Screening and verification method of prostate cancer multi-target targets based on machine learning

Abstract

The invention discloses a prostate cancer multi-target screening and verification method based on machine learning in the medical field, which utilizes LASSO and RSF models to reflect the prostate cancer based on miRNA and mRNA sequencing data, survival information and the like through parameter optimization, carries out key miRNA mining on prostate cancer characteristics, clinical prognosis and other information, and provides a prostate cancer multi-target screening and verification method. The two prognosis models are combined for use, so that the accuracy can be greatly improved, after key miRNA is screened out, the screening accuracy is further improved and verified by constructing a risk scoring coefficient, re-grouping prostate cancer patients, screening differential genes and carrying out GSEA and GO analysis on the differential genes, clues are provided for the biological significance of the miRNA, and the application prospect is wide. Based on the analysis of the RNA expression profile and clinical prognosis characteristics of the disease, a certain determined protein target of the disease does not need to be known, and low-cost, high-efficiency and high-accuracy drug screening is realized based on the corresponding relation of miRNA and mimics or inhibitors thereof.

Description

Technical Field

[0001] This invention belongs to the medical field, specifically a method for screening and validating target markers for the treatment of prostate cancer. Background Technology

[0002] Currently, cancer, including prostate cancer, is a polygenic disease. Its pathogenesis often depends on mutations, abnormal expression levels, and copy number changes of several, or even a dozen, proto-oncogenes or tumor suppressor genes, most notably p53, Ras, and PTEN. Due to the dysfunction and inability to maintain balance within the cellular regulatory network, uncontrolled cell proliferation ultimately occurs. Therefore, current clinical treatments targeting single genes or proteins are often ineffective for cancer treatment; thus, treatment targeting multiple targets is necessary.

[0003] Multi-target drug therapy can overcome the multifactorial nature and drug resistance of malignant tumors. Due to the controllable target design and diverse composition of nucleic acid drugs, they have become ideal candidates for multi-target drugs. In fact, research on drug combinations and nucleic acid or peptide drugs targeting multiple targets is progressing rapidly. However, for many cancer types, including prostate cancer, there is still a lack of target screening methods and targeted therapies for multi-target treatment in clinical practice. Another technical limitation is that multi-target drug interventions are often designed based on past medication experience and combination drug screening methods, including CRISPR (gene editing technology) screening. However, due to the relatively limited experimental conditions and the wide variety of patient genetic characteristics, reproducibility cannot be well guaranteed, and the increasingly extensive biomedical databases are not fully utilized. Some experiments are also far removed from the clinical setting, making it difficult to effectively transfer significant therapeutic effects to real-world clinical applications.

[0004] miRNAs (microRNAs) are RNA molecules approximately 21 to 23 nucleotides long that regulate the expression of other genes. miRNAs inhibit post-transcriptional gene expression by specifically binding to mRNAs (messenger RNAs), thus serving as effective drug targets for some diseases in a "medium" manner. Since both inhibitors and mimics targeting miRNAs are relatively short RNA fragments (around 20 bp), they are easily loaded using modern nanoparticle drug delivery technologies, achieving multi-target intervention on miRNAs in target cells.

[0005] Regarding multi-target target discovery and drug use, the method provided in Chinese Patent Publication No. CN113066525A, entitled "A Multi-Target Drug Screening Method Based on Ensemble Learning and Hybrid Neural Networks," includes: acquiring data; docking processing and obtaining candidate drugs based on docking scores; identifying multi-target proteins and docking candidate drugs with them to obtain the number of target proteins targeted by the candidate drugs; predicting the activity of candidate drugs based on a pre-set ensemble learning regression model to obtain predicted activity values; predicting the binding force between pathogenic target proteins and candidate drugs based on a pre-set hybrid neural network framework to obtain binding force scores between candidate drugs and target proteins; and comprehensively determining the final candidate drugs. The problem with this model is that it presupposes that the target disease has relatively clear protein targets and that there are good targeted drugs, but these conditions are difficult to meet in many diseases.

