Method and system for identifying drug resistance of anti-tumor inhibitor

Abstract

The invention belongs to the technical field of intelligent medical treatment, and particularly relates to a method and a system for identifying drug resistance of an anti-tumor inhibitor. Acquiring gene mutation data and copy number variation data of the lung adenocarcinoma patient; carrying out drug resistance label definition on the data; preprocessing the data, and judging the heterogeneity of the lung adenocarcinoma patient; on the basis of the judged heterogeneity, sBiosNet is constructed; performing parameter setting on the sBiosNet; and carrying out anti-tumor drug inhibitor drug resistance identification on the preprocessed data by using the network model with set parameters. The method is used for solving the problem that in an existing prediction model, a black box problem lacks interpretability. The number of patients with drug-resistant labels is small, so that the over-fitting problem is very easy to occur; the problem that the risk of overfitting is increased due to the high-dimensional characteristic of omics data, and the prediction performance is affected due to the fact that heterogeneity possibly exists among different queues is solved.

Description

Technical Field

[0001] The present invention belongs to the field of smart medical technology, and specifically relates to a method and system for identifying drug resistance of anti-tumor inhibitors. Background Art

[0002] As one of the most commonly used immunotherapy methods, PD-1 inhibitors can block the binding of PD-L1 on the surface of tumor cells to the receptor protein PD-1 on the surface of T cells, prompting T cells to kill tumor cells and achieve the purpose of anti-tumor treatment. The U.S. Food and Drug Administration has approved the clinical use of PD-1 inhibitors to treat a variety of cancers. However, due to the heterogeneity of tumors and the complexity of the microenvironment, the objective response rate of tumor patients to PD-1 inhibitors is usually low. For example, when PD-1 inhibitors are used to treat melanoma, only 30%-40% of patients respond, and the objective response rate of patients with non-small cell lung cancer to PD-1 inhibitor treatment is only 15%-20%. Currently, biomarkers used to identify PD-1 inhibitor resistance, such as tumor mutation load, PD-L1 expression level, CD8+T cell level, etc., lack standard cutoff values, high cost of use, and limited predictive performance. Some studies have used omics data to establish deep neural network prediction models to identify PD-1 inhibitor resistance, but most of these studies construct prediction models based on supervised learning, which have limited generalization capabilities and are difficult to make biological explanations for the prediction results.

[0003] Although PD-1 inhibitors have made significant progress in the clinical treatment of tumors, only a small number of patients can truly benefit from PD-1 inhibitor treatment due to tumor heterogeneity and the complexity of the microenvironment. Therefore, accurately identifying the resistance of tumor patients to PD-1 inhibitors before clinical treatment and selecting sensitive patients for treatment can not only achieve the best treatment effect, but also avoid secondary damage to resistant patients due to toxic side effects. Currently known biomarkers related to PD-1 inhibitor resistance, such as tumor mutation load and PD-L1 expression level, have no standard cutoff values ​​in practical applications and have limited predictive ability. Summary of the invention

[0004] The present invention provides a method and system for identifying drug resistance to anti-tumor inhibitors, which are used to solve the problems in existing prediction models, such as the "black box" problem that makes it lack of interpretability; the small number of patients with drug resistance labels that makes it very easy to overfit; the "high dimensionality" characteristics of omics data that increase the risk of overfitting; and the possible heterogeneity between different cohorts that affects its prediction performance.

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

[0006] A method for identifying drug resistance of an anti-tumor inhibitor, the method comprising the following steps:

[0007] Step 1: Obtain gene mutation data and copy number variation data of patients with lung adenocarcinoma;

[0008] Step 2: Define the drug resistance label for the data in step 1;

[0009] Step 3: Preprocess the data from step 1 and determine the heterogeneity of lung adenocarcinoma patients;

[0010] Step 4: Based on the heterogeneity determined in step 3, construct sBiosNet;

[0011] Step 5: Set parameters for sBiosNet in step 4;

[0012] Step 6: Use the network model with parameters set in step 5 to identify the resistance of anti-tumor drug inhibitors to the data pre-processed in step 3.

[0013] Furthermore, the step 4 of constructing a semi-supervised biological sparse neural network model specifically includes the following steps:

[0014] Step 4.1: Use biological information from the Reactome database to sparse the structure of sBiosNet;

[0015] Step 4.2: sBiosNet generates pseudo labels with high confidence;

[0016] Step 4.3: Generate mixed samples using MixMatch algorithm;

[0017] Step 4.4: Construct a loss function based on the prediction results of the forward propagation;

[0018] Step 4.5: Update the model parameters using the back-propagation algorithm based on gradient descent.

[0019] Furthermore, the step 4.1 is specifically that the constructed sBiosNet includes a new input layer, a new gene hidden layer, a new sub-pathway hidden layer, a new parent pathway hidden layer and a new output layer;

[0020] The new input layer includes an input layer, a Dropout layer in front of the input layer, a Dropout layer in the back of the input layer, and a ReLU activation function;

[0021] The new first hidden layer includes the first hidden layer, a Dropout layer in front of the first hidden layer, a Dropout layer in the back of the first hidden layer, and a ReLU activation function;

[0022] The new second hidden layer includes a second hidden layer, a Dropout layer in front of the second hidden layer, a Dropout layer in the back of the second hidden layer, and a ReLU activation function;

[0023] The new third hidden layer includes a third hidden layer, a Dropout layer in front of the third hidden layer, a Dropout layer in the back of the third hidden layer, and a ReLU activation function;

[0024] The new output layer includes an output layer combined with a Softmax function;

[0025] The input layer is used to receive multi-omics data of tumor patients, such as gene mutation data and copy number variation data. Each neuron in this layer corresponds to the variation of a gene.

[0026] In the first hidden layer, each neuron corresponds to a specific gene, which is used to integrate the multi-omics data of the patient and extract the characteristics of the patient at the gene level;

[0027] In the second hidden layer, each neuron corresponds to a specific sub-pathway, and the characteristics of the tumor patient at the gene level are integrated into the sub-pathway level to obtain the characteristics of the tumor patient at the sub-pathway level;

[0028] In the third hidden layer, neurons represent biological parent pathways, and this layer further integrates the features of sub-pathways to obtain activation features of tumor patients at the parent pathway level;

[0029] The output layer uses the characteristics of tumor patients at the parent pathway level to identify their resistance to PD-1 inhibitors.

[0030] Furthermore, the step 4.2 is specifically to use L={(x l,1 ,y l,1 ),…,(x l,n ,y l,n )} represents a set of n tumor patients with drug resistance labels, where x l,n represents the multi-omics data of a labeled tumor patient n, yl,n represents the drug resistance label of the tumor patient; using U = {(x u,1 ),…,(x u,m )} represents m tumor patients without drug resistance labels, where x u,m represents the multi-omics data of unlabeled patient m; using f θ represents sBiosNet, θ represents its parameters to be learned;

[0031] Using the prediction model f θ Predict the drug resistance of unlabeled tumor patient u and get the predicted value Will The maximum probability value in is recorded as the confidence score of sample u, that is:

[0032]

[0033] right Perform the Sharpen operation to obtain the pseudo label y′ of the unlabeled sample u u , expressed using the formula:

[0034]

[0035] Among them, the parameter T is used to control the pseudo label y′ u Entropy of

[0036] Using a predefined confidence threshold p0, of is called a high-confidence pseudo-label. of is called a low-confidence pseudo-label; therefore, the set of c unlabeled samples with high-confidence pseudo-labels is U1={(x u,1 ,y′ u,1 ),…,(x u,c ,y′ u,c )}, the set of mc unlabeled samples with low confidence pseudo labels is U2 = {(x u,c+1 ,y′ u,c+1 ),…,(x u,m ,y′ u,m )}.

