Single-cell drug causal gene discovery method and system based on adversarial transfer learning
By constructing an adversarial transfer learning model, the drug sensitivity tags in cell line data are migrated to single-cell data, and combined with potential result filling and causal analysis, the causal genes of single-cell drugs are identified, which solves the problem of insufficient drug sensitivity prediction data at the single-cell level, improves prediction accuracy and has an in-depth understanding of the mechanism of drug action.
Patent Information
- Application Number
- CN202510045085.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-13
- Publication Date
- 2025-04-29
- Estimated Expiration
- 2045-01-13
AI Technical Summary
The prior art lacks drug sensitivity labeling data when predicting drug sensitivity at the single-cell level, and deep learning models are not accurate enough when predicting drug sensitivity in high mutation-rate diseases, making it difficult to effectively analyze complex heterogeneity.
Adopting an adversarial transfer learning method, by constructing an adversarial transfer model including source classification model and target domain discriminator, using drug sensitivity tags in cell line data, knowledge is transferred to single-cell data with the aid of adversarial learning mechanism, and combining potential result filling and causal analysis to identify the causal genes of single-cell drugs.
It effectively solves the problem of lack of drug sensitivity labeling data in single-cell analysis, improves the accuracy of single-cell drug sensitivity prediction, provides important clues for a deep understanding of the mechanism of drug action, and can better understand the mechanism of drug resistance in cancer cells.
Smart Images
Figure CN119446281B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of drug sensitivity prediction, and in particular to a method and system for discovering single-cell drug causal genes based on adversarial transfer learning. Background Art
[0002] Drug resistance and the resulting ineffectiveness of drug treatments account for up to 90% of cancer-related deaths. In addition to some tumor subpopulations that are inherently resistant to selected therapies, cancer cells acquire drug resistance in a variety of ways. Some tumor subpopulations themselves are inherently resistant to selected therapies, while other cancer cells can escape the killing effects of drugs through complex mechanisms such as drug activation, target alternation, and drug efflux. Therefore, a deep understanding of the heterogeneity of cancer cell subpopulations is crucial to improving the effectiveness of cancer treatment, and the prediction of drug sensitivity of independent cancer subpopulations and the identification of causal genes for cancer drug sensitivity at the single-cell level have become key research directions in this field.
[0003] The emergence of single-cell RNA sequencing (scRNA-seq) technology has brought new opportunities for cancer research. It can analyze transcriptome profiles at the level of individual cells, thereby more accurately studying the heterogeneity of cancer. Through single-cell RNA sequencing, researchers can obtain gene expression information for each cell, gain an in-depth understanding of the differences between cancer cell subpopulations, and then infer the source of drug resistance. However, due to the relative complexity of single-cell level research, it is difficult to obtain accurate and large-scale drug sensitivity annotation data, which makes drug sensitivity prediction and related research at the single-cell level lack of sufficient data support, which seriously limits the development and in-depth research in this field.
[0004] Deep learning (DL) has made some progress in predicting the sensitivity of cells to drug treatment regimens with its powerful data analysis capabilities. Despite this, deep learning models still have many shortcomings in predicting drug sensitivity. First, the decision-making process of deep learning models is like a "black box" and lacks transparency. The mechanism behind its prediction is difficult to explain and often needs to be verified through in vitro and in vivo experiments, which is not only time-consuming and labor-intensive, but also increases research costs. Secondly, when predicting drug sensitivity of diseases with high mutation rates (such as HIV and cancer), the lack of sufficient relevant data to train the model makes it difficult for the model to form generalized reasoning capabilities. For these highly heterogeneous diseases, the diversity of disease states greatly reduces the accuracy of DL models. In addition, many current deep learning models rely on batch sequencing data for training. The limitation of cell resolution makes it difficult to effectively analyze complex heterogeneity and accurately capture biological characteristics and changes at the single-cell level, thus affecting the accurate prediction of drug sensitivity. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides a method for discovering single-cell drug causal genes based on adversarial transfer learning. Through supervised learning, adversarial learning, potential result filling and causal analysis, the present invention solves the key problem of lack of drug sensitivity annotation data in single-cell analysis on the one hand, and successfully discovers causal genes on the other hand, providing important clues for in-depth understanding of the drug action mechanism.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] In a first aspect, the present invention provides a method for discovering single-cell drug causal genes based on adversarial transfer learning, comprising:
[0008] Obtain source domain datasets, including cell line gene expression data and drug sensitivity labels, and obtain target domain datasets, including single-cell gene expression data;
[0009] Constructing an adversarial migration model including a source classification model and a target domain discriminator, wherein the source classification model outputs a drug sensitivity prediction result of a cell line according to an input source domain dataset, and the target domain discriminator is used to judge the domain category; training and optimizing the adversarial migration model according to the source domain dataset and the target domain dataset, and obtaining a single cell drug sensitivity prediction result according to the trained and optimized adversarial migration model;
[0010] The k-nearest-neighbor matching algorithm is used to find matching cells in different single-cell drug sensitivity prediction results, and the potential results are estimated by Poisson regression to obtain the counterfactual value of the single-cell gene expression data; based on the single-cell gene expression data and the counterfactual value of the single-cell gene expression data, the variational Bayesian inference is used to alternately update the posterior distribution of the model parameters until convergence to obtain the confounding effect; the confounding effect is fixed as its posterior mean, and the differential effect is calculated; the causal effect is calculated based on the confounding effect and the differential effect, and the causal gene of the single-cell drug is screened out.
