standard protein expression spectrum using an autoencoder includes: inputting the preprocessed protein expression spectrum data into the autoencoder for nonlinear conversion, and calculating the loss value of the input data and the reconstructed data, when the loss value is smaller, the bottleneck layer node has a stronger explanatory power for the original data, and outputting the bottleneck layer node at this time; the process of nonlinear conversion of the protein expression spectrum data using an autoencoder includes: inputting the standardized protein expression spectrum into the autoencoder, converting it through a relu activation function in the first hidden layer to obtain a hidden node; inputting the hidden node into the bottleneck layer for conversion to obtain a converted bottleneck node; inputting the bottleneck node into the first hidden layer of the decoder, and converting it through a relu activation function to obtain a hidden node of the first decoder hidden layer, inputting the hidden node into the reconstruction layer, and converting it through a sigmiod function to obtain data reconstructed by the autoencoder.

[0014] Furthermore, the objective function of the autoencoder is:

[0015]

[0016] Among them, MSE represents the objective function, x i Represents the input data of the current sample, x i ` represents the output value of the autoencoder, and n represents the total number of samples.

[0017] Preferably, the process of using a consistency clustering algorithm to consistently cluster the features extracted by the automatic encoder includes: the consistency clustering algorithm is a Euclidean k-means clustering algorithm; setting the maximum number of clusters, the ratio of repeated sampling, the number of samplings, and the number of iterations; sampling and clustering the samples according to the set ratio of repeated sampling, and recording the clustering distance between samples in each iteration; when the iteration is completed, calculating the average value of the clustering distance between samples, and using the average value as the final distance matrix; calculating the similarity between samples according to the distance matrix, and obtaining the clustering result according to the sample similarity calculation result; each cluster in the clustering result represents a subtype label.

[0018] Preferably, the classifier adopts a random forest classifier.

[0019] Preferably, the process of selecting proteins in the training set using the univariate Cox hazard ratio model includes: inputting the proteins in the training set into the univariate Cox hazard ratio model for risk assessment, outputting the statistical significance value and risk value of each protein on the prognosis; setting a threshold, comparing the statistical significance value of each protein on the prognosis with the set threshold, and retaining the protein when it is less than the set threshold, otherwise removing the protein from the training set; the formula for processing protein data using the univariate Cox hazard ratio model is:

[0020] h(t)=h0(t)*exp(b1x1+b2x2+…+b n x n )

[0021] Among them, t represents time, h(t) represents the risk of death at that time point, x represents the factor, exp(b i ) represents the HR of the factor.

[0022] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, it implements any of the above-mentioned gastric cancer proteomic typing framework identification methods based on deep learning feature extraction.

[0023] To achieve the above-mentioned purpose, the present invention also provides a gastric cancer proteomic classification framework identification device based on deep learning feature extraction, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the gastric cancer proteomic classification framework identification device based on deep learning feature extraction can perform any of the above-mentioned gastric cancer proteomic classification framework identification methods based on deep learning feature extraction.

[0024] Beneficial effects of the present invention:

[0025] 1) Using a deep learning autoencoder to perform consistent clustering on feature extraction of stage II and stage III gastric cancer patients from multiple centers, molecular subtypes with significant survival differences were obtained, and the subtypes showed good prognosis, chemotherapy benefit, and poor prognosis, chemotherapy ineffectiveness; 2) The patients predicted by the random forest classifier also showed good prognosis, chemotherapy benefit, and poor prognosis, chemotherapy ineffectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flowchart of the gastric cancer proteomic typing framework identification method based on deep learning feature extraction of the present invention;

[0027] Figure 2 It is the structure diagram of the automatic encoder of the present invention;

[0028] Figure 3 It is a schematic diagram of a KM curve of subtype-associated clinical information obtained by consistent clustering of the present invention;

[0029] Figure 4 It is a schematic diagram of the KM curve of the subtype-associated chemotherapy status obtained by consistent clustering of the present invention;

[0030] Figure 5 A schematic diagram of the survival difference of the subtype to which the classifier of the present invention predicts a new sample;