[0037] Furthermore, the step 4.3 is specifically as follows: for the labeled sample set L = {(x l,1 ,y l,1 ),…,(x l,n ,y l,n )} and the unlabeled sample set U1={(x u,1 ,y′ u,1 ),…,(x u,c ,y′ u,c )} Perform random Mixup operation to generate mixed samples For two samples (x i ,y i ) and (x j ,y j ), the Mixup operation is:

[0038]

[0039] Among them, the parameter λ is randomly taken from Beta(mix α ,mix α) distribution, set λ≥0.5, when y i is the true label of the labeled sample, As the true label; when y i When y% is a pseudo label of an unlabeled sample, y% is also regarded as the pseudo label of the unlabeled sample;

[0040] Use the Mixup operation to traverse all samples of the above two sample sets L and U1, and the mixed labeled sample set and a mixed set of unlabeled samples The mixed unlabeled sample set and the unlabeled sample set U2 with low confidence pseudo labels, and the combined sample set is re-recorded as the unlabeled sample set U = {(x u,1 ,y′ u,1 ),…,(x u,m ,y′ u,m )}.

[0041] Furthermore, the step 4.4 is specifically to calculate the loss function using the drug resistance probability of the tumor patient obtained by the forward propagation process and the drug resistance label of the patient, and the loss function is composed of two parts: supervised loss and unsupervised loss; wherein the supervised loss is the predicted value using the mixed labeled samples Mixing tags with it The cross entropy loss is calculated, and the unsupervised loss is the predicted value f using the unlabeled sample θ (x u ) and its pseudo label y′ u The mean square error is calculated, that is:

[0042]

[0043] Among them, CELoss and MSELoss represent the cross entropy loss function and the mean square error loss function respectively, t is the number of iterations of model training, and w(t) is the weight function of unsupervised loss;

[0044] Construct the following w(t) function:

[0045]

[0046] Among them, w max is the maximum weight of the unsupervised loss; in the initial t0 training process, w(t) = 0, at which time only the supervised loss is used to update the parameter θ of sBiosNet, so that it can quickly learn the information of tumor patients' resistance to PD-1 inhibitors; in the next t1 training process, w(t) increases according to the trend of the Gaussian curve until it stabilizes at w maxThe weight of the unsupervised loss is gradually increased to improve the prediction accuracy and stability of sBiosNet for unlabeled samples.

[0047] Furthermore, the step 4.5 is specifically that after the loss function is calculated through the forward propagation process, the parameter θ of sBiosNet is updated using the back propagation algorithm based on gradient descent. The updating process of the parameter θ is:

[0048]

[0049] Among them, η is the learning rate, t is the number of iterations; in order to speed up the convergence of sBiosNet during training, a fixed step size of n is used. η The learning rate decay strategy dynamically adjusts the learning rate η, that is, every n η After training for several iterations, the learning rate η becomes the original η γ times.

[0050] A system for identifying drug resistance to anti-tumor inhibitors, the system using the above-mentioned method for identifying drug resistance to anti-tumor inhibitors, the system comprising:

[0051] Data acquisition module, which obtains gene mutation data and copy number variation data of patients with lung adenocarcinoma;

[0052] The drug resistance label definition module defines drug resistance labels based on gene mutation data and copy number variation data of patients with lung adenocarcinoma;

[0053] The data preprocessing and judgment module preprocesses the gene mutation data and copy number variation data of lung adenocarcinoma patients and judges the heterogeneity of lung adenocarcinoma patients;

[0054] The network model building module builds a semi-supervised biological sparse neural network model sBiosNet based on the heterogeneity of lung adenocarcinoma patients;

[0055] Parameter setting module, which sets the parameters of the semi-supervised biological sparse neural network model sBiosNet;

[0056] Step 6: Use the network model with parameters set in step 5 to identify the resistance of anti-tumor drug inhibitors to the data processed in step 3.

[0057] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the above method is implemented.

[0058] The beneficial effects of the present invention are:

[0059] The sBiosNet of the present invention can not only accurately identify the resistance of lung adenocarcinoma patients to PD-1 inhibitors, but also identify the survival of patients under PD-1 inhibitor treatment. At this time, sBiosNet is a prediction model for the survival of lung adenocarcinoma patients, rather than a prognostic model.

[0060] The present invention is used to identify the drug resistance of tumor patients to PD-1 inhibitors. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic flow chart of the method of the present invention.

[0062] Figure 2 It is a schematic diagram of the algorithm flow of the present invention, wherein, Figure 2 A is a schematic diagram of the sBiosNet network structure sparseness. Figure 2 B is a schematic diagram of the pseudo-label generation process. Figure 2 C is a schematic diagram of the MixMatch process. Figure 2 D is a schematic diagram of the forward propagation process and the back propagation process.

[0063] Figure 3 It is a flow chart of lung adenocarcinoma sample screening of the present invention.

[0064] Figure 4 It is a schematic diagram of the intersection of genes in the gene mutation spectrum and copy number variation spectrum of the present invention.

[0065] Figure 5 is a distribution diagram between the Rizvi cohort and Miao cohort samples of the present invention, wherein, Figure 5 A is the distribution diagram of PCA between Rizvi cohort and Miao cohort samples. Figure 5 B is the t-SNE distribution diagram between the Rizvi cohort and Miao cohort samples. Figure 5 C is the UMAP distribution diagram between Rizvi cohort and Miao cohort samples.

[0066] Figure 6 is a bar graph of the number of neurons and the amount of parameters to be learned in the sBiosNet of the present invention, wherein: Figure 6 A is a bar chart of the number of neurons in each neural layer in the Embedding module of sBiosNet. Figure 6 B is a bar chart of the parameter quantities of sBiosNet and a fully connected neural network with the same structure as sBiosNet.

[0067] Figure 7 is a graph showing the changing trend of cross entropy loss and AUROC during the pre-training process of the present invention, wherein: Figure 7 A is the trend graph of the change of cross entropy loss with the number of training times. Figure 7 B is the trend chart of AUROC changing with the number of training times.

[0068] Figure 8 Schematic diagram of the recognition ability of each prediction model of the present invention for the drug resistance of lung adenocarcinoma patients in the Miao cohort, wherein: Figure 8 A is the ROC curve of each prediction model for identifying the resistance of Miao cohort patients to PD-1 inhibitors. Figure 8 B is the PR curve of each prediction model identifying the resistance of Miao cohort patients to PD-1 inhibitors. Figure 8 C is the accuracy and F1 score of each prediction model in identifying drug resistance in Miao cohort patients, Figure 8 D is the confusion matrix diagram of sBiosNet’s prediction results for tumor patients in the Miao cohort.

[0069] Fig. 9 is the stability of each prediction model of the present invention in identifying the drug resistance of Miao cohort samples, wherein, Fig. 9 A is the box plot of AUROC of each prediction model in the ten-fold cross validation process. Fig. 9 B is the box plot of AUPR of each prediction model in the ten-fold cross validation process. Fig. 9 C is a box plot of the accuracy of each prediction model in the ten-fold cross validation process. Fig. 9 D is a box plot of the F1 score of each prediction model during the ten-fold cross validation process.

[0070] Fig.10 is the effect of identifying the drug resistance of Miao cohort samples in each ablation experiment of the present invention, wherein, Fig.10 A is the ROC curve and AUROC of each ablation experiment. Fig.10 B is the PR curve and AUPR of each ablation experiment. Fig.10 C is a bar chart of the accuracy and F1 score in each ablation experiment.

