Method for predicting curative effect of immunotherapy of ovarian cancer patient and application

Abstract

According to the invention, 22 related genes of lactic acid metabolism are analyzed and researched, the 22 related genes are divided into 3 ovarian cancer subgroups, and immune cells, tumor purity, stromal cells and the like infiltrated by the 3 ovarian cancer subgroups in a tumor microenvironment are obviously different and have statistical differences. On the basis, a neural network algorithm is introduced, 22 lactic acid modified genes are used as prediction indexes, and a prediction model is established to evaluate the sensitivity degree of an ovarian cancer patient in receiving anti-PD-1 treatment. For example, 22 related genes of lactic acid metabolism mentioned in the invention are collected and arranged from previous literatures and reports, and are derived from a medical library (pubard, website: https: / / pubmec.ncbi.nlm.nih.gov / ) subordinate to the National Institute of Health, 'lactate / lactation 'is used as a keyword to search, and queried related researches are manually filtered and screened. According to the present invention, the 22 lactic acid modification genes comprise STC1, PLOD2, VCAN, HBB, SLC2A3, ATF3, DUSP1, AURKA, RRM2, SLC1A4, TXNIP, HELLS, NFS1, ABCC1, SLC16A3, SLC5A8, SLC5A12, NDRG3, HIF1A, LDHA, LDHB and SLC16A1;

Description

Technical Field

[0001] The invention relates to bioinformatics and biomedicine, and in particular to the preparation of an ovarian cancer prediction system. Background Art

[0002] Ovarian cancer is a malignant tumor with a high diagnosis rate in the female reproductive system and the second most common cause of gynecological cancer death in women around the world. There are 239,000 new cases (3.6% of all cancer cases) and 152,000 deaths (4.3% of all cancer deaths) each year worldwide. According to relevant research data from Europe and the United States, one in every six women with the disease dies within 90 days of diagnosis. In 2022, the National Cancer Center released the latest cancer statistics. The number of new cases of ovarian cancer in my country is 57,200 per year, with a crude incidence rate of 8.47 / 100,000; the number of deaths per year is 27,200, and the crude mortality rate is 4.04 / 100,000. Both the incidence and mortality rates are higher than the world standard rates (5.59 / 100,000 and 2.45 / 100,000, respectively). This puts higher demands on Chinese doctors engaged in gynecological tumors, especially ovarian cancer diagnosis and treatment. The early onset of ovarian cancer is hidden, without any obvious manifestations, resulting in most of them being in the middle and late stages when they first visit the doctor. Some patients may accidentally feel a mass in the lower abdomen or discover it during a gynecological examination. They may often feel discomfort in the lower abdomen, but generally have no obvious abdominal pain. When complications such as pedicle torsion, rupture, and infection occur, lower abdominal pain may occur, and some may experience menstrual disorders or amenorrhea. If the tumor is incarcerated in the pelvic cavity, it may cause frequent urination and constipation; when a huge ovarian tumor compresses the diaphragm or pleural effusion occurs, it may manifest as dyspnea and palpitations; if it is a malignant tumor, the patient often has digestive tract symptoms such as loss of appetite and indigestion, and abdominal distension may occur when there is ascites. In the late stage of ovarian cancer, symptoms such as fatigue, weight loss, and anemia may occur.

[0003] In recent years, with the improvement of surgical concepts and techniques, the advancement of genetic testing technology, and the development of new drugs, the treatment model of ovarian cancer has basically formed a full-process management system including surgery, chemotherapy, and maintenance therapy. Surgery and chemotherapy based on platinum drugs have always been the main treatment for ovarian cancer. With the deepening of research, targeted therapy and immunotherapy have gradually been applied to the treatment of ovarian cancer in recent years. Immunotherapy (such as anti-PD-1 therapy) can activate cytotoxic T cells by blocking immune checkpoints, enabling them to directly kill tumors, bringing hope to tumor patients, especially those with advanced tumors. Immunotherapy has been widely used in tumors such as melanoma and lung cancer. Some patients have achieved clinical complete remission (CR) and tumor disappearance by immunotherapy alone. However, the application of immunotherapy in ovarian cancer is still relatively small. The main reason is that ovarian cancer patients have limited sensitivity to immunotherapy and lack effective prediction methods. Considering the high cost and potential side effects of immunotherapy, it can be seen that early diagnosis and early treatment have a positive effect on improving the poor prognosis of ovarian cancer patients.

