Multimodal critical boundary biomarker identification method

Abstract

The application discloses a multi-modal critical edge biomarker identification method, takes an individual cancer patient as a dynamic network system, combines a dynamic network theory biomarker theory and a multi-modal evolutionary algorithm, performs hidden space search by using a restricted Boltzmann machine on the basis of an MMPDNB model, and is a new multi-modal PDENB identification model. Firstly, a PEN of the cancer individual patient is constructed. Then, an optimization objective function is designed. Finally, a multi-modal optimization algorithm is used to search for a PDENB set. The application can not only promote the researches on a mathematical model and an algorithm design of the PDENB identification problem, but also help to understand the individual heterogeneity of the cancer, and realize the early diagnosis and treatment of the cancer individual patient.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cancer individual biomarker identification, and particularly relates to a multi-modal critical edge biomarker identification method. BACKGROUND

[0002] With the gradual acceleration of population aging, industrialization and urbanization in China, accompanied by a series of reasons such as unhealthy lifestyle and environmental exposure, cancer has become one of the main causes of death of Chinese residents in recent years. The development process of cancer can be roughly divided into normal state, critical state and disease state. Among them, the critical state is also called critical state or pre-disease state, which widely exists in various systems. A large number of studies have shown that in the progress of complex diseases, deterioration is not smooth but very sudden, and can cause the system to mutate from one state to another at the critical point. Traditional molecular biomarkers (a group of genes or proteins) usually detect whether patients are in a normal state or a disease state based on the differential expression of genes in normal samples and disease samples, ignoring the detection of critical state. The main reason for detecting the critical state of cancer development is that once the development of cancer cells approaches a critical point or is after the critical point, it usually causes the spread and metastasis of cancer cells. At the same time, it is very difficult to predict the critical state of cancer because from the perspective of gene expression, the change of the tissue is slow and weak before the tissue gradually transitions from the normal state to the critical threshold. Therefore, identifying a marker for detecting the critical state of cancer patients for early diagnosis and treatment has important clinical significance. Medical research shows that cancer individual heterogeneity is one of the important reasons why cancer treatment cannot make substantial breakthroughs, that is, the expression degree of cancer mutation genes is different in the same period of different individual patients with the same cancer, and the genes as biomarkers are different. Therefore, the mining and analysis of individual biomarkers of cancer patients are helpful to understand the individual heterogeneity of cancer disease and provide technical support for precision medicine of cancer.

[0003] At present, many studies take the network biomarkers of individual cancer patients as the basis for individual disease judgment, and quantify the biological network markers of patient samples to detect the critical state in the development of cancer. Related studies have proved that multi-modal problems widely exist in the field of biomedicine, and the existence of multi-modal network biomarkers can provide more choices for decision makers.

[0004] In recent years, researchers have realized that cancer is a complex disease caused by the co-mutation of many essential genes, and have begun to focus on edge biomarker (i.e. the interaction of two genes) identification. In particular, the synthetic lethality of edge biomarkers not only helps individualize cancer treatment, but also helps explore the anti-cancer potential of other disease drugs. As an edge biomarker, the synthetic lethality edge biomarker is defined as two non-lethal genes, which will cause cell death if they are simultaneously mutated or not expressed in a cell. Due to its selective lethal effect on cancer cells containing specific gene mutations, targeting synthetic lethal genes is expected to broaden the way of precision medicine compared with cytotoxic chemotherapy. Therefore, it is necessary to pay more attention to the identification of personalized dynamic edge-network biomarkers (PDENB) to detect early warning signals of the critical state of individual patients and promote the understanding of tumor heterogeneity in the system level of cancer precision medicine. However, for complex biological networks, the size of the edge network is generally much larger than that of the node network, which will cause the existing multi-modal optimization algorithm to easily fall into local optimum without large-scale optimization strategy. Therefore, how to introduce appropriate large-scale processing technology and design appropriate multi-modal optimization algorithm to solve the problem of multi-modal PDENB identification has important significance for promoting early diagnosis and individualized treatment of complex diseases. SUMMARY

[0005] In order to solve the above multi-modal PDENB identification problem, the present application provides a multi-modal critical edge biomarker identification method.

[0006] The technical scheme adopted by the present application is:

[0007] The multi-modal critical edge biomarker identification method comprises the following steps:

[0008] S1, constructing the PEN of the individual patient with cancer;

[0009] S2, designing an optimization objective function;

[0010] S3, using a multi-modal optimization algorithm to find a set of PDENB.