[0031] Figure 6 This is the KM curve of the chemotherapy status associated with the subtype to which the new sample belongs predicted by the classifier of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] The present invention mainly combines deep learning autoencoders to perform feature extraction (dimensionality reduction) for proteomic molecular typing of gastric cancer patients with tumor-node-metastasis staging (TNM stage) in stage II and stage III. The gastric cancer samples are clustered consistently using the features extracted by the autoencoder in order to obtain subtypes with survival differences. In addition, typing gastric cancer samples using the features extracted by the autoencoder can distinguish people who benefit from chemotherapy or who are ineffective or even harmful to chemotherapy. Subsequently, the present invention uses unique features (such as differential proteins) between subtypes to construct a classifier, which can be used to predict samples of unknown subtypes, and obtain output labels for samples that have a good prognosis, are recommended for chemotherapy, or have a poor prognosis, and are not recommended for chemotherapy. According to the prediction results, the treatment of new samples can be guided.

[0034] A gastric cancer proteomic classification framework identification method based on deep learning feature extraction, such as Figure 1 As shown, the method includes: obtaining a protein expression spectrum to be identified and preprocessing the protein expression spectrum; inputting the preprocessed protein expression spectrum into a trained autoencoder for feature extraction, and obtaining subtype labels based on consistency clustering; establishing a classifier according to the classification label to classify the protein expression spectrum, and completing the identification of the gastric cancer proteomic classification framework.

[0035] The process of training a classifier includes:

[0036] S1: Obtain the patient's protein expression profile dataset and preprocess the protein expression profile in the dataset;

[0037] S2: Use autoencoder to extract features from the preprocessed protein expression profile;

[0038] S3: Use a consistent clustering algorithm to consistently cluster the features extracted by the autoencoder to obtain subtype labels; divide the subtype labels into a training set and a test set, where the training set is used to train the classifier and the test set is used to test the trained classifier;

[0039] S4: The univariate Cox proportional hazards model was used to select proteins in the training set, and the selected proteins were standardized;

[0040] S5: Input the standardized protein into the classifier for training to obtain a trained classifier.

[0041] Preprocessing of protein expression profiles includes: screening out high-confidence proteins in the protein expression profile data set; removing high-abundance proteins and redundant proteins from the high-confidence proteins; performing comprehensive quantitative data standardization on the high-confidence proteins after removing the high-abundance proteins and redundant proteins, and removing proteins with identification frequencies lower than 10% of the total number of samples to obtain a standard protein expression profile.

[0042] like Figure 2 As shown in the figure, the autoencoder is a butterfly-shaped symmetrical structure network, which includes two parts: encoder (input layer to bottleneck layer part) and decoder (bottleneck layer's next layer to reconstructed layer part). The encoder is mainly used to learn the implicit features of the input data (this implicit feature is expressed as the nodes of the bottleneck layer), and the decoder is mainly used to reconstruct the input data using the implicit features obtained by the encoder.

[0043] The process of using an autoencoder to extract bottleneck layer features of a standard protein expression spectrum includes: inputting the preprocessed protein expression spectrum data into the autoencoder for nonlinear conversion, and calculating the loss value of the input data and the reconstructed data. When the loss value is smaller, the bottleneck layer node has a stronger explanatory power for the original data, and the bottleneck layer node at this time is output; the process of using an autoencoder to perform nonlinear conversion on the protein expression spectrum data includes: inputting the standardized protein expression spectrum into the autoencoder, converting it through the relu activation function in the first hidden layer to obtain a hidden node; inputting the hidden node into the bottleneck layer for conversion to obtain a converted bottleneck layer node; inputting the bottleneck layer node into the first hidden layer of the decoder, and converting it through the relu activation function to obtain a hidden node of the first hidden layer of the decoder, inputting this hidden node into the reconstruction layer, and converting it through the sigmiod function to obtain the data reconstructed by the autoencoder.

[0044] A specific implementation method of extracting bottleneck layer features from a standard protein expression spectrum using an autoencoder includes inputting the preprocessed protein expression spectrum data into the autoencoder for nonlinear conversion, calculating the loss value of the input data and the reconstructed data, and outputting the bottleneck layer node at this time when the loss value is smaller. First, the standardized protein expression spectrum is input into the autoencoder, and the first hidden layer is converted by the relu activation function (here, all values ​​less than 0 are converted to 0, and values ​​greater than 0 remain unchanged), and finally 500 hidden nodes are output from the first hidden layer to the bottleneck layer, and the bottleneck layer converts these 500 nodes into 100 nodes. Next, the 100 nodes converted by the bottleneck layer are input into the first hidden layer of the decoder, and the result is also converted by the relu activation function, and 500 nodes are output to the reconstruction layer. Finally, the reconstruction layer outputs nodes with the same dimension as the input expression spectrum and sigmiod conversion.