[0004] PD-1 is a member of the CD28 family and is expressed on the surface of activated T cells, B cells, dendritic cells, and macrophages. Its function is to inhibit the activation of immune cells, which is a normal stabilization mechanism of the body's immune system. Tumors can induce infiltrating T cells to highly express PD-1 molecules, or they themselves highly express PD-L1 (PD-1's ligand), resulting in continuous activation of the PD-1 / PD-L1 pathway in the tumor microenvironment, inhibiting T cell activity and thus failing to play a tumor-killing role. This is the tumor cell immune escape mechanism.

[0005] In the human body, normal cells almost all break down glucose through aerobic respiration, and produce ATP for energy metabolism through oxidative phosphorylation of tricarboxylic acid (TCA) cycle, while 85% of the energy metabolism of tumor cells is produced through anaerobic glycolysis. This phenomenon is called the Warburg effect. Although the ATP produced by the Warburg effect is less than that of aerobic respiration, it is faster and can better meet the metabolic needs of unlimited proliferation of tumor cells. While tumor cells are anaerobic glycolysis, they will produce a large amount of lactic acid, inducing microenvironment acidification. This acidified tumor microenvironment inhibits the infiltration of immune cells, thereby allowing tumor cells to escape immune surveillance and facilitate their growth.

[0006] Neural networks are one of the most studied algorithms in recent years. Compared with traditional algorithms, they are more suitable for big data analysis such as genes, and can better fit the complex nonlinear relationship between predictor variables and outcome events in the real world. Summary of the invention

[0007] The present invention detects the expression of 22 genes related to lactic acid modification in normal ovarian tissue and ovarian tumors, and uses a dimensionality reduction algorithm to prove that the expression of the 22 genes related to lactic acid modification can reflect the immune infiltration in the tumor microenvironment of ovarian cancer patients. Based on this, the present invention introduces a neural network algorithm, uses 22 lactic acid modification related genes as prediction indicators, and establishes a prediction model to evaluate the sensitivity of ovarian cancer patients to anti-PD-1 treatment.

[0008] In a first aspect, the present invention provides a sensitivity analysis system for ovarian cancer patients after receiving immunotherapy, wherein the system is obtained by the following steps:

[0009] S01 Data acquisition: Obtain tumor samples from test subjects (resection specimens or clinical puncture specimens) and perform next-generation sequencing (NGS) to obtain TPM results of 22 lactylation-modified genes;

[0010] S02 constructs a model: inputs the TPM values ​​of the lactic acid modification-related genes screened above into the input layer of the neural network model, and then performs data fitting in the hidden layer of the neural network model; outputs a set of numbers between 0 and 1 through the output layer; and forms a sensitivity analysis system for immunotherapy of ovarian cancer patients after training and optimization;

[0011] S03 System performance evaluation: System performance evaluation is based on AUC, specificity, sensitivity, and accuracy to evaluate system performance.

[0012] Furthermore, the immunotherapy is selected from PD-L1 therapy.

[0013] Furthermore, the lactic acid modification-related genes include STC1, PLOD2, VCAN, HBB, SLC2A3, ATF3, DUSP1, AURKA, RRM2, SLC1A4, TXNIP, HELLS, NFS1, ABCC1, SLC16A3, SLC5A8, SLC5A12, NDRG3, HIF1A, LDHA, LDHB and / or SLC16A1.

[0014] Furthermore, the TPM value is the expression value of 22 lactic acid modified genes from the pathological tissue of tumor samples (resection specimens or clinical puncture specimens) of ovarian cancer patients obtained by second-generation sequencing transcriptome.

