Constrained multi-objective optimization method for detecting drug targets in individual cancer patients

Abstract

The present invention discloses a constrained multi-objective optimization method for detecting drug targets in individual cancer patients, comprising: constructing a personalized gene interaction network (PGIN) from the genomic data of the individual patient; forming a structural network control model based on multi-objective optimization; and using a constrained multi-objective evolutionary algorithm to search for a personalized drug target set, i.e., a driver gene set. The present invention can effectively identify drug targets in individual cancer patients, explore cancer heterogeneity, and provide a new perspective for understanding tumor heterogeneity in precision medicine.

Description

Technical Field

[0001] The present invention relates to the technical field of individualized cancer drug target identification, in particular to a constrained multi-objective optimization method for detecting drug targets in individual cancer patients. Background Art

[0002] In recent years, with the gradual acceleration of the aging of my country's population, industrialization, and urbanization, coupled with a series of reasons such as unhealthy lifestyles and environmental exposure, cancer has become one of the main causes of death among Chinese residents in recent years.

[0003] Numerous clinical studies have shown that cancer has strong individual heterogeneity. Due to the differences in the microenvironment of different individual cancers, the individual heterogeneity of cancer patients makes it difficult to quickly identify the right drugs for the treatment, thus preventing substantial breakthroughs. In the process of combined treatment of diseases, it is necessary to consider individual heterogeneity. However, due to the limited effective sample size of individual cancer patients, it is difficult for traditional statistical methods to extract effective information. Therefore, in the era of precision medicine, it is quite challenging to predict individual targeted combination drugs, rather than traditional patient cohort targeted combination drugs.

[0004] With the development of network science, many tools based on sample-specific gene interaction networks, including CPGD, PNC, SCS, and DawnRank, have been developed to identify personalized drug targets (PDTs) by identifying individual driver genes targeted by drugs in personalized gene interaction networks (PGINs). Among these methods, the network structure control method is one of the effective methods for identifying PDTs. It aims to use the principles of network structure control to identify optimized driver gene sets from PGINs to achieve network state transitions. These driver gene sets are candidate drug targets for drug stimulation signals, and their state transitions can change the entire network state from the initial state to the desired state.

[0005] It is worth noting that the latest research results show that network structure control can be transformed into a constrained optimization problem, thereby selecting the optimal set of individual drug target nodes. However, current network structure control methods face two major problems. First, these methods are mostly NP-hard problems in large-scale networks with nonlinear dynamics, and it is computationally difficult to find the optimal set of driver nodes. The other is that they are based on single-objective optimization to identify the minimum set of driver nodes, without considering the non-uniqueness of the optimal targeted driver gene (drug target) set, that is, there may be other drug target sets that do not have the minimum number of drug targets, but these drug target sets contain the largest effective prior information of drug targets. Therefore, it is necessary to design more accurate constrained optimization algorithms to overcome the above shortcomings. Summary of the Invention

[0006] To solve the problems existing in the prior art, the purpose of the present invention is to provide a constrained multi-objective optimization method for detecting drug targets in individual cancer patients. The present invention can effectively identify personalized driver genes for individual cancer patients through constrained multi-objective optimization.

[0007] To achieve the above objectives, the present invention adopts a technical solution: a constrained multi-objective optimization method for detecting drug targets in individual cancer patients, comprising the following steps:

[0008] Step 1: Build a personalized gene interaction network (PGIN) from the genomic data of individual patients;

[0009] Step 2: Form a structural network control model based on multi-objective optimization;

[0010] Step 3: Use the constrained multi-objective evolutionary algorithm to find a personalized drug target set, that is, a driver gene set.

[0011] As a further improvement of the present invention, the step 1 is specifically as follows:

[0012] First, a co-expression network based on tumor samples and normal samples of a single patient is established; then, in the normal sample network and the tumor sample network, the P value of the edge between gene i and gene j is used to determine whether to use this edge to construct a personalized gene interaction network (PGIN), where the P value is obtained based on the Z score of the difference in the Pearson correlation coefficient between gene i and gene j.

