Single-cell drug sensitivity prediction method fusing cross-attention and domain adaptation

Abstract

The application relates to the technical field of deep learning, and is an application of deep learning technology in drug research and development, in particular to a single-cell drug sensitivity prediction method fusing cross attention and domain self-adaption, which comprises the following steps: data preprocessing, a self-encoder, cross attention, domain self-adaption and single-cell drug sensitivity prediction.The application first applies domain self-adaption based on cross attention to the problem of predicting single-cell drug sensitivity, and combines gene expression data and drug structure data to construct a model, and has achieved a performance of 0.85 AUC in predicting single-cell drug sensitivity.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically the application of deep learning technology in drug development, and particularly to the application of cross-attention mechanisms and domain adaptation in single-cell drug sensitivity prediction. Background Technology

[0002] High-throughput drug screening technologies have generated massive amounts of drug response data, covering thousands of tumor cell lines, which has facilitated the development of computational prediction of cell line drug responses. Currently, much of the cell line sequencing data only reflects the average signal of gene transcription, masking the differences between different cell types or subpopulations. The development of single-cell sequencing technology allows us to explore the complexity and spatiotemporal variability of tumor cell components, providing an opportunity to understand drug-tumor cell interactions. However, due to cost and technological limitations, current single-cell level drug response data only includes a limited number of cancer types and drugs. The insufficiency of single-cell sequencing data and drug sensitivity data restricts the development of predictive models. Some works have already modeled drug-induced cell line sequencing data, using deep transfer learning to transfer cell line drug response knowledge to single cells, becoming an effective means to overcome the lack of training data in the single-cell domain. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies and reduce the cost of experimental verification. This invention proposes a single-cell drug sensitivity prediction method that integrates cross-attention and domain adaptation, providing a novel approach for predicting single-cell drug sensitivity using cell line drug sensitivity.

[0004] The specific technical solution of this invention is: a single-cell drug sensitivity prediction method integrating cross-attention and domain adaptation, comprising the following steps:

[0005] 1) Data preprocessing:

[0006] The datasets required for the 1-1 model are cell line gene expression dataset, single cell gene expression dataset, cell line drug sensitivity dataset, and drug structure dataset corresponding to the drug.

[0007] In step 1-1, the drug SMILES descriptor is used as drug structure data.

[0008] 1-2 Binarize the z-score values ​​in the cell line drug sensitivity dataset to 0 and 1, obtaining a binarized cell line drug sensitivity dataset A. Since the cell line gene expression dataset and the single-cell gene expression dataset contain the same genes, perform the same standardization on both datasets. The standardized cell line gene expression dataset includes data with rows named as gene names and columns named as cell line names. The standardized single-cell gene expression dataset includes data with rows named as gene names and columns named as single cell names. The drug structure dataset includes one column of drug names and one column of drug structure data. Subsequently, we extract the single-cell name column data from the standardized single-cell gene expression dataset, add a column named "single cell name" to dataset A, and repeatedly add the single-cell name data to the single-cell name column of dataset A. This results in a new binarized cell line drug sensitivity dataset A containing one column of cell line names, one column of drug names, one column of single-cell names, and one column of binarized cell line drug sensitivity data as label data.

[0009] In steps 1-2, the method for binarizing the z-score values ​​is as follows: based on the z-score values ​​in the cell line drug sensitivity dataset, z-score values ​​greater than 0 are converted to 1, and z-score values ​​less than or equal to 0 are converted to 0.

[0010] 1-3 Use 80% of dataset A as training dataset A train 20% as validation dataset A val Before inputting data into the model, we constructed a training dataset A. train and validation dataset A val There are four datasets in total, including the source domain and the target domain. The source domain dataset is the training dataset. Training dataset for the target domain Validation dataset of the source domain Validation dataset for the target domain

[0011] 1-4 Training dataset for the source domain The construction process is as follows: We will use the standardized cell line gene expression profile dataset, drug structure dataset, and dataset A train We read the data and iterate through the drug structure dataset row by row. For each row of drug structure data, we use the RKDit tool to generate a molecular graph, where nodes represent atoms and edges represent chemical bonds. At the end of the iteration, we form a dictionary of drug names as the bond values ​​and molecular graph data as the values. Then we iterate through the row-by-row dataset A... trainWe obtain the cell line name, drug name, and label data for each row. Based on the cell line name, we extract the corresponding gene feature data from the standardized cell line gene expression dataset. Based on the drug name, we extract the corresponding molecular map data from the dictionary. Then, we save the gene feature data, molecular map data, and label data. After the iteration is complete, we save the data to a file. For the source domain validation dataset... The construction process is similar, taking dataset A as an example. train Replace with dataset A val For the target domain training dataset Replace the normalized cell line gene expression dataset with the normalized single-cell gene expression dataset. For the target domain validation dataset... Replace the standardized cell line gene expression dataset with the standardized single-cell gene expression dataset, and set dataset A to... train Replace with dataset A val .