[0015] Furthermore, the neural network has a total of 9 layers, with the initial input being a linear layer of 22×16 nodes, followed by data fitting, and then a dropout layer that randomly silences 20% of the neurons, and passes it to the next linear layer of 16×4 nodes, and again data fitting and a dropout layer that silences 20% of the neurons, and finally passes through a linear layer of 4×1 nodes, and outputs a set of numbers between 0 and 1 through data fitting.

[0016] Furthermore, the data fitting is performed using batch training and batch normalization data.

[0017] Furthermore, the batch training preferably includes 30-100 data points for training each time.

[0018] Furthermore, the data fitting includes activation, optimization, normalization, weighting, verification and conversion.

[0019] Furthermore, the data activation is preferably a relu function and / or a sigmoid function.

[0020] Furthermore, the optimizer preferably uses the adam function, and the learning rate is set to 0.01.

[0021] Furthermore, the normalization refers to subtracting the mean of the "TPM" of the data from the TPM values ​​of the 22 lactic acid modification-related genes obtained and dividing by the standard deviation. The mean standard deviation of the data refers to the average value and standard deviation of the TPM expression values ​​of the 22 lactic acid modification genes, and finally the normalized values ​​of the 22 lactic acid modification-related genes are obtained respectively.

[0022] Furthermore, the weights of the 22×16-node linear layer can be selected from the following:

[0023] Table 1. Weights of the linear layer with 22×16 nodes

[0024]

[0025]

[0026]

[0027] Furthermore, the weights of the linear layer of 16×4 nodes can be selected from the following:

[0028] Table 2 Weighted values ​​of the linear layer of 16×4 nodes

[0029]

[0030]

[0031] Furthermore, the weights of the linear layer of 4×1 nodes can be selected from the following:

[0032] Table 3 Weighted values ​​of the linear layer of 4×1 nodes

[0033]

[0034] Furthermore, the verification of the data fitting is preferably an early stopping function, that is, a check is performed every 10-50 rounds, and the training is automatically terminated when the model performance is not significantly improved.

[0035] Furthermore, the data conversion preferably uses a sigmoid function to convert the output 0-1 number into a treatment sensitivity probability value.

[0036] Furthermore, the output probability value is the probability value of the sensitivity of ovarian cancer patients to immunotherapy.

[0037] Furthermore, when the probability value of detecting the sensitivity of an ovarian cancer patient is less than 50%, it indicates that the ovarian cancer patient is insensitive to immunotherapy and is not suitable for immunotherapy; when the probability value of detecting the sensitivity of an ovarian cancer patient is greater than or equal to 50%, it indicates that the patient is sensitive to immunotherapy.

[0038] Furthermore, the model evaluation is to evaluate the model performance based on AUC>85%, specificity>70%, sensitivity>98%, and accuracy>80%.

[0039] Furthermore, in the model evaluation, the closer the AUC is to 1, the specificity is greater than 70%, the sensitivity is greater than 98%, and the accuracy is greater than 80%, and the stronger the prediction efficiency of the model is.

[0040] In a second aspect, the present invention provides a method for constructing a sensitivity analysis system for predicting ovarian cancer patients after receiving immunotherapy, the method being as follows:

[0041] S01 Data acquisition: Obtain tumor samples from test subjects (resection specimens or clinical puncture specimens) and perform next-generation sequencing (NGS) to obtain TPM results of 22 lactylation modification-related genes;

[0042] S02 Model construction: input the TPM values ​​of lactic acid modification-related genes of the second-generation sequencing transcriptome of the ovarian cancer pathological tissue of the above-mentioned test subject into the input layer of the neural network model, and then perform data fitting in the hidden layer of the neural network model; output a set of numbers between 0 and 1 through the output layer; after training and optimization, form a sensitivity analysis model for immunotherapy of ovarian cancer patients;

[0043] S03 System performance evaluation: The system evaluation is based on AUC>85%, specificity>70%, sensitivity>98%, and accuracy>80% to evaluate model performance.

