Cervical cancer biomarker screening method based on multi-disease correlation analysis

Abstract

The invention provides a cervical cancer biomarker screening method based on multi-disease correlation analysis. The method comprises the following steps: carrying out preliminary prediction by utilizing a multilayer perceptron in combination with a bidirectional encoder model and a multi-disease module overlapping method; a prediction result is verified through differential expression gene analysis, and the reliability of the prediction method is ensured; determining an optimal prediction result through network proximity analysis, immune infiltration analysis and function analysis; and finally, verifying the result by using the HPA database. Compared with a traditional method, from the occurrence and development process of cervical cancer, the causal relationship of the diseases and the regulation and control relationship between the diseases are considered, and a new thought is provided for development of cervical cancer biomarkers.

Description

Technical Field

[0001] The present invention belongs to the field of disease biomarker screening, and specifically is a biomarker screening method for inducing cervical cancer based on vaginal microecology-enhanced persistent infection of human papillomavirus (HPV). Background Art

[0002] Cervical cancer is a common malignant tumor that harms women. According to data from the World Health Organization, there will be about 660,000 new cases and about 350,000 deaths in 2022, with a mortality rate of over 50%. It has been confirmed that persistent HPV infection is the main cause of cervical cancer. The existing HPV vaccine cannot eliminate persistent HPV infection or block its progression to malignant tumors. The precise prevention and treatment of cervical cancer in HPV-infected people is still a hot issue.

[0003] Persistent HPV infection is closely related to the imbalance of vaginal microecology, which is accompanied by inflammation and a decline in local immune capacity, which will affect the clearance of HPV and increase the possibility of its progression to malignant tumors. Therefore, strengthening the screening of biomarkers for HPV persistent infection-induced cervical cancer based on vaginal microecology is of great significance for the precise prevention and treatment of cervical cancer. Summary of the invention

[0004] The present invention aims to provide a biomarker screening method based on the occurrence and development process of cervical cancer. This method comprehensively considers the causal relationship in disease progression and the mutual regulation between different diseases, and makes preliminary predictions through a multi-layer perceptron combined with a bidirectional encoder model and a multi-disease module overlapping method; differentially expressed gene analysis is used to preliminarily verify the prediction results to ensure the reliability of the prediction method; the best prediction results are determined through network proximity analysis, immune infiltration analysis, and functional analysis; and finally the results are verified using the HPA database. The present invention can screen out potential cervical cancer biomarkers at low cost, high efficiency, and accuracy, providing a new idea for the development of cervical cancer biomarkers.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] The method for screening biomarkers for cervical cancer based on multi-disease association analysis includes the following steps:

[0007] (1) Data collection and processing

[0008] Gene expression profiles of patients with cervical cancer, vaginitis, and HPV infection were downloaded from the GEO database. Disease risk genes of cervical cancer, vaginitis, and HPV infection were collected from the GeneCards database and the OMIM database. Disease gene association data were obtained from the NCG database, COSMIC database, OncoKB database, CTD database, and literature search. Disease-gene networks were constructed using the CTD database, and gene-protein networks were extracted from the UniProt database. Protein-protein interaction networks were constructed using the STRING database, vaginal flora metabolite targets were collected from the MiMeDB database, drug targets were collected from the DrugBank database, and immune cell markers were collected from the TISIDB database. The collected drug targets, disease risk genes, and vaginal flora metabolite targets were standardized in the Uniprot database.

[0009] (2) Construction of Multi-layer Perceptron Combined with Bidirectional Encoder Model

[0010] The disease-gene network, gene-protein network, protein-protein network and protein-disease network were integrated into a biological heterogeneous network, and the disease semantic similarity matrix and gene function similarity matrix were introduced as auxiliary networks. The random walk algorithm was used for path sampling, and the Transformer encoder based on the self-attention mechanism was used for feature extraction. A multi-layer feedforward neural network was used to predict cervical cancer biomarkers.

[0011] (3) Multi-disease module overlapping method

[0012] The collected disease risk genes for cervical cancer, vaginitis and HPV infection were mapped to the human interaction group network respectively to form a cervical cancer disease module, a vaginitis disease module and an HPV infection disease module. They were then simultaneously mapped to the human interaction group network to obtain the regulatory relationship between the risk genes in the three modules, and the mutually regulated risk genes in the three modules were used as potential biomarkers for cervical cancer.