[0011] As a further technical solution, the source classification model is optimized by minimizing the classification loss, and the source classification model is constrained by a hierarchical equation. The formula for minimizing the classification loss is:
[0012] = - ;in, For cell line gene expression data, is the drug sensitivity label, K is the number of categories, Gene expression data for cell lines The mapping of
[0013] As a further technical solution, the hierarchical equality constraint is: , ) ;in, For mapping of single-cell gene expression data, is the constraint function, For hierarchical indexing, , Gene expression data for cell lines in Mapping in layers, For single-cell gene expression data The mapping in the layer, and = .
[0014] As a further technical solution, the adversarial transfer model is trained according to the source domain dataset and the target domain dataset, specifically including: training and optimizing the target domain discriminator according to the supervised adversarial loss function, and training and optimizing the mapping of the single-cell gene expression data according to the target mapping loss function, wherein the supervised adversarial loss function = ), where D is the target domain discriminator, For single-cell gene expression data, and represent the expectations of cell line gene expression data and single cell gene expression data, respectively. ) is the maximum mean difference; target mapping loss function ; .
[0015] As a further technical solution, the specific method for obtaining the counterfactual value of the single-cell gene expression data is: ;in, To obtain single-cell drug sensitivity prediction results, i.e. drug sensitivity labels, against migration models, Single cell gene expression data with drug sensitivity labels, , is the regression coefficient, is the error term, represents a feature matrix, ; recalculate Obtain counterfactual values for single-cell gene expression data.
[0016] As a further technical solution, the specific method of obtaining the confounding effect and the difference effect is: first construct a first likelihood function: ;in, is a confounding effect; using variational Bayesian inference to alternately update the posterior distribution of model parameters until convergence, we get: ; ; Fixed confounding effects Its posterior mean , by constructing the second likelihood function: ,in is the difference effect, and then calculate the expectation of the difference effect: , and finally through Get the difference effect.
[0017] As a further technical solution, the calculation of the causal effect is to calculate the average causal effect for each gene: ; Among them, for each gene, , represents the average result of highly expressed genes, Represents the average result for genes with low expression.
[0018] In a second aspect, the present invention provides a single-cell drug causal gene discovery system based on adversarial transfer learning, comprising:
[0019] The data acquisition module is configured to: acquire a source domain dataset including cell line gene expression data and drug sensitivity labels, and acquire a target domain dataset including single cell gene expression data;
[0020] The adversarial migration module is configured to: construct an adversarial migration model including a source classification model and a target domain discriminator, wherein the source classification model outputs a drug sensitivity prediction result of a cell line according to an input source domain data set, and the target domain discriminator is used to judge a domain category; train and optimize the adversarial migration model according to the source domain data set and the target domain data set, and obtain a single cell drug sensitivity prediction result according to the trained and optimized adversarial migration model;
[0021] The causal gene discovery module is configured as follows: using the k-nearest neighbor matching algorithm to find matching cells in different single-cell drug sensitivity prediction results, estimating potential results through Poisson regression, and obtaining counterfactual values of single-cell gene expression data; using variational Bayesian inference to alternately update the posterior distribution of model parameters according to the single-cell gene expression data and the counterfactual values of the single-cell gene expression data until convergence, to obtain the confounding effect; fixing the confounding effect as its posterior mean, and calculating the differential effect; calculating the causal effect based on the confounding effect and the differential effect, and screening out the causal genes of single-cell drugs.