[0011] Further, in step S1, the method for constructing the PEN of the individual patient with cancer is:

[0012] S11, using the gene expression data of the normal sample and the tumor sample of the individual patient with cancer and the gene expression data of the reference sample normal tissue and the somatic mutation data, using a paired single-sample network to construct a PGIN;

[0013] S12, select gene pairs with strong correlation in PGIN as candidate edges to construct PEN; select genes with inter-gene pPCC greater than the average pPCC of the entire network as candidate gene pairs to construct edge networks.

[0014] S13, Calculate candidate gene pairs that are linked in PGIN ( <x i ,x j >, <x k ,x l The individualized high-order Pearson correlation coefficient (phPCC) between the two factors is as follows:

[0015]

[0016]

[0017] Where shPCC is the higher-order Pearson correlation coefficient for a single sample. V is the higher-order average expression level of gene i in the reference sample. r (x i ) represents the higher-order variance of gene i's expression on the reference sample; shPCC Normal shPCC is the higher-order Pearson correlation coefficient for a single sample of normal tissue. Tumor It is the higher-order Pearson correlation coefficient of a single sample of tumor tissue;

[0018] S14 uses candidate gene pairs as "nodes" in the edge network and phPCC between candidate gene pairs as "edges" in the edge network to construct PENs for individual cancer patients at different stages of cancer.

[0019] Furthermore, in step S2, the objective function is optimized as follows:

[0020]

[0021]

[0022] Where, the solution X = (x1, x2, ..., x n ,) is a binary decision vector, where n represents the number of gene pairs in PEN; X can represent which gene pairs are selected as molecules in the module or are not selected. If gene pair i is selected as a molecule in the module, then x i =1, otherwise x i =0; pDEin(X) represents the standard deviation of phPCC between gene pairs within the PDENB module; |phPCC in(X)| represents the absolute value of the average phPCC between gene pairs within the PDENB module; |phPCC out (X)| is the absolute value of the average phPCC between gene pairs within the PDENB module and gene pairs connected with outside.

[0023] Further, in step S3, the multi-modal optimization algorithm is:

[0024] S31, an initial population P0 of size N is randomly generated from PEN using the same population initialization strategy as MMPDNB;

[0025] S32, iterative optimization is performed;

[0026] S33, a set of PDENB of the cancer individual is output.

[0027] Further, in step S32, the iterative optimization process is:

[0028] S321, the objective function of the solutions in the initial population is calculated and the Pareto front on which the solutions are located is calculated using non-dominated sorting;

[0029] S322, N / 2 parents are generated from the current population Pt (t=0,1,2) using the tournament selection method based on decision space niching;

[0030] S323, offspring are generated; in the process of generating offspring, two parents are randomly selected, if the judgment condition is met, the offspring are generated using hidden space search, the RBM model is trained using non-dominated solutions in the current population, then the two parents are mapped to the hidden space through the RBM model, the offspring are generated in the hidden space using crossover and mutation operators, and finally the offspring generated in the hidden space are reconstructed to the original space; if the condition is not met, the offspring are generated in the original space using crossover and mutation operators; after the offspring are generated, the objective function of the offspring is calculated, and singular solutions are corrected, the offspring and the parents are combined into a combined population R with a size of 2N;

[0031] S324, N good quality solutions are selected from the combined population as the population of the next iteration; if the termination condition is met, the solutions with a Pareto front of 1 in the current population are output as multiple PDENB of the individual patient; otherwise, optimization is continued.

[0032] The present application has the beneficial effects that: the present application not only promotes the research on mathematical models and algorithm design for driving PDENB identification problems, but also helps to understand the heterogeneity of cancer individuals, and realizes early diagnosis and treatment of cancer individual patients. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 An illustrative explanation of hidden space search;

[0034] Figure 2(a) is the framework of MMPDENB-RBM model;

[0035] Figure 2(b) is the flowchart of MMPDENB-RBM model;

[0036] Figure 3(a) is the average score of PDENB of BRCA patients in different stages;

[0037] Figure 3(b) is the average score of PDENB of LUSC patients in different stages;

[0038] Figure 3(c) is the average score of PDENB of LUAD patients in different stages;

[0039] Figure 4(a) is the enrichment of synthetic lethal biomarkers in PDENB identified by each algorithm;

[0040] Figure 4(b) is the enrichment of driver genes in PDENB identified by each algorithm;

[0041] Figure 4(c) is the enrichment of Marker in PDENB identified by each algorithm;

[0042] Figure 5 Target spots and effective drugs provided by MMPDENB-RBM for patients in early stages of three cancers;

[0043] Figure 6 Multi-modal performance comparison of four evolutionary algorithms. DETAILED DESCRIPTION

[0044] The application will be further described below in conjunction with the drawings.