[0044] Furthermore, the closer the AUC is to 1, the stronger the prediction performance is.

[0045] Furthermore, the immunotherapy is selected from PD-L1 therapy.

[0046] Furthermore, the lactic acid modification-related gene is selected from STC1, PLOD2, VCAN, HBB, SLC2A3, ATF3, DUSP1, AURKA, RRM2, SLC1A4, TXNIP, HELLS, NFS1, ABCC1, SLC16A3, SLC5A8, SLC5A12, NDRG3, HIF1A, LDHA, LDHB and / or SLC16A1.

[0047] Furthermore, the TPM value is the expression value of 22 lactic acid modification genes obtained by second-generation sequencing transcriptome from lactic acid modification-related genes in pathological tissues of tumor samples (resection specimens or clinical puncture specimens) of ovarian cancer patients.

[0048] Furthermore, the neural network has a total of 9 layers, with the initial input being a linear layer of 22×16 nodes, followed by data fitting, and then a dropout layer that randomly silences 20% of the neurons, and passes it to the next linear layer of 16×4 nodes, and again data fitting and a dropout layer that silences 20% of the neurons, and finally passes through a linear layer of 4×1 nodes, and outputs a set of numbers between 0 and 1 through data fitting.

[0049] Furthermore, the data fitting is performed using batch training and batch normalization data.

[0050] Furthermore, the batch training preferably includes 30-100 data points for training each time.

[0051] Furthermore, the data fitting includes activation, optimization, normalization, weighting, verification and conversion.

[0052] Furthermore, the data activation is preferably a relu function and / or a sigmoid function.

[0053] Furthermore, the optimizer preferably uses the adam function, and the learning rate is set to 0.01.

[0054] Furthermore, the normalization refers to subtracting the mean of the "TPM" of the data from the TPM values ​​of the 22 lactic acid modification-related genes obtained and dividing by the standard deviation. The mean standard deviation of the data refers to the average value and standard deviation of the TPM expression values ​​of the 22 lactic acid modification genes, and finally the normalized values ​​of the 22 lactic acid modification-related genes are obtained respectively.

[0055] Furthermore, the linear layer weights of the 22×16 nodes may be selected from Table 1.

[0056] Furthermore, the linear layer weights of the 16×4 nodes may be selected from Table 2.

[0057] Furthermore, the linear layer weights of the 4×1 nodes may be selected from Table 3.

[0058] Furthermore, the verification of the data fitting is preferably an early stopping function, that is, a check is performed every 10-50 rounds, and the training is automatically terminated when the model performance is not significantly improved.

[0059] Furthermore, the data conversion preferably uses a sigmoid function to convert the output 0-1 number into a treatment sensitivity probability value.

[0060] Furthermore, the output probability value is the probability value of the sensitivity of ovarian cancer patients to immunotherapy.

[0061] Furthermore, when the predicted probability value of ovarian cancer patients is detected to be less than 50%, it indicates that the ovarian cancer patients are insensitive to immunotherapy and are not suitable for immunotherapy; when the predicted probability value of ovarian cancer patients is detected to be ≥50%, it indicates that the patients are sensitive to immunotherapy.

[0062] In a third aspect, the present invention provides an application of a prediction model for predicting the sensitivity of ovarian cancer patients after immunotherapy in distinguishing benefit strategies for ovarian cancer patients, wherein the strategy is the method described in the first and second aspects of the present invention.

[0063] Furthermore, when the predicted probability value of the sensitivity prediction model is ≥50%, it means that the patient is sensitive to immunotherapy and can continue to receive immunotherapy, which means that the patient is a beneficiary population;

[0064] When the predicted probability value of the sensitivity prediction model is less than 50%, it indicates that the ovarian cancer patient is insensitive to immunotherapy and is not suitable for immunotherapy, that is, is a non-benefiting population.