[0006] The method described in Chinese Patent Publication No. CN104965998A, entitled "Screening Method for Multi-Target Drugs and / or Drug Combinations," first searches a drug target database, compiling data on drug-produced targets, investigational targets, and corresponding drugs for each target to obtain data on target-drug correspondences. Then, it uses systems genetics to screen for associated targets and target combinations. Based on the obtained target-drug correspondence data and associated target combinations, it screens multi-target drugs and / or drug combinations. However, this method has two main problems: first, it does not effectively utilize the correlation between patients' clinical prognostic information and their genotype or expression profile information; second, the drug-target relationship is complex and fraught with uncertainty, and drug screening methods based on this principle risk amplifying these errors. Summary of the Invention

[0007] The purpose of this invention is to solve the above-mentioned problems in existing prostate cancer treatments and to provide a machine learning-based method for screening and validating multiple miRNA targets for prostate cancer.

[0008] To achieve the above objectives, the technical solution of the machine learning-based multi-miRNA target screening method for prostate cancer of the present invention includes the following steps:

[0009] Step 1): Download RNAseq data of normal prostate tissue from publicly available biomedical databases as normal group data and RNAseq data of prostate cancer including patient prognostic information as tumor group data;

[0010] Step 2): Using the normal group data and tumor group data as input information for the Deseq2 package in R language, calculate the expression difference value logFC and statistical significance p value of each miRNA between the tumor group and the normal group, and screen out j miRNA data that meet the condition |logFC|>1.5 and p<0.05;

[0011] Step 3): Normalize the j miRNA data to obtain normalized j miRNA data; use the normalized j miRNA data and the patient prognostic information corresponding to these j miRNA data as input to the LASSO model, and the LASSO model outputs a risk score polynomial, where each term in the risk score polynomial contains a target.

[0012] Step 4): Using the j miRNA data and corresponding patient prognostic information mentioned in Step 2) as input to the RSF model, the j miRNA data are simultaneously screened to obtain a distribution map with the importance of each miRNA on the horizontal axis and the numerical distribution map of each miRNA on the vertical axis. From this distribution map, the top k miRNAs are obtained.

[0013] Step 5): Based on k miRNAs, retain the terms containing k miRNAs in the risk scoring polynomial described in Step 3), delete the remaining terms, and obtain the rewritten risk scoring polynomial. All target terms in the rewritten risk scoring polynomial are the selected multi-target terms.

[0014] Furthermore, the risk scoring polynomial based on LASSO mentioned in step 3) is RS = (A1*B1) + (A2*B2) + ... + (A... i *B i ..., A1, A2, A i For the corresponding term coefficients, B1, B2, B i , where i represents the number of target points.

[0015] Furthermore, the rewritten risk scoring polynomial is RS = A f *B f +……+A m *B m A f A m For the corresponding term, the coefficient is B, and for multi-target points, the target is B. f , ..., B m B f and B m It is a miRNA that is contained in i miRNAs and also a miRNA that is contained in k miRNAs; it is the intersection of i miRNAs and k miRNAs.

[0016] The present invention provides a method for validating a screening method, comprising the following steps:

[0017] The risk score RS value for each prostate cancer patient was calculated based on the rewritten risk score polynomial, and all patients were divided into a high RS group and a low RS group based on the RS value.

[0018] By combining existing prognostic information of prostate cancer patients, Kaplan-Meier survival curves were obtained to determine whether the risk score derived from RS values ​​was associated with poor prognosis in prostate cancer patients.

[0019] Furthermore, prostate cancer patients in the biomedical database were sorted according to their RS values, with the first quarter being the RS-high group and the last quarter being the RS-low group. The R language package limma was used as input, taking the mRNA expression data corresponding to the RS-high and RS-low groups as input, and outputting the logFC and p values ​​of the mRNAs in the two groups. Based on the condition that |logFC|>1 and p<0.05, n mRNAs with significant expression differences were selected, and the remaining mRNAs were discarded. Using the n mRNAs as input, gene function enrichment analysis was performed on the bioinformatics platform.

[0020] The advantages of this invention using the above technical solution are:

[0021] 1. This invention utilizes LASSO (Least Absolute Value Convergence and Selection Operator) and RSF (Random Survival Forest) models. Through parameter optimization, it mines key miRNAs reflecting the characteristics and clinical prognosis of prostate cancer based on its miRNA and mRNA (messenger RNA) sequencing data and survival information. Compared with simple cancer-to-near-tumor alignment and the use of a single prognostic model to screen prognostic-related key miRNAs, the combined use of the two prognostic models can greatly increase accuracy, and the screened key miRNAs have the potential to become drug targets.

