Method and device for accurately predicting drug sensitivity response of clinical patient by using decoupling synthesis method

Abstract

The invention discloses a training method for decoupling into a drug response migration prediction model, and the method comprises the following steps: a drug specificity training step, extracting, using a drug-specific encoder, a feature shared between the first unlabeled cell line sample data constituting the source domain and the first unlabeled clinical sample data constituting the target domain and associated with the particular drug response; a decoupling forming step: a decoupling forming step; a second encoder capable of extracting a trained source domain of features independent of the particular drug response in second unlabeled cell line sample data and a third encoder capable of extracting a trained target domain of features independent of the particular drug response in second unlabeled clinical sample data by the drug-specific encoder the encoder is used for splicing the output of the drug specificity encoder and the output of the third encoder of the trained target domain, so as to obtain synthetic sample data; the synthetic data is then used to train the drug-specific encoder, thereby obtaining the trained drug-specific encoder, the method and apparatus decouple the cell line and clinical inputs into features related and independent of drug response and innovatively utilize the synthetic data to augment the dataset, and the method and apparatus can be used to identify the drug-specific encoder. And the drug sensitivity of clinical cancer patients is accurately predicted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological information and deep learning, and particularly relates to a method and device for accurately predicting clinical patient drug sensitivity response by using decoupling synthesis method. BACKGROUND

[0002] Predicting patient response to a given drug using patient's omics features (also known as pharmacogenomics) is a key task in precision oncology. However, for real clinical patients, data with drug response labels are usually scarce or hard to obtain publicly, on one hand, due to the extremely small sample size caused by the nature of the trial itself; on the other hand, the trial and treatment methods for patients in the clinic are very sensitive, and it is difficult to obtain the desired data publicly. This directly leads to the fact that researchers have difficulty in directly using clinical data to achieve good prediction results.

[0003] This has prompted researchers to create and collect large-scale preclinical resources, such as patient-derived xenografts or in vitro cultured cancer cell lines, as intermediates for understanding patient drug sensitivity information. In these preclinical resources, researchers can relatively easily conduct drug response experiments and thus obtain a large amount of response data. Although this method partially alleviates the problem of sample scarcity, the number of samples in the preclinical dataset is still less than the large number of omics features in the clinical environment, making it difficult to capture the complete distribution of clinical samples. In addition, due to the lack of tumor microenvironment and immune system, there is a big difference between the distribution of preclinical and clinical samples.

[0004] Currently, there are mainly the following methods for predicting the drug sensitivity of clinical patients: (1) The first type directly uses simple machine learning models or multi-layer neural networks to learn the response labels of cell lines and clinics to predict the drug response of test patients. This kind of method is mostly simple, and the prediction result is also poor; (2) The second type uses the transfer learning principle in deep learning to learn the potential representation of response information from large-scale preclinical and clinical data, but these methods actually achieve this transfer by learning the response label distribution in the clinical domain, ignoring the distribution difference between the test set and the clinical deployment environment. In the prediction of out-of-sample test samples, the result is often poor. This is called inductive learning in transfer learning, that is, by learning the label distribution in the test sample, it is transferred to unknown tasks and data; with the development of deep learning, the problems mentioned in the second type of method have been proposed in the last 2-3 years. In order to solve these problems, researchers have begun to try to achieve the learning of clinical response prediction without learning the distribution of clinical response labels. However, this kind of method has been published for a short time, and is limited by the number of samples, making it difficult to achieve ideal prediction results.

[0005] Therefore, there is an urgent need in the art to develop a method and device for accurately predicting the drug sensitivity response of clinical patients by using decoupling synthesis method, which decouples cell lines and clinical inputs into features related and unrelated to drug response, and innovatively uses synthetic data to expand the data set to accurately predict the drug sensitivity of cancer clinical patients. SUMMARY

[0006] The purpose of the present application is to provide a method and device for accurately predicting the drug sensitivity response of clinical patients by using decoupling synthesis method, which decouples cell lines and clinical inputs into features related and unrelated to drug response, and innovatively uses synthetic data to expand the data set to accurately predict the drug sensitivity of cancer clinical patients.

[0007] The first aspect of the present application provides a method for training a decoupling synthesis drug response migration prediction model, comprising the following steps:

[0008] (a) inputting first unlabeled cell line sample data from a source domain and first unlabeled clinical sample data from a target domain as input features to pre-train a first encoder, thereby obtaining a pre-trained first encoder, wherein the cell line sample data comprises omics expression of an in-vitro cultured cell line, and the clinical sample data comprises omics expression of a patient;

[0009] (b) inputting labeled cell line sample data of a specific drug from the source domain into a first prediction model composed of the pre-trained first encoder obtained in step (a) and a task classifier, performing drug-specific supervised training on the pre-trained first encoder and the task classifier, thereby obtaining a drug-specific trained first encoder and a task classifier, and the output of the drug-specific trained task classifier is a drug response value of the cell line sample to the specific drug;

[0010] (c) merging the parameters in the drug-specific trained first encoder and the task classifier obtained in step (b) into a drug-specific encoder, then fixing the drug-specific encoder, inputting second unlabeled cell line sample data from the source domain into the drug-specific encoder and a second encoder, training the second encoder, thereby obtaining a trained second encoder, inputting second unlabeled clinical sample data from the target domain into the drug-specific encoder and a third encoder, training the third encoder, thereby obtaining a trained third encoder, wherein the trained second encoder can extract features unrelated to the response of the specific drug in the second unlabeled cell line sample data, and the trained third encoder can extract features unrelated to the response of the specific drug in the second unlabeled clinical sample data;

[0011] (d) inputting a third unlabeled cell line sample data from the source domain into the drug-specific encoder, inputting a third unlabeled clinical sample data from the target domain into the trained third encoder obtained in step (c), concatenating the output of the drug-specific encoder and the output of the trained third encoder, thereby obtaining synthetic sample data;

[0012] (e) inputting the synthetic sample data and the labeled cell line sample data for the specific drug into the step (b) first prediction model, performing drug-specific supervised training on the pre-trained first encoder and task classifier, thereby obtaining a new drug-specific trained first encoder and task classifier;

[0013] (f) then, iteratively performing steps (c)-(e) in turn, thereby achieving training of the first encoder and task classifier.

[0014] In another preferred embodiment, the second unlabeled cell line sample data and the third unlabeled cell line sample data are the same.

[0015] In another preferred embodiment, the second unlabeled cell line sample data and the third unlabeled cell line sample data are the same and are from the Genomics of Drug Sensitivity in Cancer (GDSC).

[0016] In another preferred embodiment, the second unlabeled cell line sample data and the third unlabeled cell line sample data are unlabeled cell line sample data for the specific drug.