[0071] Fig.11 is the prediction stability of the model in each ablation experiment of the present invention, where Fig.11 A is the box plot of AUROC of each ablation experiment in the ten-fold cross validation process. Fig.11 B is the box plot of AUPR of each ablation experiment in the ten-fold cross validation process. Fig.11 C is a box plot of the accuracy of each ablation experiment in the ten-fold cross validation process. Fig.11 D is the box plot of the F1 score of each ablation experiment during the ten-fold cross validation process.

[0072] Fig.12 It is a visualized schematic diagram of the sBiosNet forward propagation process of the present invention.

[0073] Fig.13 It is a schematic diagram of the survival curve of patients in high and low risk groups identified by sBiosNet of the present invention, wherein: Fig.13 A is the OS of high- and low-risk patients in the Miao cohort. Fig.13 B is the PFS of high- and low-risk patients in the Miao cohort. Fig.13 C is the OS of high- and low-risk patients in the TCGA cohort. Fig.13 D is the PFS of high- and low-risk patients in the TCGA cohort. Fig.13 E is the DFI of high- and low-risk patients in the TCGA cohort. Fig.13 F is the DSS of high- and low-risk patients in the TCGA cohort. DETAILED DESCRIPTION

[0074] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0075] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0076] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include plural forms unless the context clearly indicates otherwise.

[0077] The following is attached to this application specification Figure 1-13 , the technical solutions in the embodiments of the present application are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0078] In the following description, many specific details are set forth to facilitate a full understanding of the present application, but the present application may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0079] Implementation Method 1

[0080] This embodiment provides a method for identifying drug resistance of an anti-tumor inhibitor, such as Figure 1 As shown, the method comprises the following steps:

[0081] Step 1: Obtain gene mutation data and copy number variation data of patients with lung adenocarcinoma;

[0082] Step 2: Define the drug resistance label for the data in step 1;

[0083] Step 3: Preprocess the data from step 1 and determine the heterogeneity of lung adenocarcinoma patients;

[0084] Step 4: Based on the heterogeneity determined in step 3, a semi-supervised biological sparse neural network model sBiosNet is constructed;

[0085] Step 5: Set parameters for sBiosNet in step 4;

[0086] Step 6: Use the network model with parameters set in step 5 to identify the resistance of anti-tumor drug inhibitors to the data pre-processed in step 3.

[0087] Furthermore, if Figure 2 As shown, the step 4 of constructing a semi-supervised biological sparse neural network model specifically includes the following steps:

[0088] Step 4.1: Use biological information from the Reactome database to sparse the structure of sBiosNet;

[0089] Step 4.2: The semi-supervised biological sparse neural network model sBiosNet generates pseudo labels with high confidence;

[0090] Step 4.3: Generate mixed samples using MixMatch algorithm;

[0091] Step 4.4: Construct a loss function based on the prediction results of the forward propagation;

[0092] Step 4.5: Update the model parameters using the back-propagation algorithm based on gradient descent.

[0093] Furthermore, the step 4.1 is specifically that the constructed sBiosNet includes a new input layer, a new gene hidden layer, a new sub-pathway hidden layer, a new parent pathway hidden layer and a new output layer;

[0094] The new input layer includes an input layer, a Dropout layer in front of the input layer, a Dropout layer in the back of the input layer, and a ReLU activation function;

[0095] The new first hidden layer includes the first hidden layer, a Dropout layer in front of the first hidden layer, a Dropout layer in the back of the first hidden layer, and a ReLU activation function;

[0096] The new second hidden layer includes a second hidden layer, a Dropout layer in front of the second hidden layer, a Dropout layer in the back of the second hidden layer, and a ReLU activation function;

[0097] The new third hidden layer includes a third hidden layer, a Dropout layer in front of the third hidden layer, a Dropout layer in the back of the third hidden layer, and a ReLU activation function;

[0098] The new first hidden layer corresponds to a new gene hidden layer, and the new gene hidden layer includes a gene hidden layer, a Dropout layer in front of the gene hidden layer, a Dropout layer in the back of the gene hidden layer, and a ReLU activation function;

[0099] The new second hidden layer corresponds to a new sub-pathway hidden layer, and the new sub-pathway hidden layer includes a sub-pathway hidden layer, a Dropout layer in front of the sub-pathway hidden layer, and a Dropout layer and a ReLU activation function in the back of the sub-pathway hidden layer;

[0100] The new third hidden layer corresponds to the new parent pathway hidden layer, and the new parent pathway hidden layer includes the parent pathway hidden layer, a Dropout layer in front of the parent pathway hidden layer, and a Dropout layer and a ReLU activation function in the back of the parent pathway hidden layer;

[0101] The new output layer includes an output layer combined with a Softmax function; outputting the probability of drug resistance of tumor patients to PD-1 inhibitors;

[0102] The input layer is used to receive multi-omics data of tumor patients, such as gene mutation data and copy number variation data. Each neuron in this layer corresponds to the variation of a gene, such as EGFR mutation, FGFR copy number amplification, etc.;

[0103] In the first hidden layer, each neuron corresponds to a specific gene, which is used to integrate the multi-omics data of the patient and extract the characteristics of the patient at the gene level; for example, the neuron corresponding to the EGFR gene performs weighted integration on the characteristics of the neurons corresponding to the EGFR mutation and copy number amplification in the input layer, that is, the characteristics of the tumor patient on the EGFR gene are obtained;

[0104] In the second hidden layer, each neuron corresponds to a specific sub-pathway, and the characteristics of the tumor patient at the gene level are integrated into the sub-pathway level to obtain the characteristics of the tumor patient at the sub-pathway level;

[0105] In the third hidden layer, neurons represent parent pathways with more complex biological functions. This layer further integrates the characteristics of sub-pathways to obtain activation characteristics of tumor patients at the parent pathway level.

[0106] The output layer uses the characteristics of tumor patients at the parent pathway level to identify their resistance to PD-1 inhibitors.

[0107] Just as neurons correspond to specific genes or pathways, each edge of sBiosNet also has a specific biological meaning. For example, it is known that gene G1 only exists in subpathway P. c,1 and P c,m-1 Therefore, in the network structure of sBiosNet, the only connection edge starting from G1 is G1→P c,1 and G1→P c,m-1 , that is, the characteristics of gene G1 are only transmitted to subpathway P c,1 and P c,m-1 , and will not be transmitted to other sub-paths; similarly, the only connecting edge starting from the sub-path is P c,1 →P p,1 and P c,1 →P p,p , represents the subpath P c,1 The features are only enriched in the parent pathway P p,1 and P p,p superior.