[0022] 2. After screening out key miRNAs, this invention constructs risk scoring coefficients, regroups prostate cancer patients, and screens differentially expressed genes. GSEA and GO analyses of these differentially expressed genes help to further improve and verify the accuracy of the screening, while also providing clues to the biological significance of these miRNAs.

[0023] 3. Compared to existing methods that screen drug combinations based on protein targets of a specific disease through molecular docking, this invention is based on the analysis of the RNA expression profile and clinical prognostic characteristics of the disease, and does not require knowledge of a specific protein target of the disease. Furthermore, this method is not based on protein-small molecule drug docking prediction, which has a certain margin of error, but rather on the correspondence between miRNAs and their mimics or inhibitors, a correspondence that is relatively certain.

[0024] 4. In the drug target screening process, this invention combines sequencing data with prognostic analysis and incorporates two prognostic models, LASSO and RSF, for key target screening. This approach is not only highly efficient but also reflects both the prognosis of prostate cancer patients and the sequencing characteristics of the samples themselves, exhibiting high accuracy and applicability to other cancer types.

[0025] 5. This invention utilizes the correlation between a patient's clinical prognostic information and their genotype or expression profile information to achieve low-cost, high-efficiency, and high-accuracy drug screening from the perspectives of multi-target analysis, ensemble learning, and hybrid neural networks. Attached Figure Description

[0026] Figure 1 Volcano plot of differentially expressed mRNA between cancerous and adjacent normal tissues in TCGA prostate cancer patients;

[0027] Figure 2 A heatmap of differentially expressed mRNA between cancerous and adjacent normal tissues in TCGA prostate cancer patients.

[0028] Figure 3 The graph shows the calculation results of the LASSO model. The x-axis represents the penalty system, and the y-axis represents the likelihood bias.

[0029] Figure 4 The graph shows the results of the RSF model calculation, with the horizontal axis representing the importance of each miRNA and the vertical axis representing each miRNA.

[0030] Figure 5 To calculate the RS score plot for each patient based on the RS equation obtained from LASSO, the patients were divided into two groups, one high and one low, and survival curves were calculated based on this.

[0031] Figure 6 Heatmap of differential mRNA between RS-high and RS-low groups;

[0032] Figure 7 Volcano plot showing the differential mRNAs between the RS-high and RS-low groups;

[0033] Figure 8 KEGG functional enrichment plot of differentially expressed genes between RS-high and RS-low groups;

[0034] Figure 9 GSEA analysis of differentially expressed genes between the RS-high and RS-low groups shows a negative correlation with the p53 signaling pathway;

[0035] Figure 10 GSEA analysis of differentially expressed genes between the RS-high and RS-low groups shows a negative correlation with apoptosis;

[0036] Figure 11 The GSEA analysis plot of differentially expressed genes between the RS-high and RS-low groups shows a negative correlation with apical junctions. Detailed Implementation

[0037] This invention starts by exploring multiple targets for prostate cancer, using differentially expressed miRNAs between cancerous and adjacent tissues as candidate targets for screening key targets in prostate cancer. The specific method is as follows:

[0038] RNA-seq (transcriptome sequencing) data from N normal prostate tissue samples and RNA-seq data from M prostate cancer samples were downloaded from the publicly available biomedical database TCGA (Cancer Genome Atlas). M > 9N, N > 50. These RNA-seq data included mRNA (messenger RNA) and miRNA (microRNA), along with corresponding clinical information, including the survival or death status and timing of prostate cancer patients within a specific timeframe. The RNA-seq data from the N normal prostate tissue samples were used as the normal control group, and the RNA-seq data from the M prostate cancer samples were used as the tumor group. Duplicate cases in the M prostate cancer RNA-seq data were removed by calculating the average value.

[0039] The Deseq2 package in R is used as input to calculate the expression difference (logFC value) and statistical significance (p value) of each miRNA between the tumor group and the normal group.

[0040] The logFC and p values ​​were used to select j miRNAs that met the condition |logFC|>1.5 and p<0.05. These j miRNAs were considered to be miRNAs with significant differences between the tumor group and the normal group. If j>50, the remaining miRNAs that did not meet the condition were discarded. At this point, the expression matrix of j miRNAs of M prostate cancer patients was constructed.