[0065] Furthermore, the immunotherapy is selected from PD-L1 therapy. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Dimensionality reduction algorithm diagram of 22 lactic acid modification related genes;

[0067] Figure 2 ESITMATE algorithm score diagram of 22 lactic acid modification related genes;

[0068] Figure 3 Ovarian cancer model prediction flow chart;

[0069] Figure 4 Neural network shape graph;

[0070] Figure 5 Training curve graph;

[0071] Figure 6 Schematic diagram of the prediction of sensitivity to immunotherapy in patients with ovarian cancer. DETAILED DESCRIPTION

[0072] In the present invention, by analyzing the expression difference of the 22 lactic acid modification related genes in cancer and normal ovarian tissue, the results show that the expression of the 22 lactic acid modification related genes in normal ovarian tissue or ovarian tumor is not the same, and it is statistically significant (P value <0.05). Therefore, in the present invention, the 22 lactic acid modification related genes are used as the input variables of this prediction tool to predict the sensitivity analysis of immunotherapy of ovarian cancer anti-PD-1, and the 22 lactic acid modification related genes are STC1, PLOD2, VCAN, HBB, SLC2A3, ATF3, DUSP1, AURKA, RRM2, SLC1A4, TXNIP, HELLS, NFS1, ABCC1, SLC16A3, SLC5A8, SLC5A12, NDRG3, HIF1A, LDHA, LDHB and SLC16A1.

[0073] The term "TPM" as used herein refers to Trans Per Kilobase of exon model per Million mapped reads (Trans Per Kilobase of exon model per Million mapped reads); obtained by next-generation sequencing (NGS), that is, TPM data is the standardized expression value of a gene.

[0074] The term "neural network BP (backpropagation)" used in this article is a multi-layer feedforward network trained by error back propagation (abbreviated as error back propagation). Its algorithm is called BP algorithm. Its basic idea is gradient descent method, which uses gradient search technology to minimize the loss value between the actual output value and the expected output value of the network.

[0075] The term "normalization" used in this article refers to subtracting the mean of the "TPM" of the data from the TPM values ​​of the 22 lactic acid modification-related genes obtained and dividing by the standard deviation. The mean standard deviation of the data refers to the average value and standard deviation of the TPM expression values ​​of the 22 lactic acid modification genes, and finally the normalized values ​​of the 22 lactic acid modification-related genes are obtained (see Table 4).

[0076] Table 4 Mean and standard deviation of TPM expression values ​​of 22 lactylation-related genes

[0077]

[0078]

[0079] The “dimensionality reduction algorithm” described in this article can classify ovarian cancer patients into three categories based on 22 lactylation-related genes (see Figure 1), where the ESITMATE algorithm showed that the stromal cell scores, infiltrating immune cell scores, algorithm ESTIMATE scores, and tumor purity scores of the three types of patients were significantly different (see Figure 2 ), the ESTIMATE algorithm was developed by the top international tumor research institution, the Anderson Cancer Center in the United States. The stromal cell score obtained by the algorithm can reflect the proportion of matrix components in the tissue (mainly non-tumor stromal cells, etc.), and the infiltrating immune cell score can estimate the infiltration ratio of immune cells in tumor tissue (patients with more immune cell infiltration usually have a better prognosis). The ESIMATE score is used to further infer tumor purity, and the tumor purity score can evaluate the proportion of tumor cells in the microenvironment (the more tumor cells, the worse the prognosis). Therefore, in the present invention, 22 lactic acid modification-related genes are collected and obtained from the bioinformatics website. Ovarian cancer patients are divided into three categories according to the 22 lactic acid modification-related genes. The three ovarian cancer subgroups separated have significantly different infiltrating immune cells, tumor purity, stromal cells, etc. in the tumor microenvironment, and there are statistical differences, indicating that the expression of 22 lactic acid modification-related genes can reflect the immune infiltration in the tumor microenvironment of ovarian cancer patients.

[0080] The "lactylation modification" mentioned in this article refers to the modification that mainly occurs on histones. After histone lactylation modification, metabolites such as lactyl-CoA will be produced, thereby regulating life activities (such as promoting M1-like polarization in M2 macrophages).

[0081] The experimental methods in the following examples are all conventional methods unless otherwise specified.