[0017] In another preferred embodiment, the first unlabeled cell line sample data is all cell line omics expression sample in the Genomics of Drug Sensitivity in Cancer.

[0018] In another preferred embodiment, the labeled cell line sample data for the specific drug is from the Genomics of Drug Sensitivity in Cancer.

[0019] In another preferred embodiment, the second unlabeled cell line sample data, the third unlabeled cell line sample data, and the labeled cell line sample data for the specific drug are all cell line sample data for the specific drug from the Genomics of Drug Sensitivity in Cancer, wherein the second unlabeled cell line sample data and the third unlabeled cell line sample data are unlabeled cell line sample data for the specific drug.

[0020] In another preferred embodiment, the labeled cell line sample data for the specific drug comprises omics expression of a cell line and response value of the cell line to the drug.

[0021] In another preferred embodiment, the labeled cell line sample data of the specific drug comprises the experimental cell line sample data that have been experimented for sensitivity to the specific drug and the drug response values of the experimental cell line sample data to the specific drug, and the experimental cell line sample data in the labeled cell line sample data of the specific drug is a subset of the first unlabeled cell line sample data.

[0022] In another preferred embodiment, the first unlabeled clinical sample data, the second unlabeled clinical sample data and the third unlabeled clinical sample data are the same.

[0023] In another preferred embodiment, the first unlabeled clinical sample data, the second unlabeled clinical sample data and the third unlabeled clinical sample data are all from The Cancer Genome Atlas (TCGA) and comprise all patient omics expression sample data in TCGA.

[0024] In another preferred embodiment, the drug-specific encoder obtained in step (c) is capable of extracting features shared between the first unlabeled cell line sample data and the first unlabeled clinical sample data and related to the response to the specific drug.

[0025] In another preferred embodiment, in step (c), the input of the second unlabeled cell line sample data from the source domain is divided into features in the second unlabeled cell line sample data related to the drug response and features in the second unlabeled cell line sample data unrelated to the drug response by fixing the drug-specific encoder and training the second encoder.

[0026] In another preferred embodiment, in step (c), the input of the second unlabeled clinical sample data from the target domain is divided into features in the second unlabeled clinical sample data related to the drug response and features in the second unlabeled clinical sample data unrelated to the drug response by fixing the drug-specific encoder and training the third encoder.

[0027] In another preferred embodiment, the omics features of the cell line samples of the first unlabeled cell line sample data constitute the source domain.

[0028] In another preferred embodiment, the omics features of the patients of the first unlabeled clinical sample data constitute the target domain.

[0029] In another preferred embodiment, in step (a), an adversarial domain discriminator is further included, and the training of the adversarial domain discriminator is such that the output of the first encoder cannot distinguish whether the first unlabeled cell line sample data is from the source domain or the first unlabeled clinical sample data is from the target domain.

[0030] In another preferred embodiment, the adversarial domain discriminator maps the samples of the two domains (the target domain and the source domain) to a shared latent space.

[0031] In another preferred embodiment, in step (c), the second encoder is trained by: the merged output of the output of the drug-specific encoder and the output of the second encoder contains all the features of the input second unlabeled cell line sample data, and the drug-specific encoder and the second encoder remain orthogonal, and / or the third encoder is trained by: the merged output of the output of the drug-specific encoder and the output of the third encoder contains all the features of the input second unlabeled clinical sample data, and the drug-specific encoder and the third encoder remain orthogonal.

[0032] In another preferred embodiment, in step (c), the drug-specific encoder extracts data features and the second encoder extracts data features through the decoder output.

[0033] In step (c), the drug-specific encoder extracts data features and the third encoder extracts data features through the decoder output.

[0034] In another preferred embodiment, in step (d), the output of the drug-specific encoder and the output of the trained third encoder are spliced together, and the synthetic sample data is obtained through the decoder output.

[0035] In another preferred embodiment, in the first training process, in step (d), the number of cell line samples of the third unlabeled cell line sample data is C1, the number of patient samples of the third unlabeled clinical sample data is C2, and the number of synthetic samples is at most C1xC2, and the label of the synthetic sample is from the drug response value of the cell line sample to the specific drug.

[0036] In another preferred embodiment, the number of iterations is 1, or 2, or 3, or 4 or more.

[0037] The second aspect of the present application provides a training device for decoupling a drug response migration prediction model, comprising:

[0038] a memory for storing computer executable instructions; and,

[0039] a processor coupled to the memory for implementing the steps of any of the above-described methods when the computer-executable instructions are executed.

[0040] A third aspect of the present application provides a computer-readable storage medium, having stored therein computer-executable instructions, which, when executed by a processor, implement any of the steps of the above-described methods.

[0041] It should be understood that all the technical features described above and the technical features described in detail below (such as the embodiments) can be combined with each other within the scope of the present application, thereby constituting new or preferred technical solutions. Due to the limited space, they will not be listed one by one here. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. It should be understood that the drawings described below are only some of the embodiments of the present application, and those skilled in the art can also obtain other embodiments from these drawings without creative labor.

[0043] Figure 1 is a schematic diagram of the method for training the disentangled synthesis drug response migration prediction model according to the first embodiment of the present application;

[0044] Figure 2a The drug-specific training phase in the training method of the disentangled synthesis drug response migration prediction model of the present application is shown, including model pre-training including domain adaptation and drug-specific training process;

[0045] Figure 2b The schematic diagram of the feature disentanglement phase and data synthesis in the training method of the disentangled synthesis drug response migration prediction model is shown;

[0046] Figure 3 is a schematic diagram of the detailed training steps of the disentangled synthesis drug response migration prediction model (DiSyn model);

[0047] Figure 4 is the result of the average AUROC of the model under different amounts of synthesized data changing with the number of iterations;

[0048] Figure 5is the distribution of AUROC scores of the DiSyn model of the present application on 16 drugs of the TCGA clinical data with the amount of synthetic data and the number of iterations;

[0049] Figure 6a is the embedding result of the drug-specific encoder on cell lines and patient samples before adding synthetic data for training.

[0050] Figure 6b is the embedding result of the drug-specific encoder on cell lines and patient samples after adding synthetic data for training.

[0051] Figure 7 is the comparison result of the DiSyn model according to an embodiment of the present application and other baseline models on five indicators (AUROC, AUPRC, ACC, F1 and APS) of TCGA data.