[0013] As a further improvement of the present invention, the specific conditions for constructing the PGIN personalized gene interaction network are as follows:

[0014] If the P value of the tumor sample network is lower than a certain threshold and the P value of the normal sample network is greater than a certain threshold, or the P value of the tumor sample network is greater than a certain threshold and the P value of the normal sample network is lower than a certain threshold, then this edge is retained to construct a personalized gene interaction network PGIN.

[0015] As a further improvement of the present invention, the Pearson correlation coefficient ΔPCC of gene i and gene j is (ij, k ) and its Z score Z (ij,k) The calculation method is as follows:

[0016]

[0017]

[0018] Where n represents the number of reference samples, k represents the kth patient in the perturbation network, is the PCC of genes i and j on n reference samples, represents the PCC of genes i and j for individual patient k on n+1 reference samples; the P value is based on Z ij,k Calculated from the standard normal distribution.

[0019] As a further improvement of the present invention, by integrating somatic mutation data across cancer type-specific data into the personalized gene interaction network PGIN, the personalized pPCC score of the edge of the personalized gene interaction network PGIN is calculated as follows:

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] Among them, comutation means comutation, coexpression means coexpression, represents the personalized Pearson coefficient of genes i and j in patient k sample, i,j represents the Jaccard coefficient of genes i and j, represents the Pearson correlation coefficient between genes i and j in the tumor sample of individual patient k, represents the Pearson correlation coefficient of genes i and j in the normal samples of individual patient k, S(i) and S(j) are the sets of tumors with mutated genes i and j, respectively, after examining somatic mutations in a given cancer dataset, and D 10 It means that after sorting a set of data in ascending order, 10% of the data are in D 10 under.

[0026] As a further improvement of the present invention, in step 2, the structure network control model is as follows:

[0027]

[0028]

[0029]

[0030] Among them, f1 and f2 are two optimized objective functions; g(x) represents the constraint condition, which makes the network controllable; x=(x1,x2,…,x n) represents a binary decision variable; V represents the vertex set of the network; D and DT represent the identified drug target set and the known original drug target set, respectively. If the gene is selected, then x i =1, otherwise, x i =0; l = (l1, l2, ..., l n ) represents a label vector with binary encoding. If the gene is within a known drug target, l i =1, otherwise, l i = 0; n represents the number of genes in PGIN, and m represents the number of constraints.

[0031] As a further improvement of the present invention, the step 3 specifically includes the following steps:

[0032] Step 3.1. Initialize a main population P0 of size N and two auxiliary populations P1 and P2 of size N1, where N1 = N * 30%. The main population optimizes the original constrained multi-objective optimization problem CMOP, and the auxiliary populations optimize the corresponding constrained single-objective optimization problems CSOP.

[0033] Step 3.2: Use the tournament selection strategy to select N / 2 parents from the current population P0, and N1 / 2 parents from P1 and P2 respectively;

[0034] Step 3.3: Use the crossover mutation operator to generate a child population Q0 of size N / 2 and child populations Q1 and Q2 of size N1 / 2, respectively. Merge the main population, auxiliary population, and all child populations R0 = P0∪Q0∪Q1∪Q2, R1 = P1∪Q0∪Q1∪Q2, and R2 = P2∪Q0∪Q1∪Q2.

[0035] Step 3.4: Select N solutions from the merged population R0 as population P0 according to the ε constraint method, and select N1 solutions from the merged populations R1 and R2 as populations P1 and P2 according to a ranking-based fitness function, which will serve as the population for the next iteration.

[0036] Step 3.5: Output multiple driver gene sets until the termination condition of the iteration is met, that is, the maximum number of function evaluations is met.

[0037] As a further improvement of the present invention, in step 3.1, a multi-task optimization idea is adopted. The main population optimizes the original constrained multi-objective optimization problem, and the auxiliary population optimizes the corresponding derived constrained single-objective optimization problem. The evolution of these populations is regarded as multiple different but related optimization tasks; the main population is committed to global search while the auxiliary population focuses on local search, developing areas that have not been searched by the main population, helping the main population to cross the infeasible domain to ensure its diversity.

[0038] As a further improvement of the present invention, during environment selection, a ranking-based fitness function method is used to calculate the fitness value of the auxiliary population.