[0041] The expression values ​​of j miRNAs were normalized using the Deseq2 package in R language to obtain normalized j miRNA data.

[0042] The normalized j miRNA data and the corresponding patient prognostic information (survival status information) of these j miRNA data are used as inputs, and the glmnet package in R language is used to fit and build a LASSO model.

[0043] For the established LASSO model, different values ​​of the regularization parameter α were tried to reduce the LASSO model while retaining significant factors. A higher value of the regularization parameter α will remove more miRNA terms that are generally not related to the prognosis of prostate cancer and retain important miRNA terms. However, the regularization parameter α should not be too high, otherwise miRNA terms with important coefficients will also be removed. In this invention, the value of the regularization parameter α was finally determined to be 0.63 to reduce the multicollinearity of the model and retain the important information of the miRNA terms that are most related to the prognosis of prostate cancer. Based on the regularization parameter α = 0.63, the LASSO model solves the coefficients by minimizing the sum of squared residuals and the sum of regularization terms, thereby constructing a polynomial equation and outputting the following risk score RS polynomial (1), each term containing a target:

[0044] RS =(A1*B1)+(A2 *B2)+……+(A i *B i )…… (1)

[0045] In equation (1), A1, A2, A i The coefficients of the corresponding terms are B1, B2, and B. i Let i be the corresponding target point, where i is the number of target points and also the number of terms in the polynomial.

[0046] The j miRNA data and corresponding patient prognostic information were used as input to the RSF model. The RandomForestSRC package in R language was called to simultaneously screen the j miRNA data using the RSF algorithm. The parameters of the RandomForestSRC package were set as follows: ntree=50, nsplit=1, na.action="na.impute", tree.err=TRUE, importance=TRUE, block.size=1, where a single random split point was used. In the output results, miRNAs with large positive importance values ​​were considered to be closely related to the RSF model, while miRNAs with importance close to or less than 0 were considered to be insignificant. A distribution map with the importance of each miRNA on the horizontal axis and the numerical distribution of each miRNA on the vertical axis was obtained. The top k miRNAs were obtained from the distribution map, k>1. Based on this, the expression of the polynomial equation of the risk score RS was rewritten. The term containing k miRNAs was retained in the risk score RS polynomial, and the rest were deleted. The rewritten risk score RS polynomial (2) is as follows:

[0047] RS = A f *B f +…… +A m *B m (2)

[0048] All target points included in the polynomial of this rewritten risk score RS are the selected multi-target targets, i.e., the multi-target targets are B. f , ..., B m , where A f A m B is the coefficient of the corresponding term. f and B m It is a miRNA that is contained in both i miRNAs and k miRNAs, and is the intersection of the i miRNAs and k miRNAs mentioned above.

[0049] The following verifies the clinical significance of the risk score RS expression: The RS risk score for each prostate cancer patient is calculated based on the rewritten RS polynomial. Patients are then halved into a high RS group and a low RS group based on their RS values. Combined with existing prognostic information (including survival time) for prostate cancer patients, the `survival` and `survminer` packages in R are used to obtain the Kaplan-Meier analysis results, i.e., the Kaplan-Meier survival curves. Kaplan-Meier analysis can determine whether the risk score derived from RS is associated with poor prognosis in prostate cancer patients.

[0050] To further elucidate the biological significance of this risk score (RS value), TCGA prostate cancer patients were ranked according to their RS values, with the top quartile classified as the RS-high group and the bottom quartile as the RS-low group. The R package Limma was then used, taking the mRNA expression data corresponding to the RS-high and RS-low groups as input, and outputting the logFC and p-values ​​of the mRNAs in both groups. Based on this, n mRNAs with significantly different expression were selected using the criteria |logFC|>1 and p<0.05, discarding the remaining mRNAs. Using these n mRNAs as input, KEGG (gene function enrichment) analysis was performed on the bioinformatics platform https: / / www.omicshare.com / to check for enrichments related to tumor-related pathways.

