A reinforcement learning gated single-cell drug response prediction method

Abstract

The application discloses a reinforced learning gate single-cell drug reaction prediction method and relates to the technical field of biological information.The application is aimed at solving the problems of low prediction accuracy and poor real-time performance of the existing single-cell drug reaction prediction method.The application comprises the following steps: obtaining a standardized expression matrix of a single-cell cancer sample to be predicted;obtaining a standardized communication feature matrix of the single-cell cancer sample to be predicted by using the standardized expression matrix of the single-cell cancer sample to be predicted;obtaining a to-be-predicted feature matrix by using the standardized communication feature matrix of the single-cell cancer sample to be predicted;inputting the to-be-predicted feature matrix into a final drug sensitivity prediction model to obtain a drug reaction prediction probability;comparing the drug reaction prediction probability with a classification threshold value;if the drug reaction prediction probability is greater than or equal to the classification threshold value, it is indicated that the single-cell cancer sample to be predicted is sensitive to a specified drug;otherwise, it is indicated that the single-cell cancer sample to be predicted is resistant to the specified drug.The application is used for predicting single-cell drug reactions.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a reinforcement learning-gated single-cell drug response prediction method. Background Technology

[0002] Cancer is a highly heterogeneous and complex disease; the same anticancer drug can elicit significantly different treatment responses among different patients and different cell subpopulations within the same tumor. Intratumoral heterogeneity can lead to the survival, proliferation, and ultimately recurrence and metastasis of pre-existing drug-resistant cells or treatment-induced drug-resistant cells under drug stress. Therefore, predicting the sensitivity or resistance status of cancer cells to drugs at single-cell resolution is crucial for identifying potential drug-resistant cell subpopulations, understanding treatment failure mechanisms, and supporting precision medicine.

[0003] Traditional bulk RNA-seq can characterize the overall transcriptome state of a sample, but it reflects the average expression level of a mixed cell population and struggles to reveal differences between individual cells. Single-cell RNA sequencing (scRNA-seq) can analyze the transcriptome characteristics of individual cells, providing an important technological foundation for studying tumor heterogeneity, cell state transitions, and drug-resistant cell subpopulations. However, single-cell data tagged with drug responses are usually limited in quantity and cover a narrow range of cancer types and drugs, making it difficult to directly train stable single-cell drug response prediction models.

[0004] Currently, cancer drug response prediction often relies on large-scale drug screening and pharmacogenomics resources. The Genomics of Drug Sensitivity in Cancer (GDSC) database provides a wealth of bulk expression profiles and drug sensitivity information for cancer cell lines, serving as source domain data for learning drug response-related transcriptome patterns. Resources such as the Cancer Cell Line Encyclopedia (CCLE) and the Cancer Therapeutics Response Portal (CTRP) also provide multi-omics and drug screening data for cancer drug response modeling. In contrast, the target domain data available for single-cell drug response studies is relatively limited, mainly derived from public transcriptome databases such as the Gene Expression Omnibus (GEO). For example, GSE149215 in the GEO database contains scRNA-seq data of the PC9 cell line before and after Etoposide treatment, and GSE108383 contains scRNA-seq data of the A375 and 451Lu cell lines under PLX4720-related conditions.

[0005] Besides expression profiling data, ligand-receptor interaction databases in single-cell analysis are also important data sources for characterizing intercellular communication. ConnectomeDB is a ligand-receptor interaction database used to organize and characterize human cell communication relationships. It provides a large amount of ligand-receptor pair information with human biological background, which can be used to deduce potential cell communication patterns from transcriptome expression data. Based on the ligand-receptor prior relationships in ConnectomeDB, ligand-receptor pairs in which both the ligand and receptor genes exist in the same gene set of the source and target domains can be screened. Furthermore, biologically interpretable ligand-receptor communication features can be constructed, thereby helping to characterize the potential link between cell communication status and single-cell drug response heterogeneity.

[0006] The DoRothEA database provides prior regulatory relationships between transcription factors and their target genes. A transcription factor and its corresponding set of target genes are typically defined as a regulator. The DoRothEA regulator database integrates multiple sources of evidence, including literature, ChIP-seq peaks, transcription factor binding site motifs, and regulatory relationships inferred from gene expression. It categorizes the confidence level of the interaction between transcription factors and target genes into five levels, A to E, based on the degree of evidence support, where A represents the highest confidence level and E represents a relatively low confidence level. DoRothEA can be used for transcription factor activity analysis in bulk RNA-seq and scRNA-seq data. This invention utilizes the prior regulatory relationships of transcription factors provided by DoRothEA to estimate the transcriptional regulatory activities of ligand and receptor genes, thereby providing additional reliability constraints for ligand-receptor communication characteristics.

[0007] Early methods for predicting cancer drug response were largely based on traditional machine learning. Matrix factorization methods constructed potential response matrices using the similarity between drugs and cell lines; recommendation system methods transformed drug response prediction into a drug-cell line preference inference problem; support vector machine models learned drug sensitivity and resistance boundaries through maximum margin classification; and random forest models used multiple decision trees to fuse multi-omics features to improve prediction stability. These methods are simple to implement and offer some interpretability, but they struggle to fully represent the complex nonlinear patterns in high-dimensional transcriptome data.

[0008] Deep learning methods address this deficiency by enhancing the feature representation capabilities of drug response prediction. DeepCDR utilizes graph neural networks to model drug molecular structures and integrates cell line omics features; GraphDRP extracts drug graph representations through graph neural networks and combines them with cell transcriptome features to complete predictions; GraphCDR introduces graph modeling and contrastive learning mechanisms to enhance the representation of drug-cell relationships; CODE-AE employs a context-aware decontamination autoencoder to transfer drug response knowledge learned in cell lines to patients or target samples, mitigating the impact of differences in data distribution.

[0009] For the problem of single-cell drug response prediction from bulkRNA-seq to scRNA-seq, transfer learning and domain adaptation methods have gradually gained attention. scDEAL aligns bulk and single-cell representations using Maximum Mean Discrepancy (MMD) to predict single-cell drug responses; SCAD employs adversarial domain adaptation, enabling the feature extractor to learn domain-invariant representations that are difficult to distinguish between the source and target domains; SSDA4Drug utilizes semi-supervised few-shot transfer learning, improving target domain discrimination through a small number of target domain labels, conditional entropy optimization, and adversarial perturbations; while methods such as Beyondcell and scRank rely more on manually designed sensitivity scores or prior networks to rank cell populations for drug sensitivity.

[0010] While the aforementioned methods have advanced single-cell drug response prediction, several shortcomings remain. First, many models primarily align source and target domains at the gene expression distribution level, lacking explicit modeling of prior intercellular communication. This makes it difficult to reflect the potential link between ligand-receptor interactions and drug response heterogeneity, resulting in low accuracy in single-cell drug response prediction. Second, existing single-cell drug response prediction methods typically use gene expression features or their low-dimensional representations as primary inputs, lacking modular expression of receptor-level communication structures and their relationship with drug response. Consequently, it is difficult to identify key receptor communication modules related to drug sensitivity or resistance, leading to low accuracy in single-cell drug response prediction. Third, domain adaptation strength is often manually preset, making dynamic adjustment based on changes in source domain prediction performance and the degree of source-target domain alignment difficult, resulting in poor real-time performance of single-cell drug response prediction. Summary of the Invention

[0011] This invention addresses the problems of low prediction accuracy and poor real-time performance in existing single-cell drug response prediction methods by proposing a reinforcement learning-gated single-cell drug response prediction method.

[0012] A reinforcement learning-gated single-cell drug response prediction method includes:

[0013] S1. Obtain the standardized expression matrix of the single-cell cancer sample to be predicted;

[0014] S2. Obtain the standardized communication feature matrix of the single-cell cancer sample to be predicted using the standardized expression matrix of the single-cell cancer sample to be predicted.

[0015] S3. Obtain the feature matrix to be predicted using the standardized communication feature matrix of the single-cell cancer sample to be predicted;

[0016] S4. Input the feature matrix to be predicted into the final drug sensitivity prediction model. The drug response prediction probability is obtained and compared with the classification threshold. If the drug response prediction probability is greater than or equal to the classification threshold, it means that the single-cell cancer sample to be predicted is sensitive to the specified drug; otherwise, it means that the single-cell cancer sample to be predicted is resistant to the specified drug.

[0017] Furthermore, the acquisition of the standardized expression matrix of the single-cell cancer sample to be predicted in S1 specifically involves:

[0018] S101. Obtain single-cell RNA sequencing expression data of the single-cell cancer sample to be predicted, and obtain the common genes of the single-cell cancer sample to be predicted. The expression values ​​on the matrix are used to obtain the expression matrix to be predicted. ;

[0019] in, Indicates the first A common gene, , Indicates the number of common genes. This represents the number of single-cell cancer samples to be predicted. It is the set of real numbers;

[0020] S102. Standardize the expression matrix to be predicted to obtain the standardized expression matrix of the sample to be predicted. .

[0021] Further, in step S2, obtaining the standardized communication feature matrix of the single-cell cancer sample to be predicted using the standardized expression matrix of the single-cell cancer sample to be predicted specifically involves:

[0022] S201. Obtain the first expression matrix of the single-cell cancer sample to be predicted using the normalized expression matrix. The original communication scores of ligand-receptor pairs in a single-cell cancer sample to be predicted are as follows:

[0023]

[0024]

[0025] in, and They are the first Ligand genes in a single-cell cancer sample to be predicted and receptor genes The standardized expression value, It is the original communication score. It is the total number of ligand-receptor pairs. It is the first Ligand genes in ligand-receptor pairs It is the first Receptor genes in ligand-receptor pairs It is a set of ligand-receptor pairs;

[0026] S202. Calculate the standardized communication feature matrix of the sample to be predicted using the original communication score. :

[0027]

[0028] in, It is the first The first single-cell cancer sample to be predicted corresponds to the first... Standardized communication characteristics of ligand-receptor pairs It is the upper limit of communication feature truncation. This means truncating the value to a range. , It is the first Static reliability weights for each ligand-receptor pair It is the mean of the raw communication scores of the ligand-receptor pair. It is the standard deviation of the original communication score of the ligand-receptor pair.