[0052] Figure 8 is the distribution of the six drugs with the highest improvement of the DiSyn model according to an embodiment of the present application compared with the baseline model, the box plot is drawn with all test results of five independent experiments, the baseline model is a machine learning model support vector machine (SVM), and the baseline model is CODE-AE, and the six drugs are paclitaxel, docetaxel, cisplatin, methotrexate, doxorubicin and bleomycin;

[0053] Figure 9 is a schematic diagram of TCGA clinical data omics expression data used for training of the DiSyn model according to an embodiment of the present application, each row is a patient, and each column is a gene feature;

[0054] Figure 10 is a schematic diagram of all GDSC cell line omics expression data used for training of the DiSyn model according to an embodiment of the present application, each row is a cell line sample, and each column is a gene feature;

[0055] Figure 11 is a schematic diagram of expression data of a cell line with a certain drug sensitivity response used for training of the DiSyn model according to an embodiment of the present application, each row is a cell line sample, and each column is a gene feature.

[0056] Figure 12 is Figure 11 is the drug sensitivity response value of the corresponding cell line sample in the table.

[0057] Figure 13The expression data of the cell lines with certain drug sensitivity response used for training of the DiSyn model is shown as input method in the drug-specific training process.

[0058] Figure 14 The expression data of the clinical patients with certain drug sensitivity response used for result evaluation is shown as a schematic diagram, with each row representing a patient and each column representing a gene feature.

[0059] Figure 15 The expression data of the clinical patients with certain drug sensitivity response used for result evaluation is shown as a schematic diagram, with each row representing a patient and each column representing a gene feature. Figure 14 The sample drug sensitivity response values corresponding to the clinic in the above table. DETAILED DESCRIPTION

[0060] The inventors have developed, for the first time, a decoupled drug response migration prediction model, in which a unique clinical patient drug sensitivity prediction migration framework is proposed, making full use of pre-clinical tumor-related resources and clinical data to predict the response of clinical patients to specific drugs. Specifically, the cell lines and clinical inputs are decoupled into drug response-related and unrelated features, and the data set is expanded innovatively using synthetic data, making accurate predictions of drug sensitivity of cancer clinical patients. The drug sensitivity model of the present application predicts drugs including 16 drugs such as Temozolomide, Bicalutamide, Bleomycin, etc. on patients and 25 drugs on tumor xenografts. In addition, five independent repeated experiments show that the prediction results exceed the highest level of all published models.

[0061] TERMS

[0062] As used herein, the term "factors related to a specific drug response are referred to as drug-specific factors, and factors unrelated to a specific drug response are referred to as non-specific factors";

[0063] As used herein, the terms "omics features" and "omics expression" are used interchangeably.

[0064] As used herein, the terms "drug-specific encoder" and "specific drug response-related information encoder" are used interchangeably.

[0065] As used herein, the terms "drug response-unrelated feature encoder" and "drug non-specific encoder" are used interchangeably.

[0066] It should be noted that in the present application, relational terms such as first and second are used merely to distinguish one entity or action from another entity or action, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions.

[0067] The present application has at least one of the following advantages

[0068] (a) The drug sensitivity prediction model obtained by the training method of the drug sensitivity model of the present application makes full use of limited cell lines and clinical patient data, and achieves the highest prediction level of actual clinical patient drug response among all published models;

[0069] (b) The training method of the drug sensitivity model of the present application proposes a unique transfer learning method, which decouples the preclinical and clinical inputs into features related and unrelated to drug response, and innovatively uses the response-unrelated features to synthesize mixed data (synthetic samples), significantly improving the prediction performance of the drug prediction model;

[0070] (c) The drug sensitivity prediction model obtained by the training method of the drug sensitivity model of the present application has never seen the labels on the test samples during the training process, which greatly improves the generalization ability of the model for different application scenarios;

[0071] (c) The drug sensitivity model provided by the present application aims to assist in providing personalized medication guidance, improving treatment effect, and ultimately improving patient quality of life.

[0072] In the following description, many technical details are presented in order to enable the reader to better understand the present application. However, those skilled in the art can understand that the technical solutions claimed by the present application can be realized even without these technical details and based on various changes and modifications of the following embodiments.

[0073] Structure design of the disentangled synthesis drug response transfer prediction model (DiSyn model)

[0074] The present application provides a disentangled synthesis drug response transfer prediction model, which is divided into two stages of training and prediction, and the training process is further divided into two steps of drug specificity training and decoupling iteration. Each part of the model is described in detail below.

[0075] Referring to Figure 2a and Figure 2b It can be seen that the main structure of the drug response transfer prediction model, the main idea of the model design is to decouple the inputs of the two domains (source domain and target domain) from the encoder E drug_specific and the encoder of the features unrelated to drug response in each domain Better transfer prediction in the shared subspace of drug response related features. S, T represent cell lines (Source) and clinical patients (Target) respectively.

[0076] Note that the dual denotation (encoder of drug response related features shared between domains) can be interpreted as the drug response related features can also be unshared, such as the immune microenvironment around cancer cells in patients is missing in cell lines cultured in vitro, and the impact of this environment on drugs is unique to patients. However, because the response label on the patient cannot be used during training, this part of the information cannot be known in theory, so only the shared part of the response information can be considered here.

[0077] In which, Figure 2a and 2b also describes the synthesis steps. Figure 2a For the schematic diagram, extract the part of the feature related to the drug response (drug response specific) from the sample with the drug response label on the cell line and recombine it with the feature unrelated to the drug response in the clinical sample (non-specific feature) to generate a synthetic sample. The label of the synthetic sample corresponds to the label of the cell line sample used.

[0078] This sample contains patient information and drug response labels on cell lines, which can be further used in the drug response transfer prediction model training shown in Figure 2b , greatly expanding the number of labeled samples.

[0079] In the drug response transfer prediction model training process, first, the drug specific training stage is carried out. Unlabeled cell line data and unlabeled clinical data are used as input, including gene expression profiles and drug reaction information. The subnetwork E drug_specific is used to learn the latent features of the training samples, and the activation of E drug_specific is used to train the classification of cell line drug response for a given cell line input.

[0080] In the decoupling process, E drug_specific remains fixed. As shown in Figure 2b , in order to learn the features unrelated to the drug in the input, we use a new network architecture. First, we use the decoder D and the drug independent feature encoder on each domain to reconstruct the input. We use an orthogonality loss to maximize the distance between the response related and unrelated features in each domain. After dividing the input into response related and unrelated features, we can use attribute exchange and decoupling synthesis to generate synthetic data.

[0081] As shown in Figure 2b, the labels of synthetic data and response-related features come from source domain, and response-irrelevant features come from clinical domain. Next, we can use synthetic data to further participate in drug-specific training phase, so as to improve the performance of E drug_specific . Then synthetic data further trains response-irrelevant feature encoder to ensure that it does not contain information related to the response of a specific drug. The original motivation behind this iterative process is to generate more accurate synthetic data with an effective feature encoder, and the more accurate synthetic data, in turn, can further promote the training of the encoder. When accurate synthetic data is introduced into the training process, it can help improve the performance of the model by providing additional examples and variants that may not be fully represented in the original data set. We show the performance improvement of each iteration in the results.