[0082] Example 1 Acquisition and correlation analysis of lactic acid genes

[0083] 1. Comparison of the expression of lactic acid modification-related genes in cancer and normal ovarian tissues detected by Wilcoxon method

[0084] Lactate genes included: STC1, PLOD2, VCAN, HBB, SLC2A3, ATF3, DUSP1, AURKA, RRM2, SLC1A4, TXNIP, HELLS, NFS1, ABCC1, SLC16A3, SLC5A8, SLC5A12, NDRG3, HIF1A, LDHA, LDHB, and SLC16A1.

[0085] Download the next-generation sequencing data of ovarian cancer patients from the Cancer Genome Atlas (TCGA, https: / / www.cancer.gov / ccg / research / genome-sequencing / tcga). The TCGA-OV sub-database collects the next-generation sequencing data of cancer tissues from 379 ovarian cancer patients. Download the next-generation sequencing data of normal ovarian tissues from 88 volunteers from the GETx project (https: / / gtexportal.org / home / ). Compare the expression of 22 lactylation-related genes in cancer and normal tissues.

[0086] Table 5Comparison of lactylation-related genes expression between cancer and normal ovarian tissue.

[0088]

[0089]

[0090] IQR,interquartile range.SD,standard deviation.*P<0.05,**P<0.01,***P<0.001.

[0091] The results showed that the expression of 22 lactylation modification-related genes was different in normal ovarian tissue and ovarian tumors.

[0092] At the same time, we found that only 22 lactylation-related genes can further classify ovarian cancer patients in TCGA-OV. The dimensionality reduction algorithm suggests that ovarian cancer patients can be divided into three categories based on the 22 lactylation-related genes (see Figure 1 ), the ESITMATE algorithm showed that the stromal cell scores, infiltrating immune cell scores, algorithm ESTIMATE scores, and tumor purity scores of the three types of patients were significantly different (see Figure 2 ). Further analysis of 22 lactylation-related genes collected from bioinformatics websites revealed that ovarian cancer patients can be grouped according to the expression of the 22 lactylation-related genes. The three ovarian cancer subgroups identified had significantly different infiltrating immune cells in the tumor microenvironment, tumor purity, and stromal cells, with statistically significant differences.

[0093] Therefore, it was demonstrated that the expression of 22 lactylation modification-related genes could reflect the immune infiltration in the tumor microenvironment of ovarian cancer patients.

[0094] Example 2 Establishment and evaluation of sensitivity analysis model for ovarian cancer patients after immunotherapy

[0095] Patient inclusion criteria in the model: Patients diagnosed with epithelial ovarian cancer who relapsed or progressed less than 6 months after the last platinum chemotherapy and received guadecitabine and pembrolizumab (anti-PD-1 drug) in the Northwestern University clinical trial Guadecitabine and Pembrolizumab in Treating Patients With Recurrent Ovarian, Primary Peritoneal, or Fallopian Tube Cancer (clinical trial registration number: NCT02901899) were included. The original sequencing data (TPM, etc.) were stored on the National Institutes of Health Gene Expression Omnibus (number: GSE188249).

[0096] According to the process of the prediction model of ovarian cancer patients using the neural network algorithm (see Figure 3 ), according to the internationally recognized solid tumor treatment evaluation standard RECIST 1.1, the treatment effect of the patient is evaluated, and the effectiveness of the patient after anti-PD-1 treatment is detected (the probability of complete remission CR or partial remission PR). The model operation steps are as follows:

[0097] 1. Users should send ovarian cancer patient tumor samples (resection specimens or clinical puncture specimens) to a third-party company for next-generation sequencing (NGS) to detect the TPM results of 22 lactylation-modified genes and enter the test results into the software.

[0098] 2. After the TPM data is entered, the user can click the "Predict" button, and the software will automatically call the built-in pre-trained neural network algorithm to predict the probability of the patient being sensitive to anti-PD-1 treatment (complete remission CR or partial remission PR). The predicted results will be displayed in the "Result" area of ​​the software.