[0045] The preprocessed gastric cancer protein expression profile is standardized and used as the input of the autoencoder, and the nodes of the bottleneck layer of the autoencoder are taken as new features. The autoencoder has 3 hidden layers, with 500, 100, and 500 nodes respectively. The hidden layer uses relu as its activation function, but the bottleneck layer does not set any activation function. The calculation formula of the relu activation function is as follows:

[0046] y=f(x)=max(0,x)

[0047] Among them, x represents the input data of the current sample.

[0048] For the reconstruction layer, sigmoid is selected as its activation function, and the expression of the activation function is:

[0049]

[0050] For the autoencoder network, the mean square error (MSE) is selected as its objective function (loss function) to evaluate the error between the input X and the output X'. The MSE calculation formula is as follows:

[0051]

[0052] Among them, MSE represents the objective function, x i Represents the input data of the current sample, x` i Represents the output value of the autoencoder, and n represents the total number of samples.

[0053] After extracting the nodes of the bottleneck layer of the autoencoder, the R package ConsensusClusterPlus was used to perform consistency clustering. The clustering algorithm uses k-means based on Euclidean distance. The proportion of samples selected for each resampling is 80%. Specifically, the consistency clustering algorithm is a k-means clustering algorithm based on Euclidean distance; set the maximum number of clusters to 5, the proportion of repeated sampling to 80%, and the number of sampling to 200; sample and cluster the samples according to the set proportion of repeated sampling, and record the cluster distance between samples in each iteration; when the iteration is completed, calculate the average value of the cluster distance between samples, and use the average value as the final distance matrix; calculate the similarity between samples based on the distance matrix, and obtain the clustering result based on the sample similarity calculation result; each cluster in the clustering result represents the subtype label.

[0054] The Kaplan-Meier method was used to perform association analysis between the subtypes obtained by consistent clustering and the corresponding clinical information of the patients, explaining the prognostic differences between subtypes and the patients' responses to chemotherapy drugs (chemotherapy benefit, chemotherapy ineffectiveness, and chemotherapy harm).

[0055] The process of selecting proteins in the training set using the univariate Cox hazard ratio model includes: inputting the proteins in the training set into the univariate Cox hazard ratio model for risk assessment, outputting the statistical significance value and risk value of each protein on prognosis; setting a threshold, comparing the statistical significance value of each protein on prognosis with the set threshold, and retaining the protein if it is less than the set threshold, otherwise removing the protein from the training set; the formula for processing protein data using the univariate Cox hazard ratio model is:

[0056] h(t)=h0(t)*exp(b1x1+b2x2+…+b n x n )

[0057] Among them, t represents time, h(t) represents the risk of death at that time point, x represents the factor, exp(b i ) represents the HR of the factor. (HR=1: no effect, HR<1: reduced risk, HR>1: increased risk) The CoxPHFitter function of Python is used to calculate the risk value and p-value of each protein, and finally the statistical matrix (partial statistical results) in the following format is output:

[0058] coef exp(coef) p -log2(p) lower0.95 upper 0.95 protein 0.000104 1.000104 0.977497 0.032835 -0.007123 0.007331

[0059] Among them, exp(coef) is the HR of the protein, p is the statistical significance of the protein for prognosis, and lower and upper are the confidence intervals of HR.

[0060] The data in the gastric cancer proteomic typing framework recognition method based on deep learning feature extraction of the present invention comes from 833 cases of surgical resection samples of stage II and stage III gastric cancer fixed in formalin and embedded in paraffin (FFPE) from multiple centers. First, the 833 samples were divided into a discovery set and an independent validation set, and then the discovery set was randomly divided into a training set and a test set at a ratio of 7:3. Using this framework, it was found that about 43% of the patients in the discovery set benefited from adjuvant chemotherapy, and the prognosis of this group of patients was significantly better than those who did not benefit from chemotherapy. That is, the five-year survival rate of patients with a better prognosis who received chemotherapy was 12 percentage points higher than that of patients who did not receive chemotherapy, and the five-year survival rate of patients with a poorer prognosis who received chemotherapy and those who did not receive chemotherapy was not significantly improved.