[0045] The application regards individual cancer patients as a dynamic network system, combines a dynamic network biomarker theory and a multi-modal evolutionary algorithm, performs hidden space search on the basis of an MMPDNB model by using a restricted Boltzmann machine (RBM), and proposes a new multi-modal PDENB identification model, namely MMPDENB-RBM. The input of the MMPDENB-RBM is gene expression data of paired samples of individual patients, gene expression data and gene mutation data of reference samples (all normal tissues or tumor tissues), and the output of the MMPDENB-RBM is multiple PDENB (namely a group of Pareto optimal solutions) of individual patients. In short, the MMPDENB-RBM mainly includes two parts: (1) gene expression data of known individual patients, gene expression data and gene mutation data of reference samples, and the specification is constructed according to a paired single sample network and individualized edge network analysis; and (2) for each PEN, the application increases a hidden space search mechanism on the basis of the MMPDNB, and proposes a multi-modal multi-objective evolutionary algorithm based on hidden space search, so as to obtain PDENB of a single patient.

[0046] The multi-modal critical edge biomarker identification method provided in the application comprises the following steps:

[0047] S1, constructing a PEN of a cancer individual patient, and the specific steps are:

[0048] S11, obtaining gene expression data of normal samples and tumor samples of the cancer individual patient, gene expression data of reference sample normal tissues, and somatic mutation data, and using a PGIN constructed by a paired single sample network as a node network of an edge network.

[0049] S12, selecting a gene pair with strong correlation in the PGIN as a candidate edge to construct the PEN; and selecting a gene pair with a pPCC greater than an average pPCC of the entire network as a candidate gene pair to construct the edge network.

[0050] S13, calculating a personalized high-order Pearson correlation coefficient (phPCC) between candidate gene pairs (<x i ,x j >, <x k ,x l >) existing in the PGIN:

[0051]

[0052]

[0053] Where shPCC is the higher-order Pearson correlation coefficient for a single sample. V is the higher-order average expression level of gene i in the reference sample. r (x i ) represents the higher-order variance of gene i's expression on the reference sample; shPCC Normal shPCC is the higher-order Pearson correlation coefficient for a single sample of normal tissue. Tumor It is the higher-order Pearson correlation coefficient of a single sample of tumor tissue.

[0054] S14 uses candidate gene pairs as "nodes" in the edge network and phPCC between candidate gene pairs as "edges" in the edge network to construct PENs for individual cancer patients at different stages of cancer.

[0055] S2, Design and optimize the objective function;

[0056] This invention defines the multimodal PDENB identification problem as a discrete multimodal multi-objective optimization problem, and then identifies multimodal PDENBs that can be used to detect the critical state of individual cancer patients.

[0057] MMPDENB-RBM identifies PDENB from PEN by optimizing the following two objective functions:

[0058]

[0059]

[0060] Where, the solution X = (x1, x2, ..., x n ) is a binary decision vector, where n represents the number of gene pairs in PEN; X can represent which gene pairs are selected as molecules in the module or are not selected. If gene pair i is selected as a molecule in the module, then x i =1, otherwise x i =0; pDEin(X) represents the standard deviation of phPCC between gene pairs within the PDENB module; |phPCC in (X)| represents the absolute value of the average phPCC between gene pairs within the PDENB module; |phPCC out (X)| is the absolute value of the phPCC mean between gene pairs within the PDENB module and gene pairs connected externally.

[0061] S3, using a multimodal optimization algorithm to find the PDENB set;

[0062] Based on the improved multi-modal multi-objective evolutionary algorithm MMPDNB framework, the hidden space search strategy (i.e. using RBM model to cross and mutate in hidden space, and then reducing the search space) is introduced. The multi-modal multi-objective evolutionary algorithm based on hidden space search is used to identify multi-modal PDENB on PEN. The steps include:

[0063] S31, an initial population P0 (i.e. solution set) of size N is randomly generated from PEN using the same population initialization strategy as MMPDNB;

[0064] S32, iterative optimization is performed; during the iteration process:

[0065] S321, the objective function of the solutions in the initial population is calculated, and the Pareto front on which the solutions are located is calculated using non-dominated sorting;