[0039] As a further improvement of the present invention, the ranking-based fitness function method ranks each individual in two ways: one biased towards constraints and the other biased towards the objective function, and the two rankings are weighted together by a weight parameter:

[0040] First, for the ranking-based fitness function that is biased towards constraints, the individuals in the population are sorted in ascending order according to the ε constraint method, and the ranking number of each individual is recorded as R1; then, in order to use the information of the objective function, each individual is sorted based on its objective function to obtain the ranking number R2 of each individual; finally, the ranking-based fitness function combines R1 and R2 into a new formula that represents the fitness of the individual for the balanced objective and constraints, as shown below:

[0041] Fitness=w*R1+(1-w)*R2

[0042]

[0043] Among them, P f Indicates the proportion of feasible solutions in this generation, G is the current generation, G max is the maximum algebra.

[0044] The beneficial effects of the present invention are:

[0045] The constrained multi-objective optimization method based on network structure control theory of the present invention can effectively identify drug targets for individual cancer patients, explore cancer heterogeneity, and provide a new perspective for understanding tumor heterogeneity in precision medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is the overall flow chart of Example 1;

[0047] Figure 2 Schematic diagram for explaining the multi-task framework in Example 1;

[0048] Figure 3 The performance of the clinical combination drug discovered in Example 1 with the traditional network control method;

[0049] Figure 4 The performance of the constrained multi-objective evolutionary method in discovering clinical combination drugs in Example 1;

[0050] Figure 5 The performance of Example 1 in discovering clinical combination drugs with other methods based on cancer-specific driver genes;

[0051] Figure 6 Schematic diagram of the results of the driver gene set enriched in a given CGC dataset in Example 1;

[0052] Figure 7 Schematic diagram of PF for individual patients at different stages on the BRCA cancer dataset under the MDS, NCUA, and DFVS frameworks in Example 1. DETAILED DESCRIPTION

[0053] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0054] Example 1

[0055] This example provides a large-scale constrained variables based multi-tasking cooperative evolutionary algorithm (LSCV-MCEA). The model inputs gene expression data from paired samples of individual cancer patients and outputs a set of driver genes for each individual patient. The described LSCV-MCEA consists of three main components: constructing a PGIN from the individual patient's genomic data; forming a multi-objective optimization-based structural network control model (MONCP); and designing a constrained multi-objective evolutionary strategy for identifying driver genes.

[0056] like Figure 1 and Figure 2 As shown, this embodiment provides a constrained multi-objective optimization method for detecting drug targets in individual cancer patients, which includes the following steps:

[0057] Step 1: Construct PGIN for individual cancer patients:

[0058] Paired samples of individual cancer patients (i.e., normal samples and tumor samples from the same patient) were filtered and obtained from the TCGA data portal. The PGIN of individual cancer patients was constructed using the paired single sample network method, following the following principles. The first step is to establish a co-expression network based on tumor samples and normal samples of a single patient. Then, it is necessary to determine whether to use this edge to construct the PGIN based on the P value of the edge between gene i and gene j in the normal sample network and the tumor sample network. The specific conditions are as follows: If the P value of the tumor sample network is lower than 0.05 (the co-expression relationship between the interaction of the two genes is significant) and greater than 0.05 in the normal sample network (not significant), or vice versa, this edge is retained to constitute the PGIN. In addition, the ΔPCC value of the edge can be obtained by calculating the Z-value of the edge. The ΔPCC of an edge between gene i and gene j and its Z-value score can be calculated:

[0059]

[0060]

[0061] where n represents the number of reference samples and k represents the kth patient in the perturbation network. is the PCC of genes i and j on n reference samples, represents the PCC of genes i and j for individual patient k on n+1 reference samples; the P value can be based on Z i The standard normal distribution of j and k is calculated; in addition, by integrating somatic mutation data across cancer type-specific data into the PGIN, a method for scoring personalized pPCC at the edge of the PGIN is calculated as follows:

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] Among them, comutation means comutation, coexpression means coexpression, represents the personalized Pearson coefficient of genes i and j in patient k sample, i,j represents the Jaccard coefficient of genes i and j, represents the Pearson correlation coefficient between genes i and j in the tumor sample of individual patient k, represents the Pearson correlation coefficient of genes i and j in the normal samples of individual patient k, S(i) and S(j) are the sets of tumors with mutated genes i and j, respectively, after examining somatic mutations in a given cancer dataset, and D 10 It means that after sorting a set of data in ascending order, 10% of the data are in D 10 under.