[0061] Subsequently, the classifier was constructed using the molecular features of the two subtypes, and the classifier was trained on the training set combined with 10-fold cross-validation, and the classifier performance was tested on the test set. Finally, the stability of the performance of molecular typing to predict the prognosis of gastric cancer patients and the number of people who benefit from chemotherapy was verified on the independent validation set by extracting features through the autoencoder. Finally, the subtypes of the samples in the independent validation set were predicted by constructing a classifier, and the subtype labels corresponding to the samples in the independent validation set were output and associated with clinical information. We found that the subtype labels predicted by the samples in the independent validation set also had survival differences, and the five-year survival rate of patients with a good prognosis who received chemotherapy was 25 percentage points higher than that of patients who did not receive chemotherapy. Similarly, there was no significant difference between patients with a poor prognosis who received chemotherapy and those who did not receive chemotherapy.

[0062] like Figure 3 As shown, the present invention can distinguish gastric cancer patients with different prognoses by extracting features through an automatic encoder. Prognosis mainly concerns the patient's survival status and survival time. By drawing a KM curve, the prognosis of a subtype can be more intuitively displayed. The information obtained from the figure is that subtype 1 patients have a better prognosis than subtype 2, and have a longer survival time after surgery.

[0063] In addition to distinguishing patients with different prognoses, the present invention also hopes to distinguish people who benefit from chemotherapy. According to clinical manifestations, it is expected that patients will have a better prognosis due to the therapeutic effect of chemotherapy, or that patients will have a worse prognosis due to the toxic side effects of chemotherapy. Therefore, the subtypes are separately associated with chemotherapy status analysis, and the prognosis of patients after chemotherapy is also intuitively displayed through the KM curve. Figure 4 The left side shows the survival curve of subtype 1 patients associated with chemotherapy status. It can be seen from the figure that the survival rate of patients in subtype 1 who received chemotherapy was significantly higher than that of patients who did not receive chemotherapy. The five-year survival rate of patients who received chemotherapy was 65.3%, while that of patients who did not receive chemotherapy was 52.6%, an overall increase of 12 percentage points. Figure 4On the right is the survival curve of subtype II patients associated with chemotherapy status. The five-year survival rate of patients who received chemotherapy was 54.2%, while that of patients who did not receive chemotherapy was 51.1%, which showed no significant improvement.

[0064] The subtypes obtained by consistent clustering on the discovery set ultimately need to establish a classifier for prediction. Only when the classifier predicts that the samples also show good prognosis and chemotherapy benefit for subtype 1, and poor prognosis and chemotherapy ineffectiveness for subtype 2, can it be said that the subtypes obtained above are reliable and the method is reliable as a whole. Figure 5 Figure 6 Basically the same Figure 4 Figure 5 ,only Figure 5-6 is the result on the prediction set.

[0065] In one embodiment of the present invention, the present invention also includes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned methods for identifying a gastric cancer proteomic classification framework based on deep learning feature extraction.

[0066] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to the computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other media that can store program codes.

[0067] A gastric cancer proteomic classification framework recognition device based on deep learning feature extraction, comprising a processor and a memory; the memory is used to store computer programs; the processor is connected to the memory and is used to execute the computer programs stored in the memory, so that the gastric cancer proteomic classification framework recognition device based on deep learning feature extraction performs any of the above-mentioned gastric cancer proteomic classification framework recognition methods based on deep learning feature extraction.

[0068] Specifically, the memory includes: ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk, etc., which can store program codes.

[0069] Preferably, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0070] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A gastric cancer proteomic typing framework recognition method based on deep learning feature extraction, characterized in that: include: Obtaining a protein expression profile to be identified and preprocessing the protein expression profile; The preprocessed protein expression profile is input into the autoencoder to extract the bottleneck layer features, and the extracted bottleneck layer features are clustered by the consistency clustering algorithm to obtain the subtype label; Input the subtype labels into the classifier to obtain the classification results of the subtype labels, and complete the identification of the gastric cancer proteomics classification framework; The process of training a classifier includes: S1: Obtain the patient's protein expression profile dataset and preprocess the protein expression profile in the dataset; S2: An automatic encoder is used to extract bottleneck layer features from the preprocessed protein expression profile; S3: Use a consistent clustering algorithm to consistently cluster the features extracted by the autoencoder to obtain subtype labels; divide all subtype labels into a training set and a test set, where the training set is used to train the classifier and the test set is used to test the trained classifier; S4: The univariate Cox proportional hazards model was used to select proteins in the training set, and the selected proteins were standardized; S5: Input the standardized protein into the classifier for training to obtain a trained classifier.