[0066] S322, N / 2 parents are generated from the current population Pt (t = 0, 1, 2) using the tournament selection method based on decision space niching;

[0067] S323, offspring are generated; during the generation of offspring, two parents are randomly selected, if the judgment condition is met, the offspring are generated using hidden space search, the RBM model is trained using the non-dominated solutions in the current population, then the two parents are mapped to the hidden space through the RBM model, the offspring are generated in the hidden space using crossover and mutation operators, and finally the offspring generated in the hidden space are reconstructed to the original space; if the condition is not met, the offspring are generated in the original space using crossover and mutation operators; after the offspring are generated, the objective function of the offspring is calculated, and the singular solutions are corrected, the offspring and the parents are combined into a combined population R of size 2N;

[0068] S324, N good quality solutions are selected from the combined population as the population for the next iteration; if the termination condition is met, the solutions with Pareto front of 1 in the current population are output as the multiple PDENB of the individual patient; otherwise, the optimization continues.

[0069] S33, a set of PDENB of the cancer individual is output.

[0070] The framework of the proposed MMPDENB-RBM is given in Algorithm 1.

[0071] Algorithm 1: Multi-modal multi-objective evolutionary algorithm based on hidden space search in MMPDENB-RBM

[0072] Input: gene expression data I of tumor sample and normal sample of cancer individual patient D ;

[0073] Gene mutation data M of tumor sample of cancer individual patient D ;

[0074] Reference sample normal tissue gene expression data RN D ;

[0075] Reference gene interaction network Net

[0076] Population size NP

[0077] Iteration condition (maximum number of function evaluations) T.

[0078] Output: multiple PDENB of individual cancer patients

[0079]

[0080]

[0081] Discrimination condition using hidden space search strategy: the present application sets a condition, that is, when fe>1 / 2maxFE and rand<ρ are met, offspring are generated using hidden space. Wherein fe is the number of function evaluations so far, maxFE is the maximum number of function evaluations, which is also the iteration termination condition. When fe>1 / 2maxFE, it means that the current population evolution process has consumed half of the computing resources, the algorithm has explored a wider search space, and the quality of the solution has been greatly improved compared to the early stage. It can be considered to use the non-dominated solution in the current population to train the RBM model to better approximate the Pareto hidden space. rand is a random number, and ρ is the probability of generating offspring using hidden space, which represents the proportion of the number of offspring generated by hidden space search that are successfully retained to the next generation in all offspring retained to the next generation. The larger this value is, the better the quality of the offspring generated using hidden space, and the greater the probability of using hidden space generation strategy. In the present application, the update of parameter ρ is consistent with the MOEA / PSL algorithm. Therefore, after the early optimization, the quality of the non-dominated solution used to train the RBM model is improved, not only enabling the RBM model to approximate the Pareto hidden space, but also reducing the computing time of the entire algorithm.

[0082] Step of generating offspring using hidden space search: first, select two parents P1 and P2 as input, map to hidden space (i.e. the hidden layer representation of the RBM model) using the RBM model to obtain and Then, a child is generated in the hidden space using the crossover operator and the add or delete strategy Reconstruction to the original space. It should be noted that the size of the hidden layer k of the RBM needs to be determined. The present application refers to the calculation method of k in the MOEA / PSL algorithm. This method points out that the size of the hidden layer is determined by the sparsity of the non-dominated solution, which meets the characteristics of biomarkers not requiring a large number of molecules. Generally, the size of the hidden layer k is much smaller than the input layer n, so the search space is converted from the original search space 2n to the Pareto hidden space 2k, greatly reducing the search space. MMPDENB-RBM uses Pareto hidden space search to generate offspring, which can greatly reduce the search difficulty, and uses a special multi-modal solution generation strategy to mine the multi-modal nature of PDENB. In summary, MMPDENB-RBM with large-scale optimization capability and multi-modal optimization capability can effectively solve the identification work of multi-modal PDENB.

[0083] Experimental setup

[0084] The present application will verify the MMPDNB method proposed in three cancer data. They are 112 breast invasive carcinoma (Breast Invasive Carcinoma, BRCA) data, 49 lung squamous cell carcinoma (lung squamous cell carcinoma, LUSC) data, and 57 lung adenocarcinoma (lung adenocarcinoma, LUAD) data.

[0085] Parameter setting

[0086] The parameters of MMPDENB-RBM are as follows:

[0087] (1) Population size N: The population size of each patient data is set to 300 in the present application.