[0012] 2) Cell line drug sensitivity prediction:

[0013] 2-1 Input data consists of n source domains We have labeled samples, where d and c represent drug structure data and cell line gene expression data, respectively, and y represents the sensitivity label of the drug to the cell line. We process the dataset using methods 1-4 and then input it into the model.

[0014] 2-2 Using a drug encoder, drug descriptors are encoded into l-dimensional vectors.

[0015] 2-3 Using a cell encoder, cell line gene expression data are encoded into an l-dimensional latent vector.

[0016] 2-4 To fully extract the knowledge between the drug and cells, we utilize the multi-head cross-attention module in the Transformer. Based on the drug and cell line features obtained in 2-2 and 2-3 respectively, the model derives an l-dimensional vector.

[0017] In sections 2-4, a two-layer, eight-head cross-attention network model is used as the encoder for cell-drug interactions.

[0018] 2-5 A fully connected multilayer perceptron is used as a classifier to assess the correlation between cell line pharmacogenomics information and drug response at the cell line level, based on the results derived in 2-4. We use it as input to a classifier to predict drug sensitivity at the cell line level;

[0019] In sections 2-5, the model uses a linear layer with 128 input dimensions and 2 output dimensions as a classifier.

[0020] 3) Domain adaptation based on cross-attention:

[0021] 3-1 Input data consists of n target domains Unlabeled samples, where d and s represent drug structure data and single-cell gene expression data, respectively. We process the dataset using methods 1-4 and then input it into the model.

[0022] 3-2 Using the same methods as steps 2) 2-2, 2-3, and 2-4, obtain the l-dimensional vector of single-cell drug interaction.

[0023] 3-3 For domain adaptation, a domain discriminator helps all encoders extract domain-invariant cell-drug interaction knowledge. The drug encoder, cell encoder, and cell-drug interaction encoder are trained adversarially as model feature extractors and the domain discriminator: the training objective of the domain discriminator is to maximize the applicability of the input features a i The feature extractor correctly identifies the domain (source or target domain), but the domain discriminator cannot correctly determine feature a. i They come from which domain, thus forming an adversarial relationship;

[0024] 3-4. Adversarial learning is used to jointly train and update the feature extractor, classifier, and domain discriminator in steps 2) and 3).

[0025] 4) Single-cell drug sensitivity prediction:

[0026] 4-1 Test Dataset Preparation: The required dataset for the test is: the structure data of the drug to be tested, and the standardized single-cell gene expression dataset from step 1.

[0027] 4-2 After obtaining the well-trained feature extractor and classifier assembly, the test dataset is input into the model to predict the sensitivity information of single cells to drugs.

[0028] The entire framework can be viewed as two parts:

[0029] Supervised learning: Building a model to predict drug response label classification in cell lines;

[0030] Unsupervised learning: Transferring drug cell knowledge at the cell line level and using the model to predict drug response labels for single cells.

[0031] The present invention has the following beneficial effects:

[0032] 1. A single-cell drug sensitivity prediction method that integrates cross-attention and domain adaptation uses cell line response data for multiple drugs during training, which solves the problem of insufficient training data in single-drug models, resulting in insufficient model generalization ability. This makes it possible to learn about drug-cell interactions using the cross-attention mechanism.

[0033] 2. A single-cell drug sensitivity prediction method that integrates cross-attention and domain adaptation has extremely high accuracy in predicting whether a single cell is sensitive to a certain drug on multiple single-cell datasets. It can be used in related tumor heterogeneity studies to provide a certain reference for experimenters, thereby saving manpower and resources and reducing experimental costs. Attached Figure Description

[0034] Figure 1 This is a model framework diagram of the present invention;

[0035] Figure 2 The ROC curve and AUC index of this invention are used to predict the sensitivity of head and neck cancer single-cell JUH006 to the drug Gefitinib. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0037] The cell line gene expression data and cell line drug sensitivity data from the GDSC database were downloaded from https: / / www.cancerrxgene.org / downloads / bulk_download; the drug structure data were downloaded from the ZINC database, specifically the SMILES descriptors for the relevant drug names, located at https: / / zinc.docking.org / ; and the single-cell gene expression data were downloaded from the GEO database, located at https: / / www.ncbi.nlm.nih.gov / geo / .

