Tumor immunotherapy efficacy prediction model based on arl6ip5 gene and associated genes

Abstract

The present application relates to bioinformatics and biomedical. Specifically relates to the present application provides a kind of based on neural network algorithm's sensitivity prediction method of immunotherapy of tumor patient, the actual diagnostic value of the utilization ARL6IP5 gene and its associated gene (total 853 genes) and patient immunotherapy curative effect is used as prediction index, to establish prediction model, to evaluate the sensitivity of tumor clinical immunotherapy;When detecting that tumor patient receives immunotherapy, its prediction probability value <50%, it is explained that the prognosis of the patient to immunotherapy is poor, is not suitable for taking immunotherapy;When detecting that tumor patient receives immunotherapy, its prediction probability value ≥50%, it is explained that the prognosis effect of the method of the patient to immunotherapy is good, is suitable for receiving immunotherapy.

Description

Technical Field

[0001] This invention relates to bioinformatics and biomedicine, specifically to the preparation of a tumor immunotherapy efficacy prediction model. Background Technology

[0002] Current research indicates that cancer is the second leading cause of death after cardiovascular disease. Malignant tumors are among the major diseases threatening human health in the new century and a significant global public health problem, imposing a tremendous disease burden on society. Cancer is a genetic disease caused by the long-term combined effects of multiple factors. When the body is exposed to carcinogens such as chemicals, physical agents, and viruses in the environment, and / or due to factors such as genetics, endocrine function, sex, and age, a series of abnormal gene changes occur, leading to the formation of a tumor. The clinical manifestations of cancer patients are related to factors such as the nature, location, and growth rate of the tumor. In the early stages of the disease, there are often no significant symptoms. As the tumor grows, local symptoms such as swelling, pain, bleeding, and obstruction appear, as well as systemic symptoms such as fever, fatigue, anemia, and weight loss.

[0003] Various cancer treatments include chemotherapy, targeted therapy, surgery, radiotherapy, and biotherapy, which have provided some benefits to cancer patients. However, some errors still exist in cancer treatment strategies. These errors not only cause patients considerable suffering during treatment but also fail to improve prognosis in some cancers. Furthermore, the high cost of treatment ultimately leads to a decline in patients' quality of life, and the treatment effect does not meet expectations. In addition, the inherent biological characteristics of tumors and individual differences are ignored, and the sensitivity of tumors to treatment and the patient's genetic background are not fully considered, resulting in different prognoses for patients receiving the same treatment. Predictive biomarkers for cancer treatment efficacy include proteins, messenger RNA, DNA, microRNA, metabolites, or integration of these biomolecules. Chinese Patent 107976442A discloses a cancer prognostic predictive biomarker—lymph node tissue—and a method for detecting lymph nodes in predicting tumors. However, the detection process for lymph node predictive biomarkers is time-consuming and cumbersome. Therefore, there is an urgent need for a simple, reliable, and highly accurate cancer prediction system.

[0004] In recent years, a new treatment approach, immunotherapy, has brought hope to cancer patients, especially those with advanced-stage cancer. Immunotherapy drugs such as immune checkpoint inhibitors (ICIs) significantly prolong patient survival by blocking immune checkpoints like PD-L1, and some patients have even achieved complete pathological remission with immunotherapy combined with chemotherapy. However, currently only 20-40% of cancer patients are sensitive to immunotherapy. Current screening methods for potential beneficiaries of immunotherapy mainly include tests for programmed cell death 1 ligand 1 (PD-L1), tumor mutational burden (TMB), microsatellite instability (MMI), and lymphocyte infiltration. These tests, based on invasive procedures obtained from lesions, are often invasive, relying on pathological tissue. Their clinical application is limited by factors such as the size and location of clinical specimens, and the high cost of invasive procedures increases patient discomfort. Meanwhile, due to the unique biological principles of immunotherapy, the aforementioned methods have limited predictive power for the efficacy of combined immunotherapy and chemotherapy. The sensitivity and specificity of these detection methods are mostly between 60% and 70%, and there is room for improvement in their predictive efficacy for the overall population. Furthermore, these tests can only provide a reference for the potential efficacy of immunotherapy regimens in clinical diagnosis and treatment; they cannot provide additional treatment pathways, improve the efficacy of immunotherapy, or even reverse primary drug resistance in patients. They also consume significant medical resources, resulting in a limited overall cost-effectiveness ratio. In addition, current single-marker, single-dimensional detection methods can only infer imaging or pathological responses and cannot distinguish between patients with objective radiological remission but not major pathological remission, presenting significant limitations. Currently, there is also a lack of universally recognized and effective predictive indicators or tools for immunotherapy efficacy. Considering the high cost, long treatment duration, varying prognoses due to individual differences, and various side effects of immunotherapy, the design of tools that effectively predict immunotherapy efficacy can identify individuals sensitive or tolerant to immunotherapy in advance, reducing unnecessary medical costs and adverse drug events. Summary of the Invention