[0068] Therefore, the pPCC of an individual patient does not simply distinguish the differences in gene expression between normal and tumor samples, but can reflect the connections between genes during the dynamic process of cancer development by considering somatic mutation data.

[0069] Step 2: Form a structural network control model MONCP based on multi-objective optimization:

[0070] Considering the limitations of traditional network control methods (MMS, MDS, NCUA, DFVS):

[0071] (1) In large-scale networks with nonlinear dynamics, it is difficult to find an optimal set of driving nodes that can achieve an approximately effective solution;

[0072] (2) Another limitation is that these methods only focus on controlling the system through the minimum set of driver nodes, which is a control based on a single-objective optimization and ignores the previous drug target identification of the optimal driver gene. Therefore, under the framework of DFVS, NCUA, and MDS, the system is controlled by considering the minimum set of driver nodes and the maximum pre-known drug target information. At the same time, various constraints on nodes / variables need to be satisfied to ensure that the network is controllable, as follows:

[0073]

[0074]

[0075]

[0076] Among them, f1 and f2 are two optimized objective functions; g(x) represents the constraint condition, which makes the network controllable; x=(x1,x2,…,x n ) represents a binary decision variable; V represents the vertex set of the network; D and DT represent the identified drug target set and the known original drug target set, respectively. If the gene is selected, then x i =1, otherwise, x i =0; l = (l1, l2, ..., l n ) represents a label vector with binary encoding. If the gene is within a known drug target, l i =1, otherwise, l i = 0; n represents the number of genes in PGIN, and m represents the number of constraints.

[0077] Step 3: Use constrained multi-objective evolutionary algorithm to find personalized driver gene sets:

[0078] Based on the evolutionary multi-tasking framework, a new multi-tasking mechanism is introduced, and a new constrained multi-objective evolutionary algorithm is proposed. The process is as follows: First, a main population P0 of size N and two auxiliary populations (P1 and P2) of size N1 are initialized, where N1 = N * 30%. The main population optimizes the original constrained multi-objective optimization problem (CMOP), and the auxiliary populations optimize the corresponding constrained single-objective optimization problems (CSOP). A tournament selection strategy is then used to select N / 2 (N1 / 2) parents from the current population P0 (P1 and P2). A crossover mutation operator is then used to generate an N / 2 offspring population Q0 and N1 / 2 offspring populations Q1 and Q2, respectively. The main and auxiliary populations are then merged with all offspring populations to form R0 = P0∪Q0∪Q1∪Q2, R1 = P1∪Q0∪Q1∪Q2, and R2 = P2∪Q0∪Q1∪Q2. N solutions are then selected from the merged population R0 as population P0 using the ε constraint. A ranking-based fitness function is then used to select N1 solutions from the merged populations R1 and R2 as populations P1 and P2, which serve as the population for the next iteration. Finally, multiple sets of driver genes are output until the iteration terminates, i.e., the maximum number of function evaluations is met. The proposed LSCV-MCEA framework is given in Algorithm 1.

[0079]

[0080]

[0081] Multi-task framework: For decision space constraint problems, the constraint landscape is relatively complex due to the large number of decision space constraints. However, current constrained multi-objective evolutionary algorithms mainly focus on solving constraint problems in the target space. The constraints and targets are positively correlated, and the optimization target is equivalent to the optimization constraint. Therefore, existing constrained multi-objective evolutionary algorithms are difficult to directly apply to the process of identifying driver genes. Based on this, a multi-task optimization concept is adopted. The main population optimizes the original constrained multi-objective optimization problem, and the auxiliary population optimizes the corresponding derived constrained single-objective optimization problem. The evolution of these populations is regarded as multiple different but related optimization tasks. The main population is committed to global search, while the auxiliary population focuses on local search, exploring areas not searched by the main population, helping the main population to cross the infeasible region and maintain its diversity.