[0022] One or more technical solutions of the present invention have the following beneficial effects:
[0023] (1) This paper constructs an adversarial transfer model including a source classification model and a target domain discriminator, cleverly utilizes the existing drug sensitivity labels in the cell line data, and uses the adversarial learning mechanism to successfully transfer the knowledge in the cell line to the single-cell data, effectively solving the key problem of the lack of drug sensitivity annotation data in single-cell analysis. This innovative strategy makes it possible to predict drug sensitivity at the single-cell level, and the prediction results have a high accuracy, providing favorable data support for the discovery of causal genes.
[0024] (2) The present invention combines the single-cell drug sensitivity signature predicted by the anti-migration model with the counterfactual framework, calculates the confounding effect and differential effect, analyzes and identifies and controls the confounding factors, and calculates the difference in the results of sensitive and resistant cells after controlling the confounding variables. This differential effect can be regarded as an estimate of the causal effect of the gene on drug sensitivity, which represents the net effect of gene expression on drug sensitivity after controlling all known confounding factors. Based on this, researchers can more clearly understand how cancer cells develop drug resistance, conduct precise intervention on the drug resistance mechanism, and overcome the drug resistance of cancer cells. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0026] Figure 1 is a flow chart of the method of the present invention;
[0027] Figure 2 It is the framework diagram of the anti-migration of the present invention;
[0028] Figure 3 This is the causal reasoning framework diagram of the present invention. DETAILED DESCRIPTION
[0029] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.
[0030] Embodiment 1
[0031] This embodiment provides a single-cell drug causal gene discovery method based on adversarial transfer learning, which is generally divided into four parts: (1) supervised learning, establishing a model for predicting response label classification at the source domain batch level (cell line); (2) adversarial learning, the discriminator determines whether the sample is the source domain or the target domain, and the target mapping makes the mapping space of the target domain as close as possible to the mapping space of the source domain. The target mapping and the classifier are transferred to the single-cell level for label prediction to predict the drug response at the single-cell level; (3) potential result filling: based on counterfactuals, the k-nearest neighbor (k-NN) matching algorithm is used to find matching cells under different results, and the potential results are estimated by Poisson regression; (4) causal analysis: based on the obtained potential results, the confounding effects are identified, the confounding bias is eliminated, the single-cell data without confounding bias is obtained, and the causal effect is calculated to obtain the causal gene of the drug.
[0032] The specific steps are as follows:
[0033] S1: Obtain source domain datasets, including cell line gene expression data and drug sensitivity labels, and obtain target domain datasets, including single-cell gene expression data;
[0034] S2: constructing an adversarial migration model including a source classification model and a target domain discriminator, wherein the source classification model outputs a drug sensitivity prediction result of a cell line according to an input source domain dataset, and the target domain discriminator is used to judge the domain category; training and optimizing the adversarial migration model according to the source domain dataset and the target domain dataset, and obtaining a single cell drug sensitivity prediction result according to the trained and optimized adversarial migration model;
[0035] S3: Use the k-nearest neighbor matching algorithm to find matching cells in different single-cell drug sensitivity prediction results, estimate the potential results through Poisson regression, and obtain the counterfactual value of the single-cell gene expression data; according to the single-cell gene expression data and the counterfactual value of the single-cell gene expression data, use variational Bayesian inference to alternately update the posterior distribution of the model parameters until convergence to obtain the confounding effect; fix the confounding effect as its posterior mean and calculate the differential effect; calculate the causal effect based on the confounding effect and the differential effect, and screen out the causal gene of the single-cell drug.
[0036] In step S1, the source domain is distributed Extracting cell line gene expression data and drug sensitivity labels , from the target distribution Extracting single-cell gene expression data .
[0037] In step S2, a source classification model is first constructed , source classification model It is composed of a four-layer neural network ((500, 512), (512, 256), (256, 128), (128, 1)), drop_prob = 0.1, the hidden layer uses the Leakyrelu activation function, the output layer uses the Sigmod activation function, source classification model can be directly applied to the target representation, so there is no need to build a target classification model, so set = = .