[0099] 3. Output prediction results: The output value of the neural network is a probability value, which is the probability that the system predicts the patient's sensitivity to the treatment effect (complete remission CR or partial remission PR) after receiving anti-PD-1 treatment.

[0100] Training method: The expression values ​​of the 22 genes initially input are used as nodes of the input layer. Batch normalization is enabled, dropout layer (randomly silence some neurons, here set to silence 20%), batch training (i.e. only select a part of the data instead of all the data for training each time, this time set to select 64 data each time), select relu function as activation function (convert linear layer to nonlinear layer) and select adam function as optimizer (learning rate set to 0.01), earlystopping function is turned on (this function automatically monitors the training process, and automatically stops training if the effect is still not significantly improved after several rounds of training, here set to 30 rounds), the output item is a single unbiased item (only a number is returned in the end instead of a formula), and the output item is finally converted by sigmoid function (the output number is converted into a probability value).

[0101] Neural network shape (see Figure 4 ): This neural network has a total of 9 layers. The initial input is a linear layer of 22×16 nodes, which is then activated by the relu function and normalized by the standardization layer (minus the "TPM" mean of the data and divided by the standard deviation). Then the dropout layer randomly silences 20% of the neurons and passes it to the next 16×4 node linear layer, which is activated by the relu function again, and normalized and silenced by the standardization layer and dropout layer again. 20% of the neurons, finally pass through a 4x 1 node linear layer, after activation and conversion by the sigmoid function, a set of values ​​between 0 and 1 is output. The training process is recorded in the training curve (see Figure 5 ). The horizontal axis of the curve is a training round (called "epoch"), and the vertical axis is the difference between the predicted value and the true value, called "loss". As the training progresses, the loss will theoretically decrease. Since we have enabled the early stopping function that automatically monitors and terminates training, the function will continue to monitor the training process. When the model performance continues for 30 epochs and the effect is still not significantly improved, the training will be automatically terminated to prevent overfitting.

[0102] The area under the receiver operating curve (AUC) is used to evaluate model performance. The model has predictive value only when AUC is greater than 0.5, and the closer it is to 1, the stronger the predictive efficiency.

[0103] After 30 rounds of training, the early stopping function automatically terminated the training. On the Train cohort training set, the AUC was 1.0000, and on the Validation cohort validation set, the AUC was 0.8750, the specificity was 0.7500, the sensitivity was 1.0000, and the accuracy was 0.8333. The closer the AUC value is to 1, the stronger the prediction efficiency is (see Table 6). Therefore, the establishment of this model can more accurately predict the immune efficacy of patients with ovarian cancer immunotherapy and provide reference for clinicians.

[0104] Table 6 Performance of deep learning models for predicting immunotherapy response in ovarian cancer.

[0105] Table Performance ofdeep learningmodelusedtopredictovarian cancer'simmunotherapy

[0106] response.

[0107]

[0108] AUC,area under the receiver operator characteristic curve.NPV,negative predictive value.PPV,

[0109] positive predictive value.

[0110] Finally, the test found that by randomly testing the immunotherapy sensitivity of a certain ovarian cancer patient, the model predicted that the probability of the ovarian cancer patient being sensitive to immunotherapy was 52% (see Figure 6 ), indicating that the ovarian cancer patient is suitable for immunotherapy.

Claims

1. A system for predicting the sensitivity of ovarian cancer patients after receiving immunotherapy, the system is obtained by the following steps: S01 Data acquisition: Obtain tumor samples from test subjects (resection specimens or clinical puncture specimens) and perform next-generation sequencing (NGS) to obtain TPM results of 22 lactylation-modified genes; S02 constructs a model: inputs the TPM values ​​of the lactic acid modification-related genes screened above into the input layer of the neural network model, and then performs data fitting in the hidden layer of the neural network model; outputs a set of numbers between 0 and 1 through the output layer; and forms a sensitivity analysis system for immunotherapy of ovarian cancer patients after training and optimization; S03 System performance evaluation: System performance evaluation is based on AUC, specificity, sensitivity, and accuracy to evaluate system performance.