[0108] Furthermore, the step 4.2 is specifically to use L={(x l,1 ,y l,1 ),…,(x l,n ,y l,n )} represents a set of n tumor patients with drug resistance labels, where x l,n represents the multi-omics data of a labeled tumor patient n, yl,n represents the drug resistance label of the tumor patient; using U = {(x u,1 ),…,(x u,m )} represents m tumor patients without drug resistance labels, where x u,m represents the multi-omics data of unlabeled patient m; using f θ represents sBiosNet, θ represents its parameters to be learned;

[0109] Using the prediction model f θ Predict the drug resistance of unlabeled tumor patient u and get the predicted value Will The maximum probability value in is recorded as the confidence score of sample u, that is:

[0110]

[0111] In order to ensure that the prediction model f θ The prediction stability of unlabeled samples is Perform the Sharpen operation to obtain the pseudo label y′ of the unlabeled sample u u , see Figure 2 B, expressed using the formula:

[0112]

[0113] Among them, the parameter T is used to control the pseudo label y′ u entropy; as the value of T decreases, y′ u It will gradually tend to a one-hot distribution, and its entropy will continue to decrease;

[0114] Using a predefined confidence threshold p0, of is called a high-confidence pseudo-label. of is called a low-confidence pseudo-label; therefore, the set of c unlabeled samples with high-confidence pseudo-labels is U1={(x u,1 ,y′ u,1 ),…,(x u,c ,y′ u,c )}, the set of mc unlabeled samples with low confidence pseudo labels is U2 = {(x u,c+1 ,y′ u,c+1 ),…,(x u,m ,y′ u,m )}

[0115] Furthermore, the step 4.3 is specifically as follows: for the labeled sample set L = {(x l,1 ,y l,1 ),…,(x l,n ,y l,n )} and the unlabeled sample set U1={(x u,1 ,y′ u,1 ),…,(x u,c ,y′ u,c )} Perform random Mixup operation to generate mixed samples See Figure 2 B; For two samples (x i ,y i ) and (x j ,y j), the Mixup operation is:

[0116]

[0117] Among them, the parameter λ is randomly taken from Beta(mix α ,mix α ) distribution, which helps to improve the model f θ The prediction stability of The "authenticity" is more inclined to y i , that is, when y i is the true label of the labeled sample, As the true label; when y i When y% is a pseudo label of an unlabeled sample, y% is also regarded as the pseudo label of the unlabeled sample;

[0118] Use the Mixup operation to traverse all samples of the above two sample sets L and U1 to obtain two new mixed sample sets: mixed labeled sample sets and a mixed set of unlabeled samples Then, the mixed unlabeled sample set and the unlabeled sample set U2 with low confidence pseudo labels, and the combined sample set is re-recorded as the unlabeled sample set U = {(x u,1 ,y′ u,1 ),…,(x u,m ,y′ u,m )}.

[0119] Furthermore, the step 4.4 is specifically to calculate the loss function using the drug resistance probability of the tumor patient obtained by the forward propagation process and the drug resistance label of the patient, such as Figure 2 As shown in C, the loss function consists of two parts: supervised loss and unsupervised loss; among them, supervised loss is the predicted value using mixed labeled samples Mixing tags with The cross entropy loss is calculated, and the unsupervised loss is the predicted value f using the unlabeled sample θ (x u ) and its pseudo label y′ u The mean square error of the calculation is:

[0120]

[0121] Among them, CELoss and MSELoss represent the cross entropy loss function and the mean square error loss function respectively, t is the number of iterations of model training, and w(t) is the weight function of unsupervised loss;

[0122] In the initial stage of training, it is difficult for sBiosNet to accurately identify the resistance of tumor patients to PD-1 inhibitors. At this time, the pseudo-labels generated by sBiosNet for unlabeled samples are "unreliable", that is, in the initial stage of model training, it is necessary to give a very small weight to the unsupervised loss; as the number of training times increases, the accuracy of the pseudo-labels generated by sBiosNet continues to improve, and the proportion of unsupervised loss can be gradually increased to improve the prediction accuracy of the model for unlabeled samples; therefore, based on the above ideas, the present invention constructs the following w(t) function:

[0123]

[0124] Among them, w max is the maximum weight of the unsupervised loss; in the initial t0 training process, w(t) = 0, at which time only the supervised loss is used to update the parameter θ of sBiosNet, so that it can quickly learn the information of tumor patients' resistance to PD-1 inhibitors; in the next t1 training process, w(t) increases according to the trend of the Gaussian curve until it stabilizes at w max The weight of the unsupervised loss is gradually increased to improve the prediction accuracy and stability of sBiosNet for unlabeled samples.

[0125] Furthermore, the step 4.5 is specifically to update the parameter θ of sBiosNet using the back propagation algorithm based on gradient descent after calculating the loss function through the forward propagation process, as detailed in Figure 2 D. The update process of parameter θ is:

[0126]

[0127] Among them, η is the learning rate and t is the number of iterations. At the same time, in order to speed up the convergence of sBiosNet during training, a fixed step size of n is used. η The learning rate decay strategy dynamically adjusts the learning rate η, that is, every n η After training for several iterations, the learning rate η becomes the original η γ times.

[0128] Specifically,

[0129] 1 Study Cohort and Data Preprocessing

[0130] 1.1 Screening of research subjects

[0131] Specifically, the step 1 includes three cohort data in the present invention, namely, the Rizvi cohort and the Miao cohort composed of tumor patients treated with PD-1 inhibitors, and the TCGA lung adenocarcinoma cohort that did not receive PD-1 inhibitors. Among them, the Rizvi cohort includes a total of 240 tumor patients who received PD-1 inhibitors, excluding 54 tumor patients without lung adenocarcinoma and 1 lung adenocarcinoma patient without gene mutation data. Finally, 185 lung adenocarcinoma patients with gene mutation data, copy number variation data and related clinical characteristics were screened out, see Figure 3 It was randomly divided into training set, validation set and hyperparameter selection set in a ratio of 8:1:1. The training set was used to train and update the parameters of sBiosNet. The validation set was combined with the early stopping strategy to select the best model. The hyperparameter selection set was combined with the grid search strategy to determine the best hyperparameter combination.

[0132] Similarly, the Miao cohort included 249 tumor patients treated with PD-1 inhibitors, excluding 202 patients with non-lung adenocarcinoma tumors. Finally, 47 patients with lung adenocarcinoma with gene mutation data, copy number variation data, and clinical characteristics were screened out. For details, see Figure 3 The 47 patients with lung adenocarcinoma were used as a test set to evaluate the recognition ability of sBiosNet for PD-1 inhibitors. During the training process of sBiosNet, the test set samples were used as unlabeled samples to participate in the update process of model parameters.

[0133] At the same time, in the screening process of the above-mentioned tumor patients, a total of 257 solid tumor patients treated with PD-1 inhibitors in the Rizvi cohort and the Miao cohort were not selected. After excluding one lung adenocarcinoma patient who lacked gene mutation data, the present invention used the remaining 256 solid tumor patients as a pre-training sample set to pre-train sBiosNet.

[0134] 1.2 Defining drug resistance labels

[0135] Specifically, in the Miao and Rizvi cohorts, the 1.1 version of the response evaluation criteria in solid tumors (RECIST) was used to determine the efficacy of PD-1 inhibitors in the treatment of solid tumors, and the efficacy was divided into four categories: complete response (CR), partial response (PR), stable disease (SD) and progressive disease (PD). The two cohorts also provided the overall survival (OS) and progression-free survival (PFS) of tumor patients under PD-1 inhibitor treatment.

[0136] RECIST response results of CR, PR, and SD with PFS>6 months are regarded as durable clinical response (DCB), and PD and SD with PFS≤6 months are regarded as non-durable clinical response (NDCB). In the present invention, DCB is regarded as sensitive to PD-1 inhibitor treatment, and NDCB is regarded as resistant to PD-1 inhibitor treatment.

[0137] 1.3 Omics data preprocessing

[0138] Specifically, in step 3, the gene mutation and copy number variation data of lung adenocarcinoma patients in the Rizvi cohort were detected using MSK-IMPACT sequencing technology

[90] , and the gene mutation and copy number variation data of tumor patients in the Miao cohort were obtained from whole-exome sequencing (WES).

[91] For gene mutation data, non-synonymous mutation genes of the two cohort samples were screened, and the mutant gene was assigned a value of 1 and the wild-type gene was assigned a value of 0. In order to avoid the interference of genes with low mutation rates on the performance of the prediction model, genes with mutation rates not exceeding 1% were excluded.