2. According to claim 1, a gastric cancer proteomic typing framework recognition method based on deep learning feature extraction is characterized in that: Preprocessing of protein expression profiles includes: screening out high-confidence proteins in the protein expression profile data set; removing high-abundance proteins and redundant proteins from the high-confidence proteins; performing comprehensive quantitative data standardization on the high-confidence proteins after removing the high-abundance proteins and redundant proteins, and removing proteins with identification frequencies lower than 10% of the total number of samples to obtain a standard protein expression profile.

3. The method for identifying a gastric cancer proteomics typing framework based on deep learning feature extraction according to claim 1, characterized in that: The process of using an autoencoder to extract bottleneck layer features of a standard protein expression spectrum includes: inputting the preprocessed protein expression spectrum data into the autoencoder for nonlinear conversion, and calculating the loss value of the input data and the reconstructed data. When the loss value is smaller, the bottleneck layer node has a stronger explanatory power for the original data, and the bottleneck layer node at this time is output; the process of using an autoencoder to perform nonlinear conversion on the protein expression spectrum data includes: inputting the standardized protein expression spectrum into the autoencoder, converting it through the relu activation function in the first hidden layer to obtain a hidden node; inputting the hidden node into the bottleneck layer for conversion to obtain a converted bottleneck node; inputting the bottleneck node into the first hidden layer of the decoder, and converting it through the relu activation function to obtain a hidden node of the first decoder hidden layer, inputting this hidden node into the reconstruction layer, and converting it through the sigmiod function to obtain reconstructed data.

4. The method for identifying a gastric cancer proteomic typing framework based on deep learning feature extraction according to claim 3, characterized in that: The loss formula for input data and reconstructed data is: Among them, MSE represents the objective function, x i Represents the input data of the current sample, x i ` represents the output value of the autoencoder, and n represents the total number of samples.

5. The method for identifying a gastric cancer proteomic typing framework based on deep learning feature extraction according to claim 1, characterized in that: The process of using a consistent clustering algorithm to consistently cluster the features extracted by the autoencoder includes: the consistent clustering algorithm is a Euclidean k-means clustering algorithm; setting the maximum number of clusters, the ratio of repeated sampling, and the number of samplings; sampling and clustering the samples according to the set ratio of repeated sampling, and recording the clustering distance between samples in each iteration; when the iteration is completed, calculating the average value of the clustering distance between samples, and using the average value as the final distance matrix; calculating the similarity between samples based on the distance matrix, and obtaining the clustering result based on the sample similarity calculation result; each cluster in the clustering result represents a subtype label.

6. The method for identifying a gastric cancer proteomic typing framework based on deep learning feature extraction according to claim 1, characterized in that: The classifier used is random forest classifier.

7. The method for identifying a gastric cancer proteomic typing framework based on deep learning feature extraction according to claim 1, characterized in that: The process of selecting proteins in the training set using the univariate Cox hazard ratio model includes: inputting the proteins in the training set into the univariate Cox hazard ratio model for risk assessment, outputting the statistical significance value and risk value of each protein on prognosis; setting a threshold, comparing the statistical significance value of each protein on prognosis with the set threshold, and retaining the protein if it is less than the set threshold, otherwise removing the protein from the training set; the formula for processing protein data using the univariate Cox hazard ratio model is: h(t)=h0(t)*exp(b1x1+b2x2+…+b n x n ) Among them, t represents time, h(t) represents the risk of death at that time point, x represents the factor, exp(b i ) represents the HR of the factor.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the method for identifying a gastric cancer proteomic typing framework based on deep learning feature extraction according to any one of claims 1 to 7.

9. A gastric cancer proteomics classification framework recognition device based on deep learning feature extraction, characterized in that: It comprises a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the gastric cancer proteomic classification framework recognition device based on deep learning feature extraction executes any one of the gastric cancer proteomic classification framework recognition methods based on deep learning feature extraction in claims 1 to 7.

Images