2. The system for predicting the sensitivity of ovarian cancer patients to immunotherapy according to claim 1, characterized in that: The 22 lactate modification-related genes are selected from: STC1, PLOD2, VCAN, HBB, SLC2A3, ATF3, DUSP1, AURKA, RRM2, SLC1A4, TXNIP, HELLS, NFS1, ABCC1, SLC16A3, SLC5A8, SLC5A12, NDRG3, HIF1A, LDHA, LDHB, and SLC16A1.

3. The system for predicting the sensitivity of ovarian cancer patients after receiving immunotherapy according to claim 1, characterized in that: The TPM value is the expression value of 22 lactic acid modification related genes obtained from the pathological tissue of tumor samples (resection specimens or clinical puncture specimens) of ovarian cancer patients through the second generation sequencing transcriptome.

4. The system for predicting the sensitivity of ovarian cancer patients after receiving immunotherapy according to claim 1, characterized in that: The neural network has a total of 9 layers. The initial input is a linear layer of 22×16 nodes, followed by data fitting, and then a dropout layer randomly silences 20% of the neurons and passes it to the next linear layer of 16×4 nodes. Data fitting and dropout layer silence 20% of the neurons again, and finally a linear layer of 4×1 nodes outputs a set of numbers between 0 and 1 through data fitting.

5. The system for predicting sensitivity analysis of ovarian cancer patients after receiving immunotherapy according to claim 1, characterized in that: The data fitting is performed by batch training and batch normalization data, and preferably 30-100 data are involved in the batch training each time.

6. The system for predicting sensitivity analysis of ovarian cancer patients after receiving immunotherapy according to claim 1, characterized in that: The data conversion is preferably a sigmoid function, which converts the output 0-1 number into a treatment sensitivity probability value, and the output probability value is the probability value of the sensitivity of ovarian cancer patients to immunotherapy.

7. The system for predicting sensitivity analysis of ovarian cancer patients after receiving immunotherapy according to claim 1, characterized in that: When the probability value of detecting the sensitivity of an ovarian cancer patient is less than 50%, it indicates that the ovarian cancer patient is insensitive to immunotherapy and is not suitable for immunotherapy; when the probability value of detecting the sensitivity of an ovarian cancer patient is greater than or equal to 50%, it indicates that the patient is sensitive to immunotherapy.

8. A method for constructing a sensitivity analysis system for predicting ovarian cancer patients after receiving immunotherapy, the method being as follows: S01 Data acquisition: Obtain tumor samples from test subjects (resection specimens or clinical puncture specimens) and perform next-generation sequencing (NGS) to obtain TPM results of 22 lactylation modification-related genes; S02 Model construction: input the TPM values ​​of lactic acid modification-related genes of the second-generation sequencing transcriptome of the ovarian cancer pathological tissue of the above-mentioned test subject into the input layer of the neural network model, and then perform data fitting in the hidden layer of the neural network model; output a set of numbers between 0 and 1 through the output layer; after training and optimization, form a sensitivity analysis model for immunotherapy of ovarian cancer patients; S03 System performance evaluation: The system evaluation is based on AUC>85%, specificity>70%, sensitivity>98%, and accuracy>80% to evaluate model performance.

9. Use of the sensitivity analysis system for predicting ovarian cancer patients after receiving immunotherapy as described in any one of claims 1 to 8 in distinguishing benefit strategies for ovarian cancer patients.

10. The use of the sensitivity analysis system for predicting ovarian cancer patients after receiving immunotherapy according to claim 9 in distinguishing benefit strategies for ovarian cancer patients, characterized in that: When the predicted probability value of the sensitivity prediction model is ≥50%, it means that the patient is sensitive to immunotherapy and can continue to take immunotherapy, that is, the beneficiary population; when the predicted probability value of the sensitivity prediction model is <50%, it means that the ovarian cancer patient is not sensitive to immunotherapy and is not suitable for immunotherapy, that is, the non-benefiting population.

Images