[0082] In the prediction phase, we use the specific drug response-related information encoder E drug_specific to predict the omics features of the test sample, and output the drug response prediction.

[0083] A method for decoupling the training and prediction of a synthetic drug response migration prediction model

[0084] Referring to Figure 3 , the first embodiment of the present application provides a method for decoupling the training of a synthetic drug response migration prediction model, comprising the following steps:

[0085] (a) inputting first unlabeled cell line sample data from a source domain (Source) and first unlabeled clinical sample data from a target domain (Target) as input features, and pre-training a first encoder E task_specific , so as to obtain a pre-trained first encoder E task_specific ; the cell line sample data includes the omics expression of an in vitro cultured cell line, and the clinical sample data includes the omics expression of a patient;

[0086] In an embodiment, the first unlabeled cell line sample data comes from the GDSC database, and the first unlabeled cell line sample data has 966 samples and corresponding 1363 genes, which are also irrelevant to drug response (see Figure 10 ); the first unlabeled clinical sample data comes from the TCGA database, and the first unlabeled clinical sample data has 10237 samples. Each patient uses 1363 genes, because the omics data of these patients is irrelevant to their drug response labels (see Figure 9 ).

[0087] (b) inputting labeled cell line sample data of a specific drug from the source domain into the pre-trained first encoder E task_specifica first prediction model composed of the pre-trained first encoder E task_specific a task classifier, and the task classifier is drug-specific supervised trained, thereby obtaining a drug-specific trained first encoder E task_specific a task classifier, and the output of the drug-specific trained task classifier is a drug response value of the cell line sample to the specific drug;

[0088] (c) inputting the drug-specific trained first encoder E task_specific and the task classifier into the drug-specific encoder E drug_specific (i.e. a specific drug response related information encoder, which can also be called a feature extractor), and then inputting the drug-specific encoder E drug_specific fixed, respectively inputting the second unlabeled cell line sample data from the source domain into the drug-specific encoder E drug_specific and a second encoder (i.e. a drug response independent feature encoder or a drug non-specific encoder in the source domain), and training the second encoder , thereby obtaining a trained second encoder respectively inputting the second unlabeled clinical sample data from the target domain into the drug-specific encoder E drug_specific and a third encoder (i.e. a drug response independent feature encoder or a drug non-specific encoder in the target domain), and training the third encoder , thereby obtaining a trained third encoder The trained second encoder can extract features in the second unlabeled cell line sample data that are independent of the specific drug response, and the trained third encoder can extract features in the second unlabeled clinical sample data that are independent of the specific drug response.

[0089] (d) inputting the third unlabeled cell line sample data from the source domain into the drug-specific encoder E drug_specific , and inputting the third unlabeled clinical sample data from the target domain into the trained third encoder obtained in step (c), and concatenating the output of the drug-specific encoder E drug_specific and the output of the trained third encoder , thereby obtaining synthetic sample data.

[0090] In an embodiment, the second and third unlabeled cell line sample data are drug-specific.

[0091] (e) inputting the synthetic sample data and the labeled cell line sample data of the specific drug into the step (b) first prediction model, and performing drug-specific supervised training on the pre-trained first encoder E task_specific and task classifier, thereby obtaining a new drug-specific trained first encoder E task_specific and task classifier; at this time, the task classifier obtains a more effective task classifier than the first step training;

[0092] (f) then, iteratively performing steps (c)-(e) in turn, thereby realizing the training of the first encoder E task_specific and task classifier.

[0093] In an embodiment, the first unlabeled cell line sample data is all cell line expression sample data in the Genomics of Drug Sensitivity in Cancer (GDSC) database.

[0094] The second unlabeled cell line sample data and the third unlabeled cell line sample data are the same; the second unlabeled cell line sample data, the third unlabeled cell line sample data, and the labeled cell line sample data of the specific drug are cell line sample data from the GDSC database for a specific drug, wherein the second unlabeled cell line sample data and the third unlabeled cell line sample data are unlabeled cell line sample data for the specific drug. For example, the labeled cell line sample data of the specific drug is labeled (drug response value) cell line sample data for drug A from the Genomics of Drug Sensitivity in Cancer database, and the second unlabeled cell line sample data and the third unlabeled cell line sample data are unlabeled cell line sample data for drug A.

[0095] Preferably, the cell line sample data used in steps (b)-(e) are all for a certain specific drug (for example, drugs Temozolomide, Bleomycin, Docetaxel, and Gemcitabine in the Genomics of Drug Sensitivity in Cancer database).

[0096] In an embodiment, the first unlabeled clinical sample data, the second unlabeled clinical sample data, and the third unlabeled clinical sample data are all from the Cancer Genome Atlas (TCGA) database; the first unlabeled clinical sample data, the second unlabeled clinical sample data, and the third unlabeled clinical sample data are all patient genomic expression sample data in the TCGA database.

[0097] Preferably, in the first training process, in the step (d), the number of cell line samples of the third unlabeled cell line sample data is C1 (for example, in the GDSC database, the number of Bleomycin is C1 equal to 843), the number of patient samples of the third unlabeled clinical sample data is C2 (for example, in the TCGA database, C2 is equal to 10237), after multiple iterations, the number of synthetic samples is at most C1xC2 (i.e. 843x10237=8629791), and the label of the synthetic sample is from the drug response value of the cell line sample to the specific drug. For example, taking Bleomycin as an example, in the first training, when the synthetic data amount coefficient is 2, the model randomly selects twice the number of clinical samples of Bleomycin labeled sample amount (i.e. 1686 clinical samples) from the TCGA database for data synthesis, and generates 1686 synthetic samples.

[0098] The application also provides a training method for decoupling a drug response migration prediction model, comprising the following steps:

[0099] Input the omics characteristics of the test sample into the first prediction model, and the task classifier of the first prediction model outputs the drug response value.

[0100] Method for establishing the DiSyn model of the application

[0101] In the present application, factors related to the response to a specific drug are referred to as drug-specific factors, and factors unrelated to the response to a specific drug are referred to as non-specific factors.

[0102] The whole process is divided into two steps and iterated with each other in subsequent training.