[0013] (4) Analysis of differentially expressed genes

[0014] The differentially expressed genes between the three diseases of cervical cancer, vaginitis and HPV infection and the normal group were calculated respectively, and the above three groups of differentially expressed genes were simultaneously mapped in the human interaction group network to obtain the regulatory relationship among the three groups. The prediction results based on the multi-layer perceptron combined with the bidirectional encoder model and the multi-disease module overlapping method were statistically analyzed to determine whether they were differentially expressed genes that were mutually regulated among the three diseases, so as to determine the reliability of the above prediction method.

[0015] (5) Screening of candidate markers

[0016] The interactome network for predicting biomarkers was constructed, and the importance of biomarkers was ranked using eight algorithms (Subgraph, Degree, Eigenvector, Betweenness, Information, LAC, closed, and Network) in Cytoscape 3.10.2 software. The top 30 markers in any four algorithms were selected as candidate markers.

[0017] (6) Network Proximity Analysis

[0018] Network proximity analysis was used to reveal whether the candidate markers had good network proximity with vaginal flora metabolites and HPV therapeutic drugs.

[0019] (7) Immune infiltration analysis

[0020] Immune infiltration analysis was used to evaluate the effects of candidate markers on the immune characteristics in the tumor microenvironment of cervical cancer based on the infiltration levels of 28 immune cells.

[0021] (8) Gene enrichment analysis

[0022] The GeneMANIA database was used to query proteins interacting with candidate biomarkers, and gene enrichment analysis was performed using the DAVID database to explore the specific mechanism of action of candidate biomarkers and determine the final prediction results.

[0023] (9) Validation of biomarkers

[0024] Using the immunohistochemistry data provided in the HPA database, the expression differences of the final predictive markers between normal cervical tissue and cervical cancer tissue were compared to ensure that the screened biomarkers have high clinical application potential.

[0025] The standardization in step (1) of the present invention refers to converting the Uniprot IDs of all drug targets, disease risk genes and vaginal flora metabolite targets into standard gene names.

[0026] The disease semantic similarity matrix in step (2) of the present invention is calculated based on a directed acyclic graph (DAG) constructed based on disease grid descriptors obtained from the MeSH database. In DAG (d), the contribution of a node n to disease d is calculated by formula 1, where N (d) represents the set of all ancestral nodes containing disease d. The similarity between disease di and disease dj is calculated by formula 2. The gene function similarity matrix is ​​calculated based on gene ontology (GO). After obtaining GO annotations using the GOSemSim package in R software, it is calculated by aggregating different semantic similarity scores of multiple GO terms related to the gene. The Transformer self-attention model is based on three learnable weight matrices {W Q ,W K ,W V}, map the sequence X to obtain Q = W Q X, K = W K X and V = W V X, and then perform self-attention operation according to Formula 3.

[0027]

[0028]

[0029]

[0030] The mutually regulated risk genes in the three disease modules in step (3) of the present invention refer to 1) risk genes shared by the three modules, 2) risk genes shared by two modules and connected to the third module, and 3) risk genes that can be connected between the three modules, including risk genes that can be connected to two groups and risk genes that can be connected to three groups.

[0031] The differentially expressed genes that are mutually regulated among the three diseases in step (4) of the present invention refer to: 1) differentially expressed genes shared by the three groups, 2) differentially expressed genes shared by two groups and capable of being connected to the third group, and 3) differentially expressed genes that can be connected to the three groups, including differentially expressed genes that can be connected to the two groups and to the three groups.

[0032] The network proximity in step (6) of the present invention involves the nearest neighbor distance (dc) between the candidate marker (A) and the drug / metabolite target (T), and the relative average shortest distance (Z dc ), dc is calculated by formula 4, Z dc Calculated by formula 5. If Z dc <0 means that the candidate marker has an influence on HPV infection and vaginal microenvironment and has the potential to become a biomarker for cervical cancer.