[0139] For copy number variation data, the R software package CNTools (version 1.42.0) was used to annotate the copy number variation data. According to the gene copy ratio (copyratio, Cr), log2(Cr)>0.5 is called copy number variation amplification, and log2(Cr)<-0.5 is called copy number variation deletion. In the copy number variation amplification spectrum, copy number variation amplification is marked as 1, and non-amplification is marked as 0; in the copy number variation deletion spectrum, copy number variation deletion is marked as 1, and non-deletion is marked as 0. Genes with a variation rate of no more than 5% in the copy number variation amplification spectrum and the copy number variation deletion spectrum were eliminated.

[0140] The above gene mutation spectrum, copy number variation amplification spectrum and copy number variation deletion spectrum were screened using genes in the Reactome database. A total of 140 mutant genes, 602 copy number variation amplification genes and 13 copy number variation deletion genes were included in the present invention, see Figure 4 .

[0141] 1.4 Heterogeneity among Tumor Patients

[0142] Step 3 also includes the present invention using principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP)

[94] Three dimensionality reduction methods were used to extract the characteristics of lung adenocarcinoma patients in the Rizvi and Miao cohorts, and the distribution of tumor patients in these two cohorts was visualized. The results showed that there was no significant heterogeneity between lung adenocarcinoma patients in the Rizvi and Miao cohorts, see Figure 5 .

[0143] 2 Prediction model and its parameter settings

[0144] 2.1sBiosNet structural characteristics and parameter settings

[0145] Furthermore, the step 5 sets the parameters of the model specifically by using the gene-subpathway and subpathway-parent pathway relationships in the Reactome database to sparse the Embedding module of sBiosNet. This module consists of an input layer and three hidden layers, wherein the input layer is used to receive multi-omics data of lung adenocarcinoma patients, and the hidden layers are used to extract the characteristics of patients at the gene level, subpathway level, and parent pathway level. The number of neurons in the input layer and the three hidden layers are 755, 744, 1630, and 615, respectively, see Figure 6 A. To better extract and fit the features of lung adenocarcinoma patients, the input layer and hidden layer are connected to a Dropout layer, a BatchNorm layer, and a ReLU activation function. The output layer of sBiosNet is a fully connected neural layer connected to a Softmax function, containing two output neurons. The output layer receives the features of lung adenocarcinoma patients at the parent pathway level and outputs the probability of resistance to PD-1 inhibitors.

[0146] After using the gene-subpathway and subpathway-parent pathway relationships in the Reactome database to sparsely structure sBiosNet, the number of parameters to be learned is 14635. For a fully connected neural network with the same number of neural layers and neurons as sBiosNet, the number of parameters to be learned is 2781111, which is more than 190 times different (see Figure 6 B), showing the degree to which the Reactome biological information contributes to the sparsification of the neural network.

[0147] In order to better train sBiosNet, the batch size is set to 64, the total number of iterations is 400, the number of early stops is 40, the Adam optimizer is used to update the model parameters, and the initial learning rate η0 is 0.01. The learning rate is dynamically adjusted using a learning rate decay strategy with a fixed step size of 30, and the decay weight is η γ . Set the parameter T of the Sharpen operation to 0.95 and the parameter t0 of the weight function w(t) of the unsupervised loss to 10.

[0148] Through grid search, other hyperparameters of sBiosNet are selected, mainly including the learning rate decay weight η γ ∈{0.75,0.9}, the confidence threshold of the pseudo label p0∈{0.6,0.7}, the parameter Mix of the Beta distribution in the MixMatch process α ∈{0.15,0.35}, the probability p of the Dropout layer dropout ∈{0.1,0.2}, weight decay rate w in Adam optimizer decay ∈{1×e -3 ,1×e -4}, the parameters of the unsupervised loss weight function w(t) are t1∈{20,30} and w max ∈{2.5,3.5,4.5}.

[0149] 2.2 Comparison prediction model and its parameter settings

[0150] The present invention sets the following prediction models as controls: a Logistic regression model (L2-Logit) based on L2 regularization, a random forest (RF), a support vector machine based on radial basis kernel function and linear kernel function (Rbf-SVM and Linear-SVM), a tumor mutation load combined with a Logistic regression model, and a neural network with the same structure as sBiosNet, but trained using supervised learning. Model BiosNet.

[0151] Use grid search to select the best hyperparameters for these models, specifically the regularization parameter C for L2-Logit logit ∈{1.0,2.0,3.0,4.0}, the number of decision trees in RF N tree ∈{100,500,1000,2000}, the regularization parameter C of Rbf-SVM Rbf-SVM ∈{1.0,2.0,3.0,4.0} and the regularization parameter C of Linear-SVM Linear-SVM ∈{1.0,2.0,3.0,4.0}. The hyperparameter settings and search range of BiosNet are the same as those of sBiosNet.

[0152] 2.3 Ablation experiment and its parameter setting

[0153] The present invention further sets up the following ablation experiment to explore the effects of Reactome pathway information, neural network structure sparsification, multi-omics fusion and transfer learning on the prediction performance of sBiosNet. The fully connected neural network with the same neural layer and number of neurons as sBiosNet is denoted as sDNet1. The fully connected neural network with the same neural layer and parameter amount to be learned as sBiosNet is denoted as sDNet2, and the number of neurons in its hidden layer is 19, 10 and 6 respectively. The sBiosNet that only uses copy number variation data and gene mutation data is denoted as OnlyCNV and OnlyMut respectively. The sBiosNet that does not use transfer learning is denoted as No-transfer. In the ablation experiment, the hyperparameter settings of all prediction models are the same as those of sBiosNet.

[0154] 2.4 Prediction performance evaluation indicators

[0155] The present invention uses the receiver operating curve (ROC) and its area under the ROC curve (AUROC), the precision-recall (PR) curve and its area under the PR curve (AUPR), accuracy and F1 score to evaluate the ability of each prediction model to identify lung adenocarcinoma patients' resistance to PD-1 inhibitors. The closer the values ​​of the above indicators are to 1, the better the prediction performance of the model.

[0156] 3 Model Pre-training

[0157] sBiosNet, sDNet1, and sDNet2 were pre-trained using the pre-training sample set. In pre-training, the batch size was set to 8 and the training was iterated 30 times. The cross entropy loss between the model prediction value and the drug resistance label was calculated, and the parameters were updated using the Adam optimizer. The learning rate was set to 0.01 and the weight decay rate was set to 1×e -4 .

[0158] The present invention observes and records the cross entropy loss and AUROC of sBiosNet, sDNet1 and sDNet2 on pre-training samples during the pre-training process. Figure 7 A and 9B show the changes of cross entropy loss and AUROC of the three prediction models with the number of training times. The results show that with the increase of training times, the accuracy of sBiosNet, sDNet1 and sDNet2 in identifying the resistance of tumor patients to PD-1 inhibitors in the pre-training set continues to improve, and the difference between the model prediction value and the patient's resistance label continues to decrease and stabilize.

[0159] 4 Results

[0160] The present invention uses the gene mutation data and copy number variation data of lung adenocarcinoma patients in the training set and validation set to train and select the best prediction model, uses the hyperparameter selection set to screen the best hyperparameter combination, and uses the test set samples to evaluate the ability of each prediction model, with the evaluation indicators being AUROC, AUPR, accuracy and F1 score.