[0103] 1.1 Domain Introduction

[0104] First, a domain includes two components: feature space and marginal probability distribution P(X), wherein Given a specific domain D = {X, P(X)}, the task includes two components: label space and target prediction function f(·). In the present application, taking patient data as an example, all potential omics characteristics from cell lines and patients constitute the source domain and the target domain, respectively and All samples x sampled from the two domains constitute data sets D S and D T . More specifically, the labeled data set in the source domain is denoted as wherein denotes N sthe i-th instance of the gene expression sample, and as its drug response label. Similarly, we denote the unlabeled dataset in the target domain as where is the N T gene expression sample.

[0105] 1.2 Domain adaptation (specific task training)

[0106] using a drug-specific encoder E drug_specific to learn the latent features of the training samples for the specific drug response. First, the drug-specific encoder E drug_specific is trained to disentangle the latent features of the source and target domains, and then the activations of E drug_specific are trained to classify the cell line drug response of a given cell line .

[0107] The loss of this step

[0108] L = L class + γL DA , where γ is a hyperparameter weight. L class is the negative log-likelihood of the classifier prediction;

[0109] where, is the classifier output. L DA aims to absorb the common features of the source and target domains. We choose the Domain Separation Network (Bousmalis, K., Trigeorgis, G., Silberman, N., Krishnan, D. & Erhan, D. Domain Separation Networks) as the backend, which has been proven to perform well in domain adaptation.

[0110] In the Domain Separation Network, the input samples from each domain are divided into two subspaces: features private within each domain and features shared across domains. The loss function can be defined as

[0111] L DA = L recon + αL diff + βL sim

[0112] where α, β are hyperparameter weights, L recon is used to reconstruct the input data, usually defined as the output of the decoder and the input Root Mean Square Error (RMSE) between.

[0113]

[0114] L diff is an orthogonal constraint to separate the common and private features,

[0115]

[0116] is the square of the Frobenius norm, H is a metric whose rows are the hidden vectors from source and target samples.

[0117] L sim aims to encourage the common features of the two domains to be as close as possible. This loss can be the Maximum Mean Discrepancy (MMD) (Gretton, A., Borgwardt, K. M., Rasch, M. J., Schlkopf, Y. & Smola, A. A kernel two-sample test. J. Mach. Learn. Res. 13, 723-773 (2012)) or an adversarial loss, we report model results using both ways. The MMD loss is a kernel distance function measure between the samples.

[0118]

[0119] where κ(·, ·) is a positive definite kernel function, and are the common features extracted from the input samples and respectively.

[0120] ​Similarly, the adversarial similarity loss aims to absorb the common features between two domains. We define a discriminator to distinguish the hidden features of source and target domain data. Through the adversarial training setting, the discriminator is encouraged to confuse the feature generator as much as possible. We employ a gradient reversal layer (GRL) network to achieve this (Ganin, Y. & Lempitsky, V. Unsupervised domain adaptation by backpropagation. in Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37 1180-1189 (JMLR.org, Lille, France, 2015)). The GRL unit is a layer that simply reverses the direction of the gradient passing through it. The loss function of the discriminator can be defined as the binomial cross-entropy of the prediction task.

[0121]

[0122] where represents the input sample x i the domain label prediction, y i is the true value.

[0123] Through this process, we have captured the drug-specific features in the common features of the input.

[0124] 1.3 Decoupling

[0125] In the decoupling step, E drug_specific is kept fixed. To learn the drug non-specific features, we use a new network architecture. First, we employ a decoder D and a drug-agnostic feature encoder E drug_unspecific on each domain to reconstruct the input. Inspired by the domain disentanglement network, we employ an orthogonality loss to maximize the distance between the specific and non-specific features in each domain. The loss function can be represented as

[0126] L = L recon + σL diff

[0127] where σ is a hyperparameter weight, where L rec and L diff are consistent with the previous definition.

[0128] After separating the input into drug-specific and non-specific features, we can employ attribute swapping and decoupling synthesis to generate synthetic data. As Figure 2bAs shown, the labels of synthetic data and drug-specific specificity come from the source domain, and non-specific features come from the target domain.

[0129]

[0130]

[0131] where concat(·) is the vector concatenation operation.

[0132] Then, we can further train the non-specific feature extractor using synthetic data to further remove information related to specific drug response.

[0133] The loss function can be expressed as

[0134] L = L recon + σL diff - λ*L adv_class

[0135] where σ, λ are hyperparameter weights.

[0136] L recon and L diff are in the same form as the previous definition, but count synthetic samples.

[0137]

[0138] L adv_class is the adversarial loss of drug response classification.

[0139]

[0140] where y i is the actual value of the drug response label of sample i, is the predicted value of the corresponding label.

[0141] 1.4 Iteration

[0142] The synthetic data generated in the decoupling step can further participate in the training phase of the specific task. The initial motivation behind this iterative process is that an effective feature extractor can generate more accurate synthetic data, and more accurate synthetic data can further promote the training of the extractor.

[0143] The training process is similar to the first step, except that the loss function of the domain adaptation stage is slightly modified.

[0144]

[0145] where G(·) is the feature classifier, Represents the mix of real and synthetic data. To enhance the generalization performance beyond the cell line dataset, we mixed real data and synthetic data for model training.

[0146] 2.1 Dataset

[0147] In this application, we investigate the classification performance of the DiSyn model on three publicly available datasets: TCGA (Hutter, C. & Zenklusen, J. C. The Cancer Genome Atlas: Creating Lasting Value beyond Its Data. Cell 173, 283-285 (2018), NIBR PDXE (Gao, H. et al. High-throughput screening using patient-derived tumor xenografts to predict clinical trial drug response. Nat. Med. 21, 1318-1325 (2015)) Mulligan et al. data (Mulligan, G. et al. Gene expression profiling and correlation with outcome in clinical trials of the proteasome inhibitor bortezomib. Blood 109, 3177-3188 (2007)). Gene expression profiles of cancer cell lines and patient tumors were used for pre-training. We also investigated the results using copy number variation and genomic mutations as input, which showed that expression profiles produced the best performance. We applied max-min normalization to the expression features. Subsequently, we performed feature selection on the top 1000 genes on the input features according to the method (Clayton, E. A. Leveraging TCGA gene expression data to build predictive models for cancer drug response. (2020).). The specific genes we selected can be found in the supplementary material. For the training phase of the training for the specific task, we collected the labeled cell line dataset from the GDSC database. We selected labeled samples from the target domain as the test dataset. Considering the sample size, we selected 24 drugs for testing, of which 16 drugs were from TCGA, 7 drugs were from PDX, and 1 drug was from Mulligan et al. data. For each drug, we trained the model separately and compared it with the baseline.

[0148] 2.2 Training process

[0149] The DiSyn model of the present application starts with a pre-training process, whose training set is the gene expression profiles from the source and target domains. The number of epochs in this process and other hyperparameters are determined by a grid search process.