[0005] This invention utilizes the ARL6IP5 gene and its associated genes as predictive indicators to establish a predictive model for evaluating the efficacy of clinical immunotherapy for tumors.

[0006] In a first aspect, the present invention provides a method for preparing a predictive model of immunotherapy sensitivity in cancer patients, the method comprising the following steps: S01 Data Acquisition: Acquire the TPM expression values ​​of the ARL6IP5 gene and its associated genes in the next-generation sequencing transcriptome of pathological tissues from immunotherapy patients, as well as the actual diagnostic values ​​of the patients' immunotherapy efficacy. S02 Model Construction: The above two values ​​are used as prevariable data input layers to the neural network model, and data fitting is performed in the hidden layers of the neural network model; a set of values ​​between 0 and 1 is output through the output layer; after training and optimization, a prognostic model for immunotherapy of tumor patients is formed. S03 Evaluate model performance: The area under the receiver operating curve (AUC) is used to evaluate model performance; when the AUC is greater than 0.5, the model has predictive value, and the closer it is to 1, the stronger the predictive power.

[0007] Furthermore, the tumor patient is preferably a patient with bladder cancer (transitional cell carcinoma of the urinary tract).

[0008] Furthermore, the immunotherapy is preferably PD-L1 therapy.

[0009] Furthermore, the ARL6IP5 gene and its associated genes (a total of 853 genes) are shown in the table below: Table 1. List of ARL6IP5 genes and their associated genes

[0010] Furthermore, the TPM value is the expression value of the ARL6IP5 gene and its associated genes obtained by next-generation sequencing transcriptome analysis from the pathological tissue of immunotherapy patients (resected by puncture or surgery).

[0011] Furthermore, the actual diagnostic value of the patients' immunotherapy efficacy is assessed by senior clinicians at the associate chief physician level or above according to the recist 1.1 criteria, and their treatment outcome (BOR) is evaluated. The BOR includes immunotherapy response (sensitive) (CR+PR) and immunotherapy ineffectiveness (insensitive) (SD+PD).

[0012] Furthermore, the immunotherapy response (sensitive) (CR+PR) is recorded as 1; the immunotherapy ineffective (insensitive) (SD+PD) is recorded as 0.

[0013] Furthermore, the neural network has a total of 10 layers. The initial input is a linear layer with 853×8 nodes. After data fitting, a dropout layer randomly silences 30% of the neurons and passes them to the next 8×4 node linear layer. Data fitting and another dropout layer silence 30% of the neurons. Finally, it passes through a 4×1 node linear layer and outputs a set of values ​​between 0 and 1 through data fitting.

[0014] Furthermore, the data fitting is performed using batch training and batch standardized data.

[0015] Furthermore, each batch training session selects 30-100 data points for training.

[0016] Furthermore, the data fitting includes activation, optimization, normalization, weighting, verification, and transformation.

[0017] Furthermore, the data activation is preferably performed using the ReLU function and / or the sigmoid function.

[0018] Furthermore, the optimizer preferably uses the Adam function, with a learning rate set to 0.01.

[0019] Furthermore, the normalization refers to subtracting the mean of the data from the standardized value and dividing by the standard deviation.

[0020] Furthermore, the linear layer weighting of the 853×8 nodes can be selected from the following: Table 2. Linear layer weighting values ​​for 853×8 nodes

[0021] Furthermore, the linear layer weighting of the 8×4 nodes can be selected from the following: Table 3. Linear Layer Weighting Values ​​for 8×4 Nodes

[0022] Furthermore, the linear layer weighting of the 4×1 nodes can be selected from the following: Table 4. Linear Layer Weighting Values ​​for 4×1 Nodes

[0023] Furthermore, the data fitting verification is preferably performed using an early stopping function, which checks the model every 10-50 rounds and automatically terminates training when the model performance improvement is not significant.