[0088] (2) Termination condition: The termination criterion is the maximum number of function evaluations for each independent run, which is set to 30,000 times. Similarly, in order to reduce the uncertainty of the results, MMPDENB-RBM is independently run 30 times on the PEN of individual patients, and the Pareto optimal solution of the union of the 30 independent running results is selected as the multiple PDENB of the individual patient with cancer.

[0089] Comparison algorithm

[0090] MMPDENB-RBM is compared with other 5 methods, which are divided into the following two categories according to their characteristics:

[0091] (1) Multi-modal multi-objective evolutionary algorithms: MMPDNB, MOEA / PSL and MP-MMEA. The parameters of the three multi-modal evolutionary algorithms can be divided into two parts: the same parameters as MMPDENB-RBM (i.e., the size of the population and the maximum number of function evaluations) and the specific parameters of these algorithms. For the sake of fairness, the shared parameters of all algorithms should be consistent. For the specific parameters of other algorithms, the specific parameter settings of other algorithms are consistent with the original text. In addition, the number of independent runs of these algorithms on each individual patient is also 30 times.

[0092] (2) DNB algorithms: Clustering-DNB (Clustering-DNB) and Landscape-DNB (Landscape-DNB, L-DNB) methods can be used to identify the PDENB of a single PEN patient. The Clustering-DNB method decomposes the PEN into multiple modules by hierarchical clustering, and the module with the highest warning score is taken as the PDENB module. The L-DNB method calculates the first-order neighborhood network warning signal score of each node (i.e., gene pair) on the PEN, and takes it as the score of each node. All nodes are ranked in descending order of score, and the top K gene pairs are taken as the PDENB module of the PEN of a single patient.

[0093] Performance indicators

[0094] The present application quantifies the performance of different algorithms using the following indicators:

[0095] (1) Average warning score: The average score of the highest warning score of individual patients in each stage is calculated, and this value is taken as the warning score of each stage, and the dynamic change curve in the process of cancer development is obtained, which is used to evaluate the ability of each algorithm to detect critical states.

[0096] (2) Hypervolume (HV) index: Since evolutionary algorithms will provide a set of Pareto optimal solutions, the HV index can evaluate the diversity of Pareto optimal solutions in the target space and the convergence of evolutionary algorithms. It should be noted that the calculation of the HV value requires a reference point, which is set to 1.2 times the maximum value of the two objective functions.

[0097] (3) F-score: In addition, in order to evaluate the effectiveness of MMPDENB-RBM model and other methods in identifying cancer tissue-specific edge network biomarkers, the present application calculates the F-score of each algorithm in terms of PDNB containing synthetic lethal edge biomarkers, driver genes and biomarker genes.

[0098] Experimental results

[0099] MMPDENB-RBM can effectively detect early warning signals in the early stages of cancer development.

[0100] As shown in FIG. 3, the peaks of BRCA, LUSC and LUAD patients appear in stages IIIB, IIA and IIA, respectively, which are considered as critical stages of the three cancers. In addition, the PDENB identified by MMPDENB-RBM has a higher early warning score than the PDENB detected by other algorithms, and the PDENB early warning score of the MMPDENB-RBM at the critical stage is obviously different from the score at other stages.

[0101] Compared with other multi-modal multi-objective evolutionary algorithms, MMPDENB-RBM and MMPDNB perform better in early warning signal detection. It should be noted that the early warning signals detected by MMPDENB-RBM and MMPDNB in LUSC data are similar, because the two algorithms use the same offspring generation strategy. However, the average score of MMPDNB is lower than that of MMPDENB-RBM.

[0102] Compared with the Clustering-DNB method and the L-DNB algorithm, the MMPDENB-RBM model with optimization capability can obtain a larger PDENB early warning signal score, indicating that MMPDENB-RBM can more effectively detect the early warning signals of patients with the three cancers.

[0103] The PDENB identified by MMPDENB-RBM has a higher cancer driver gene enrichment.

[0104] In order to verify whether the algorithms such as MMPDENB-RBM can effectively identify cancer tissue-specific synthetic lethal edge biomarkers, driver genes and marker genes, the present application takes known cancer tissue-specific synthetic lethal edge biomarkers, driver genes and marker genes as the standard to calculate the average F-score of all PDENB. FIG. 3 shows that the individualized edge network biomarkers identified by MMPDENB-RBM have a higher enrichment in the three cancer data compared with other algorithms.

[0105] The PDENB identified by MMPDENB-RBM contains effective drug targets.