[0150] After the pre-training process, we introduce a drug-specific training phase to learn the response information specific to each drug. The activations of the feature extractor are used for drug-specific classification. In this process, we define a classifier to learn the response of cell line samples to specific drugs.

[0151] The cell line samples are split into five folds according to drug sensitivity for five-fold cross-validation. In each run, we use 4 / 5 of the folds to train the model, while the remaining 1 / 5 is used as a validation set to search for the best parameters. It is worth mentioning that, in order to ensure fairness, the five-fold split setting is consistent in all cross-validation baseline models. We use the AUROC result on the validation set as the stopping criterion. When the training is over, we consider that the encoder has learned the underlying information about the response to a specific drug.

[0152] In the next phase, to obtain a drug non-specific encoder for each domain, we inherit and fix the specific encoder obtained in the previous step and retrain the remaining components in a new network structure. The number of training epochs and other hyperparameter weights for this step are also determined by parameter search.

[0153] The synthetic data generated by the drug non-specific encoder obtained in this step can further participate in subsequent drug-specific training. This training process is like the first step, but we incorporate synthetic data. The scale of synthetic data and other hyperparameters are also determined by parameter search. For all labeled GDSC cell line samples used in data synthesis and drug-specific training, we follow a consistent five-fold split setting to avoid any data leakage problems related to data set leakage.

[0154] With the end of drug-specific training, we should have obtained a more effective classifier. The model goes through three such cycles of change for each drug. Finally, we record the performance of each fold on the test set and report the average performance. Finally, in order to obtain more accurate and fair results, we performed five independent experiments for each drug.

[0155] 2.3 Baseline

[0156] In the present application, DiSyn is compared with various baseline models, including state-of-the-art deep learning methods and many other machine learning methods, to demonstrate its effectiveness.

[0157] In line with DiSyn, we performed five independent replicates. In each independent run, we split the training data into five folds according to drug sensitivity for a fair performance evaluation. For the CODE-AE baseline model (A context-aware deconfounding autoencoder for robust prediction of personalized clinical drug response from cell-line compound screening. 4, (2022)), we used the training code it provided and had the same cross-validation setup as us. For a fair comparison, we also used the same parameter ranges in DiSyn's grid search as possible. For the AITL baseline model (Sharifi-Noghabi, H., Peng, S., Zolotareva, O., Collins, C. C. & Ester, M. AITL: Adversarial Inductive Transfer Learning with input and output space adaptation for pharmacogenomics. Bioinformatics 36, i380-i388 (2020)), we used the training code provided in its GitHub project. Also due to AITL's stratification strategy, we could not complete the experiments for certain drugs. Given that AITL is an inductive model, we also used the labeled TCGA samples to train the MTL (Multi-task Learning) network. We split the training samples into five folds and reported the average scores. For the multi-drug prediction model TUGDA baseline model (Peres Da Silva, R., Suphavilai, C. & Nagarajan, N. TUGDA: task uncertainty guided domain adaptation for robust generalization of cancer drug response prediction from in vitro to in vivo settings. Bioinformatics 37, i76-i83 (2021)) we adopted the hyperparameter search based on its GitHub project and reported the best performance.For the TRANSACT baseline model (Mourragui, S. M. C. et al. Predicting patient response with models trained on cell lines and patient-derived xenografts by nonlinear transfer learning. Proc. Natl. Acad. Sci. 118, e2106682118 (2021)) and the PRECISE baseline model (Mourragui, S., Loog, M., Van De Wiel, M. A., Reinders, M. J. T. & Wessels, L. F. A. PRECISE: a domain adaptation approach to transfer predictors of drug response from pre-clinical models to tumors. Bioinformatics 35, i510-i519 (2019)), we used the best hyperparameter settings provided by the authors, as the original papers did not specify parameters related to random state specificity. For the baseline models of Random Forest, SVM, ElasticNet and AdaBoost, we used the Scikit-learn package for five-fold cross-validation stratification and used the same training and parameter search procedure. The specific parameter ranges we used for grid search can be found in the supplement.

[0158] 2.4 Evaluation method

[0159] For each target domain, we performed five independent repetitions. In each independent experiment, we performed five-fold cross-validation to ensure the fairness of the evaluation. In each fold, we recorded the area under the receiver operating characteristic curve (AUROC) on the test set as the primary indicator of evaluating the classification performance of the model, as it is stable to label distribution changes and has been evaluated in various classification methods. In addition, we also reported the area under the precision-recall curve (AUPRC), accuracy (ACC), F1 score and average precision score (APS) of the model. The average of the five test folds was used for comparison with the baseline model. All models were trained specifically for cell line data and the best parameters were selected, followed by testing on the reserved labeled test set to report the performance results.

[0160] 3. Comparison of model prediction results

[0161] We compared our model with the current state-of-the-art methods on AUROC, AUPRC and other five metrics on 16 drugs (Temozolomide, Bicalutamide, Bleomycin, etc.) and 25 drugs on tumor xenografts in TCGA database. Five independent repeated experiments show that, among the 16 drugs on clinical samples, our model improves about 15.4% on average compared to the drug sensitivity prediction method published in Nature Machine Intelligence in 2022.

[0162] On patient-derived xenograft tissue (PDX) and Mulligan et al. data from patients, our model also has about 5% improvement compared to the second place.

[0163] In addition, we also did the change of model performance with the number of iterations under different synthetic data volume (see Figure 4 and Figure 5 ).

[0164] It can be seen that the prediction performance of the model is significantly improved with the number of iterations, and the use of synthetic data brings significant performance improvement to the model.

[0165] In order to make the purpose, technical scheme and advantages of the present application clearer, one embodiment of the present application will be further described in detail below with reference to the drawings. It should be understood that these only provide some examples that the present application can take, but are not intended to limit the scope of the present application.

[0166] Embodiment

[0167] In this embodiment, a specific training process of the DiSyn model of the present application is given, in which the training set is the gene expression profile from the Genomics of Drug Sensitivity in Cancer (GDSC database) and the American Cancer Genome Atlas (TCGA database).

[0168] Training process

[0169] The DiSyn model includes a pre-training process, and the training set of the process is the gene expression profile from GDSC and TCGA. The number of training rounds and other hyperparameters in the process are determined by a grid search process.