[0029] Furthermore, in step S3, obtaining the feature matrix to be predicted using the standardized communication feature matrix of the single-cell cancer sample to be predicted specifically involves:

[0030] First, for any The gating communication characteristics are obtained by utilizing standardized communication features, specifically:

[0031]

[0032] in, It is the first One receptor central module, It is a communication characteristic after gating. It is the first The final gating weights of each receptor central module;

[0033] Then, the first Standardized expression vectors of a single-cell cancer sample to be predicted Communication feature vector after gating The vectors are concatenated to obtain the input vector to be predicted.

[0034]

[0035] in, This represents vector concatenation. It is the input vector to be predicted;

[0036] Finally, the feature matrix to be predicted is constructed using all the input vectors to be predicted:

[0037]

[0038] in, It is the feature matrix to be predicted. It is the first One feature vector to be predicted, It is the matrix transpose.

[0039] Furthermore, the final drug sensitivity prediction model is obtained in the following way:

[0040] Step 1: Based on the source domain bulk RNA-seq drug response data and the target domain scRNA-seq drug response data, obtain the source domain expression matrix, the target domain training expression matrix, and the target domain test expression matrix. Use the source domain expression matrix and the target domain training expression matrix to obtain standardized parameters, and then standardize the source domain expression matrix, the target domain training expression matrix, and the target domain test expression matrix based on these standardized parameters to obtain the source domain standardized expression matrix. Target domain training standardized representation matrix and target domain test normalized expression matrix Specifically:

[0041] Step 11: Obtain the source domain expression matrix and target domain expression matrix based on source domain bulk RNA-seq drug response data and target domain scRNA-seq drug response data, specifically as follows:

[0042] First, we acquired drug response data based on source domain bulk RNA-seq and target domain scRNA-seq, obtaining the common gene set of the source and target domains. ;

[0043] in, Indicates the first A common gene, , Indicates the number of common genes;

[0044] Then, the expression vectors of the source domain data and the target domain data on the common genes are retained, thus forming the source domain expression vectors. and target domain representation matrix ;

[0045] Steps 1 and 2: Divide the target domain representation matrix into a target domain training index set. and target domain test index set Based on the index set, the target domain training representation matrix is ​​obtained. and target domain test expression matrix Let the source domain representation matrix be... The source domain tag is denoted as ;

[0046] Step 13: Represent the matrix using the source domain and target domain training representation matrix Obtain for each common gene corresponding mean and standard deviation Using the mean and standard deviation The expression values ​​of the source domain expression matrix and the target domain expression matrix are standardized to obtain the source domain standardized expression matrix. Target domain training standardized representation matrix and target domain test normalized expression matrix :

[0047]

[0048] in, It is a data field. , It is the source domain. It is the target domain. Shared genes The corresponding mean expression value, Shared genes The corresponding standard deviation of the expression value, It is the sample index in the data domain. It is a data domain medium sample Chinese gene The corresponding expression value, yes The standardization results It is a constant;

[0049] Step 2: Obtain the ligand-receptor pair set Based on source domain normalized representation matrix Target domain training standardized representation matrix and target domain test normalized expression matrix The communication characteristics of each ligand-receptor pair are obtained, thus yielding the source domain communication characteristic matrix. Communication feature matrix with target domain Specifically:

[0050] Step 2: 1. Retrieve ligand-receptor pair information from the ConnectomeDB database, retaining those where both the ligand gene and receptor gene belong to the specified category. The ligand-receptor pairs are obtained to form a set of ligand-receptor pairs. ;

[0051] in, It is the first One ligand-receptor pair, , It is a ligand-receptor pair index. It is the total number of ligand-receptor pairs. It is a ligand gene. It is a receptor gene;

[0052] Step 22: Obtaining Genes Regulatory activity fraction Specifically:

[0053]

[0054] in, It's a gene. Regulatory activity fraction , yes The corresponding set of regulatory transcription factors, It's a gene. transcription factor genes, This indicates taking the absolute value. Genes in the source domain sample The standardized representation vector, Genes in the source domain sample The standardized representation vector, This represents the Pearson correlation coefficient;

[0055] Step 23: Obtain the regulatory activity fraction of each gene in the ligand-receptor pair obtained in Step 21, and calculate the static reliability weight of each ligand-receptor pair using the regulatory activity fraction of each gene in the ligand-receptor pair, specifically as follows:

[0056]

[0057] in, It is the first Static reliability weights for each ligand-receptor pair It is the slope parameter of the Sigmoid function. It is a ligand gene The regulatory activity fraction, It is a receptor gene The regulatory activity fraction;

[0058] Step 24: Obtain the raw communication score for each ligand-receptor pair, specifically:

[0059]

[0060] in, It is a sample Middle ligand gene The standardized expression value, It is a sample Middle receptor gene The standardized expression value, It is the raw communication score of the ligand-receptor pair. It takes the maximum value;

[0061] Step 25: Obtain the mean and standard deviation of the raw communication scores of ligand-receptor pairs in the source domain samples, specifically as follows:

[0062]

[0063]

[0064] in, It is the mean of the raw communication scores of ligand-receptor pairs in the source domain samples. It is the first The nth source domain sample The original communication score of each ligand-receptor pair. It is the standard deviation of the raw communication score of ligand-receptor pairs in the source domain sample;

[0065] Step 26: Using the mean and standard deviation of the original communication scores of ligand-receptor pairs in the source domain samples obtained in Step 25 and the static reliability weights of each ligand-receptor pair obtained in Step 23, obtain the source domain communication feature matrix and the target domain communication feature matrix, specifically:

[0066]

[0067] in, It is a sample The A standardized communication feature, It is the upper limit of communication feature truncation. This indicates that the value will be truncated to... interval;

[0068] Step 3: Based on ligand-receptor pair sets Obtain the connectivity of each receptor and construct a receptor central module based on the receptor connectivity.

[0069] Step 4: Initialize the iteration rounds , will strengthen learning state vector Initialize to ;

[0070] Step 5: Transfer the reinforcement learning state vector Input into a reinforcement learning network to obtain standardized actions ;

[0071] in, express The gating weights of each receptor central module, This represents the standardized action used to adjust the strength of the domain adaptation.

[0072] Step 6: Apply the standardized actions obtained in Step 5 to the source domain communication feature matrix obtained in Step 2. Communication feature matrix with target domain Gating is performed on the standardized communication features in the data, for any The gated communication feature matrix is ​​obtained as follows:

[0073]

[0074] in, It is the first Communication characteristics after gating in round iteration It is a standardized communication feature. It is the first The receptor central module in the first Gating weights in round iteration;

[0075] Step 7: Use the source domain to normalize the representation matrix Target domain training standardized representation matrix A total training set is constructed using the corresponding gated post-communication feature matrix. The drug sensitivity prediction model is then trained in the e-th round using the total training set to obtain the trained drug sensitivity prediction model.

[0076] Step 8: Determine if the current training round has reached the preset total training rounds. If it has, use the drug sensitivity prediction model trained in round e as the final drug sensitivity prediction model and save the final drug sensitivity prediction model. If the current gating weight vector is used as the final gating weight vector, then freeze the drug sensitivity prediction model trained in round e, and construct a reinforcement learning state vector based on the drug sensitivity prediction model trained in round e. ,make Then return to step five.

[0077] Furthermore, the ligand-receptor pair-based set in step three... Obtain the connectivity of each receptor, and construct the receptor central module based on the receptor connectivity, specifically as follows:

[0078] Step 31: Based on ligand-receptor pair sets The connectivity of each receptor is obtained as follows:

[0079]

[0080]

[0081]

[0082] in, It is a receptor connectivity, It is a receptor The corresponding ligand-receptor pair index set, Indicates the first One ligand-receptor pair, It is a set of ligand-receptor pairs. yes Any receptor in the middle;

[0083] Step 32, the receptor According to connectivity Sort from highest to lowest, and place the first... Each receptor is treated as a highly connected receptor, and a receptor central module is constructed for each highly connected receptor:

[0084]

[0085] in, Indicated by receptor The corresponding ligand-receptor pair forms the first One receptor central module, It is a high-connectivity receptor label. , It is the preset number of highly connected receptors. It is the first A highly connected receptor;

[0086] Step 33: Obtain the set of ligand-receptor pair indices corresponding to all receptor central modules:

[0087]

[0088] in, It is the set of ligand-receptor pair indices corresponding to all receptor central modules.

[0089] Furthermore, the reinforcement learning network in step five includes: an Actor policy module and a Critic module;

[0090] The Actor policy module is used to determine the reinforcement learning state vector. The output obtains standardized actions;

[0091] The Actor strategy module includes: an action mean generation unit, a strategy sampling unit, and an action mapping unit;

[0092] The action mean generation unit employs a multilayer perceptron (MLP) to generate values ​​based on the state vector. The mean of the output action distribution is as follows:

[0093]

[0094] in, It is the mean of the action distribution. It is the action mean generation unit. These are the network parameters of the action mean generation unit;

[0095] The strategy sampling unit is used to sample the original actions based on the mean of the action distribution using a Gaussian strategy. :

[0096]

[0097] in, It is the original action. It is the identity matrix. The standard deviation is the preset value for the Gaussian strategy.

[0098] The action mapping unit is used to perform Sigmoid mapping on the original action to obtain a standardized action. Specifically:

[0099]

[0100]

[0101] in, express The gating weights of each receptor central module, , It is the first The gating weights of each receptor central module, This represents the standardized action used to adjust the strength of the domain adaptation.

[0102] The Critic module employs a multilayer perceptron (MLP) to obtain the value of the reinforcement learning state vector, specifically:

[0103]

[0104] in, It is the Critic network mapping function. This is the value of reinforcement learning state vectors.