[0024] Furthermore, the data conversion is preferably performed using the sigmoid function, which converts the output number into a value between 0 and 1.

[0025] Furthermore, the output set of values ​​between 0 and 1 represents the predicted probability of a cancer patient's sensitivity to immunotherapy.

[0026] Furthermore, when the predicted probability value of a tumor patient is detected to be <50%, it indicates that the tumor patient is not sensitive to immunotherapy and is not suitable for immunotherapy; when the predicted probability value of a tumor patient is detected to be ≥50%, it indicates that the patient is sensitive to immunotherapy.

[0027] Secondly, the present invention provides a sensitivity prediction and analysis model for immunotherapy in cancer patients, which is obtained by the following method: S01 Data Acquisition: Acquire the TPM expression values ​​of the ARL6IP5 gene and its associated genes in the next-generation sequencing transcriptome of pathological tissues from immunotherapy patients, as well as the actual diagnostic values ​​of the patients' immunotherapy efficacy. S02 Model Construction: The above two values ​​are used as prevariable data input layers to the neural network model, and data fitting is performed in the hidden layers of the neural network model; a set of values ​​between 0 and 1 is output through the output layer; after training and optimization, a prognostic model for immunotherapy of tumor patients is formed. S03 Evaluate model performance: The area under the receiver operating curve (AUC) is used to evaluate model performance; when the AUC is greater than 0.5, the model has predictive value, and the closer it is to 1, the stronger the predictive power.

[0028] Furthermore, the tumor patient is preferably a patient with bladder cancer (transitional cell carcinoma of the urinary tract).

[0029] Furthermore, the immunotherapy is preferably PD-L1 therapy.

[0030] Furthermore, the TPM value is the expression value of the ARL6IP5 gene and its associated genes obtained by next-generation sequencing transcriptome analysis from the pathological tissue of immunotherapy patients (resected by puncture or surgery).

[0031] Furthermore, the actual diagnostic value of the patients' immunotherapy efficacy is assessed by senior clinicians at the associate chief physician level or above according to the recist 1.1 criteria, and their treatment outcome (BOR) is evaluated. The BOR includes immunotherapy response (sensitive) (CR+PR) and immunotherapy ineffectiveness (insensitive) (SD+PD).

[0032] Furthermore, the immunotherapy response (sensitive) (CR+PR) is recorded as 1; the immunotherapy ineffective (insensitive) (SD+PD) is recorded as 0.

[0033] Furthermore, the neural network has a total of 10 layers. The initial input is a linear layer with 853×8 nodes. After data fitting, a dropout layer randomly silences 30% of the neurons and passes them to the next 8×4 node linear layer. Data fitting and another dropout layer silence 30% of the neurons. Finally, it passes through a 4×1 node linear layer and outputs a set of values ​​between 0 and 1 through data fitting.

[0034] Furthermore, the data fitting is performed using batch training and batch standardized data.

[0035] Furthermore, each batch training session selects 30-100 data points for training.

[0036] Furthermore, the data fitting includes activation, optimization, normalization, weighting, verification, and transformation.

[0037] Furthermore, the data activation is preferably performed using the ReLU function and / or the sigmoid function.

[0038] Furthermore, the optimizer preferably uses the Adam function, with a learning rate set to 0.01.

[0039] Furthermore, the normalization refers to subtracting the mean of the data from the standardized value and dividing by the standard deviation.

[0040] Furthermore, the linear layer weighting of the 853×8 nodes can be selected from Table 2.

[0041] Furthermore, the linear layer weighting of the 8×4 nodes can be selected from the table.

[0042] Furthermore, the linear layer weighting of the 4×1 nodes can be selected from Table 3.

[0043] Furthermore, the data fitting verification is preferably performed using an early stopping function, which checks the model every 10-50 rounds and automatically terminates training when the model performance improvement is not significant.

[0044] Furthermore, the data conversion is preferably performed using the sigmoid function, which converts the output number into a value between 0 and 1.