[0161] 4.1 Prediction Accuracy of Each Prediction Model

[0162] The performance of each prediction model in the present invention in identifying the resistance of lung adenocarcinoma patients to PD-1 inhibitors in the test set is as follows: Fig.12 As shown. Among them, Figure 8A shows the ROC curves of various prediction models and their area under the curve AUROC. It can be found that the AUROC of sBiosNet is 0.888, which is much higher than all other prediction models and has the best prediction performance. At the same time, the AUROC of BiosNet trained using supervised learning on the test set samples is 0.726, which is significantly lower than the sBiosNet model of semi-supervised learning, indicating that semi-supervised learning does help improve the ability of sBiosNet to identify tumor patients in the test set who are resistant to PD-1 inhibitors. Among all traditional prediction models, the TMB combined with the Logistic regression model has the best predictive performance (AUROC of 0.762), which is better than RF, Rbf-SVM, L2-Logit and Linear-SVM, proving the possibility of TMB as a predictive biomarker for lung adenocarcinoma patients' resistance to PD-1 inhibitors, but in the present invention, the optimal cutoff value of TMB is also unknown. From Figure 8 B shows the PR curves of various prediction models and their area under the curve AUPR. Figure 8 From the accuracy and F1 score shown in Figure 3, sBiosNet has the largest score and performs best in the task of identifying resistance to PD-1 inhibitors in patients with lung adenocarcinoma.

[0163] The present invention further uses sBiosNet to construct a confusion matrix for the prediction results of tumor patients in the test set and visualizes it. Figure 8 D. It can be seen that sBiosNet has a sensitivity of 90.48% in identifying lung adenocarcinoma patients' resistance to PD-1 inhibitors, and a specificity of nearly 85% in identifying sensitivity to PD-1 inhibitors, which has good predictive accuracy and demonstrates the potential of this model for clinical practice.

[0164] From the above results, we can see that when the number of training set samples is small, traditional prediction models and neural network models based on supervised learning cannot effectively deal with the problem of overfitting. sBiosNet uses transfer learning, Dropout neural layers, semi-supervised learning, and adding weight decay to the optimizer to significantly reduce the neural network model's demand for the number of training set samples and improve the prediction ability of sBiosNet.

[0165] 4.2 Prediction stability of each prediction model

[0166] The present invention further uses 10-fold cross validation to evaluate the predictive stability of models such as sBiosNet, that is, the lung adenocarcinoma patients in the Rizvi cohort are randomly divided into 10 folds, each fold includes a training set (90%) and a validation set (10%), ensuring that there is no cross-overlapping of validation set samples in each fold. The lung adenocarcinoma patients in the Miao cohort are used as the test set to evaluate the model prediction performance.

[0167] Fig. 9 The prediction performance of each prediction model in the ten-fold cross validation is shown. Taking AUROC as an example, sBiosNet not only has the best prediction accuracy (mean value is 0.868), but also has the best prediction stability (standard deviation is 0.027), and performs best among all prediction models. RF and BiosNet also have good prediction accuracy and stability. Figure 8 The results shown are consistent. However, the prediction accuracy and stability of Rbf-SVM in the ten-fold cross validation are poor, indicating that Rbf-SVM is not suitable for prediction tasks with a small number of training set samples. Similarly, the AUPR, accuracy and F1 scores of each prediction model show the same trend as AUROC ( Fig. 9 B- Fig. 9 D). sBiosNet performed best in all evaluation indicators, further demonstrating that sBiosNet has an absolute advantage in the task of identifying lung adenocarcinoma patients with resistance to PD-1 inhibitors.

[0168] 4.3 Prediction Accuracy of Models in Each Ablation Experiment

[0169] Through ablation experiments, the present invention further explored the relationship between Reactome pathway information, neural network structure sparsification, multi-omics data fusion and transfer learning and sBiosNet prediction performance. The results are as follows Fig.10 shown.

[0170] In the task of identifying lung adenocarcinoma patients with resistance to PD-1 inhibitors, the predictive capabilities of sBiosNet and sDNet2 were significantly better than sDNet1 (the AUROCs of sBiosNet, sDNet1 and sDNet2 were 0.888, 0.616 and 0.816, respectively), indicating that whether it is to reduce the connection edges in the neural network to make it sparse, or directly reduce the number of hidden layer neurons in the neural network, it can significantly reduce the number of neural network parameters, avoid overfitting of the neural network, and improve its predictive ability. At the same time, the predictive performance of sBiosNet is also significantly better than sDNet2, indicating that combining the biological knowledge in the Reactome database with the feature extraction process of the neural network can further improve the predictive ability of sBiosNet.

[0171] At the same time, in the ablation experiments using only copy number variation data and only gene mutation data, the AUROC of OnlyCNV and OnlyMut were 0.646 and 0.757, respectively, which were significantly lower than sBiosNet. This result shows that sBiosNet has better prediction accuracy by integrating multi-omics data of lung adenocarcinoma patients and learning patients' resistance to PD-1 inhibitors from multiple dimensions. In addition, OnlyMut has better prediction performance than OnlyCNV, indicating that in this task, the gene mutation spectrum has more information than the gene copy number variation spectrum.

[0172] The present invention also sets up an ablation experiment without using transfer learning and directly performs random initialization of parameters on sBiosNet. This model is recorded as No-transfer to explore the impact of transfer learning on the prediction performance of sBiosNet. Fig.10 The results shown show that the AUROC is only 0.661 without transfer learning, indicating that transfer learning does help improve the ability of sBiosNet to identify lung adenocarcinoma patients with resistance to PD-1 inhibitors.

[0173] 4.4 Prediction stability of the model in each ablation experiment

[0174] The present invention further set up a ten-fold cross validation based on the ablation experiment, and the results are as follows: Fig.11 The results show that sBiosNet still has the best prediction accuracy and stability in AUROC, AUPR, accuracy and F1 score (sBisoNet has the largest mean and the smallest standard deviation). At the same time, the above research results also show that sDNet2 has better prediction accuracy and stability than sDNet1. OnlyMut has better prediction accuracy than OnlyCNV, but its prediction stability is worse than OnlyCNV.

[0175] 4.5 Visualizing sBiosNet

[0176] Based on the test set of tumor patients, the DeepLIFT algorithm is used to calculate the contribution score AdjC of the jth neuron in the i-th neural layer in sBiosNet. i,j According to AdjC i,j , sort all neurons in the i-th neural layer, judge the relative importance of each neuron, and give the visualization result of sBiosNet. Fig.12 According to the type of omics data, the contribution scores of the input layer neurons are integrated into gene mutation, gene copy number variation amplification and copy number variation deletion. Fig.12Only the top 10 neurons with the largest contribution scores are shown in Figure 1, and they are arranged in descending order of contribution scores. The connection edges between neurons represent the feature propagation process, and the thickness of the connection edges corresponds to the size of the trainable parameters in sBiosNet.

[0177] from Fig.12 It can be seen that in the input layer of sBiosNet, the contribution score of the neurons corresponding to gene mutation data is higher than that of the neurons corresponding to copy number variation data, indicating that gene mutation data makes a greater contribution in identifying lung adenocarcinoma patients' resistance to PD-1 inhibitors and provides more information. This result is consistent with the fact that OnlyMut's prediction performance is better than OnlyCNV in the aforementioned ablation experiment.

[0178] Among all 744 genes, TP53 mutation had the highest contribution score in identifying PD-1 inhibitor resistance in lung adenocarcinoma. TP53 mutation was found to be associated with PD-L1 expression and increased TMB, and can be used as a predictive marker to identify the response of patients with non-small cell lung cancer to PD-1 / PD-L1 inhibitor treatment. Fig.12 It can also be found that TP53 mutations mainly play a role through Pre-NOTCH expression and NOTCH signaling pathway, affecting the resistance of lung adenocarcinoma patients to PD-1 inhibitors. It was found that mutations in NOTCH family genes can improve the response of non-small cell lung cancer patients to immunotherapy and prolong the survival of patients.