[0170] Specifically, referring to Figure 3 , the pre-training step (a) is to pre-train E task_specific using the unlabeled cell lines and the omics features of patients;

[0171] Drug-specific training step (b) uses the drug-specific supervised training on the cell line samples contained in the drug and their response labels. From here on, everything is done separately for each drug;

[0172] Feature decoupling step (c) merges the partial parameters in the two encoders into E drug_specific and fixes the training of the two domain-specific response-agnostic information encoders where the DiSyn model design ensures that it is orthogonal to E drug_specific The merging contains all the information of the input, while maintaining orthogonality (independence) between the two;

[0173] Step (d) obtains E drug_specific and the respective After that, the synthetic samples can be recombined using the cell line and patient information. The synthetic sample labels come from the drug response values of the cell lines corresponding to the drugs;

[0174] Step (e) further participates in the supervised training of E drug_specific to obtain a more effective feature encoder than the previous step;

[0175] Step (f) feature decoupling, training response-agnostic feature encoders. A more effective relevant feature encoder can in turn train a more effective irrelevant feature encoder, thereby improving the quality of synthetic data for the next iteration.

[0176] According to the drug sensitivity, the labeled samples are divided into five folds for cross-validation. That is, each time we run, we use 4 / 5 of the cell line labeled data to train the model, and the remaining 1 / 5 is used as the validation set. It is worth mentioning that the five-fold partitioning setting is consistent in all models with cross-validation. During drug-specific training, the prediction results of the validation set are used as the model training stopping criterion. We believe that by the end of training, the encoder E drug_specific has learned the latent information related to drug response.

[0177] In order to obtain drug non-specific encoders E drug_unspecific from each domain, the drug-specific encoders in the previous step will be fixed and the remaining components will be retrained in a new network structure. The number of training rounds and other hyperparameter weights are also determined by parameter search. The drug non-specific encoders obtained from this step can in turn generate synthetic data to further participate in subsequent drug-specific training. The training process is similar to the first step, but we add synthetic data. Among them, the synthetic data and other hyperparameters are also determined by parameter search. With the end of the second drug-specific training, we have theoretically obtained a more effective task classifier than the first training.

[0178] The model went through three such training cycles for each drug. Finally, we report the performance of the drug-specific classifiers on the TCGA test set as the test results.

[0179] Ablation experiment results

[0180] Figure 4 and Figure 5 The AUROC scores of our model on the 16 drugs in the TCGA clinical data (see Table 1) are shown in Figure 2 as a function of the amount of synthetic data and the number of iterations. The amount of synthetic data = 0, 1, 2, 3 refers to the ratio of the amount of synthetic samples to the amount of labeled cell line data samples for the drug at the beginning of the iteration. 0 means only the labeled cell line samples are used for the iteration.

[0181] Table 1: The number of samples for the 16 different drugs

[0182] Drug name Sample size Temozolomide 892 Sorafenib 392 5-Fluorouracil 896 Methotrexate 930 Bicalutamide 948 Cisplatin 940 Pemetrexed 888 Paclitaxel 391 Vinblastine 928 Etoposide 867 Doxorubicin 953 Bleomycin 843 Docetaxel 930 Gemcitabine 952 Tamoxifen 908 Vinorelbine 868

[0183] As can be seen, the use of synthetic data significantly improves the training results. Specifically, with the same number of iterations, the addition of one times the training samples of synthetic data brings a very large improvement compared to the use of only the original cell line samples. Similarly, with the increase of the number of iterations, the average performance of the model also gradually improves.

[0184] DiSyn model training prediction results

[0185] During the training process, the drug response related encoder E drug_specific is reduced in dimension, Figure 6a and Figure 6b represent the related encoder of the first iteration and the best result, respectively.

[0186] R represents Response, and NR represents Not-Response. TCGA test refers to the test set samples used for model testing. As can be seen, the samples of the two domains of clinical patients (TCGA) and cell lines (GDSC) can be well embedded in the same distribution; and with the iteration of the model, the encoder can better classify the response and non-response samples.

[0187] Model results on clinical samples

[0188] Referring to Figure 7 , we report the results of the model on the five indicators. Among them, AUROC is used as the main indicator for our model training early stopping strategy and performance report. In order to measure fairness and impartiality, we conducted five independent repeated experiments and used five-fold cross-validation in each experiment. In the training steps common to the models, we used the same parameter range.

[0189] We achieved the first place in 14 out of 16 drugs. Calculating the average AUROC score of 16 drugs, we exceeded CODE-AE published in Nature Machine Intelligence in 2022 by about 14.2%.

[0190] AUROC mainly considers the proportion of true positives in all true positives in the predicted positive examples, and PRC mainly considers the proportion of true positives in all predicted positive examples. The main difference between the two is that AUPRC is greatly affected by the distribution of the proportion of true samples. Taking Sorafenib as an example, there are only 2 positive examples (true positives) in the test set of 15 samples, which should be the reason for its low AUPRC result, and the performance in AUROC is more stable.

[0191] It should be noted that the baseline models (Ada, PRECISE, Random Forest, TRANSACTSVM, ElasticNet, TUGDA, AITL, CODE-AE, VAEN, MOLI, TGSA-CNV, PaccMann, TGSA-EXP, DrugCell, TGSA, TGSA-MUT and CRDNN) of the present application are all recorded in the prior art.

[0192] See Figure 8 , specifically we show the boxplot of AUROC distribution of six groups of drugs and calculate the relative improvement (Relative Improvement) of the average score of AUROC based on the second-ranked model CODE-AE

[0193]

[0194] This indicator represents the percentage improvement of a new model over a baseline model in a certain indicator. On the top five drugs with the largest improvement, our model exceeded CODE-AE by an average of 26.80%. Bleomycin works by binding to DNA to cause nucleic acid chain breaks, and compared to CODE-AE, our model also achieved an improvement of about 13.31% on this drug.

[0195] Supplementary notes on the data set used

[0196] (1) The "unlabeled clinical data" mentioned in the application and the examples refers to the patient and omics expression contained in the Cancer Genome Atlas data set, each row is a patient sample, and each column is a gene. We have a total of 10237 patient samples, and each patient uses 1363-dimensional genes. Because the omics data of these patients is not related to their drug response label, we only learn the data distribution, so it is called "unlabeled clinical data" (see Figure 9 ).

[0197] (2) The cell line data used in pre-training refers to the omics expression of cancer cell lines cultured in vitro in the GDSC data set, a total of 966 samples and corresponding 1363-dimensional genes. This part is also unrelated to drug response (see Figure 10 ).

[0198] (3) After pre-training, we use "labeled cell line data" to train the model for drug-specific training, which refers to the cell line samples that have conducted sensitivity experiments and their drug sensitivity results. The area under the survival curve of the drug-treated cells is used as a measure. Take Bleomycin as an example, not every cell line has done the experiment, and the cell line sample with experimental data is 843. We use this part as input to calculate drug response and compare it with the real response to train the model (see Figure 11 ).