[0045] Furthermore, the output set of values ​​between 0 and 1 represents the predicted probability of a cancer patient's sensitivity to immunotherapy.

[0046] Furthermore, when the predicted probability value of a tumor patient is detected to be <50%, it indicates that the tumor patient is not sensitive to immunotherapy and is not suitable for immunotherapy; when the predicted probability value of a tumor patient is detected to be ≥50%, it indicates that the patient is sensitive to immunotherapy.

[0047] Thirdly, the present invention provides an application in a strategy for identifying cancer patients as sensitive to or beneficiaries of immunotherapy, the strategy being the method described in the first and second aspects of the present invention.

[0048] Furthermore, the application is further defined as follows: when the predicted probability value of a tumor patient is detected to be <50%, it indicates that the tumor patient is not sensitive to immunotherapy and is not suitable for immunotherapy; when the predicted probability value of a tumor patient is detected to be ≥50%, it indicates that the patient is sensitive to immunotherapy and needs to undergo immunotherapy.

[0049] Furthermore, the tumor patient is preferably a patient with bladder cancer (transitional cell carcinoma of the urinary tract).

[0050] Furthermore, the immunotherapy is preferably PD-L1 therapy. Attached Figure Description

[0051] Figure 1 Schematic diagram of tumor neural network Figure 2 The early stopping function automatically monitors the training process curve. Figure 3 Schematic diagram of methods for predicting the sensitivity of cancer patients to immunotherapy. Figure 4 Performance evaluation of immunotherapy sensitivity prediction models for cancer patients Figure 5 Schematic diagram of immunotherapy sensitivity prediction in cancer patients Detailed Implementation

[0052] This invention analyzes and extracts all genes related to ARL6IP5 expression from internationally renowned bioinformatics websites (statistical P value < 0.05) as input variables for this prediction tool to predict the efficacy of immunotherapy.

[0053] The term "TPM" as used herein refers to Trans per Kilobase of exon model per Million mapped reads (TPM). Obtained through next-generation sequencing (NGS), TPM data represents the standardized expression value of a gene. In one embodiment, the ARL6IP5 gene and its associated genes are extracted from pathological tissue of cancer patients undergoing immunotherapy. Next-generation sequencing is used to obtain the TPM data of the ARL6IP5 gene and its associated genes transcriptome from these patients. This TPM expression value is used as the standardized value for the input layer of a neural network, and a pre-trained neural network algorithm is used to predict the probability of the patient's sensitivity / benefit from immunotherapy. In this invention, the ARL6IP5 gene and its associated genes from cancer patients undergoing immunotherapy are obtained as raw data and standardized values ​​are obtained. These standardized values ​​refer to the TPM expression values ​​of the ARL6IP5 gene and its associated genes from the 192 selected cancer patients in this invention.

[0054] Normalization refers to subtracting the mean of the data from the standardized value and dividing by the standard deviation. The mean and standard deviation of the data refer to the average and standard deviation of the TPM expression values ​​of the ARL6IP5 gene and its related genes in 192 enrolled cancer patients.

[0055] The actual diagnostic value for immunotherapy efficacy in cancer patients: The cancer patients described in this article are bladder cancer (transitional cell carcinoma of the urogenital tract) patients who received PD-L1 immunotherapy. The actual diagnostic value for immunotherapy efficacy (sensitivity) in these patients is determined according to the CREIST 1.1 criteria, assessed by senior clinicians at the associate chief physician level or above (best of response, BOR). BOR includes: CR (complete response); PR (partial response); SD (stabeldisease); PD (progressive disease). A BOR value of 0 or 1 indicates that CR+PR is defined as immunotherapy efficacy (sensitivity), recorded as 1; SD+PD is defined as immunotherapy ineffectiveness (insensitivity), recorded as 0.

[0056] In one implementation, immunotherapy refers to 192 bladder cancer patients who received PD-L1 immunotherapy, with their TPM expression values ​​from next-generation sequencing (NGS) data and the actual diagnostic value of the immunotherapy efficacy in tumor patients as prevariables.