[0038] After building the above source classification model Finally, the optimization is performed by minimizing the classification loss, and the source classification model is constrained by the hierarchical equality. Specifically, the formula for minimizing the classification loss is: = - ;in, For cell line gene expression data, is the drug sensitivity label, K is the number of categories, in this embodiment, K=2, there are only two cases: sensitive and insensitive. Gene expression data for cell lines For the mapping between source domain dataset and target domain dataset, a hierarchical representation is considered, where each layer parameter is represented as or , for a given set of equivalent layers Then, the constraint space can be described by hierarchical equality constraints as: , ) ;in, For mapping of single-cell gene expression data, is the constraint function, For hierarchical indexing, , Gene expression data for cell lines in Mapping in layers, For single-cell gene expression data Mapping in layers, each individual layer can be constrained independently, the constraint form in this example is that the source and target layers are equal: = ).
[0039] Secondly, the target domain discriminator D is constructed. The target domain discriminator D is composed of a four-layer neural network ((500, 512), (512, 256), (256, 128), (128, 1)), drop_prob=0.2, the hidden layer uses the Leakyrelu activation function, and the output layer uses the Sigmod activation function. The target domain discriminator D discriminates whether the data point is obtained from the source domain dataset or the target domain dataset, and the target mapping makes the target domain dataset as close to the source domain dataset mapping as possible, and the two are updated through adversarial training. The target domain discriminator is optimized according to the supervised adversarial loss function, and the mapping of single-cell gene expression data is optimized according to the target mapping loss function. The supervised adversarial loss function is defined as:
[0040] = ), where D is the target domain discriminator, For single-cell gene expression data, and represent the expectations of cell line gene expression data and single cell gene expression data, respectively. ) is the maximum mean difference.
[0041] The target mapping loss function is defined as:
[0042] ；
[0043] ；
[0044] MMD (Max mean discrepancy) is mainly used to measure the distance between two different but related distributions. The distance between two distributions is defined as:
[0045] ;in, represents the source domain data vector, Represents the target domain data vector, and H indicates that this distance is measured by φ(.) mapping the data into the Reproducing Kernel Hilbert Space (RKHS).
[0046] Mapping of single-cell gene expression data The feature space of single-cell data is made as similar as possible to the feature space of cell lines, while the target domain discriminator identifies whether the feature space of the sample comes from the source domain dataset or the target domain dataset as much as possible. In this way, the two are trained adversarially. Finally, the target domain discriminator cannot identify whether the sample comes from the source domain dataset or the target domain dataset, indicating that the mapping learning of the target domain is successful and the source domain dataset and the target domain dataset can be confused. Therefore, the source classification model can be applied to the target domain to predict the drug sensitivity labels of single cells. By optimizing the adversarial migration model according to the above method, the drug sensitivity prediction results of single cells can be accurately obtained.
[0047] In this example, a causal gene is defined as a gene that affects or is affected by the drug sensitivity status independently of other confounding variables.
[0048] In this example, we consider that single-cell gene expression data may be confounded by unknown biological and technical covariates that may lead to confounding of their drug sensitivity label assignment. These confounders confuse actual drug sensitivity-specific effects with other effects of unknown covariates, leading to false discoveries and weakening the statistical power of differential expression analysis. Rubin's potential outcome framework attempts to distinguish the effects that actually lead to drug sensitivity or resistance from other effects by asking the following counterfactual question:
[0049] In a binary control study, each cell observed is either sensitive or resistant to the drug, and the other value is not observed (indicated by the "?" symbol). For cells If the drug sensitivity label assignment variable is used, the potential expression of the gene cannot be directly observed in the data:
[0050] ;in is the single-cell gene expression data, and 0 or 1 represents the single-cell drug sensitivity prediction result obtained by the anti-migration model.
[0051] In this example, a variable is defined that can affect the drug sensitivity label ( ) and gene expression ( ), if it is associated with drug sensitivity and gene expression variables ( ), which remained associated with expression even after drug sensitivity conditioning ( Unless this example adjusts for an adequate set of confounding variables, the observed gene expression changes are not necessarily causal effects of the drug sensitivity mechanism. In order to make causal inferences, it is crucial to state the contingency hypothesis:
[0052] 1. (Consistency hypothesis) Stable cell drug sensitivity effect: Assuming the potential expression of cell j, that is, and , which is not affected by the expression of other cells .
[0053] 2. Conditional independence assumption: For non-causal genes, by definition, gene expression is independent of drug sensitivity. Therefore, this example does not need to make assumptions about this issue. However, for causal genes, this example assumes that a cell's potential (counterfactual) causal gene expression is independent of its drug sensitivity signature when sufficient confounding variables are controlled.