[0105] Furthermore, in step seven, the source domain normalized representation matrix is ​​used... Target domain training standardized representation matrix A total training set is constructed using the corresponding gated post-communication feature matrix. The e-th round of training is then used to train the drug sensitivity prediction model, resulting in a well-trained model. Specifically:

[0106] Step 71: Utilize the source domain normalized representation matrix Target domain training standardized representation matrix The total training set is constructed using the corresponding gated post-communication feature matrix, specifically as follows:

[0107] First, normalize the source domain representation matrix. Target domain training standardized representation matrix The first dataset is composed of standardized representation vectors. The standardized representation vectors in the first dataset are then concatenated with the corresponding gated post-communication feature vectors to obtain the training input vector, specifically:

[0108]

[0109] in, Training rounds The input vector, It is a standardized representation vector. It is the feature vector of post-gated communication;

[0110] Then, the source domain training input vector and its corresponding drug response label and the target domain training input vector are combined to form the total training set;

[0111] Step 72: Use the total training set to train the drug sensitivity prediction model for the e-th round of training to obtain the well-trained drug sensitivity prediction model for the e-th round.

[0112] The drug sensitivity prediction model includes: a dual-branch encoder and a classification predictor;

[0113] The dual-branch encoder includes: a gene expression coding branch, a communication feature coding branch, and a feature fusion module;

[0114] The gene expression coding branch is used to encode the standardized expression vector in the input vector to obtain the gene expression code:

[0115]

[0116] in, It is gene expression encoding, and These are gene expression coding branching parameters. Indicates batch normalization. This represents the activation function. This indicates a random deactivation operation. It is a standardized representation vector;

[0117] The communication feature encoding branch is used to encode the gated communication feature vector in the input vector to obtain the communication feature code, specifically:

[0118]

[0119] in, It is the feature vector of post-gated communication. It is a communication feature encoding. and Indicates the communication feature coding branch parameters;

[0120] The feature fusion module is used to combine gene expression coding and communication feature coding to obtain a fused feature representation, specifically:

[0121]

[0122] in, It is a fusion feature representation;

[0123] The classification predictor uses fused features to obtain the predicted probability of drug sensitivity, specifically:

[0124] =

[0125]

[0126] in, Indicates sample Predicted probability of sensitivity to a specified drug. It is a hidden representation. It is a predictor. These are predictor parameters. and Represents the latent space mapping parameters. This represents the last hidden representation of the predictor. and This represents the output layer parameters.

[0127] Furthermore, in step eight, it is determined whether the current training round has reached the preset total training rounds. If the total training rounds have been reached, the drug sensitivity prediction model trained in the e-th round is used as the final drug sensitivity prediction model, and the final drug sensitivity prediction model is saved. If the current gating weight vector is used as the final gating weight vector, then freeze the drug sensitivity prediction model trained in round e, and construct a reinforcement learning state vector based on the drug sensitivity prediction model trained in round e. ,make Then return to step five, which is as follows:

[0128] Step 81: Determine whether the current training round has reached the preset total training round. If the total training round has been reached, use the drug sensitivity prediction model trained in the e-th round as the final drug sensitivity prediction model and save the final drug sensitivity prediction model. Use the current gating weight vector as the final gating weight vector. Otherwise, freeze the drug sensitivity prediction model trained in the e-th round and execute step 82.

[0129] Step 82: Calculate the AUPR of the e-th round drug sensitivity prediction model on the source domain training samples using the source domain normalized expression matrix and corresponding drug sensitivity labels, and obtain the change in source domain AUPR using the source domain AUPR. Specifically:

[0130]

[0131] in, It is the AUPR of the drug sensitivity prediction model trained in the e-th round on the source domain training samples. It is the AUPR of the drug sensitivity prediction model trained in the (e-1)th round on the source domain training samples. It is the change in AUPR of the source domain;

[0132] Step 83: Extract target domain training samples from the overall training set, add perturbations to the target domain training samples, and input the target domain training samples and the perturbated target domain training samples into the drug sensitivity prediction model trained in the e-th round to obtain the original prediction probabilities. And the predicted probability after adding perturbations This achieves consistency in the target domain, specifically:

[0133]

[0134] in, It is the predicted probability after adding perturbations. It is target domain consistency. It is the total number of training samples in the target domain. These are the training sample labels for the target domain;

[0135] Step 84: Extract the fused feature representation obtained in the e-th training round, construct the DeepCORAL loss based on the fused feature representation, and use the DeepCORAL loss to obtain the domain alignment score. Specifically:

[0136]

[0137]

[0138]

[0139]

[0140] in, It is the domain alignment score. It is the first Deep Coral loss, Denotes the Frobenius norm. Indicates training rounds, Indicates the number of samples in the source domain. This represents the number of training samples in the target domain. Indicates the fusion feature dimension. It is the source domain fusion feature matrix centered according to the sample dimension. It is a target domain fusion feature matrix centered according to the sample dimensions. It is the source domain covariance matrix. It is the target domain covariance matrix. It is transpose;

[0141] Step 85: Combine the source domain AUPR, source domain AUPR change, domain alignment score, and target domain consistency obtained in round e to form the state vector for the next iteration. and order Return to step five.

[0142] Furthermore, the drug sensitivity prediction model employs source domain supervised classification loss. Domain Adaptation Loss The training, specifically:

[0143] Determine whether the domain adaptation update condition is met. If the current round does not meet the domain adaptation update condition, then use the source domain supervised classification loss. Update the parameters of the bi-branch encoder and the classification predictor; if the current round meets the domain adaptation update condition, then first apply the source domain supervised classification loss. Update the parameters of the dual-branch encoder and the classification predictor, then freeze the classification predictor parameters and apply the domain adaptation loss. Update the bi-branch encoder parameters and the domain adaptation loss. ;

[0144] The source domain has supervised classification loss. as follows:

[0145]

[0146] in, It is the number of samples in the source domain. It is the first The true drug response label of each source domain sample It is the first The drug sensitivity prediction model for the first round Drug sensitivity prediction probability for each source domain sample It is a supervised classification loss in the source domain;

[0147] The domain adaptation loss Specifically:

[0148]

[0149]

[0150] in, It is a domain-adaptive weight. It is a preset base domain adaptation weight. It is the minimum scaling factor. It is the maximum scaling factor;

[0151] Reinforcement learning network parameters use Actor loss and Critic loss Update, Actor loss and Critic loss Specifically:

[0152]

[0153]

[0154]

[0155]

[0156]

[0157]

[0158] in, It is a smoothed L1 loss function. It is a Critic loss. It is an Actor loss. It is the joint logarithmic probability of the original action. It is a component index. It is the original action The One portion, yes The One portion, , It is the advantage value of standardized movements. The value of reinforcement learning state vectors, Is to execute the first Rewards earned from standardized movements, Indicates the domain alignment score weights. Indicates the consistency weight of the target domain prediction. This represents the average gating strength penalty weight. It is the first The average gating strength of the wheel, It is the first Round Gating weights of each receptor central module, This represents the total number of receptor central modules.

[0159] The beneficial effects of this invention are as follows:

[0160] This invention transfers the rich drug response knowledge from bulk RNA-seq of source domain cancer cells to scRNA-seq data of the target domain. It can output the predicted probability of drug response for single-cell samples in the target domain at single-cell resolution, which helps analyze the differences in drug response among different cell subpopulations within the tumor and improves the accuracy of single-cell drug response prediction. This invention introduces ConnectomeDB ligand-receptor prior relationships and DoRothEA transcription factor regulation prior relationships, expanding ordinary gene expression features into communication features with cell communication implications and transcriptional regulation evidence. This improves the ability of the drug sensitivity prediction model to characterize drug response-related biological processes and enhances the interpretability, thereby improving the accuracy of single-cell drug response prediction. This invention uses a receptor-centric module construction strategy, designating ligand-receptor pairs corresponding to highly connected receptors as receptor-centric communication modules. It utilizes a reinforcement learning Actor-Critic mechanism to dynamically adjust the gating weights of different receptor modules, enabling the drug sensitivity prediction model to adaptively highlight key communication modules related to drug response based on the training state, reducing the redundancy caused by fixed communication feature inputs. This invention combines bi-branch feature encoding, Deep CORAL domain adaptation, and reinforcement learning gating control, enabling the drug sensitivity prediction model to simultaneously learn drug response discrimination information from labeled samples in the source domain and source-target domain shared representations. Simultaneously, the domain adaptation strength is dynamically adjusted by the Actor policy network based on source domain AUPR, source domain AUPR change, source-target domain alignment score, and target domain prediction consistency. This mitigates the data distribution differences between bulkRNA-seq and scRNA-seq, improves the real-time performance of single-cell drug response prediction, and enhances the stability of prediction results for unlabeled single-cell samples in the target domain. Furthermore, this invention does not use target domain test labels in model training, parameter updates, or threshold determination during training and testing. Instead, after training, the target domain training set threshold is determined based on the target domain training samples, and this threshold is used for binary classification of drug response prediction results for single-cell samples in the target domain. This improves the accuracy of single-cell drug response prediction and ensures the independence, correctness, and logical integrity of the transfer learning prediction process. Attached Figure Description

[0161] Figure 1 This is a flowchart of the present invention;

[0162] Figure 2 This is a comparison chart of the significance analysis of the method of this invention with other comparative methods on the Etoposide dataset;

[0163] Figure 3 This is a comparison chart of the significance analysis of the method of the present invention with other comparative methods on the PLX4720_A375 dataset;

[0164] Figure 4 This is a comparison chart of the significance analysis of the method of the present invention with other comparative methods on the PLX4720_451Lu dataset. Detailed Implementation

[0165] To address the problems of existing single-cell cancer drug response prediction methods, such as significant distribution differences between source domain bulk RNA-seq and target domain scRNA-seq, scarce single-cell drug response annotations, insufficient utilization of ligand-receptor cell communication priors, difficulty in dynamically adjusting the reliability of communication features, and fixed domain adaptation strength, this invention proposes a reinforcement learning-gated single-cell drug response prediction method. This invention utilizes source domain bulk RNA-seq drug response data to learn drug sensitivity discrimination knowledge, maintains single-cell resolution using target domain scRNA-seq data, constructs cell communication features through a ligand-receptor database, calculates communication reliability weights through transcription factor-regulated priors, achieves modular expression of ligand-receptor pairs through receptor central module construction, and further utilizes the Actor-Critic reinforcement learning mechanism to adaptively learn receptor module gating weights and domain adaptation loss weights, ultimately completing the target domain single-cell drug response classification prediction. The invention will be described below with specific implementation details.