[0199] (4) See Figure 12 , which shows the response label of the cell line sample in the sensitivity experiment for the drug:

[0200] This part of the labeled sample is largely missing in the clinical data, so cell line samples are used for transfer learning, which is the biggest difficulty of this task. With the introduction of synthetic data, the difficulty caused by the lack of labeled / tagged data is alleviated. See Figure 13 , Figure 13 shows the expression data of the cell line with a certain drug sensitivity response used by the DiSyn model training in the drug-specific training process.

[0201] See Figure 14 , Figure 14 is a schematic diagram of the expression data of the clinical patient with a certain drug sensitivity response used for result evaluation, each row is a patient, and each column is a gene feature. In the result evaluation of the DiSyn model of the present application, the labeled clinical samples not involved in the whole model training are used, and this part of the sample is very small and only used as a test set. Take Bleomycin as an example, this part of the sample is only 52.

[0202] SeeFigure 15 , which shows its true response label. We use the omics features of this part of samples to embed in drug response related encoder and make prediction, calculate each index with its corresponding drug response label, and report the results.

[0203] Embodiments of the present application also provide a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions are executed by a processor to implement the method embodiments of the present application. The computer readable storage medium includes permanent and non-permanent, removable and non-removable media, which can be realized by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition herein, the computer readable storage medium does not include transitory computer readable media, such as modulated data signals and carriers.

[0204] Further, the embodiments of the present application also provide a training device for decoupling a drug response migration prediction model, comprising a memory for storing computer executable instructions, and a processor; the processor is configured to implement the steps of the above-mentioned method embodiments when executing the computer executable instructions in the memory. The processor can be a central processing unit (CPU), a graphic processing unit (GPU), a digital signal processor (DSP), a microcontroller unit (MCU), a neural processing unit (NPU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, etc. The memory can be a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk, a solid state disk, etc. The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware processor execution, or a combination of hardware and software modules in the processor.

[0205] The sequence numbers used in describing the steps of the method do not constitute any limitation on the sequence of the steps. For example, the step with a larger sequence number does not have to be executed after the step with a smaller sequence number, but can be executed before the step with a smaller sequence number, or can be executed in parallel, as long as such an execution sequence is reasonable for a person skilled in the art. For another example, the plurality of steps with consecutive sequence numbers (for example, step a, step b, step c, etc.) do not limit that other steps can be executed therebetween, for example, other steps can be executed between step a and step b.

[0206] All the documents mentioned in the present specification are considered to be included in the disclosure of the present application as a whole, so as to be used as a modification if necessary. In addition, it should be understood that the above description is merely preferred embodiments of the present specification, and is not intended to limit the protection scope of the present specification. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of the present specification shall be included in the protection scope of one or more embodiments of the present specification.

Claims

1. A training method for decoupling a drug response migration prediction model, characterized in that, The method comprises the following steps: (a) pre-training a first encoder by taking first unlabeled cell line sample data from a source domain and first unlabeled clinical sample data from a target domain as input features, so as to obtain a pre-trained first encoder, wherein the cell line sample data comprises omics expression of an in-vitro cultured cell line, and the clinical sample data comprises omics expression of a patient; (b) inputting labeled cell line sample data of a specific drug from the source domain into a first prediction model composed of the pre-trained first encoder obtained in step (a) and a task classifier, performing drug-specific supervised training on the pre-trained first encoder and the task classifier, so as to obtain a drug-specific trained first encoder and a task classifier, and the output of the drug-specific trained task classifier is a drug response value of the cell line sample to the specific drug; (c) combining parameters in the drug-specific trained first encoder and the task classifier obtained in step (b) into a drug-specific encoder, then fixing the drug-specific encoder, inputting second unlabeled cell line sample data from the source domain into the drug-specific encoder and a second encoder respectively, training the second encoder, so as to obtain a trained second encoder, inputting second unlabeled clinical sample data from the target domain into the drug-specific encoder and a third encoder respectively, training the third encoder, so as to obtain a trained third encoder, wherein the trained second encoder can extract features irrelevant to the response of the specific drug in the second unlabeled cell line sample data, and the trained third encoder can extract features irrelevant to the response of the specific drug in the second unlabeled clinical sample data; (d) inputting third unlabeled cell line sample data from the source domain into the drug-specific encoder, inputting third unlabeled clinical sample data from the target domain into the trained third encoder obtained in step (c), and splicing the output of the drug-specific encoder and the output of the trained third encoder together, so as to obtain synthetic sample data; (e) inputting the synthetic sample data and the labeled cell line sample data of the specific drug into the first prediction model in step (b), performing drug-specific supervised training on the pre-trained first encoder and the task classifier, so as to obtain a new drug-specific trained first encoder and a task classifier; (f) then, iteratively performing steps (c)-(e) in sequence, so as to train the first encoder and the task classifier.

2. The method of claim 1, wherein, The second unlabeled cell line sample data and the third unlabeled cell line sample data are the same.

3. The method of claim 1, wherein, The first unlabeled clinical sample data, the second unlabeled clinical sample data and the third unlabeled clinical sample data are the same.

4. The method of claim 1, wherein, The drug-specific encoder obtained in step (c) is capable of extracting features shared between the first unlabeled cell line sample data and the first unlabeled clinical sample data and associated with the specific drug response.

5. The method of claim 1, wherein, In step (a), an adversarial domain discriminator is further included, and training of the adversarial domain discriminator makes the output of the first encoder unable to distinguish whether the first unlabeled cell line sample data is from a source domain or the first unlabeled clinical sample data is from a target domain.

6. The method of claim 1, wherein, In step (c), the second encoder is trained in such a manner that a merged output obtained by merging the output of the drug-specific encoder and the output of the second encoder contains all features of the input second unlabeled cell line sample data, and the drug-specific encoder and the second encoder remain orthogonal, and / or the third encoder is trained in such a manner that a merged output obtained by merging the output of the drug-specific encoder and the output of the third encoder contains all features of the input second unlabeled clinical sample data, and the drug-specific encoder and the third encoder remain orthogonal.

7. The method of claim 6, wherein, In the first training process, in step (d), the number of cell line samples of the third unlabeled cell line sample data is C1, the number of patient samples of the third unlabeled clinical sample data is C2, the number of synthetic samples is at most C1xC2, and the label of the synthetic sample is from the drug response value of the cell line sample to the specific drug.

8. The method of claim 1, wherein, The number of iterations is 1, or 2, or 3, or 4. 9.A training device for decoupling a drug response migration prediction model, characterized in that, Comprising: a memory for storing computer executable instructions; and, a processor coupled to the memory for implementing the steps in the method of any one of claims 1 to 8 when the computer executable instructions are executed by the processor.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions, and the computer executable instructions are executed by the processor to implement the steps in the method of any one of claims 1 to 8.

Images