[0051] Gene set enrichment analysis (GSEA) was performed on the RS-high and RS-low groups. The input was the expression values ​​of all mRNAs in the RS-high and RS-low groups. The GSEA software was downloaded from "https: / / www.gsea-msigdb.org / gsea / index.jsp", and the reference gene set was set to "h.all.v7.2.symbols.gmt". The output curve showed that if the entire curve was below 0, the signaling pathway was negatively correlated with high RS; otherwise, it was positively correlated. The enrichment of tumor-related pathways such as cell proliferation, apoptosis, junctions, and migration was examined. An embodiment of the present invention is provided below:

[0052] Step 1: Download RNA-seq data from 51 normal prostate tissue samples and 484 prostate cancer samples from the TCGA database. The clinical information includes prognostic data for prostate cancer patients within a specific timeframe. The RNA-seq data from the 51 normal prostate tissue samples (adjacent to cancerous tissue) are used as the normal control group, and the RNA-seq data from the 484 prostate cancer samples are used as the tumor group. Duplicate cases are removed from the 484 prostate cancer RNA-seq data samples after calculating the average value.

[0053] Step 2: Use the Deseq2 package in R language, taking all the normal group and tumor group data as input to the Deseq2 package, and directly calculate the expression difference value (logFC value) and statistical significance value (p value) of each miRNA between the tumor group and the normal group. Figure 1 The image shows a volcano plot of differentially expressed mRNA between cancerous and adjacent tissues in TCGA prostate cancer patients. Figure 2 The image shows a heatmap of differentially expressed mRNA between cancerous and adjacent tissues in TCGA prostate cancer patients.

[0054] Step 3: Evaluate the logFC and p-values. If |logFC| > 1.5 and p < 0.05, select 59 miRNAs that meet the criteria, discarding the rest. These 59 miRNAs represent miRNAs with significant differences between the tumor and normal groups. This allows for the creation of a class matrix of miRNAs from 484 tumor tissues and 59 adjacent normal tissues of prostate cancer patients. Table 1 below shows a schematic of the class matrix of some miRNAs from some prostate cancer patients. The first row of Table 1 contains the prostate cancer patient ID, the first column contains the miRNA name, and the numbers in Table 1 represent the expression level of a specific miRNA in that prostate cancer patient.

[0055] Table 1

[0056]

[0057] Step 4: Based on the 59 selected miRNA data and the corresponding patient prognostic information (survival data) of these 59 miRNA data, the expression values ​​of the 59 miRNA data are normalized using the Deseq2 package in R language.

[0058] Use the glmnet package in R to build a LASSO model, such as... Figure 3 The diagram shows the calculation of the LASSO model, where the horizontal axis represents the penalty system and the vertical axis represents the likelihood bias. Normalized miRNA data was used as input to the LASSO model, and different values ​​of the regularization parameter α were experimented with to reduce the LASSO model while retaining significant factors. A key characteristic of the α value selection is that a higher α value removes more miRNA terms with moderate relevance to prostate cancer prognosis, retaining more important miRNA terms. However, the α value cannot be too high, otherwise, miRNA terms with significant coefficients will also be removed. Since this invention aims to discover multi-target information, RSF will be applied below. Finally, the results of LASSO and RSF will be comprehensively analyzed, with at least two miRNAs intersecting. The α value of LASSO will be set and selected so that the number of terms in the risk score coefficient (RS) expression is at least three. Then, the intersection with the miRNAs ranked in the top 10% of "importance" in the RSF analysis will be taken. If the result is less than two, the α value of LASSO will be further reduced to make the RS expression have more terms, and the intersection with the miRNAs ranked in the top 10% of importance in the RSF analysis will be taken again to see if the requirement of at least two miRNAs is met.

[0059] First, to ensure that the RS expression after LASSO analysis has 3 terms, the value of α is set to 0.63, resulting in the following trinomial equation (3) for the initial risk score coefficient RS:

[0060] RS = (0.23101947*mir-20a)+(-0.04270738*mir-221)+(-0.06095458*mir-873) (3)

[0061] Corresponding to equation (1), in this embodiment, A1 = 0.23101947, B1 = mir-20a, A2 = -0.04270738, B2 = mir-221, A3 = -0.06095458, and B3 = mir-873.