[0033]

[0034]

[0035] The present invention has the following beneficial effects and advantages:

[0036] 1. The cervical cancer biomarker screening method based on multi-disease association analysis described in the present invention, on the basis of clarifying that vaginal microecological imbalance and enhanced HPV persistent infection are important causes of cervical cancer, takes into account the occurrence and development process of the disease and its pathological microenvironment, and predicts cervical cancer biomarkers based on the association between diseases, which increases the reliability of the screening results.

[0037] 2. Compared with other biomarker prediction methods that only start from a single disease, the multi-layer perceptron combined with the bidirectional encoder model and the multi-disease module overlapping method proposed in the present invention reveals the mutual connection between diseases, provides insights into the comprehensive nature of the pathogenesis of cervical cancer, and overcomes the problem of mechanism explanatory power lacking in traditional computational methods.

[0038] 3. Compared with traditional experiments, the present invention can obtain potential biomarkers for cervical cancer in large quantities, quickly and at low cost, and can analyze the potential mechanisms of candidate biomarkers.

[0039] 4. The present invention provides optional potential biomarkers for the precise prevention or treatment of cervical cancer, and also provides a reliable biomarker screening method for other diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0041] Figure 2 A schematic diagram of the risk genes that can be mutually regulated among the three disease modules and a diagram of the prediction results of cervical cancer biomarkers;

[0042] Figure 3-Figure 11 The immune infiltration results of IL1B, ERBB2, AKT1, KRAS, EGFR, CDH1, CTNNB1, and PIK3CA are shown in order;

[0043] Fig.12 This is a diagram of the functional analysis results of AKT1, EGFR, CDH1, CTNNB1, and PIK3CA;

[0044] Figure 13-Figure 16 The following are the immunohistochemistry results of CTNNB1, CDH1, PIK3CA, and AKT1 in the HPA database. DETAILED DESCRIPTION

[0045] The following examples are only used to help understand the method and core concept of the present invention. It should be noted that, for those skilled in the art, without departing from the principles of the present invention, the present invention can be improved and modified in several ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention. The steps of the cervical cancer biomarker screening method based on multi-disease association analysis in the present invention are as follows: Figure 1 shown.

[0046] Example 1 Cervical cancer biomarker screening method based on multi-disease association analysis

[0047] 1. Methods and Steps

[0048] 1.1 Data Collection and Processing

[0049] The GEOquery package was used to download the gene expression profile datasets of cervical cancer, vaginitis, and HPV infection from the GEO (gene expression Omnibus) database (https: / / www.ncbi.nlm.nih.gov / geo / ). Then, the limma package was used for background correction, and the probe IDs were converted to gene names using the corresponding platform annotation package. Finally, a matrix with sample names as row names and gene names as column names was obtained for further analysis.

[0050] Disease risk genes for cervical cancer, vaginitis, and HPV infection were collected from the GeneCards database (https: / / www.genecards.org / ) and the OMIM database (https: / / www.omim.org / ). Gene association data for 89 tumors and 13 gynecological diseases were obtained through the NCG database (http: / / ncg.kcl.ac.uk / ), the COSMIC database (https: / / cancer.sanger.ac.uk / cosmic), the OncoKB database (https: / / www.oncokb.org / ), the CTD database (https: / / ctdbase.org / ), and literature retrieval. Disease-gene networks were constructed using the CTD database. In addition, gene-protein networks were extracted from the UniProt database (https: / / www.uniprot.org / ), and the corresponding protein-protein interaction networks were constructed using the STRING database (https: / / cn.string-db.org / ). Vaginal flora metabolite targets were collected from the MiMeDB database (https: / / mimedb.org / ). Interferon targets for HPV infection were compiled from the DrugBank database (https: / / go.drugbank.com / ). All genes and drug targets were sorted in the same format in Uniprot using Python 3.9. In addition, 28 immune cell markers were collected from the TISIDB database (http: / / cis.hku.hk / TISIDB / ).

[0051] The Medical Subject Headings (MeSH) of various diseases were obtained through https: / / nlmpubs.nlm.nih.gov / projects / mesh / 2022 / meshtrees / , and the disease similarity was calculated.