[0166] Specific implementation method one: as follows Figure 1 As shown, this embodiment of a reinforcement learning-gated single-cell drug response prediction method includes the following steps:

[0167] S1. Obtain the standardized expression matrix of the single-cell cancer sample to be predicted, specifically:

[0168] S101. Obtain single-cell RNA sequencing expression data of the single-cell cancer sample to be predicted, according to the common gene set saved during the training phase. By analyzing the common gene sequence and preserving the expression values ​​of the single-cell cancer samples to be predicted on the common genes, the expression matrix to be predicted is obtained. ;

[0169] in, This represents the number of single-cell cancer samples to be predicted. It is the number of common genes. It is the first The first of the 10 samples of single-cell cancer to be predicted The expression values ​​of a common gene It is the set of real numbers;

[0170] S102, Read the data saved during the training phase. Mean expression value of common genes and standard deviation The expression matrix to be predicted is standardized to obtain the standardized expression matrix of the sample to be predicted:

[0171]

[0172]

[0173] in, It is the standardized representation matrix of the sample to be predicted. yes The standardization results It is a minimal constant that prevents the denominator from being zero.

[0174] S2. Obtain the communication feature matrix of the single-cell cancer sample to be predicted using the normalized expression matrix of the sample. Specifically:

[0175] S201. Read the ligand-receptor pair set saved during the training phase, and obtain the first expression matrix of the single-cell cancer sample to be predicted using the normalized expression matrix. The original communication scores of ligand-receptor pairs in a single-cell cancer sample to be predicted are as follows:

[0176] First, read the ligand-receptor pair set saved during the training phase:

[0177]

[0178] in, It is the total number of ligand-receptor pairs. It is the first Ligand genes in ligand-receptor pairs It is the first Receptor genes in ligand-receptor pairs;

[0179] Then, obtain the first The original communication scores of ligand-receptor pairs in a single-cell cancer sample to be predicted are as follows:

[0180]

[0181] in, and They are the first Ligand genes in a single-cell cancer sample to be predicted and receptor genes The standardized expression value, It is the original communication score;

[0182] S202, Read the data saved during the training phase. Static reliability weights of ligand-receptor pairs Mean of original communication score and the standard deviation of the original communication score The standardized communication feature matrix of the sample to be predicted is calculated using the original communication scores. :

[0183]

[0184] in, It is the first The first single-cell cancer sample to be predicted corresponds to the first... Standardized communication characteristics of ligand-receptor pairs It is the upper limit of communication feature truncation. This means truncating the value to a range. ;

[0185] S3. Obtain the feature matrix to be predicted using the final receptor central module gating weights and the communication feature matrix of the single-cell cancer sample to be predicted, specifically:

[0186] First, read the final receptor-centric module gating weight vector saved after training:

[0187]

[0188] in, It is the total number of receptor central modules. It is the first The final gating weights of each receptor central module;

[0189] For any ligand-receptor pair index selected into the receptor central module The communication characteristics obtained after gating are:

[0190]

[0191] in, It is the first One receptor central module, These are the communication characteristics after gating;

[0192] Then, the first Standardized expression vectors of a single-cell cancer sample to be predicted Communication feature vector after gating The vectors are concatenated to obtain the input vector to be predicted.

[0193]

[0194] in, This indicates vector concatenation;

[0195] Finally, the feature matrix to be predicted is constructed using all the input vectors to be predicted:

[0196]

[0197] in, It is the matrix transpose;

[0198] S4. Input the feature matrix to be predicted into the final drug sensitivity prediction model. The drug response prediction probability is obtained, and the drug response prediction result is obtained based on the drug response prediction probability and the classification threshold. Specifically:

[0199] S401. Read the final drug sensitivity prediction model saved after training. , will the 1 feature vector to be predicted Input the final drug sensitivity prediction model to obtain the predicted drug response probability:

[0200]

[0201] in, The parameters represent the final drug sensitivity prediction model. It is the first The probability of a single-cell cancer sample to be predicted being sensitive to a specified drug;

[0202] S402, Predicting the probability of drug response With classification threshold Compare and obtain the first Drug response prediction results for a single-cell cancer sample to be predicted:

[0203]

[0204] in, Indicates the first One single-cell cancer sample to be predicted is sensitive to the specified drug. Indicates the first One single-cell cancer sample to be predicted is resistant to a specified drug;

[0205] In this implementation, reinforcement learning iterations are no longer performed during the actual prediction process. It is not necessary to construct reinforcement learning state vectors, nor to update the parameters of the Actor network, Critic network, and the final drug sensitivity prediction model. Instead, it directly reads the final drug sensitivity prediction model, the final receptor central module gating weights, the common gene set, the standardized parameters, the ligand-receptor pair set, the communication feature parameters, and the classification threshold saved during the training phase to complete the drug response prediction for the single-cell cancer sample to be predicted.

[0206] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the final drug sensitivity prediction model is obtained through the following method:

[0207] Step 1: Based on the source domain bulk RNA-seq drug response data and the target domain scRNA-seq drug response data, obtain the source domain expression matrix, the target domain training expression matrix, and the target domain test expression matrix. Use the source domain expression matrix and the target domain training expression matrix to obtain standardized parameters, and then standardize the source domain expression matrix, the target domain training expression matrix, and the target domain test expression matrix based on these standardized parameters to obtain the source domain standardized expression matrix. Target domain training standardized representation matrix and target domain test normalized expression matrix Specifically:

[0208] Step 1: Obtain the source domain expression matrix and target domain expression matrix based on source domain bulk RNA-seq drug response data and target domain scRNA-seq drug response data. Specifically:

[0209] First, obtain drug response data based on source domain bulk RNA-seq. Drug response data with target domain scRNA-seq ;

[0210] in, It is the source domain sample label. It is the sample label of the target domain. It is the source domain sample set. It is the target domain sample set. It is the first The representation vector of each source domain sample. This is a source-domain drug response label; 1 indicates sensitivity, and 0 indicates resistance. It is the first The representation vector of a target domain sample. It is a target domain drug response label;

[0211] The source domain bulk RNA-seq drug response data used preprocessed bulk RNA-seq expression data from GDSC, and the drug response tags were derived from LOBICO binarized drug response tags. The target domain scRNA-seq drug response data used publicly available scRNA-seq datasets, where the Etoposide task corresponds to PC9 untreated samples and PC9 samples treated with 25 μM Etoposide for 3 days in the GSE149215 dataset, and the PLX4720_A375 and PLX4720_451Lu tasks correspond to A375 and 451Lu single-cell samples in the GSE108383 dataset.

[0212] Then, obtain the common gene set of the source and target domains. ;

[0213] in, Indicates the first A common gene, , Indicates the number of common genes;

[0214] Then, the expression vectors of the source domain data and the target domain data on the common genes are retained, thus forming the source domain expression vectors. and target domain representation matrix ;

[0215] in, It is the total number of data in the source domain. This represents the total number of data in the target domain.

[0216] Steps 1 and 2: Divide the target domain representation matrix into a target domain training index set. and target domain test index set Based on the index set, the target domain training representation matrix is ​​obtained. and target domain test expression matrix Let the source domain representation matrix be... The source domain tag is denoted as .

[0217] Step 13: Represent the matrix using the source domain and target domain training representation matrix Obtain for each common gene corresponding mean and standard deviation Using the mean and standard deviation The expression values ​​of the source domain expression matrix and the target domain expression matrix are standardized to obtain the source domain standardized expression matrix. Target domain training standardized representation matrix and target domain test normalized expression matrix ;

[0218] Using the mean and standard deviation Standardize each expression value in both the source domain and target domain sample sets, specifically as follows:

[0219]

[0220] in, It is a data field. , It is the source domain. It is the target domain. Shared genes The corresponding mean expression value, Shared genes The corresponding standard deviation of the expression value, It is the sample index in the data domain. It is a data domain medium sample Chinese gene The corresponding expression value, yes The standardization results It is a minimal constant that prevents the denominator from being zero;

[0221] The source domain standardized representation matrix and its corresponding label form the source domain dataset, the target domain training standardized representation matrix forms the target domain training set, and the target domain test standardized representation matrix and its corresponding label form the target domain test set. Each row of the source domain standardized representation matrix, the target domain training standardized representation matrix, and the target domain test standardized representation matrix corresponds to one sample.

[0222] Step 2: Obtain the ligand-receptor pair set Based on source domain normalized representation matrix Target domain training standardized representation matrix and target domain test normalized expression matrix The communication characteristics of each ligand-receptor pair are obtained, thus yielding the source domain communication characteristic matrix. Communication feature matrix with target domain Specifically:

[0223] Step 2: 1. Retrieve ligand-receptor pair information from the ConnectomeDB database, retaining those where both the ligand gene and receptor gene belong to the specified category. The ligand-receptor pairs are obtained to form a set of ligand-receptor pairs. ;

[0224] in, It is the first One ligand-receptor pair, , It is a ligand-receptor pair index. It is the total number of ligand-receptor pairs. It is a ligand gene. It is a receptor gene;

[0225] ConnectomeDB provides prior relationships of ligand-receptor interactions related to human cell communication, which can be used to construct biologically interpretable ligand-receptor communication features from transcriptome expression matrices;

[0226] Step 22: Obtain the regulatory prior relationships between transcription factors and target genes from the DoRothEA database, and identify common genes. The corresponding set of regulatory transcription factors Based on the set of regulatory transcription factors Obtaining genes Regulatory activity fraction Specifically:

[0227]

[0228] in, It's a gene. Regulatory activity fraction , yes The corresponding set of regulatory transcription factors, It's a gene. transcription factor genes, This indicates taking the absolute value. Genes in the source domain sample The standardized representation vector, Genes in the source domain sample The standardized representation vector, This represents the Pearson correlation coefficient;

[0229] Step 23: Obtain the regulatory activity fraction of each gene in the ligand-receptor pair obtained in Step 21, and calculate the static reliability weight of each ligand-receptor pair using the regulatory activity fraction of each gene in the ligand-receptor pair, specifically as follows:

[0230]

[0231]

[0232] in, It is the first Static reliability weights for each ligand-receptor pair It is the slope parameter of the Sigmoid function. It is a ligand gene The regulatory activity fraction, It is a receptor gene The regulatory activity fraction, It is the input variable of the Sigmoid function;

[0233] In this step, the reliability weight is used to characterize the degree to which the ligand-receptor pair is supported at the level of transcription factor regulation;

[0234] Step 24: Obtain the raw communication score for each ligand-receptor pair, specifically:

[0235]

[0236] in, It is a sample Middle ligand gene The standardized expression value, It is a sample Middle receptor gene The standardized expression value, It is the raw communication score of the ligand-receptor pair. It takes the maximum value;

[0237] This step describes the strength of cell communication by the co-expression activity of ligand and receptor genes, and reduces the impact of extreme expression products on subsequent model training by using square root operations.