[0057] Example 1: Establishment and evaluation of a neural network algorithm for predicting the efficacy of immunotherapy in patients with tumors. We compiled data from 192 patients with bladder cancer (transitional cell carcinoma of the urinary tract) who received PD-L1 immunotherapy; we collected TPM values ​​and actual diagnostic values ​​of immunotherapy efficacy in tumor patients to construct this model.

[0058] TPM value acquisition: The ARL6IP5 gene and its related genes were isolated from the tumor tissue of the selected patients participating in this model construction in advance; the TPM values ​​of 853 genes of the ARL6IP5 gene and its related genes of the patient were obtained by next-generation sequencing and saved to a file for later use.

[0059] The actual diagnostic value for effective (sensitive) immunotherapy: For tumor patients who have undergone PD-L1 immunotherapy, their treatment outcomes (best of response, BOR) are evaluated by senior clinicians at the associate chief physician level or above according to the recist 1.1 criteria. The BOR value is 0 or 1, that is, CR+PR is defined as effective (sensitive) immunotherapy and recorded as 1; SD+PD is defined as ineffective (insensitive) immunotherapy and recorded as 0.

[0060] Of the 192 bladder cancer (transitional cell carcinoma of the urotract) patients who received PD-L1 immunotherapy, 70% were randomly assigned to the Discovery group (N=134), 20% to the Validation group (N=40), and the remaining 10% to the Test group (N=18). The Discovery group was used for model training, and the Validation and Test groups were used for testing (see Table 5).

[0061] Training Method: The initial input consists of the standardized values ​​of all 853 relevant genes of ARL6IP5 and the nodes used as the input layer. Dropout is enabled (randomly silencing some neurons, here set to 30%), batch training is used (i.e., training only a portion of the data at a time instead of all data, here set to 64 data points per batch), ReLU is selected as the activation function (converting linear layers into non-linear layers), and Adam is selected as the optimizer (learning rate set to 0.01). The early stopping function is enabled (this function automatically monitors the training process and automatically stops training if the improvement is not significant after several rounds of training, here set to 30 rounds, or 30 epochs), the output is a single unbiased term (returning only a number instead of a formula), and the output is finally transformed by the sigmoid function (converting the output number into a probability value between 0 and 1).

[0062] The algorithm is built upon the survival theory of neural network algorithms. The aforementioned data serves as the input layer to the neural network model in this scheme. The model includes an input layer, hidden layers, and an output layer. The aforementioned data is input into this prediction system. This prediction system has two hidden layers with 8 and 4 nodes respectively, and also incorporates batch normalization and a 30% dropout ratio to help improve model performance. The output layer has only one layer, with no bias term, and the output value is a probability value indicating the probability that a patient is sensitive to immunotherapy.

[0063] Neural network shape (see) Figure 1 This neural network consists of 10 layers. The initial input is an 853×8 node linear layer, which is then activated by the ReLU function and normalized by a standardization layer (subtracting the mean of the data and dividing by the standard deviation). A dropout layer randomly silences 30% of the neurons and passes them to the next 8×4 node linear layer, where they are activated again by the ReLU function, normalized again by a standardization layer, and then silenced by another 30% of the neurons by a dropout layer. Finally, a 4×1 node linear layer is passed through the sigmoid function for activation and transformation, outputting a set of values ​​between 0 and 1. The training process is recorded in the training curve (see...). Figure 2The x-axis of the curve represents one training epoch, and the y-axis represents the difference between the predicted and the true values, called the "loss". The loss should theoretically decrease as training progresses. Because an early stopping function is enabled to automatically monitor and terminate training, this function continuously monitors the training process. If the model's performance remains ineffective after 30 epochs, training will automatically terminate to prevent overfitting.

[0064] Overfitting is avoided by using batch training (64 data points per batch) and batch standardization (automatically calculating the mean and standard deviation of the batch, subtracting the mean, and dividing by the standard deviation). An early stopping function is used, which checks approximately every 30 epochs and automatically terminates training when the model performance improvement is not significant. Dropout is also used to avoid overfitting; it randomly silences a portion of neurons during training to prevent some neurons from having excessively high prediction weights.

[0065] The model building process is as follows: The user inputs a selection or enters all 853 related genes of ARL6IP5 (see [link]). Figure 3 ).

[0066] The above data is used as the input layer to this prediction system. This prediction system has two hidden layers with 8 and 4 nodes respectively, and also includes batch normalization and a 30% dropout to help improve model performance.