[0052] 1.2 Construction of Multi-layer Perceptron Combined with Bidirectional Encoder Model

[0053] The inventors used the networkx package in Python 3.6 to integrate disease-gene networks, gene-protein networks, protein-protein networks, and protein-disease networks, and further introduced disease semantic similarity matrices and gene function similarity matrices as auxiliary networks. The disease semantic similarity matrix is ​​calculated based on a directed acyclic graph (DAG) constructed based on disease grid descriptors obtained from the MeSH database. In DAG (d), the contribution of a certain node n to disease d is calculated by formula 1, where N (d) represents the set of all ancestral nodes containing disease d. The similarity between disease di and disease dj is calculated by formula 2. The gene function similarity matrix is ​​calculated based on gene ontology (GO). After obtaining GO annotations (including GO term relationships and gene-to-GO mapping) using the GOSemSim package in R software, it is aggregated and calculated by different semantic similarity scores of multiple GO terms related to genes (including maximum semantic similarity scores, average semantic similarity scores, the average of the maximum similarity scores of each row or column in the GO similarity matrix, and the average of the maximum similarity scores on all rows and columns in the GO similarity matrix).

[0054]

[0055]

[0056] The inventors used a random walk algorithm to sample paths in biological heterogeneous networks and implemented a masking mechanism. The Transformer encoder was used to extract features between nodes. The Transformer self-attention model was constructed through three learnable weight matrices {W Q ,W K ,W V}, map the sequence X to obtain Q = W Q X, K = W K X and V = W V X, and then perform self-attention operation according to Formula 3. Adam optimizer is used for model optimization. After parameter tuning, the learning rate is set to 2e-5, the batch size is 32, and the dropout rate is 0.05.

[0057]

[0058] The inventors used a multi-layer feedforward neural network to predict potential biomarkers for cervical cancer. The data set was divided into a training set and a test set in a ratio of 6:4, and the labels of the training set were defined according to the disease-biomarker association score in the CTD database. To reduce random errors, the entire experiment was trained with 10 random data splits. The model evaluation parameters and formulas are shown below, where TP is true positive, FN is false negative, FP is false positive, TN is true negative, TPR is true positive rate, and FPR is false positive rate.

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] 1.3 Construction of multi-disease module overlapping method

[0065] The collected cervical cancer, vaginitis and HPV infection disease risk genes were mapped to the human interaction group network to form cervical cancer disease module, vaginitis disease module and HPV infection disease module. The size of each disease module is represented by S, and the u test is used to calculate the Z-score to define and describe the size of S. Gene sets with the same nodes and degree values ​​as the disease module were randomly selected and repeated 1000 times. The Z-score (Formula 9) was calculated by Prand(S) of the random gene set to test the statistical significance of the disease module clustering.

[0066]

[0067] The inventors used the method of overlapping multiple disease modules to predict potential biomarkers of cervical cancer. The three disease modules formed above were simultaneously mapped in the human interaction group network to obtain the regulatory relationship of risk genes among the three modules. The risk genes that can be mutually regulated in the three modules were used as potential biomarkers of cervical cancer. The following three situations are defined as mutually regulated risk genes: (1) risk genes shared by the three modules, (2) risk genes shared by two modules and connected to the third module, and (3) risk genes that can be connected between the three modules, including risk genes that can be connected to the shared risk genes of two groups and those that can be connected to the shared risk genes of three groups.

[0068] The inventors will use the biomarkers predicted by the multi-layer perceptron combined with the bidirectional encoder model and the multi-disease module overlapping method as the preliminary prediction results.

[0069] 1.4 Analysis of differentially expressed genes

[0070] To verify the reliability of the above two prediction methods, the inventors used the limma package of R software to calculate the differentially expressed genes between the cervical cancer group, the vaginitis group, the HPV infection group and the normal group, and simultaneously mapped the above three groups of differentially expressed genes in the human interaction group network to obtain the regulatory relationship between the three groups, and statistically analyzed whether the biomarkers used as the preliminary prediction results were differentially expressed genes that were mutually regulated among the three diseases. The criteria for differential gene screening are as follows: the P value after false discovery rate (FDR) correction is less than 0.05, and the absolute value of log2 (fold-change) is greater than 0.9.

[0071] 1.5 Screening of candidate biomarkers

[0072] The protein interaction network database (STRING) was used to construct a preliminary PPI network for predicting biomarkers. The eight algorithms (Subgraph, Degree, Eigenvector, Betweenness, Information, LAC, closed, and Network) in Cytoscape3.10.2 software were used to rank their importance. The top 30 markers shared by any four algorithms were selected as candidate markers for cervical cancer.