[0238] Step 25: Obtain the mean and standard deviation of the raw communication scores of ligand-receptor pairs in the source domain samples, specifically as follows:

[0239]

[0240]

[0241] in, It is the mean of the raw communication scores of ligand-receptor pairs in the source domain samples. It is the first The nth source domain sample The original communication score of each ligand-receptor pair. It is the standard deviation of the raw communication score of ligand-receptor pairs in the source domain sample;

[0242] Step 26: Using the mean and standard deviation of the original communication scores of ligand-receptor pairs in the source domain samples obtained in Step 25 and the static reliability weights of each ligand-receptor pair obtained in Step 23, obtain the source domain communication feature matrix and the target domain communication feature matrix, specifically:

[0243] For any data field The first in The nth sample, the th Standardized communication features Represented as:

[0244]

[0245] in, It is a sample The A standardized communication feature, It is the upper limit of communication feature truncation. This indicates that the value will be truncated to... interval;

[0246] This yields the source domain communication characteristic matrix:

[0247]

[0248] And the target domain communication feature matrix:

[0249]

[0250] In this matrix, rows represent samples, and columns represent filtered ligand-receptor pair communication features. This communication feature matrix is ​​used for subsequent receptor-centric module construction and reinforcement learning gating processing.

[0251] Step 3: Based on ligand-receptor pair sets The connectivity of each receptor is obtained, receptor central modules are constructed based on the receptor connectivity, and receptor central modules are used to form a set of receptor central modules, thereby obtaining the set of ligand-receptor pair indices corresponding to all receptor central modules, specifically:

[0252] Step 31: Based on ligand-receptor pair sets The connectivity of each receptor is obtained as follows:

[0253] For ligand-receptor pair sets Any receptor Acquire receptors The corresponding ligand-receptor pair index set and ligand set They are defined as follows:

[0254]

[0255]

[0256] in, Indicates the first One ligand-receptor pair, It is a receptor The corresponding ligand-receptor pair index set, It is a receptor The corresponding set of ligands, It is a collection of receptors;

[0257] Using receptors The corresponding ligand-receptor pair index set Acquiring receptors connectivity :

[0258]

[0259] in, It is a receptor The degree of connectivity, representing the degree of connection with the receptor. The number of linked ligand-receptor pairs;

[0260] Step 32, the receptor According to connectivity Sort from highest to lowest, and place the first... Each receptor is considered a high-connectivity receptor, and a receptor central module is constructed for each high-connectivity receptor, specifically as follows:

[0261] First, the receptor According to connectivity Sort from highest to lowest, and place the first... Each receptor is a highly connected receptor, specifically:

[0262]

[0263] in, It is the preset number of highly connected receptors. , It is a high-connectivity receptor label. It is the first A highly connected receptor;

[0264] Then, for each high-connectivity receptor , construct the first One receptor center Specifically:

[0265]

[0266] in, Indicated by receptor The corresponding ligand-receptor pair forms the first One receptor central module; For ligand-receptor pair set The original index in the data does not change due to the screening of highly connected receptors;

[0267] Step 33: Assemble the receptor central modules corresponding to each high-connectivity receptor into a receptor central module set, and obtain the ligand-receptor pair index set corresponding to all receptor central modules, specifically as follows:

[0268] First, the receptor central modules corresponding to each high-connectivity receptor are grouped into a receptor central module set:

[0269]

[0270] Then, obtain the set of ligand-receptor pair indices corresponding to all receptor central modules:

[0271]

[0272]

[0273] in, It is the set of ligand-receptor pair indices selected into the receptor central module. It is the total number of ligand-receptor pairs contained in the entire receptor central module;

[0274] Step 4: Initialize the iteration rounds , will strengthen learning state vector Initialize to ;

[0275] in, It is the first Wheel Source Domain AUPR, It is the change in AUPR of the source domain. It is the iteration round. It is the domain alignment score;

[0276] Step 5: Transfer the reinforcement learning state vector Input into a reinforcement learning network to obtain standardized actions ;

[0277] The reinforcement learning network includes: an Actor policy module and a Critic module;

[0278] The Actor policy module is used to determine the reinforcement learning state vector. The output obtains standardized actions;

[0279] The Actor strategy module includes: an action mean generation unit, a strategy sampling unit, and an action mapping unit;

[0280] The action mean generation unit employs a multilayer perceptron (MLP) to generate values ​​based on the state vector. The mean of the output action distribution is as follows:

[0281]

[0282] in, It is the mean of the action distribution. It is the action mean generation unit. These are the network parameters of the action mean generation unit; the multilayer perceptron (MLP) consists of two hidden layers, each containing 64 neurons, and uses the ReLU activation function;

[0283] The strategy sampling unit is used to sample the original actions based on the mean of the action distribution using a Gaussian strategy. :

[0284]

[0285] in, It is the original action. It is the identity matrix. The standard deviation is the preset value for the Gaussian strategy.

[0286] The action mapping unit is used to perform Sigmoid mapping on the original action to obtain a standardized action. Specifically:

[0287]

[0288]

[0289] in, express The gating weights of each receptor central module, , It is the first The gating weights of each receptor central module, This refers to the standardized actions used to adjust the strength of the adaptive domain.

[0290] The Critic module employs a multilayer perceptron (MLP) to obtain the value of the reinforcement learning state vector, specifically:

[0291]

[0292] in, It is the Critic network mapping function. This is the value of reinforcement learning state vectors;

[0293] Step 6: Apply the standardized actions obtained in Step 5 to the source domain communication feature matrix obtained in Step 2. Communication feature matrix with target domain The standardized communication features in the matrix are gated to obtain the gated communication feature matrix:

[0294] The communication characteristics after gating are obtained in the following way:

[0295] For any ,have:

[0296]

[0297] in, It is the first Communication characteristics after gating in round iteration It is a standardized communication feature. It is the first The receptor central module in the first Gating weights in round iteration;

[0298] In this step, the first The source domain communication feature matrix after wheel gating is denoted as , will the The target domain communication feature matrix after wheel gating is denoted as .

[0299] Step 7: Use the source domain to normalize the representation matrix Target domain training standardized representation matrix Target domain test standardized representation matrix The gated communication feature matrix is ​​used to construct the total training set and the total test set. The drug sensitivity prediction model is then trained and tested in the e-th round using the total training set and the total test set to obtain a well-trained and tested drug sensitivity prediction model. Specifically:

[0300] Step 71: Utilize the source domain normalized representation matrix Target domain training standardized representation matrix The total training set is constructed using the corresponding gated post-communication feature matrix, specifically as follows:

[0301] First, normalize the source domain representation matrix. Target domain training standardized representation matrix The first dataset is composed of standardized representation vectors. The standardized representation vectors in the first dataset are then concatenated with the corresponding gated post-communication feature vectors to obtain the training input vector, specifically:

[0302]

[0303] in, Training rounds The input vector, It is a standardized representation vector. It is the feature vector of post-gated communication. Indicates the number of common genes. This indicates the number of ligand-receptor pairs selected for inclusion in the receptor central module.

[0304] Then, the input vectors and drug response labels are combined to form the total training set;

[0305] Finally, the target domain test normalization expression matrix is ​​used. The standardized expression vector and the corresponding gated post-communication feature vector are concatenated to obtain the test input vector. The test input vector and the drug response label are combined to form the total test set.

[0306] Step 72: Train and test the drug sensitivity prediction model using the total training set and the total test set for the e-th round of training and testing to obtain the well-trained and tested drug sensitivity prediction model for the e-th round.

[0307] The drug sensitivity prediction model includes: a dual-branch encoder and a classification predictor;

[0308] The dual-branch encoder includes: a gene expression coding branch, a communication feature coding branch, and a feature fusion module;

[0309] The gene expression coding branch is used to encode the standardized expression vector in the input vector to obtain the gene expression code, specifically:

[0310]

[0311] in, It is gene expression encoding, and These are gene expression coding branching parameters. Indicates batch normalization. This represents the activation function. This indicates a random deactivation operation. It is a standardized representation vector;

[0312] The communication feature encoding branch is used to encode the gated communication feature vector in the input vector to obtain the communication feature code, specifically:

[0313]

[0314] in, It is the feature vector of post-gated communication. It is a communication feature encoding. and Indicates the communication feature coding branch parameters;

[0315] The feature fusion module is used to combine gene expression coding and communication feature coding to obtain a fused feature representation, specifically:

[0316]

[0317] in, It is a fusion feature representation;

[0318] The classification predictor is used to map the fused feature representation to the latent space to obtain the hidden representation, and to obtain the predicted probability of drug sensitivity based on the hidden representation, specifically:

[0319] =

[0320]

[0321] in, Indicates sample Predicted probability of sensitivity to a specified drug. It is a hidden representation. It is a predictor. These are predictor parameters. and Represents the latent space mapping parameters. This represents the last hidden representation of the predictor. and Indicates the output layer parameters;

[0322] The predictor in this step is a multilayer perceptron, and the terminal uses the sigmoid activation function.

[0323] Step 8: Determine if the current training round has reached the preset total training rounds. If it has, use the drug sensitivity prediction model trained and tested in round e as the final drug sensitivity prediction model and save the final drug sensitivity prediction model. If the current gating weight vector is used as the final gating weight vector, then freeze the drug sensitivity prediction model trained and tested in the e-th round, and construct a reinforcement learning state vector based on the drug sensitivity prediction model trained and tested in the e-th round. ,make Then return to step five, which is as follows:

[0324] Step 81: Determine whether the current training round has reached the preset total training round. If the total training round has been reached, use the drug sensitivity prediction model trained and tested in the e-th round as the final drug sensitivity prediction model and save the final drug sensitivity prediction model. Use the current gating weight vector as the final gating weight vector. Otherwise, freeze the drug sensitivity prediction model trained and tested in the e-th round and proceed to step 82.