[0106] By searching for drug target information in the iGMDR database, the differential genes of the multi-modal PDENB belonging to early BRCA contain 12 drug target genes, and these drug targets are targeted by 9 drugs related to BRCA. As shown in FIG. 4, the related results for LUSC and LUAD data are also shown in FIG. 5. Figure 5 Figure 5

[0107] MMPDENB-RBM can effectively identify multi-modal PDENB

[0108] As​​Figure 6 As shown in the figure, MMPDENB-RBM and other three multi-modal evolutionary algorithms identify the number of multi-modal PDENB individual patients on three cancer data sets. MMPDENB-RBM finds 53, 25 and 34 individual patients with multi-modal PDENB on BRCA, LUSC and LUAD data sets, respectively. The present application can conclude that MMPDENB-RBM has obvious advantages over other multi-modal evolutionary algorithms in detecting multi-modal PDENB.

[0109] All experimental results show that multi-modal optimization can effectively identify the personalized dynamic edge network biomarkers of individual patients with cancer, explore the multi-modality of network biomarkers, and MMPDENB-RBM can also provide more potential drug targets for early individual patients, and promote the development of precision medicine for cancer.

Claims

1. A multi-modal critical boundary biomarker identification method, characterized by, The method comprises the following steps: S1, constructing a PEN of a cancer individual patient, comprising the following steps: S11, constructing a PGIN using paired single sample network from gene expression data of normal sample and tumor sample of the cancer individual patient and reference sample normal tissue and somatic mutation data; S12, selecting the gene pairs with strong correlation in PGIN as candidate edges to construct PEN; selecting the genes with correlation greater than the average correlation of the whole network as candidate gene pairs to construct edge network; greater than the average correlation of the whole network greater than the average correlation of the whole network as candidate gene pairs to construct edge network; S13, Calculate candidate gene pairs that are connected in PGIN. The individualized high-order Pearson correlation coefficient between them ): ； ； wherein, is a single sample high order Pearson correlation coefficient, is a gene a high order mean of expression over a reference sample, is a gene a high order variance of expression over a reference sample; is a single sample high order Pearson correlation coefficient for normal tissue, is a single sample high order Pearson correlation coefficient for tumor tissue; S14, taking the candidate gene pairs as "nodes" in the edge network, and the relationships between the candidate gene pairs as "edges" in the edge network constructing the PEN of the cancer individual patient at different cancer stages as "edges" in the edge network; S2, designing an optimization objective function; S3, finding a PDENB set using a multi-modal optimization algorithm; The multi-modal optimization algorithm is: S31, randomly generate an initial population of size from PEN using the same population initialization strategy as MMPDNB ; S32, performing iterative optimization; S33, outputting a set of PDENB of the cancer individual.

2. The multi-modal critical boundary biomarker identification method of claim 1, wherein, In step S2, the optimization objective function is: ； ； where the solution is a binary decision vector, represents the number of gene pairs in the PEN; can represent which gene pairs are selected as molecules in the module or not, if gene pair i is selected as a molecule in the module, then = 1, otherwise = 0; represents the standard deviation of the average between gene pairs within the PDENB module; represents the absolute value of the average between gene pairs within the PDENB module; is the absolute value of the average between gene pairs within the PDENB module and gene pairs outside the module.

3. The multi-modal critical boundary biomarker identification method of claim 1, wherein, In step S32, the iterative optimization process is: S321, calculating the objective function of the solution in the initial population and calculating the Pareto front on which the solution is located using non-dominated sorting; S322, generating a parent from the current population using a tournament selection method based on decision space niching one parent;​ S323, generating offspring; in the process of generating offspring, two parents are randomly selected, if the judgment condition is satisfied, offspring are generated by using hidden space search, RBM model is trained by using non-dominated solutions in the current population, then the two parents are mapped to the hidden space through the RBM model, and the offspring are generated in the hidden space by using crossover and mutation operators, and finally the offspring generated in the hidden space are reconstructed to the original space; if the condition is not satisfied, offspring are generated in the original space by using crossover and mutation operators; after the generation of offspring is completed, the objective function of the offspring is calculated, and singular solutions are corrected, and the offspring and the parents are combined into a merged population with a size of . ​ S324, selecting from the merged population the solution with the best quality as the population for the next iteration; if the termination condition is met, outputting the solution with the Pareto front of 1 in the current population as the multiple PDENB of the individual patient; otherwise, continuing the optimization.

Images