[0073] 1.6 Network Proximity Analysis

[0074] Network proximity analysis was used to reveal whether the candidate markers had good network proximity with vaginal flora metabolites and HPV therapeutic drugs. Network proximity involves the nearest neighbor distance (dc) between the candidate marker (A) and the drug / metabolite target (T), and the relative average shortest distance (Z dc ), dc is calculated by formula 10, Z dc Calculated by formula 11. Z dc <0 indicates that the candidate marker has an influence on HPV infection and vaginal microenvironment and has the potential to become a biomarker for cervical cancer. Vaginal flora metabolite targets include acidic metabolite targets, alkaline metabolite targets, acidic alkaline metabolite targets and all metabolite targets.

[0075]

[0076]

[0077] 1.7 Immune infiltration analysis

[0078] To evaluate the effect of candidate biomarkers on immune features in the tumor microenvironment of cervical cancer, the inventors used the GSVA package in R software to quantify the infiltration of 28 immune cells based on the gene expression profile data of cervical cancer samples. Subsequently, the median expression of the candidate genes was calculated, and the cervical cancer samples were divided into high expression group and low expression group according to the median. The Wilcoxon rank sum test was used to calculate the difference in the infiltration of 28 immune cells between the two groups. If the difference was statistically significant, it was considered that the candidate marker had a potential effect on the infiltration of this immune cell.

[0079] 1.8 Gene enrichment analysis

[0080] In the study of gene mechanism, finding interacting proteins is an important step to deeply study gene function. The inventors used the GeneMANIA database (https: / / genemania.org / ) to query proteins that interact with candidate biomarkers, and performed gene enrichment analysis through the DAVID database (https: / / davidbioinformatics.nih.gov / ) to explore the specific mechanism of action of candidate biomarkers.

[0081] 1.9 Validation of biomarkers

[0082] The immunohistochemistry results in the HPA database (https: / / www.proteinatlas.org / ) were used to verify the expression differences of the predictive markers between cervical cancer tissues and normal cervical tissues to ensure that the screened biomarkers have high clinical application potential.

[0083] 2. Results and Discussion

[0084] 2.1 Data Collection and Processing

[0085] Two gene datasets were obtained from the GEO database, namely GSE63514 (including 24 normal samples, 28 cervical cancer samples and 76 HPV-infected samples) and GSE113771 (including 4 normal samples and 3 vaginitis samples). HPV infection is a necessary and sufficient condition for cervical intraepithelial lesions, so cervical intraepithelial lesion samples are also considered HPV-infected samples.

[0086] A total of 732 cervical cancer risk genes, 404 vaginitis risk genes, and 304 HPV infection risk genes were collected from GeneCards and OMIM databases. Association data of 102 diseases containing 2,717 genes were collected from NCG, COSMIC, OncoKB, CTD databases and literature reviews. 865 vaginal flora metabolite targets were retrieved from MiMeDB database, including 325 acidic metabolite targets and 162 alkaline metabolite targets. After merging and removing duplicates, 468 acidic and alkaline metabolite targets remained. Two interferon targets were collected from Drugbank database.

[0087] 2.2 Construction of Multi-layer Perceptron Combined with Bidirectional Encoder Model

[0088] The biological heterogeneous network integrated by the inventors contains 26,999 nodes and 10,599,027 edges, and the specific information is shown in Table 1. A total of 269,990 paths were sampled by the random walk algorithm. After aggregating the results of 10 repeated experiments, 315 markers showed high confidence, with a prediction score greater than 0.95, and have the potential to become potential biomarkers for cervical cancer. In addition, the good performance of the model further demonstrates the potential and practicality of this method in predicting potential biomarkers for cervical cancer, as shown in Table 2.