[0325] Step 82: Calculate the AUPR of the model in the e-th round on the source domain training samples using the source domain normalized expression matrix and corresponding drug sensitivity labels, and then use the source domain AUPR to obtain the change in source domain AUPR. Specifically:

[0326]

[0327] in, It is the AUPR of the drug sensitivity prediction model trained in round e. It is the AUPR of the drug sensitivity prediction model trained in round e-1. It is the change in AUPR of the source domain;

[0328] Step 83: Extract target domain training samples from the overall training set, add perturbations to the target domain training samples, and input the target domain training samples and the perturbated target domain training samples into the drug sensitivity prediction model trained in the e-th round to obtain the original prediction probabilities. And the predicted probability after adding perturbations This achieves consistency in the target domain, specifically:

[0329]

[0330] in, It is the predicted probability after adding perturbations. It is target domain consistency. It is the total number of training samples in the target domain. These are the training sample labels for the target domain.

[0331] In this step, the target domain training sample is the concatenation result of the target domain training standardized expression vector and the corresponding gated communication feature vector in the total training set; in order to enable the reinforcement learning state to reflect the stability of the prediction results of the unlabeled samples in the target domain, a Gaussian perturbation is added to the input of the target domain training sample.

[0332] Step 84: Extract the fused feature representation obtained in the e-th training round, construct the DeepCORAL loss based on the fused feature representation, and use the DeepCORAL loss to obtain the domain alignment score. Specifically:

[0333] Step 841: Obtain the source domain fusion feature matrix based on the fusion feature representation obtained in the e-th training round. Feature matrix fused with target domain The source domain covariance matrix and target domain covariance matrix are obtained using the source domain fusion feature matrix and the target domain fusion feature matrix, specifically as follows:

[0334]

[0335]

[0336] in, Indicates training rounds, Indicates the number of samples in the source domain. This represents the number of training samples in the target domain. Indicates the fusion feature dimension. It is the source domain fusion feature matrix centered according to the sample dimension. It is a target domain fusion feature matrix centered according to the sample dimensions. It is the source domain covariance matrix, representing the first... Second-order statistical information of wheel-source domain fusion characteristics. It is the target domain covariance matrix, representing the first... The second-order statistical information of the fused features in the target domain training round. It is the transpose symbol;

[0337] Source domain fusion features and target domain training fusion features are used to measure the distributional differences between the two data domains in the model representation space.

[0338] Step 842: Construct the Deep CORAL loss using the source domain covariance matrix and the target domain covariance matrix, specifically as follows:

[0339]

[0340] in, Indicates the first Deep Coral loss, This represents the Frobenius norm.

[0341] In this step, Deep CORAL loss is used to align the second-order statistical distribution of the source and target domain training samples in the fusion feature space, thereby reducing the distribution difference between the source domain bulk RNA-seq representation and the target domain scRNA-seq representation.

[0342] Step 843: Obtain domain alignment scores using Deep CORAL loss. :

[0343]

[0344] in, It is the domain alignment score; The larger the value, the closer the feature distributions of the source and target domains are;

[0345] Step 85: Combine the source domain AUPR, source domain AUPR change, domain alignment score, and target domain consistency obtained in round e to form the state vector for the next iteration. and order Return to step five.

[0346] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the drug sensitivity prediction model uses source domain supervised classification loss. Domain Adaptation Loss The training, specifically:

[0347] Determine whether the domain adaptation update condition is met. If the current round does not meet the domain adaptation update condition, then use the source domain supervised classification loss. Update the parameters of the dual-branch encoder and the classification predictor;

[0348] The domain adaptation update condition is as follows: ;

[0349] in, It is the domain adaptation update interval;

[0350] The source domain has supervised classification loss. as follows:

[0351]

[0352] in, It is the number of samples in the source domain. It is the first The true drug response label of each source domain sample It is the first The drug sensitivity prediction model for the first round Drug sensitivity prediction probability for each source domain sample It is a supervised classification loss in the source domain;

[0353] If the current round meets the domain adaptation update condition, then the source domain supervised classification loss is used first. Update the parameters of the dual-branch encoder and the classification predictor, then freeze the classification predictor parameters and apply the domain adaptation loss. Update the bi-branch encoder parameters and the domain adaptation loss. Specifically:

[0354]

[0355]

[0356] in, These are domain-adaptive weights, used to adjust the strength of the Deep CORAL loss's influence throughout the training process. It is a preset base domain adaptation weight. It is the minimum scaling factor. It is the maximum scaling factor.

[0357] This enables the Actor to dynamically adjust the domain adaptation strength between the source and target domains within a preset range.

[0358] Other steps or parameters are the same as in Specific Implementation Method 1.

[0359] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the reinforcement learning network parameters use Actor loss. and Critic loss Update, Actor loss and Critic loss Obtained through the following methods:

[0360] A1. Obtain the average gating strength using the gating weights of the receptor central module, and obtain the first... Round of rewards Specifically:

[0361] First, the average gating strength is obtained using the gating weights of the receptor central module, specifically:

[0362]

[0363] in, It is the first The average gating strength of the wheel, It is the first Round Gating weights of each receptor central module, This represents the total number of receptor central modules;

[0364] Then, using the average gating strength, the first The change in AUPR in the source domain is Domain alignment score Consistency with target domain prediction Obtain the Round of rewards :

[0365]

[0366] in, Is to execute the first Rewards earned from standardized movements, Indicates the domain alignment score weights. Indicates the consistency weight of the target domain prediction. This represents the average gating strength penalty weight;

[0367] No. AUPR variation in the source domain Used to encourage improved source domain prediction performance; Used to encourage the alignment of feature distributions between the source and target domains; target domain prediction consistency. This is used to encourage training samples in the target domain to maintain consistent predictions before and after small perturbations; This is used to constrain the average gating strength, preventing the Actor from indiscriminately opening all receptor communication modules.

[0368] A2. The value and reward of the reinforcement learning state vector output by the Critic network. The advantages of standardized movements are as follows:

[0369]

[0370] in, It is the standardized action advantage value, used to evaluate the quality of the Actor's sampled actions in the current state; when A larger value indicates that the current action is better than the Critic network's expectations; This is the value of reinforcement learning state vectors.

[0371] A3. Obtain the joint logarithmic probability of the original action. The Actor loss is constructed using the joint logarithmic probability of the original actions and the standardized action dominance value. Specifically:

[0372] First, obtain the joint logarithmic probability of the original action. :

[0373]

[0374] in, It is the joint logarithmic probability of the original action. It is a component index. It is the original action The One portion, yes The One portion, ;

[0375] Then, the Actor loss is constructed using the joint logarithmic probability of the original actions and the standardized action dominance value. :

[0376]

[0377] in, It is the Actor loss;

[0378] The Actor loss is used to increase the sampling probability of high-reward actions and decrease the sampling probability of low-reward actions. During parameter updates, the Actor network and the Critic network are updated separately according to their respective losses, avoiding mixed parameter updates.

[0379] A4. The value and reward of the reinforcement learning state vector output by the Critic network. Constructing Critic Loss :

[0380]

[0381] in, It is a smoothed L1 loss function. It is the Critic loss; the Critic loss is used to improve the Critic network's ability to estimate state values.

[0382] Other steps or parameters are the same as in specific implementation method one or two.

[0383] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the classification threshold... Obtained through the following methods:

[0384] During model training, target domain test labels are not used in model training and model selection. After training, based on the trained bi-branch encoder, classification predictor, and receptor center module gating weights, the drug response prediction probability is output for the target domain training samples, and the target domain training set threshold is determined using the target domain training samples. Let the target domain training sample index be... The predicted probability of the training samples in the target domain is The target domain training samples are labeled with real drug response labels. .based on and Calculate the F1 score of the target domain training set. The threshold that maximizes the F1 score will be used as the threshold for the target domain training set.

[0385]

[0386] This process uses only the target domain training samples to determine the classification threshold and does not use the target domain test labels, thus ensuring the independence and rationality of the target domain test prediction process.

[0387] Other steps or parameters are the same as in any of the specific implementation methods one to three.

[0388] Example: To verify the beneficial effects of the present invention, a verification experiment was conducted in this example:

[0389] To verify and evaluate the performance of this invention, this embodiment was evaluated on a test set, and the area under the ROC curve (AUC), area under the precision / recall curve (AUPR), F1 score, accuracy (ACC), and precision were used as evaluation metrics. The specific calculation formulas are as follows:

[0390]

[0391]

[0392]

[0393] Among them, TP (True Positive) represents cancer cell samples that are actually drug-sensitive but were correctly predicted as sensitive by the model; TN (True Negative) represents cancer cell samples that are actually drug-resistant but were correctly predicted as drug-resistant by the model; FP (False Positive) represents cancer cell samples that are actually drug-resistant but were incorrectly predicted as sensitive by the model; and FN (False Negative) represents cancer cell samples that are actually drug-sensitive but were incorrectly predicted as drug-resistant by the model.

[0394] To verify the robustness and generalization ability of the proposed method under different experimental environments, this embodiment conducted comparative experiments on the constructed dataset, systematically comparing the proposed RL-CoReGate method with existing state-of-the-art methods (including CODE-AE, SCAD, scDeal_noPre, scDeal, and SSDA4Drug). The experimental environment included five-fold cross-validation on Etoposide, PLX4720_A375, and PLX4720_451Lu. Detailed experimental results are shown in Tables 1-4 and... Figures 2-4 .

[0395] Table 1. Performance comparison results of five-fold cross-validation on Etoposide.

[0396]

[0397] Table 2. Performance Comparison Results of Five-Fold Cross-Validation on PLX4720_A375

[0398]

[0399] Table 3. Performance Comparison Results of Five-Fold Cross-Validation on PLX4720_451Lu

[0400]

[0401] Table 4 Ablation Experiment Results

[0402]

[0403] The results show that the method of the present invention performs better overall in terms of AUC, AUPR and other metrics on multiple datasets, and demonstrates good comprehensive prediction performance on different datasets, further verifying the robustness and prediction advantages of the present invention on multiple datasets.