[0062] Step 5: Use the RandomForestSRC package in R to build an RSF model. Use the 59 miRNA data from Step 3 and patient prognostic information as input to the RSF model. Simultaneously filter the 59 miRNA data using the RSF algorithm. The parameters of the RandomForestSRC package are set as follows: ntree=50, nsplit=1, na.action="na.impute", tree.err=TRUE, importance=TRUE, block.size=1. A single random split point is used here. In the output, miRNAs with large positive Importance values ​​are considered closely related to the RSF model, while miRNAs with importance close to or less than 0 are considered insignificant. Figure 4 The numerical distribution of miRNAs is plotted, with the horizontal axis representing the importance of each miRNA and the vertical axis representing the numerical distribution of each miRNA. mir-20a and mir-873 are the two highest-ranking miRNAs, naturally satisfying the requirement of being in the top 10% of importance, thus demonstrating their high prognostic correlation with prostate cancer patients. Therefore, the RS expression in step four is rewritten as the following two-term equation (4):

[0063] RS = (0.23101947*mir-20a)+(-0.06095458*mir-873) (4)

[0064] The two-term equation (4) serves as the final risk score coefficient (RS), and mir-20a and mir-873 contained in equation (4) are the selected multi-target targets.

[0065] Step Six: Verify the clinical significance of the screened multi-target markers miR-20A and miR-873. The RS value for each prostate cancer patient is calculated based on the final risk score coefficient expression from Step Five, and all patients are then halved into a high RS group and a low RS group according to this RS value.

[0066] Combining the prognostic information (including survival time information) of prostate cancer patients downloaded in step one, the `survival` and `survminer` packages in R are used to output the Kaplan-Meier analysis results (Kaplan-Meier survival curves), such as... Figure 5 The graph shows the time duration on the x-axis and the survival probability on the y-axis. A point on the curve represents the survival probability of the high or low RS group at that time point. Kaplan-Meier analysis showed that the risk score derived from RS was associated with poor prognosis in prostate cancer patients; higher scores indicated worse prognosis (p = 0.021).

[0067] Step 7: To further clarify the biological significance of this RS value, the TCGA prostate cancer patients in Step 1 were sorted according to the RS values ​​calculated in Step 5. The first quarter was the RS-high group, and the last quarter was the RS-low group.

[0068] Step 8: Call the R package limma, inputting the mRNA expression data corresponding to the RS-high and RS-low groups from Step 7. The output of this step is the logFC and p-value of the mRNAs in both groups. Based on this, 153 mRNAs with significant expression differences were selected using the criteria |logFC|>1 and p<0.05. Their distribution is shown in Table 2 (partial list of mRNAs). Figure 6 heatmap and Figure 7 As shown in the volcano plot, the remaining mRNAs were discarded. The first row of Table 2 contains the names of the statistical indicators, the first column contains the gene names, and the numbers in Table 2 are the values ​​of each statistical indicator.

[0069] Table 2

[0070]

[0071] Using the aforementioned 153 mRNAs as input values, KEGG (gene function enrichment) analysis was performed on the bioinformatics platform https: / / www.omicshare.com / . The analysis revealed that differentially expressed mRNAs were enriched in functional sets such as "Cancer:Overview" and "Cancer:Specific types," indicating that RS values ​​based on mir-873 and mir-20a can indeed effectively distinguish patients with high and low levels of tumor-related signaling pathways, demonstrating a potential association between mir-873 and mir-20a and tumor-related pathways. Figure 8 The vertical axis represents different biological functional pathways, and the horizontal axis represents the number of differentially expressed genes in these pathways. A higher number indicates that the differentially expressed genes are more enriched in the signaling pathway.

[0072] Step 9: Perform gene set enrichment analysis (GSEA) on both groups. Input the expression values ​​of all mRNAs in the low RS and high RS groups. The GSEA software can be downloaded from "https: / / www.gsea-msigdb.org / gsea / index.jsp". The reference gene set is set to "h.all.v7.2.symbols.gmt". The output results are shown below. Figure 9-11 If the overall result curve is below 0, it indicates that the signaling pathway is negatively correlated with high RS; otherwise, it is positively correlated. The evaluation criteria are signaling pathways with |NES| (corrected enrichment score) > 1 and FDR q-value (p-value after multiple hypothesis testing correction) < 0.25. Figure 9-11 The Enrichment Score reveals the p53 pathway ( Figure 9 ) and apoptosis pathway ( Figure 10 Biological processes, including those related to RS-high levels, were negatively correlated with the high RS level, indicating a strong proliferative trend of cancer cells in this group. Additionally, polar junctions... Figure 11 The values ​​were also negatively correlated with the high RS group, indicating that cancer cells in this group showed a tendency to migrate. These results suggest that RS values ​​based on mir-873 and mir-20a can indeed effectively distinguish patients with low apoptosis signaling, p53 signaling, and low polarity connectivity, demonstrating that mir-873 and mir-20a are potentially associated with the function of these cells.