[0054] 3. Overlap of confounding effects between sensitive and resistant cells: It is assumed that sensitive and resistant cells exist simultaneously in a homogeneous cell population stratified by confounding factors.
[0055] Attribute the missing parts of the potential results of the single-cell profiles, propagate the attribution results to the cellular gene expression levels, and decompose them into confounding effects and differential effects using a single-cell gene expression matrix , this example aims to estimate: (1) confounding factors, , (2) Difference effect, In other words, this embodiment wants to estimate the decomposition In brief, the algorithm steps are as follows: (1) Counterfactual data on single cell expression are estimated (or imputed) by matching cells under a particular condition with neighboring cells under the opposite condition. (2) By pairing the observed and counterfactual single-cell data, the gene expression shared between the two opposite conditions is estimated in a Bayesian posterior inference, treating them as putative confounding effects. (3) The conditional (or residual) average causal effect on the observed cell is measured while holding the estimated confounding effect fixed.
[0056] In step S3, it is assumed that the single-cell gene expression data conforms to the Poisson distribution, that is, ;in, are the parameters of the Poisson distribution, Indicates that each gene in the cell expression in single-cell gene expression data, Indicates gene The expression level in cells j, Represents cells The sequencing depth is assumed to be and It follows the gamma distribution, specifically .
[0057] In step S3, the k-nearest neighbor (k-NN) matching algorithm is used to find matching cells under different single-cell drug sensitivity prediction results, and the potential results are estimated by Poisson regression to obtain the counterfactual value of the single-cell gene expression data. , specifically: ;in, To obtain single-cell drug sensitivity prediction results, i.e. drug sensitivity labels, against migration models, Single cell gene expression data with drug sensitivity labels, , is the regression coefficient, is the error term, represents a feature matrix, ; recalculate Obtain counterfactual values for single-cell gene expression data.
[0058] After matching and regression, it is observed that and counterfactuals , one of which is structurally different from the other and has effects related to drug sensitivity. However, both can provide invariant information that implies potential confounding effects, using Indicates. The confounding effect is obtained The specific method is: first construct the first likelihood function: ; Use variational Bayesian inference to alternately update the posterior distribution of model parameters until convergence, and get:
[0059] ;
[0060] ;
[0061] Fixed confounding effects Its posterior mean At the same time, by constructing the second likelihood function: ,in is the difference effect, and then calculate the expectation of the difference effect: , and finally through The difference effect is obtained, where The prior for is the Gamma distribution, and Alternate updates until convergence.
[0062] By analyzing and identifying and controlling confounding factors, the difference in results between sensitive and resistant cells after controlling for confounding variables is calculated. This differential effect can be regarded as an estimate of the causal effect of the gene, which represents the net effect of the drug on gene expression after controlling for all known confounding factors. Use appropriate statistical methods to estimate causal effects, including traditional regression analysis, propensity score matching, or inverse probability weighting. Calculate the average causal effect for each gene:
[0063] ; where for each gene represents the average result of high expression of the gene, The average result of low expression of the gene is represented. Through the above detailed steps in this embodiment, the causal effect can be estimated more accurately, and the causal gene of the drug can be discovered and screened.
[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A single-cell drug causal gene discovery method based on adversarial transfer learning, characterized in that: include: Obtain source domain datasets, including cell line gene expression data and drug sensitivity labels, and obtain target domain datasets, including single-cell gene expression data; Constructing an adversarial migration model including a source classification model and a target domain discriminator, wherein the source classification model outputs a drug sensitivity prediction result of a cell line according to an input source domain dataset, and the target domain discriminator is used to judge the domain category; training and optimizing the adversarial migration model according to the source domain dataset and the target domain dataset, and obtaining a single cell drug sensitivity prediction result according to the trained and optimized adversarial migration model; The source classification model is optimized by minimizing the classification loss and constrained by the hierarchical equality. The formula for minimizing the classification loss is: = - ;in, For cell line gene expression data, is the drug sensitivity label, K is the number of categories, Gene expression data for cell lines The mapping of Use the k-nearest neighbor matching algorithm to find matching cells in different single-cell drug sensitivity prediction results, estimate potential results through Poisson regression, and obtain the counterfactual value of the single-cell gene expression data; use variational Bayesian inference to alternately update the posterior distribution of model parameters until convergence based on the single-cell gene expression data and the counterfactual value of the single-cell gene expression data to obtain the confounding effect; fix the confounding effect as its posterior mean and calculate the differential effect; calculate the causal effect based on the confounding effect and the differential effect, and screen out the causal gene of the single-cell drug; The specific method for obtaining the confounding effect and the difference effect is as follows: first construct the first likelihood function: ;in, is a confounding effect; using variational Bayesian inference to alternately update the posterior distribution of model parameters until convergence, we get: ; ; Fixed confounding effects Its posterior mean , by constructing the second likelihood function: ,in is the difference effect, and then calculate the expectation of the difference effect: , and finally through Get the difference effect.