[0082] Rank-based fitness function: The ranking-based fitness function method ranks each individual in two ways: one biased towards constraints and the other biased towards the objective function, and the two rankings are weighted together by a weight parameter. First, for the ranking-based fitness function that is biased towards constraints, the individuals in the population are sorted in ascending order according to the constraint method, and the rank number of each individual is recorded as . Then, in order to use the information of the objective function, each individual is sorted based on its objective function to obtain the rank number of each individual. Finally, the ranking-based fitness function combines and into a new formula that represents the fitness of the individual for the balanced objective and constraints, as shown below:

[0083] Fitness=w*R1+(1-w)*R2

[0084]

[0085] where represents the proportion of feasible solutions in this generation. is the current generation, and is the maximum generation. In particular, as and increase, changes from 0.5 to 1. It is worth noting that smaller is better, meaning that is more likely to advance to the next generation.

[0086] In this example, the proposed LSCV-MCEA method is validated on three types of cancer datasets: 112 breast invasive carcinoma (BRCA) datasets, 49 lung squamous cell carcinoma (LUSC) datasets, and 57 lung adenocarcinoma (LUAD) datasets.

[0087] experiment

[0088] Parameter settings

[0089] Population Settings: Due to the large-scale nature of the problem, the algorithm population size was set to 300. Number of Independent Runs and Termination Criteria: To reduce the chance of randomness in the results, LSCV-MCEA was independently run 30 times on each patient's PGIN. The results of these 30 independent runs were combined, and the Pareto optimal solution of this union was output. The termination criterion was the number of objective function evaluations, set to 100,000 for each independent run.

[0090] Comparison Algorithms

[0091] 19 algorithms were selected for comparison with LSCV-MCEA, and they were divided into three groups based on their characteristics. Constrained multi-objective evolutionary algorithms: MTCMO, CCMO, c-DPEA, CCMODE, CMME, NSGA-II-CDP. Their population size, number of independent runs, and termination conditions are consistent with LSCV-MCEA. Network and differentially expressed gene (DEG)-based methods: including CPGD, PNC, ActiveDriver, OncoDriveFM, DriverML, and Hub-genes. DEG-based methods include DEG-Folchange, DEG-p-value, and DEF-FDR. Traditional network control algorithms: Network control algorithms use the controllability and objectivity of the network to find a minimum set of nodes that can control or reconstruct the state of the entire network. The network control algorithms compared in this study are MMS, MDS, DFVS, and NCUA.

[0092] Experimental results

[0093] Performance of LSCV-MCEA in clinical combination drug discovery

[0094] Here, AUC is used to evaluate its performance in identifying personalized driver genes for each PGIN. The larger the AUC of a given method, the better its performance in discovering clinical combination drugs. LSCV-MCEA achieved higher AUCs than traditional structural network control methods on these three cancer datasets. Figure 3 The results show that the LSCV-MCEA method can detect drug targets more effectively than traditional structured network control methods to discover anticancer drug combinations.

[0095] Compared with other constrained multi-objective evolutionary algorithms, the PS of LSCV-MCEA can achieve higher AUC. Figure 4 The results show that LSCV-MCEA has better performance in discovering clinical combination drugs on these cancer datasets than other cancer patients.

[0096] Compared with other methods based on cancer-specific driver genes, LSCV-MCEA is more effective in identifying cancer driver genes for clinical combination drug discovery. Figure 5 .

[0097] LSCV-MCEA can be significantly enriched in the CGC dataset

[0098] To determine whether the solution set (i.e., the set of driver genes) is present in a given CGC dataset, the p-value of each solution on the PS of the three cancer datasets was calculated. The proportion of solutions with a p-value less than 0.05 was considered the enrichment significance score (ESS) of the given PS. Figure 6 The results under the MDS, NCUA, and DFVS frameworks are plotted and described. Some observations are obtained:

[0099] (1) The ESS score is usually greater than 0.8. This indicates that there are at least 80% valid solutions on the PS. Under MMS, NCUA, and DFVS, the CGC of LSCV-MCEA on these three cancer datasets is significantly enriched. These results indicate that most solutions in the PS of LSCV-MCEA contribute to the discovery of cancer driver genes.