[0404] To further verify the contribution of each module of this invention to the prediction performance, this application conducted ablation experiments on the Etoposide task, and the experimental results are shown in Table 4. In Table 4, "RL-CoReGate" represents the complete method of this invention, which simultaneously uses the receptor-centric communication module, reinforcement learning dynamic gating, communication reliability weights, and domain adaptation mechanisms; "Random gate" indicates that the gating weights of the receptor-centric module are randomly generated to verify whether the reinforcement learning gating strategy is superior to random selection; "Fixed gate x" indicates that the gating weights of all receptor-centric modules are fixed at x to verify the contribution of the dynamic gating mechanism; "w / o CCC" indicates that ligand-receptor communication features are removed, and only gene expression features are used for prediction; "w / o weight" indicates that the communication reliability weight is removed to verify the contribution of this weight to the modeling of communication features.

[0405] As shown in Table 4, the complete method "RL-CoReGate" outperforms all ablation models in terms of AUC, AUPR, F1-score, ACC, and Precision. The performance degradation of "Random gate" and "Fixed gate x" indicates that random gating, fixed gating, or indiscriminate input of all communication features into the model weakens the model's ability to identify key receptor communication modules. The performance degradation of "w / o CCC" demonstrates that ligand-receptor communication features can provide effective supplementary information for single-cell drug response prediction. The result of "w / o weight" is lower than that of the complete method, indicating that communication reliability weights help improve the effectiveness of communication features. In summary, the receptor-centric communication module, reinforcement learning dynamic gating, communication reliability weights, and dynamic domain adaptation mechanisms all contribute positively to the predictive performance of this invention.

[0406] The above examples of this invention are merely illustrative of the computational model and process of this invention, and are not intended to limit the implementation of this invention. Those skilled in the art can make various variations or modifications based on the above description; it is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of this invention are still within the scope of protection of this invention.

Claims

1. A reinforcement learning-gated single-cell drug response prediction method, characterized in that: The specific process of the method is as follows: S1. Obtain the standardized expression matrix of the single-cell cancer sample to be predicted; S2. Obtain the standardized communication feature matrix of the single-cell cancer sample to be predicted using the standardized expression matrix of the single-cell cancer sample to be predicted. S3. Obtain the feature matrix to be predicted using the standardized communication feature matrix of the single-cell cancer sample to be predicted; S4. Input the feature matrix to be predicted into the final drug sensitivity prediction model. The drug response prediction probability is obtained and compared with the classification threshold. If the drug response prediction probability is greater than or equal to the classification threshold, it means that the single-cell cancer sample to be predicted is sensitive to the specified drug; otherwise, it means that the single-cell cancer sample to be predicted is resistant to the specified drug.

2. The reinforcement learning-gated single-cell drug response prediction method according to claim 1, characterized in that: The standardized expression matrix of the single-cell cancer sample to be predicted in S1 is specifically obtained as follows: S101. Obtain single-cell RNA sequencing expression data of the single-cell cancer sample to be predicted, and obtain the common genes of the single-cell cancer sample to be predicted. The expression values ​​on the matrix are used to obtain the expression matrix to be predicted. ; in, Indicates the first A common gene, , Indicates the number of common genes. This represents the number of single-cell cancer samples to be predicted. It is the set of real numbers; S102. Standardize the expression matrix to be predicted to obtain the standardized expression matrix of the sample to be predicted. .

3. The reinforcement learning-gated single-cell drug response prediction method according to claim 2, characterized in that: The process in S2, which involves obtaining the standardized communication feature matrix of the single-cell cancer sample to be predicted using the standardized expression matrix of the sample, specifically involves: S201. Obtain the first expression matrix of the single-cell cancer sample to be predicted using the normalized expression matrix. The original communication scores of ligand-receptor pairs in a single-cell cancer sample to be predicted are as follows: in, and They are the first Ligand genes in a single-cell cancer sample to be predicted and receptor genes The standardized expression value, It is the original communication score. It is the total number of ligand-receptor pairs. It is the first Ligand genes in ligand-receptor pairs It is the first Receptor genes in ligand-receptor pairs It is a set of ligand-receptor pairs; S202. Calculate the standardized communication feature matrix of the sample to be predicted using the original communication score. : in, It is the first The first single-cell cancer sample to be predicted corresponds to the first... Standardized communication characteristics of ligand-receptor pairs It is the upper limit of communication feature truncation. This indicates that the value is truncated to an interval. , It is the first Static reliability weights for each ligand-receptor pair It is the mean of the raw communication scores of the ligand-receptor pair. It is the standard deviation of the original communication score of the ligand-receptor pair.

4. The reinforcement learning-gated single-cell drug response prediction method according to claim 3, characterized in that: The step S3, which involves obtaining the feature matrix to be predicted using the standardized communication feature matrix of the single-cell cancer sample to be predicted, specifically involves: First, for any The gating communication characteristics are obtained by utilizing standardized communication features, specifically: in, It is the first One receptor central module, It is a communication characteristic after gating. It is the first The final gating weights of each receptor central module; Then, the first Standardized expression vectors of a single-cell cancer sample to be predicted Communication feature vector after gating The vectors are concatenated to obtain the input vector to be predicted. in, This represents vector concatenation. It is the input vector to be predicted; Finally, the feature matrix to be predicted is constructed using all the input vectors to be predicted: in, It is the feature matrix to be predicted. It is the first One feature vector to be predicted, It is the matrix transpose.

5. The reinforcement learning-gated single-cell drug response prediction method according to claim 4, characterized in that: The final drug sensitivity prediction model was obtained through the following method: Step 1: Based on the source domain bulk RNA-seq drug response data and the target domain scRNA-seq drug response data, obtain the source domain expression matrix, the target domain training expression matrix, and the target domain test expression matrix. Use the source domain expression matrix and the target domain training expression matrix to obtain standardized parameters, and then standardize the source domain expression matrix, the target domain training expression matrix, and the target domain test expression matrix based on these standardized parameters to obtain the source domain standardized expression matrix. Target domain training standardized representation matrix and target domain test normalized expression matrix Specifically: Step 11: Obtain the source domain expression matrix and target domain expression matrix based on source domain bulk RNA-seq drug response data and target domain scRNA-seq drug response data, specifically as follows: First, we acquired drug response data based on source domain bulk RNA-seq and target domain scRNA-seq, obtaining the common gene set of the source and target domains. ; in, Indicates the first A common gene, , Indicates the number of common genes; Then, the expression vectors of the source domain data and the target domain data on the common genes are retained, thus forming the source domain expression vectors. and target domain representation matrix ; Steps 1 and 2: Divide the target domain representation matrix into a target domain training index set. and target domain test index set Based on the index set, the target domain training representation matrix is ​​obtained. and target domain test expression matrix Let the source domain representation matrix be... The source domain tag is denoted as ; Step 13: Represent the matrix using the source domain and target domain training representation matrix Obtain for each common gene corresponding mean and standard deviation Using the mean and standard deviation The expression values ​​of the source domain expression matrix and the target domain expression matrix are standardized to obtain the source domain standardized expression matrix. Target domain training standardized representation matrix and target domain test normalized expression matrix : in, It is a data field. , It is the source domain. It is the target domain. Shared genes The corresponding mean expression value, Shared genes The corresponding standard deviation of the expression value, It is the sample index in the data domain. It is a data domain medium sample Chinese gene The corresponding expression value, yes The standardization results It is a constant; Step 2: Obtain the ligand-receptor pair set Based on source domain normalized representation matrix Target domain training standardized representation matrix and target domain test normalized expression matrix The communication characteristics of each ligand-receptor pair are obtained, thus yielding the source domain communication characteristic matrix. Communication feature matrix with target domain Specifically: Step 2:

1. Retrieve ligand-receptor pair information from the ConnectomeDB database, retaining those where both the ligand gene and receptor gene belong to the specified category. The ligand-receptor pairs are obtained to form a set of ligand-receptor pairs. ; in, It is the first One ligand-receptor pair, , It is a ligand-receptor pair index. It is the total number of ligand-receptor pairs. It is a ligand gene. It is a receptor gene; Step 22: Obtaining Genes Regulatory activity fraction Specifically: in, It's a gene. Regulatory activity fraction , yes The corresponding set of regulatory transcription factors, It's a gene. transcription factor genes, This indicates taking the absolute value. Genes in the source domain sample The standardized representation vector, Genes in the source domain sample The standardized representation vector, This represents the Pearson correlation coefficient; Step 23: Obtain the regulatory activity fraction of each gene in the ligand-receptor pair obtained in Step 21, and calculate the static reliability weight of each ligand-receptor pair using the regulatory activity fraction of each gene in the ligand-receptor pair, specifically as follows: in, It is the first Static reliability weights for each ligand-receptor pair It is the slope parameter of the Sigmoid function. It is a ligand gene The regulatory activity fraction, It is a receptor gene The regulatory activity fraction; Step 24: Obtain the raw communication score for each ligand-receptor pair, specifically: in, It is a sample Middle ligand gene The standardized expression value, It is a sample Middle receptor gene The standardized expression value, It is the raw communication score of the ligand-receptor pair. It takes the maximum value; Step 25: Obtain the mean and standard deviation of the raw communication scores of ligand-receptor pairs in the source domain samples, specifically as follows: in, It is the mean of the raw communication scores of ligand-receptor pairs in the source domain samples. It is the first The nth source domain sample The original communication score of each ligand-receptor pair. It is the standard deviation of the raw communication score of ligand-receptor pairs in the source domain sample; Step 26: Using the mean and standard deviation of the original communication scores of ligand-receptor pairs in the source domain samples obtained in Step 25 and the static reliability weights of each ligand-receptor pair obtained in Step 23, obtain the source domain communication feature matrix and the target domain communication feature matrix, specifically: in, It is a sample The A standardized communication feature, It is the upper limit of communication feature truncation. This indicates that the value will be truncated to... interval; Step 3: Based on ligand-receptor pair sets Obtain the connectivity of each receptor and construct a receptor central module based on the receptor connectivity. Step 4: Initialize the iteration rounds , will strengthen learning state vector Initialize to ; Step 5: Transfer the reinforcement learning state vector Input into a reinforcement learning network to obtain standardized actions ; in, express The gating weights of each receptor central module, This represents the standardized action used to adjust the strength of the domain adaptation. Step 6: Apply the standardized actions obtained in Step 5 to the source domain communication feature matrix obtained in Step 2. Communication feature matrix with target domain Gating is performed on the standardized communication features in the data, for any The gated communication feature matrix is ​​obtained as follows: in, It is the first Communication characteristics after gating in round iteration It is a standardized communication feature. It is the first The receptor central module in the first Gating weights in round iteration; Step 7: Use the source domain to normalize the representation matrix Target domain training standardized representation matrix A total training set is constructed using the corresponding gated post-communication feature matrix. The drug sensitivity prediction model is then trained in the e-th round using the total training set to obtain the trained drug sensitivity prediction model. Step 8: Determine if the current training round has reached the preset total training rounds. If it has, use the drug sensitivity prediction model trained in round e as the final drug sensitivity prediction model and save the final drug sensitivity prediction model. If the current gating weight vector is used as the final gating weight vector, then freeze the drug sensitivity prediction model trained in round e, and construct a reinforcement learning state vector based on the drug sensitivity prediction model trained in round e. ,make Then return to step five.