[0179] The above results all confirm that combining the sBiosNet model with the DeepLIFT algorithm can effectively solve the "black box" problem of deep neural networks and provide a clear and reasonable biological explanation, which will help researchers and clinicians further understand the biological mechanism of lung adenocarcinoma patients' resistance to PD-1 inhibitors.

[0180] 4.6 Identifying the survival of patients with lung adenocarcinoma treated with PD-1 inhibitors

[0181] In clinical treatment, clinicians and patients are not only concerned about whether patients develop resistance to PD-1 inhibitors, but also pay more attention to the survival of tumor patients under PD-1 inhibitor treatment, such as overall survival (OS) and progression-free survival (PFS). The present invention further uses lung adenocarcinoma patients in the Miao cohort and TCGA cohort to explore the ability of sBiosNet to identify the survival of lung adenocarcinoma patients under PD-1 inhibitor treatment. Patients identified by sBiosNet as sensitive to PD-1 inhibitors are marked as low risk (Low-Risk), and identified resistant patients are marked as high risk (High-Risk). The Kaplan-Meier survival curves of patients in the high-risk group and the low-risk group were drawn, the OS and PFS of the two groups of patients were compared using the log-rank test, and the COX proportional hazard model was used to calculate the hazard ratio (hazard ratio, HR) of death or disease progression in patients in the high-risk group.

[0182] sBiosNet identified 21 low-risk patients and 26 high-risk patients in the Miao cohort. By comparing the OS of the two groups of patients, it was found that the death rate of lung adenocarcinoma patients in the low-risk group under PD-1 inhibitor treatment was 23.81%, while the death rate of patients in the high-risk group under the same treatment was 42.31%. The median OS of patients in the high-risk group was 11.3 months, which was significantly inferior to that of patients in the low-risk group (P=0.0231), with an HR of 3.25 and a 95% confidence interval of 1.12-9.48. For details, see Fig.13 A. The present invention also compared the PFS of the two groups of patients in the Miao cohort. The results showed that the median PFS of the high-risk group was 1.92 months, which was significantly shorter than the 14.70 months of the low-risk group (P<0.0001). Compared with the low-risk group, the HR of the high-risk group was 6.00, with a 95% confidence interval of 2.77-13.00. For details, see Fig.13 B.

[0183] The present invention further used the TCGA lung adenocarcinoma cohort patients as a control, used sBiosNet to identify the lung adenocarcinoma patients in the cohort into high-risk and low-risk groups, and compared the OS, PFS, disease-free survival (DFS) and disease-specific survival (DSS) of the two groups of patients. The results showed that there was no significant difference in OS (P = 0.9712), PFS (P = 0.3564), DFS (P = 0.7416) and DSS (P = 0.9032) between the high-risk and low-risk groups of the TCGA cohort, see Fig.13 C-13F.

[0184] Based on the above results, sBiosNet can not only accurately identify the resistance of lung adenocarcinoma patients to PD-1 inhibitors, but also identify the survival of patients under PD-1 inhibitor treatment. At this time, sBiosNet is a prediction model for the survival of lung adenocarcinoma patients, rather than a prognostic model.

[0185] Implementation Method 2

[0186] This embodiment provides a system for identifying drug resistance of an anti-tumor inhibitor, wherein the system uses a method for identifying drug resistance of an anti-tumor inhibitor as described in Embodiment 1, and the system comprises:

[0187] Data acquisition module, which obtains gene mutation data and copy number variation data of patients with lung adenocarcinoma;

[0188] The drug resistance label definition module defines drug resistance labels based on gene mutation data and copy number variation data of patients with lung adenocarcinoma;

[0189] The data preprocessing and judgment module preprocesses the gene mutation data and copy number variation data of lung adenocarcinoma patients and judges the heterogeneity of lung adenocarcinoma patients;

[0190] The network model building module builds a semi-supervised biological sparse neural network model sBiosNet based on the heterogeneity of lung adenocarcinoma patients;

[0191] Parameter setting module, which sets the parameters of the semi-supervised biological sparse neural network model sBiosNet;

[0192] Step 6: Use the network model with parameters set in step 5 to identify the resistance of anti-tumor drug inhibitors to the data processed in step 3.

[0193] Implementation Method 3

[0194] An embodiment of the present invention provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory is used to store software programs and modules, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory and the processor are connected via a bus. Specifically, the processor implements any step in the first embodiment above by running the computer program stored in the memory.

[0195] It should be understood that in the embodiments of the present invention, the processor referred to may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0196] The memory may include a read-only memory, a flash memory, and a random access memory, and provides instructions and data to the processor. A part or all of the memory may also include a nonvolatile random access memory.

[0197] It should be understood that if the above-mentioned integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned implementation method, and can also be completed by instructing the relevant hardware through a computer program. The above-mentioned computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method implementations when executed by the processor. Among them, the above-mentioned computer program includes computer program code, and the above-mentioned computer program code can be in source code form, object code form, executable file or some intermediate form. The above-mentioned computer-readable medium may include: any entity or device capable of carrying the above-mentioned computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the above-mentioned computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0198] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest range consistent with the principles and novel features disclosed herein.

[0199] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the above-mentioned device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the implementation method can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method implementation method, which will not be repeated here.

[0200] It should be noted that the methods and detailed examples provided in the above embodiments can be combined with the devices and equipment provided in the embodiments, and references can be made to each other, and no further details will be given.

[0201] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0202] In the embodiments provided by the present invention, it should be understood that the disclosed device / terminal equipment and method can be implemented in other ways. For example, the device / equipment implementation described above is only illustrative, for example, the division of the above modules or units is only a logical function division, and in actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0203] The above-mentioned embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above-mentioned embodiments, or replace some of the technical features therein by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for identifying drug resistance of an anti-tumor inhibitor, characterized in that: The method comprises the following steps: Step 1: Obtain gene mutation data and copy number variation data of patients with lung adenocarcinoma; Step 2: Define the drug resistance label for the data in step 1; Step 3: Preprocess the data from step 1 and determine the heterogeneity of lung adenocarcinoma patients; Step 4: Based on the heterogeneity determined in step 3, construct sBiosNet; Step 5: Set parameters for sBiosNet in step 4; Step 6: Use the network model with parameters set in step 5 to identify the resistance of anti-tumor drug inhibitors to the data pre-processed in step 3.

2. The method according to claim 1, characterized in that: The step 4 of constructing a semi-supervised biological sparse neural network model specifically includes the following steps: Step 4.1: Use biological information from the Reactome database to sparse the structure of sBiosNet; Step 4.2: sBiosNet generates pseudo labels with high confidence; Step 4.3: Generate mixed samples using MixMatch algorithm; Step 4.4: Construct a loss function based on the prediction results of the forward propagation; Step 4.5: Update the model parameters using the back-propagation algorithm based on gradient descent.