Claims

1. A method for screening multi-targets of prostate cancer based on machine learning, characterized in that The following steps are involved: Step 1): Download RNAseq data of normal prostate tissue as normal group data and RNAseq data of prostate cancer including patient prognosis information as tumor group data from a public biomedical database; Step 2): The normal group data and the tumor group data were used as input information of the Deseq2 package of R language to calculate the expression difference logFC and statistically significant p value of each miRNA between the tumor group and the normal group, and the j miRNA data that met |logFC|>1.5 and p<0.05 were screened out; Step 3): normalizing the j miRNA data to obtain normalized j miRNA data; using the normalized j miRNA data and the patient prognosis information corresponding to the j miRNA data as inputs of the LASSO model, and the LASSO model outputs a risk score polynomial, wherein each item in the risk score polynomial contains a target point target; Step 4): The j miRNA data and the corresponding patient prognosis information described in step 2) are used as the input of the RSF model, and the j miRNA data are synchronously screened to obtain a distribution map with the abscissa being the importance of each miRNA and the ordinate being the numerical value of each miRNA, and the top k miRNAs are obtained from the distribution map; Step 5): Based on the k miRNAs, the items containing the k miRNAs are retained in the risk score polynomial described in step 3), and the remaining items are deleted to obtain a rewritten risk score polynomial, and all target targets in the rewritten risk score polynomial are the screened multi-target targets.

2. The method for screening multi-target prostate cancer targets based on machine learning according to claim 1, characterized in that: The risk score polynomial based on LASSO in step 3) is RS = (A1*B1) + (A2*B2) + ... + (A i *B i )……, A1, A2, A i are the corresponding term coefficients, B1, B2, B i is the corresponding target, and i is the number of target.

3. The screening method according to claim 2, characterized in that: The rewritten risk score polynomial is RS = A f *B f +……+A m *B m , A f , A m is the coefficient of the corresponding term, and the multi-target target is B f , ..., B m , B f and B m It is a miRNA included in both i miRNAs and k miRNAs, and is the intersection of i miRNAs and k miRNAs.

4. The method for screening multi-targets of prostate cancer based on machine learning according to claim 1, characterized in that: In step 3), the regularization parameter of the LASSO model is 0.

63.

5. The screening method according to claim 1, characterized in that: In step 1), the RNAseq data of M cases of prostate cancer need to be averaged to remove duplicate cases.

6. The screening method according to claim 1, characterized in that: In step 2), j among the j miRNA data is greater than 50.

7. The screening method according to claim 1, characterized in that: In step 1), RNAseq data of N>50 normal prostate tissues and RNAseq data of M>9N prostate cancers were downloaded.

8. A method for verifying the screening method according to any one of claims 1 to 7, characterized in that: Calculating the risk score RS value of each prostate cancer patient according to the rewritten risk score polynomial, and dividing all patients into a high RS group and a low RS group in half according to the RS value; The Kaplan-Meier survival curve was obtained by combining the existing prognostic information of prostate cancer patients to determine whether the risk score obtained by the RS value is related to the poor prognosis of prostate cancer patients.

9. The verification method according to claim 8, characterized in that: Prostate cancer patients in the biomedical database were sorted according to RS values, with the first quarter as the RS-high group and the second quarter as the RS-low group; The R language package limma was called to take the mRNA expression data corresponding to the RS-high group and the RS-low group as input, and the logFC and p-value of the mRNA of the two groups were output; n mRNAs with significant expression differences were screened under the conditions of |logFC|>1 and p<0.05, and the rest were discarded; Taking n mRNAs as input, gene function enrichment analysis was performed on the bioinformatics platform.

10. The verification method according to claim 9, characterized in that: The result curve is output during gene function enrichment analysis. If the result curve is below 0 as a whole, it indicates a negative correlation with high RS, otherwise it is a positive correlation.

Images