[0067] Returning prediction results: After analysis, this system calculates the probability that the patient with this tumor is sensitive to or benefits from immunotherapy (see...). Figure 5 ).

[0068] The area under the receiver operating characteristic (AUC) is used to evaluate model performance. A model is considered to have predictive value only when the AUC is greater than 0.5, and the closer the AUC is to 1, the stronger the predictive power.

[0069] After 30 rounds of training, the early stopping function automatically terminates the training. On the Discovery group training set, the AUC is 0.9957; on the Validation group and Test group validation sets, the AUCs are 0.7657 and 0.8125, respectively, higher than previous PD-L1 algorithms and some other algorithms reported internationally. Therefore, this model can more accurately predict the immunotherapy efficacy in patients with tumor immunotherapy, providing clinicians with valuable references (see...). Figure 4 ).

[0070] Compared with existing prediction models TMB and PD-L1, PD-L1 has limited predictive performance for immunotherapy, with an area under the receiver operating characteristic curve (AUC) of only 0.569 and TMB has an AUC of 0.728. The performance of the model of this invention is shown in Table 5, with an AUC of 0.7657-0.9957, which is higher than that of existing models.

[0071] Table 5 Performance of the immunotherapy sensitivity prediction model

[0072] Ultimately, testing revealed that by randomly testing the immunotherapy sensitivity of a cancer patient, the model predicted a 15% probability of that patient being sensitive to immunotherapy (see...). Figure 5 This indicates that the patient with the tumor is not suitable for immunotherapy.

Claims

1. A method for preparing a predictive model of immunotherapy sensitivity in bladder cancer patients, the method comprising the following steps: S01 Data Acquisition: Acquire the TPM expression values ​​of the ARL6IP5 gene and its associated genes in the next-generation sequencing transcriptome of pathological tissues from immunotherapy patients, as well as the actual diagnostic values ​​of the patients' immunotherapy efficacy. S02 Model Construction: The two values ​​mentioned above are used as prevariable data and input into the neural network model through the input layer. Data fitting is performed in the hidden layers of the neural network model. The neural network model has a total of 10 layers. The initial input is an 853×8 node linear layer. After data fitting, a dropout layer randomly silences 30% of neurons and passes them to the next 8×4 node linear layer. Data fitting and dropout layer silence 30% of neurons again. Finally, it passes through a 4×1 node linear layer and outputs a set of values ​​between 0 and 1 through the output layer. After training and optimization, a prognostic model for immunotherapy of bladder cancer patients is formed. S03 Evaluate model performance: The area under the receiver operating characteristic (ROC) curve is used to evaluate model performance; when the ROC curve area is greater than 0.5, the model has predictive value, and the closer it is to 1, the stronger the predictive power. The ARL6IP5 gene and its associated genes comprise a total of 853 genes, as shown below: 。 2. The method for preparing a predictive model for immunotherapy sensitivity in bladder cancer patients as described in claim 1, characterized in that, The data fitting is performed using batch training and batch standardized data, with each batch training involving 30-100 data points.

3. The method for preparing a predictive model for immunotherapy sensitivity in bladder cancer patients as described in claim 1, characterized in that, The data fitting validation uses an early stopping function, which checks the model every 10-50 rounds and automatically terminates training when the model performance improvement is not significant.

4. The method for preparing a predictive model for immunotherapy sensitivity in bladder cancer patients as described in claim 1, characterized in that, The output is a set of values ​​between 0 and 1, which are the predicted probability values ​​of the patient's sensitivity to immunotherapy.

5. The application of the method as described in any one of claims 1-4 in the strategy of identifying bladder cancer patients as sensitive to or beneficiaries of immunotherapy.

6. The application of the strategy for determining whether a bladder cancer patient is sensitive to or benefits from immunotherapy as described in claim 5, characterized in that, The application is as follows: when the predicted probability value of a bladder cancer patient is detected to be <50%, it indicates that the bladder cancer patient is not sensitive to immunotherapy and is not suitable for immunotherapy; when the predicted probability value of a bladder cancer patient is detected to be ≥50%, it indicates that the patient is sensitive to immunotherapy and needs to be treated with immunotherapy.

Images