3. The method according to claim 2, characterized in that Specifically, the step 4.1 comprises: the constructed sBiosNet includes a new input layer, a new gene hidden layer, a new sub-pathway hidden layer, a new parent pathway hidden layer and a new output layer; The new input layer includes an input layer, a Dropout layer in front of the input layer, a Dropout layer in the back of the input layer, and a ReLU activation function; The new first hidden layer includes the first hidden layer, a Dropout layer in front of the first hidden layer, a Dropout layer in the back of the first hidden layer, and a ReLU activation function; The new second hidden layer includes a second hidden layer, a Dropout layer in front of the second hidden layer, a Dropout layer in the back of the second hidden layer, and a ReLU activation function; The new third hidden layer includes a third hidden layer, a Dropout layer in front of the third hidden layer, a Dropout layer in the back of the third hidden layer, and a ReLU activation function; The new output layer includes an output layer combined with a Softmax function; The input layer is used to receive multi-omics data of tumor patients, such as gene mutation data and copy number variation data. Each neuron in this layer corresponds to the variation of a gene. In the first hidden layer, each neuron corresponds to a specific gene, which is used to integrate the multi-omics data of the patient and extract the characteristics of the patient at the gene level; In the second hidden layer, each neuron corresponds to a specific sub-pathway, and the characteristics of the tumor patient at the gene level are integrated into the sub-pathway level to obtain the characteristics of the tumor patient at the sub-pathway level; In the third hidden layer, neurons represent biological parent pathways, and this layer further integrates the features of sub-pathways to obtain activation features of tumor patients at the parent pathway level; The output layer uses the characteristics of tumor patients at the parent pathway level to identify their resistance to PD-1 inhibitors.

4. The method according to claim 2, characterized in that: The step 4.2 is specifically to use L = {(x l,1 ,y l,1 ),L,(x l,n ,y l,n )} represents a set of n tumor patients with drug resistance labels, where x l,n represents the multi-omics data of labeled tumor patients n, y l,n Indicates the drug resistance label of the tumor patient; using U = {(x u,1 ),L,(x u,m )} represents m tumor patients without drug resistance labels, where x u,m represents the multi-omics data of unlabeled patient m; using f θ represents sBiosNet, θ represents its parameters to be learned; Using the prediction model f θ Predict the drug resistance of unlabeled tumor patient u and get the predicted value Will The maximum probability value in is recorded as the confidence score of sample u, that is: right Perform the Sharpen operation to obtain the pseudo label y′ of the unlabeled sample u u , expressed using the formula: Among them, the parameter T is used to control the pseudo label y′ u Entropy of Using a predefined confidence threshold p0, of is called a high-confidence pseudo-label. of is called a low-confidence pseudo-label; therefore, the set of c unlabeled samples with high-confidence pseudo-labels is U1={(x u,1 ,y′ u,1 ),L,(x u,c ,y′ u,c )}, the set of mc unlabeled samples with low confidence pseudo labels is U2 = {(x u,c+1 ,y′ u,c+1 ),L,(x u,m ,y′ u,m )}.

5. The method according to claim 2, characterized in that: The step 4.3 is specifically as follows: for the labeled sample set L = {(x l,1 ,y l,1 ),L,(x l,n ,y l,n )} and the unlabeled sample set U1={(x u,1 ,y′ u,1 ),L,(x u,c ,y′ u,c )} Perform random Mixup operation to generate mixed samples For two samples (x i ,y i ) and (x j ,y j ), the Mixup operation is: Among them, the parameter λ is randomly taken from Beta(mix α ,mix α ) distribution, set λ≥0.5, when y i is the true label of the labeled sample, As the true label; when y i When is the pseudo label of the unlabeled sample, is regarded as a pseudo label for the unlabeled sample; Use the Mixup operation to traverse all samples of the above two sample sets L and U1, and the mixed labeled sample set and a mixed set of unlabeled samples The mixed unlabeled sample set and the unlabeled sample set U2 with low confidence pseudo labels, and the combined sample set is re-recorded as the unlabeled sample set U = {(x u,1 ,y′ u,1 ),L,(x u,m ,y′ u,m )}.

6. The method according to claim 2, characterized in that Specifically, step 4.4 uses the drug resistance probability of the tumor patient obtained by the forward propagation process and the drug resistance label of the patient to calculate the loss function, and the loss function consists of two parts: supervised loss and unsupervised loss; wherein the supervised loss is the predicted value using the mixed labeled samples Mixing tags with it The cross entropy loss is calculated, and the unsupervised loss is the predicted value f using the unlabeled sample θ (x u ) and its pseudo label y′ u The mean square error is calculated, that is: Among them, CELoss and MSELoss represent the cross entropy loss function and the mean square error loss function respectively, t is the number of iterations of model training, and w(t) is the weight function of unsupervised loss; Construct the following w(t) function: Among them, w max is the maximum weight of the unsupervised loss; in the initial t0 training process, w(t) = 0, at which time only the supervised loss is used to update the parameter θ of sBiosNet, so that it can quickly learn the information of tumor patients' resistance to PD-1 inhibitors; in the next t1 training process, w(t) increases according to the trend of the Gaussian curve until it stabilizes at w max The weight of the unsupervised loss is gradually increased to improve the prediction accuracy and stability of sBiosNet for unlabeled samples.

7. The method according to claim 2, characterized in that Specifically, step 4.5 is to use a gradient descent-based back propagation algorithm to update the parameter θ of sBiosNet after the loss function is calculated through the forward propagation process. The updating process of the parameter θ is: Among them, η is the learning rate, t is the number of iterations; in order to speed up the convergence of sBiosNet during training, a fixed step size of n is used. η The learning rate decay strategy dynamically adjusts the learning rate η, that is, every n η After training for several iterations, the learning rate η becomes the original η γ times.

8. The method according to claim 2, characterized in that: The step 5 sets the parameters of the model as follows: the Embedding module of sBiosNet is sparsed using the gene-subpathway and subpathway-parent pathway relationships in the Reactome database; the number of neurons in the input layer and the three hidden layers are 755, 744, 1630 and 615 respectively, the batch size is set to 64, the total number of iterations is 400, the number of early stops is 40, the Adam optimizer is used to update the model parameters, and the initial learning rate η0 is 0.01; the learning rate is dynamically adjusted using a learning rate decay strategy with a fixed step size of 30, and the decay weight is η γ ; Set the parameter T of the Sharpen operation to 0.95 and the parameter t0 of the weight function w(t) of the unsupervised loss to 10; Other hyperparameters of sBiosNet, including the learning rate decay weight η, are selected by grid search. γ ∈{0.75,0.9}, the confidence threshold of the pseudo label p0∈{0.6,0.7}, the parameter Mix of the Beta distribution in the MixMatch process α ∈{0.15,0.35}, the probability p of the Dropout layer dropout ∈{0.1,0.2}, weight decay rate w in Adam optimizer decay ∈{1×e -3 ,1×e -4 }, the parameters of the unsupervised loss weight function w(t) are t1∈{20,30} and w max ∈{2.5,3.5,4.5}.

9. An anti-tumor inhibitor resistance recognition system, characterized in that: The system uses the method for identifying drug resistance of an anti-tumor inhibitor according to any one of claims 1 to 8, and the system comprises: Data acquisition module, which obtains gene mutation data and copy number variation data of patients with lung adenocarcinoma; The drug resistance label definition module defines drug resistance labels based on gene mutation data and copy number variation data of patients with lung adenocarcinoma; The data preprocessing and judgment module preprocesses the gene mutation data and copy number variation data of lung adenocarcinoma patients and judges the heterogeneity of lung adenocarcinoma patients; The network model building module builds a semi-supervised biological sparse neural network model sBiosNet based on the heterogeneity of lung adenocarcinoma patients; Parameter setting module, which sets the parameters of the semi-supervised biological sparse neural network model sBiosNet; Step 6: Use the network model with parameters set in step 5 to identify the resistance of anti-tumor drug inhibitors to the data processed in step 3.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

Images