[0089] Table 1 Statistics of node and edge relationships in biological networks

[0090]

[0091] Table 2 Five-fold cross validation results of multi-layer feedforward neural network

[0092]

[0093] 2.3 Multi-disease module overlapping method

[0094] Among the 732 cervical cancer risk genes, 628 (Z = 8.32, P < 0.0001) were significantly located in the human interactome network, forming a cervical cancer disease module (S = 434, Z = 4.49, P = 3.65 × 10 -6 ). 355 of the 404 vaginitis risk genes (Z = 3.14, P < 0.0001) were significantly located in the human interactome network, forming a vaginitis disease module (S = 181, Z = 9.56, P = 0.00085). 256 of the 304 HPV infection risk genes (Z = 8.05, P = 4.44 × 10 -16 ) were significantly located in the human interactome network, forming an HPV infection disease module (S = 153, Z = 4.06, P = 2.41 × 10 -5). There are 18 risk genes in common among the three modules, 88 risk genes in common between two modules and connected to the third module, and 383 risk genes that can be connected to each other among the three modules. Therefore, there are a total of 489 risk genes that can be mutually regulated in the three disease modules. At the same time, 84 biomarkers were predicted by the multi-layer perceptron combined with the bidirectional encoder model and the multi-disease module overlapping method, such as Figure 2 shown.

[0095] 2.4 Analysis of differentially expressed genes

[0096] The inventors found 2857 differentially expressed genes between the normal group and the cervical cancer group, 949 differentially expressed genes between the normal group and the HPV infection group, and 2246 differentially expressed genes between the normal group and the vaginitis group. The three groups of differentially expressed genes were mapped to the human interaction network, and it was found that 1 / 3 of the markers in the preliminary prediction results had the ability to regulate these three diseases, which shows that the multi-layer perceptron combined with the bidirectional encoder model and the multi-disease module overlapping method has a high reliability in predicting cervical cancer biomarkers.

[0097] 2.5 Screening of candidate differentially expressed genes

[0098] The inventors constructed a PPI network of 84 preliminary predictive biomarkers through the STRING database. The key core markers were screened using 8 algorithms (Subgraph, Degree, Eigenvector, Betweenness, Information, LAC, Proximity, Network) in Cytoscape 3.10.2 software. The top 30 markers selected by any 4 algorithms were identified as candidate markers for cervical cancer, totaling 23, including: BRCA1, CCND1, CDKN2A, CTNNB1, ESR1, FOS, IL1B, PIK3CA, PTEN, AKT1, BCL2, CDH1, CDKN1A, EGFR, EP300, ERBB2, KRAS, MYC, SMAD2, SMAD3, STAT3, TP53 and ATM.

[0099] 2.6 Network Proximity Analysis

[0100] The network proximity values ​​of the 23 candidate markers are shown in Table 3. Among them, the proximity values ​​of STAT3, CTNNB1, PTEN, IL1B, KRAS, PIK3CA, CDH1, CCND1, ATM, ERBB2, SMAD3, SMAD2, CDKN1A, EGFR, and AKT1 to interferon targets were all less than 0 (Z dc<0), suggesting that these 15 markers may have an impact on persistent HPV infection. In addition, the proximity values ​​of these markers to vaginal microbial metabolite targets (acidity or alkalinity) are also less than 0 (Z dc <0), indicating that they may have a potential regulatory effect on the pH value of the vaginal microenvironment. These markers may play an important role in the elimination or persistence of HPV infection by affecting the vaginal pH value and increasing or decreasing the local vaginal immune capacity.

[0101] Table 3 Network proximity value statistics of candidate markers

[0102]

[0103] 2.7 Immune infiltration analysis

[0104] The effects of the above 15 candidate markers on the infiltration levels of 28 immune cells were calculated. Among them, AKT1, EGFR, CTNNB1, PIK3CA, IL1B, KRAS, CDH1 and ERBB2, a total of 8 candidate markers, had relatively significant effects on the infiltration levels of immune cells, such as Figure 3-Figure 11 As shown in the figure, IL1B and ERBB2 affect 22 types of immune cell infiltration, AKT1 affects 21 types of infiltration, KRAS affects 19 types of infiltration, EGFR affects 16 types of infiltration, CDH1 affects 15 types of infiltration, CTNNB1 affects 14 types of infiltration, and PIK3CA affects 12 types of infiltration. This suggests that these markers may play a key role in the occurrence and development of tumors.