2. The method for discovering single-cell drug causal genes based on adversarial transfer learning as claimed in claim 1, characterized in that: The hierarchical equality constraints are: , ) ;in, For mapping of single-cell gene expression data, is the constraint function, For hierarchical indexing, , Gene expression data for cell lines in Mapping in layers, For single-cell gene expression data The mapping in the layer, and = .
3. The method for discovering single-cell drug causal genes based on adversarial transfer learning as claimed in claim 1, characterized in that: The training of the adversarial transfer model according to the source domain dataset and the target domain dataset specifically includes: training and optimizing the target domain discriminator according to the supervised adversarial loss function, and training and optimizing the mapping of the single-cell gene expression data according to the target mapping loss function, wherein the supervised adversarial loss function = , where D is the target domain discriminator, For single-cell gene expression data, and represent the expectations of cell line gene expression data and single cell gene expression data, respectively. is the maximum mean difference; target mapping loss function ; .
4. The method for discovering single-cell drug causal genes based on adversarial transfer learning as claimed in claim 1, characterized in that: The specific method for obtaining the counterfactual value of single-cell gene expression data is: ;in, To obtain single-cell drug sensitivity prediction results, i.e. drug sensitivity labels, against migration models, Single cell gene expression data with drug sensitivity labels, , is the regression coefficient, is the error term, represents a feature matrix, ; recalculate Obtain counterfactual values for single-cell gene expression data.
5. The method for discovering single-cell drug causal genes based on adversarial transfer learning as claimed in claim 1, characterized in that: The calculation of causal effect is to calculate the average causal effect for each gene: ;in, represents the average result of highly expressed genes, Represents the average result for genes with low expression.
6. A single-cell drug causal gene discovery system based on adversarial transfer learning, characterized in that: include: The data acquisition module is configured to: acquire a source domain dataset including cell line gene expression data and drug sensitivity labels, and acquire a target domain dataset including single cell gene expression data; The adversarial migration module is configured to: construct an adversarial migration model including a source classification model and a target domain discriminator, wherein the source classification model outputs a drug sensitivity prediction result of a cell line according to an input source domain data set, and the target domain discriminator is used to judge a domain category; train and optimize the adversarial migration model according to the source domain data set and the target domain data set, and obtain a single cell drug sensitivity prediction result according to the trained and optimized adversarial migration model; The source classification model is optimized by minimizing the classification loss and constrained by the hierarchical equality. The formula for minimizing the classification loss is: = - ;in, For cell line gene expression data, is the drug sensitivity label, K is the number of categories, Gene expression data for cell lines The mapping of The causal gene discovery module is configured as follows: using the k-nearest neighbor matching algorithm to find matching cells in different single-cell drug sensitivity prediction results, estimating potential results through Poisson regression, and obtaining counterfactual values of single-cell gene expression data; using variational Bayesian inference to alternately update the posterior distribution of model parameters according to the single-cell gene expression data and the counterfactual values of the single-cell gene expression data until convergence, and obtaining the confounding effect; fixing the confounding effect as its posterior mean, and calculating the differential effect; calculating the causal effect based on the confounding effect and the differential effect, and screening out the causal genes of single-cell drugs; The specific method for obtaining the confounding effect and the difference effect is as follows: first construct the first likelihood function: ;in, is a confounding effect; using variational Bayesian inference to alternately update the posterior distribution of model parameters until convergence, we get: ; ; Fixed confounding effects Its posterior mean , by constructing the second likelihood function: ,in is the difference effect, and then calculate the expectation of the difference effect: , and finally through Get the difference effect.
7. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for discovering single-cell drug causal genes based on adversarial transfer learning as described in any one of claims 1 to 5 are implemented.
8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for discovering single-cell drug causal genes based on adversarial transfer learning as described in any one of claims 1 to 5 are implemented.