[0100] (2) The performance of LSCV-MCEA varies under different network control frameworks. In general, the performance of LSCV-MCEA using the DFVS control framework is better than that of the NCUA and MDS control frameworks.

[0101] LSCV-MCEA can detect significant disease signals

[0102] Figure 7 All PFs of LSCV-MCEA for individual patients at different stages on the BRCA cancer dataset are shown under the MDS, DFVS, and NCUA control frameworks. Furthermore, the values of the two objectives in the MONCP for the solutions of traditional network control frameworks (i.e., MDS, DFVS, and NCUA) are also provided. The results show that the PFs of MDS, DFVS, and NCUA are more effective in detecting stage IIIB disease signals compared to traditional network control frameworks. Furthermore, the PF solutions of LSCV-MCEA are not dominated by traditional network control frameworks, indicating that LSCV-MCEA complements traditional network control principles for characterizing network control in BRCA cancer patients.

[0103] The set of driver genes identified by MMPDNB has drug targets.

[0104] To further demonstrate the complementary relationship between the solution of LSCV-MCEA and the traditional network control principle, this example considers the differential genes between LSCV-MCEA and the traditional network control principle of individual patients on three cancer datasets. The differential genes are in the PS of LSCV-MCEA, rather than in a set of driver genes in the traditional network control principle. By calculating the average frequency of the differential genes in the PS of all individual patients, they are ranked and the top 50 genes are selected as candidate drug targets for subsequent analysis. Finally, by searching for drug target information on drug response datasets (i.e., Genomics of Cancer Drug Sensitivity (GDSC) datasets and iGMDR databases), it is verified whether the top 50 genes can provide effective drug information for individual patients on the three cancer datasets, as shown in Table 1.

[0105] Table 1 Drug targets and effective drugs provided by LSCV-MCEA for three cancer datasets under the MDS, NCUA, and DFVS frameworks

[0106]

[0107] in conclusion

[0108] All experimental results show that constrained multi-objective optimization can effectively identify drug targets for individual cancer patients, explore cancer heterogeneity, and provide a new perspective for understanding tumor heterogeneity in precision medicine.

[0109] The above-described embodiments merely represent specific implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A constrained multi-objective optimization method for detecting drug targets in individual cancer patients, characterized in that: The following steps are involved: Step 1: Build a personalized gene interaction network (PGIN) from the genomic data of individual patients; The step 1 is specifically as follows: First, a co-expression network based on tumor samples and normal samples from a single patient was established. Then, in the normal sample network and the tumor sample network, the P-value of the edge between gene i and gene j was used to determine whether to use this edge to construct a personalized gene interaction network (PGIN). The P-value was obtained based on the Z-score of the difference in the Pearson correlation coefficient between gene i and gene j. The specific conditions for constructing the PGIN personalized gene interaction network are as follows: If the P value of the tumor sample network is lower than a certain threshold and the P value of the normal sample network is greater than a certain threshold, or the P value of the tumor sample network is greater than a certain threshold and the P value of the normal sample network is lower than a certain threshold, then this edge is retained to construct the personalized gene interaction network PGIN; Step 2: Form a structural network control model based on multi-objective optimization; In step 2, the structural network control model is as follows: s.t.g k (x)≤0,k=1,2...,m Among them, f1 and f2 are two optimized objective functions; g(x) represents the constraint condition, which makes the network controllable; x=(x1,x2,…,x n ) represents a binary decision variable; V represents the vertex set of the network; D and DT represent the identified drug target set and the known original drug target set, respectively. If a gene is selected, then x i =1, otherwise, x i =0; l = (l1, l2, ..., l n ) represents a label vector with binary encoding. If the gene is within a known drug target, l i =1, otherwise, l i = 0; n represents the number of genes in PGIN, and m represents the number of constraints; Step 3: Use constrained multi-objective evolutionary algorithm to find personalized drug target set, i.e., driver gene set; The step 3 specifically includes the following steps: Step 3.