[0105] 2.8 Gene enrichment analysis

[0106] In order to further determine the specific role of the above eight candidate markers in the occurrence and development of cervical cancer, the inventors conducted gene enrichment analysis on the proteins that interact with them. Fig.12As shown. The enrichment results of PIK3CA, CTNNB1, AKT1 and CDH1 all include HPV infection pathways (Humanpapillomavirus infection), focal adhesion pathways (Focal adhesion) and bacterial infection-related pathways (Bacterial invasion ofepithelial cells / Shigellosis). These pathways are closely related to HPV infection, bacterial invasion, cell adhesion, growth and differentiation. In addition, CTNNB1 and CDH1 are also enriched in the Hippo signaling pathway. Studies have shown that this pathway is associated with the process of HPV-infected cells transforming into cancer cells. These results show that the above biomarkers play a key role in the occurrence and development of cervical cancer. Therefore, the inventors identified PIK3CA, CTNNB1, AKT1 and CDH1 as the best biomarkers for cervical cancer screening.

[0107] 2.9 Validation of biomarkers

[0108] In order to verify the expression differences of the four biomarkers PIK3CA, CTNNB1, AKT1 and CDH1 between cervical cancer tissue and normal cervical tissue, the inventors used the immunohistochemistry data provided in the HPA database. Figure 13-Figure 16 As shown in the results, it can be clearly seen that these four biomarkers showed significant expression differences in these two different types of tissue samples.

[0109] After clarifying that vaginal microecological dysbiosis and persistent HPV infection are key factors in promoting the occurrence and development of cervical cancer, the inventors comprehensively considered the pathological microenvironment of cervical cancer and its development process, and developed a cervical cancer biomarker screening method based on multi-disease association analysis. This method combines a multi-layer perceptron with a bidirectional encoder model and multi-disease module overlapping technology as a preliminary biomarker prediction tool. The reliability of the prediction method was ensured by comparing and analyzing the differentially expressed genes between different sample groups. In order to further screen biomarkers, the inventors used network proximity analysis technology to evaluate the network distance between these markers and vaginal flora metabolite targets and HPV therapeutic drug targets, thereby quantifying their correlation. In addition, tumor microenvironment analysis was also performed to explore in depth how biomarkers affect the characteristics of the tumor microenvironment. Through protein interaction studies, the specific mechanism of action of the selected biomarkers was revealed, providing a solid theoretical basis for understanding their function in cervical cancer, and finally verified using the HPA database. This method not only helps to identify cervical cancer biomarkers with clinical application value, but also provides new ideas and directions for the early diagnosis, prevention and formulation of personalized treatment strategies for cervical cancer.

[0110] The inventors successfully identified PIK3CA, CTNNB1, AKT1 and CDH1 as potential biomarkers of cervical cancer in the critical stage of vaginal microecological dysbiosis and HPV persistent infection. From early involvement in disrupting vaginal homeostasis and perpetuating HPV infection to downstream effects on tumor immunity and cell proliferation, these genes were identified as potential biomarkers of cervical cancer, which are expected to eradicate HPV infection and prevent its progression.