6. The reinforcement learning-gated single-cell drug response prediction method according to claim 5, characterized in that: Step three is based on ligand-receptor pair sets. Obtain the connectivity of each receptor, and construct the receptor central module based on the receptor connectivity, specifically as follows: Step 31: Based on ligand-receptor pair sets The connectivity of each receptor is obtained as follows: in, It is a receptor connectivity, It is a receptor The corresponding ligand-receptor pair index set, Indicates the first One ligand-receptor pair, It is a set of ligand-receptor pairs. yes Any receptor in the middle; Step 32, the receptor According to connectivity Sort from highest to lowest, and place the first... Each receptor is treated as a highly connected receptor, and a receptor central module is constructed for each highly connected receptor: in, Indicated by receptor The corresponding ligand-receptor pair forms the first One receptor central module, It is a high-connectivity receptor label. , It is the preset number of highly connected receptors. It is the first A highly connected receptor; Step 33: Obtain the set of ligand-receptor pair indices corresponding to all receptor central modules: in, It is the set of ligand-receptor pair indices corresponding to all receptor central modules.

7. The reinforcement learning-gated single-cell drug response prediction method according to claim 6, characterized in that: The reinforcement learning network in step five includes: an Actor policy module and a Critic module; The Actor policy module is used to determine the reinforcement learning state vector. The output obtains standardized actions; The Actor strategy module includes: an action mean generation unit, a strategy sampling unit, and an action mapping unit; The action mean generation unit employs a multilayer perceptron (MLP) to generate values ​​based on the state vector. The mean of the output action distribution is as follows: in, It is the mean of the action distribution. It is the action mean generation unit. These are the network parameters of the action mean generation unit; The strategy sampling unit is used to sample the original actions based on the mean of the action distribution using a Gaussian strategy. : in, It is the original action. It is the identity matrix. The standard deviation is the preset value for the Gaussian strategy. The action mapping unit is used to perform Sigmoid mapping on the original action to obtain a standardized action. Specifically: in, express The gating weights of each receptor central module, , It is the first The gating weights of each receptor central module, This represents the standardized action used to adjust the strength of the domain adaptation. The Critic module employs a multilayer perceptron (MLP) to obtain the value of the reinforcement learning state vector, specifically: in, It is the Critic network mapping function. This is the value of reinforcement learning state vectors.

8. The reinforcement learning-gated single-cell drug response prediction method according to claim 7, characterized in that: The use of source domain normalized representation matrix in step seven Target domain training standardized representation matrix A total training set is constructed using the corresponding gated post-communication feature matrix. The e-th round of training is then used to train the drug sensitivity prediction model, resulting in a well-trained model. Specifically: Step 71: Utilize the source domain normalized representation matrix Target domain training standardized representation matrix The total training set is constructed using the corresponding gated post-communication feature matrix, specifically as follows: First, normalize the source domain representation matrix. Target domain training standardized representation matrix The first dataset is composed of standardized representation vectors. The standardized representation vectors in the first dataset are then concatenated with the corresponding gated post-communication feature vectors to obtain the training input vector, specifically: in, Training rounds The input vector, It is a standardized representation vector. It is the feature vector of post-gated communication; Then, the source domain training input vector and its corresponding drug response label and the target domain training input vector are combined to form the total training set; Step 72: Use the total training set to train the drug sensitivity prediction model for the e-th round of training to obtain the well-trained drug sensitivity prediction model for the e-th round. The drug sensitivity prediction model includes: a dual-branch encoder and a classification predictor; The dual-branch encoder includes: a gene expression coding branch, a communication feature coding branch, and a feature fusion module; The gene expression coding branch is used to encode the standardized expression vector in the input vector to obtain the gene expression code: in, It is gene expression encoding, and These are gene expression coding branching parameters. Indicates batch normalization. This represents the activation function. This indicates a random deactivation operation. It is a standardized representation vector; The communication feature encoding branch is used to encode the gated communication feature vector in the input vector to obtain the communication feature code, specifically: in, It is the feature vector of post-gated communication. It is a communication feature encoding. and Indicates the communication feature coding branch parameters; The feature fusion module is used to combine gene expression coding and communication feature coding to obtain a fused feature representation, specifically: in, It is a fusion feature representation; The classification predictor uses fused features to obtain the predicted probability of drug sensitivity, specifically: = in, Indicates sample Predicted probability of sensitivity to a specified drug. It is a hidden representation. It is a predictor. These are predictor parameters. and Represents the latent space mapping parameters. This represents the last hidden representation of the predictor. and This represents the output layer parameters.

9. The reinforcement learning-gated single-cell drug response prediction method according to claim 8, characterized in that: In step eight, if the current training round has reached the preset total training rounds, and if so, the drug sensitivity prediction model trained in round e is used as the final drug sensitivity prediction model, and the final drug sensitivity prediction model is saved. If the current gating weight vector is used as the final gating weight vector, then freeze the drug sensitivity prediction model trained in round e, and construct a reinforcement learning state vector based on the drug sensitivity prediction model trained in round e. ,make Then return to step five, which is as follows: Step 81: Determine whether the current training round has reached the preset total training round. If the total training round has been reached, use the drug sensitivity prediction model trained in the e-th round as the final drug sensitivity prediction model and save the final drug sensitivity prediction model. Use the current gating weight vector as the final gating weight vector. Otherwise, freeze the drug sensitivity prediction model trained in the e-th round and execute step 82. Step 82: Calculate the AUPR of the e-th round drug sensitivity prediction model on the source domain training samples using the source domain normalized expression matrix and corresponding drug sensitivity labels, and obtain the change in source domain AUPR using the source domain AUPR. Specifically: in, It is the AUPR of the drug sensitivity prediction model trained in the e-th round on the source domain training samples. It is the AUPR of the drug sensitivity prediction model trained in the (e-1)th round on the source domain training samples. It is the change in AUPR of the source domain; Step 83: Extract target domain training samples from the overall training set, add perturbations to the target domain training samples, and input the target domain training samples and the perturbated target domain training samples into the drug sensitivity prediction model trained in the e-th round to obtain the original prediction probabilities. And the predicted probability after adding perturbations This achieves consistency in the target domain, specifically: in, It is the predicted probability after adding perturbations. It is target domain consistency. It is the total number of training samples in the target domain. These are the training sample labels for the target domain; Step 84: Extract the fused feature representation obtained in the e-th training round, construct a Deep CORAL loss based on the fused feature representation, and use the Deep CORAL loss to obtain the domain alignment score. Specifically: in, It is the domain alignment score. It is the first Deep Coral loss, Denotes the Frobenius norm. Indicates training rounds, Indicates the number of samples in the source domain. This represents the number of training samples in the target domain. Indicates the fusion feature dimension. It is the source domain fusion feature matrix centered according to the sample dimension. It is a target domain fusion feature matrix centered according to the sample dimensions. It is the source domain covariance matrix. It is the target domain covariance matrix. It is transpose; Step 85: Combine the source domain AUPR, source domain AUPR change, domain alignment score, and target domain consistency obtained in round e to form the state vector for the next iteration. and order Return to step five.

10. The reinforcement learning-gated single-cell drug response prediction method according to claim 9, characterized in that: The drug sensitivity prediction model employs source domain supervised classification loss. Domain Adaptation Loss The training, specifically: Determine whether the domain adaptation update condition is met. If the current round does not meet the domain adaptation update condition, then use the source domain supervised classification loss. Update the parameters of the bi-branch encoder and the classification predictor; if the current round meets the domain adaptation update condition, then first apply the source domain supervised classification loss. Update the parameters of the dual-branch encoder and the classification predictor, then freeze the classification predictor parameters and apply the domain adaptation loss. Update the dual-branch encoder parameters and the domain adaptation loss. ; The source domain has supervised classification loss. as follows: in, It is the number of samples in the source domain. It is the first The true drug response label of each source domain sample It is the first The drug sensitivity prediction model for the first round Drug sensitivity prediction probability for each source domain sample It is a supervised classification loss in the source domain; The domain adaptation loss Specifically: in, It is a domain-adaptive weight. It is a preset base domain adaptation weight. It is the minimum scaling factor. It is the maximum scaling factor; Reinforcement learning network parameters use Actor loss and Critic loss Update, Actor loss and Critic loss Specifically: in, It is a smoothed L1 loss function. It is a Critic loss. It is an Actor loss. It is the joint logarithmic probability of the original action. It is a component index. It is the original action The One portion, yes The One portion, , It is the advantage value of standardized movements. The value of reinforcement learning state vectors, Is to execute the first Rewards earned from standardized movements, Indicates the domain alignment score weights. Indicates the consistency weight of the target domain prediction. This represents the average gating strength penalty weight. It is the first The average gating strength of the wheel, It is the first Round Gating weights of each receptor central module, This represents the total number of receptor central modules.

Images