1. Initialize a main population P0 of size N and two auxiliary populations P1 and P2 of size N1, where N1 = N * 30%. The main population optimizes the original constrained multi-objective optimization problem CMOP, and the auxiliary populations optimize the corresponding constrained single-objective optimization problems CSOP. Step 3.2: Use the tournament selection strategy to select N / 2 parents from the current population P0, and N1 / 2 parents from P1 and P2 respectively; Step 3.3, use the crossover mutation operator to generate a child population Q0 of size N / 2 and child populations Q1 and Q2 of size N1 / 2 respectively, merge the main population and auxiliary population with all the child populations R0 = P0∪Q0∪Q1∪Q2, R1 = P1∪Q0∪Q1∪Q2, R2 = P2∪Q0∪Q1∪Q2; Step 3.4: Select N solutions from the merged population R0 as population P0 according to the ε constraint method, and select N1 solutions from the merged populations R1 and R2 as populations P1 and P2 according to a ranking-based fitness function, which will serve as the population for the next iteration. Step 3.5: Output multiple driver gene sets until the termination condition of the iteration is met, that is, the maximum number of function evaluations is met.

2. The constrained multi-objective optimization method for detecting drug targets in individual cancer patients according to claim 1, characterized in that: Pearson correlation coefficient ΔPCC between gene i and gene j (ij,k) and the Z score of its difference (ij,k) The calculation method is as follows: Where n represents the number of reference samples, k represents the kth patient in the perturbation network, is the PCC of genes i and j on n reference samples, represents the PCC of genes i and j for individual patient k on n+1 reference samples; the P value is based on Z (ij,k) Calculated from the standard normal distribution.

3. The constrained multi-objective optimization method for detecting drug targets in individual cancer patients according to claim 2, characterized in that: By integrating somatic mutation data across cancer type-specific data into the personalized gene interaction network PGIN, the personalized pPCC of the edges of the personalized gene interaction network PGIN is scored as follows: Among them, comutation means comutation, coexpression means coexpression, represents the personalized Pearson coefficient of genes i and j in patient k sample, i,j represents the Jaccard coefficient of genes i and j, represents the Pearson correlation coefficient between genes i and j in the tumor sample of individual patient k, represents the Pearson correlation coefficient of genes i and j in the normal samples of individual patient k, S(i) and S(j) are the sets of tumors with mutated genes i and j, respectively, after examining somatic mutations in a given cancer dataset, and D 10 It means that after sorting a set of data in ascending order, 10% of the data are in D 10 under.

4. The constrained multi-objective optimization method for detecting drug targets in individual cancer patients according to claim 1, characterized in that: In step 3.1, a multi-task optimization approach is adopted, where the main population optimizes the original constrained multi-objective optimization problem, and the auxiliary population optimizes the corresponding derived constrained single-objective optimization problem. The evolution of these populations is regarded as multiple different but related optimization tasks. The main population is committed to global search while the auxiliary population focuses on local search, exploring areas that the main population has not searched, and helping the main population cross infeasible regions to ensure its diversity.

5. The constrained multi-objective optimization method for detecting drug targets in individual cancer patients according to claim 4, characterized in that: During environmental selection, a ranking-based fitness function method is used to calculate the fitness value of the auxiliary population.

6. The constrained multi-objective optimization method for detecting drug targets in individual cancer patients according to claim 5, characterized in that: The ranking-based fitness function method ranks each individual in two ways: one biased towards the constraints and the other biased towards the objective function, and the two rankings are weighted together by a weight parameter: First, for the ranking-based fitness function that is biased towards constraints, the individuals in the population are sorted in ascending order according to the ε constraint method, and the ranking number of each individual is recorded as R1; then, in order to use the information of the objective function, each individual is sorted based on its objective function to obtain the ranking number R2 of each individual; finally, the ranking-based fitness function combines R1 and R2 into a new formula that represents the fitness of the individual for the balanced objective and constraints, as shown below: Fitness=w*R1+(1-w)*R2 Among them, P f Indicates the proportion of feasible solutions in this generation, G is the current generation, G max is the maximum algebra.

Images