[0111] The above description is a detailed description of the preferred feasible embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modified changes completed under the technical spirit suggested by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A method for screening cervical cancer biomarkers based on multi-disease association analysis, comprising the following steps: (1) Data collection and processing Gene expression profile datasets of patients with cervical cancer, vaginitis and HPV infection were downloaded from the GEO database. Disease risk genes of cervical cancer, vaginitis and HPV infection were collected from the GeneCards database and the OMIM database. Disease gene association data were obtained from the NCG database, COSMIC database, OncoKB database, CTD database and literature search. Disease-gene network was constructed using the CTD database. Gene-protein network was extracted from the UniProt database. Protein-protein interaction network was constructed using the STRING database. Vaginal flora metabolite targets were collected from the MiMeDB database, drug targets were collected from the DrugBank database, and immune cell markers were collected from the TISIDB database. The collected drug targets, disease risk genes and vaginal flora metabolite targets were standardized in the Uniprot database. (2) Construction of Multi-layer Perceptron Combined with Bidirectional Encoder Model The disease-gene network, gene-protein network, protein-protein network and protein-disease network were integrated into a biological heterogeneous network, and the disease semantic similarity matrix and gene function similarity matrix were introduced as auxiliary networks. The random walk algorithm was used for path sampling, the Transformer encoder based on the self-attention mechanism was used for feature extraction, and the multi-layer feedforward neural network was used to predict cervical cancer biomarkers. (3) Multi-disease module overlapping method The collected disease risk genes of cervical cancer, vaginitis and HPV infection were mapped to the human interaction group network respectively, forming a cervical cancer disease module, a vaginitis disease module and an HPV infection disease module, which were then simultaneously mapped to the human interaction group network to obtain the regulatory relationship between the risk genes in the three modules, and the mutually regulated risk genes in the three modules were used as potential biomarkers for cervical cancer. (4) Analysis of differentially expressed genes The differentially expressed genes between cervical cancer, vaginitis and HPV infection and the normal group were calculated respectively, and the three groups of differentially expressed genes were simultaneously mapped in the human interaction group network to obtain the regulatory relationship between the three groups. The prediction results based on the multi-layer perceptron combined with the bidirectional encoder model and the multi-disease module overlapping method were statistically analyzed to determine whether they were differentially expressed genes that were mutually regulated among the three diseases, so as to determine the reliability of the above prediction method. (5) Screening of candidate markers The interactome network for predicting biomarkers was constructed, and the importance of biomarkers was ranked using eight algorithms (Subgraph, Degree, Eigenvector, Betweenness, Information, LAC, closed, and Network) in Cytoscape 3.10.2 software. The top 30 markers in any four algorithms were selected as candidate markers. (6) Network Proximity Analysis Network proximity analysis was used to reveal whether the candidate markers had good network proximity with vaginal flora metabolites and HPV therapeutic drugs; (7) Immune infiltration analysis The immune infiltration analysis method was used to evaluate the influence of candidate markers on the immune characteristics of the cervical cancer tumor microenvironment based on the infiltration degree of 28 immune cells; (8) Gene enrichment analysis The GeneMANIA database was used to query proteins that interact with candidate biomarkers, and gene enrichment analysis was performed using the DAVID database to explore the specific mechanism of action of candidate biomarkers and determine the final prediction results; (9) Validation of biomarkers Using the immunohistochemistry data provided in the HPA database, the expression differences of the final predictive markers between normal cervical tissue and cervical cancer tissue were compared to ensure that they have high potential for clinical application.

2. The method according to claim 1, characterized in that: Normalization in step (1) refers to converting the Uniprot IDs of all drug targets, disease risk genes, and vaginal flora metabolite targets into standardized gene names.

3. The method according to claim 1, characterized in that: The disease semantic similarity matrix in step (2) is calculated based on the directed acyclic graph (DAG) constructed based on the disease grid descriptors obtained from the MeSH database. In DAG (d), the contribution of a node n to disease d is calculated by formula 1, where N (d) represents the set of all ancestor nodes of disease d; the similarity between disease di and disease dj is calculated by formula 2; The gene function similarity matrix is ​​calculated based on the Gene Ontology (GO). After obtaining GO annotations using the GOSemSim package in R software, it is calculated by aggregating the different semantic similarity scores of multiple GO terms related to genes. The Transformer self-attention model uses three learnable weight matrices {W Q ,W K ,W V }, map the sequence X to obtain Q = W Q X, K = W K X and V = W V X, and then perform self-attention operation according to Formula 3.

4. The method according to claim 1, characterized in that: The mutually regulated risk genes in the three disease modules in step (3) refer to: 1) risk genes shared by the three modules, 2) risk genes shared by two modules and connected to the third module, and 3) risk genes that can be connected between the three modules, including risk genes that can be connected to the two groups and the three groups.

5. The method according to claim 1, characterized in that: The differentially expressed genes that are mutually regulated among the three diseases in step (4) refer to: 1) differentially expressed genes shared by the three groups, 2) differentially expressed genes shared by two groups and capable of being connected to the third group, and 3) differentially expressed genes that can be connected to the three groups, including differentially expressed genes that can be connected to the two groups and to the three groups.

6. The method according to claim 1, characterized in that: The network proximity in step (6) involves the nearest neighbor distance (dc) between the candidate marker (A) and the drug / metabolite target (T), and the relative average shortest distance (Z dc ), dc is calculated by formula 4, Z dc Calculated by formula 5; if Z dc <0 means that the candidate marker has an influence on HPV infection and vaginal microenvironment and has the potential to become a biomarker for cervical cancer.

Images