[0038] refer to Figure 1 The single-cell drug sensitivity prediction method of the present invention, which integrates cross-attention and domain adaptation, includes the following steps:

[0039] Includes the following steps:

[0040] 1) Data preprocessing:

[0041] The datasets required for the 1-1 model are cell line gene expression dataset, single cell gene expression dataset, cell line drug sensitivity dataset, and drug structure dataset corresponding to the drug.

[0042] In step 1-1, the dataset consists of cell line gene expression data from the GDSC database, cell line drug sensitivity data from the GDSC database, single-cell gene expression data from the GEO database, and drug structure data corresponding to drugs from the ZINC database and the GDSC database.

[0043] 1-2 Binarize the z-score values ​​in the cell line drug sensitivity dataset to 0 and 1, obtaining a binarized cell line drug sensitivity dataset A. Since the cell line gene expression dataset and the single-cell gene expression dataset contain the same genes, perform the same standardization on both datasets. The standardized cell line gene expression dataset includes data with rows named as gene names and columns named as cell line names. The standardized single-cell gene expression dataset includes data with rows named as gene names and columns named as single cell names. The drug structure dataset includes one column of drug names and one column of drug structure data. Subsequently, we extract the single-cell name column data from the standardized single-cell gene expression dataset, add a column named "single cell name" to dataset A, and repeatedly add the single-cell name data to the single-cell name column of dataset A, resulting in a new binarized cell line drug sensitivity dataset A containing one column of cell line names, one column of drug names, one column of single-cell names, and one column of binarized cell line drug sensitivity data as label data.

[0044] In steps 1-2, the method for binarizing the z-score values ​​is as follows: based on the z-score values ​​in the cell line drug sensitivity dataset, z-score values ​​greater than 0 are converted to 1, and z-score values ​​less than or equal to 0 are converted to 0.

[0045] 1-3 Use 80% of dataset A as training dataset A train 20% as validation dataset A val Before inputting data into the model, we constructed a training dataset A. train and validation dataset A val There are four datasets in total, including the source domain and the target domain. The source domain dataset is the training dataset. Training dataset for the target domain Validation dataset of the source domain Validation dataset for the target domain

[0046] 1-4 Training dataset for the source domain The construction process is as follows: We will use the standardized cell line gene expression profile dataset, drug structure dataset, and dataset A trainWe read the data and iterate through the drug structure dataset row by row. For each row of drug structure data, we use the RKDit tool to generate a molecular graph, where nodes represent atoms and edges represent chemical bonds. At the end of the iteration, we form a dictionary of drug names as the bond values ​​and molecular graph data as the values. Then we iterate through the row-by-row dataset A... train We obtain the cell line name, drug name, and label data for each row. Based on the cell line name, we extract the corresponding gene feature data from the standardized cell line gene expression dataset. Based on the drug name, we extract the corresponding molecular map data from the dictionary. Then, we save the gene feature data, molecular map data, and label data. After the iteration is complete, we save the data to a file. For the source domain validation dataset... The construction process is similar, taking dataset A as an example. train Replace with dataset A val For the target domain training dataset Replace the normalized cell line gene expression profile dataset with the normalized single-cell gene expression dataset. For the target domain validation dataset... Replace the standardized cell line gene expression dataset with the standardized single-cell gene expression dataset, and set dataset A to... train Replace with dataset A val .

[0047] 2) Cell line drug sensitivity prediction:

[0048] 2-1 Input data consists of n source domains We have labeled samples, where d and c represent drug structure data and cell line gene expression data, respectively, and y represents the sensitivity label of the drug to the cell line. We process the dataset using methods 1-4 and then input it into the model.

[0049] 2-2 Using a graph attention neural network as the drug encoder, molecular graph data is used as input to the graph attention neural network to encode drug descriptors into l-dimensional vectors.

[0050] In section 2-2, a two-layer, eight-head graph attention neural network is used as the drug encoder.

[0051] 2-3 A denoising autoencoder was used as a cell encoder to encode cell line gene expression data into an l-dimensional latent vector. The loss function is:

[0052]

[0053] In formula (1), n ​​is the number of cells. For cell line gene expression data, c′ i Gene expression data of cell lines after noise has been added. For noise distribution, E θ For encoding networks, D ψ For the decoding network, we chose the binomial distribution.

[0054] 2-4 To fully extract the knowledge between the drug and cells, we utilize the multi-head cross-attention module in the Transformer. Based on the drug and cell line features obtained in 2-2 and 2-3 respectively, the model derives an l-dimensional vector.

[0055] In sections 2-4, a two-layer, eight-head cross-attention network model is used as the encoder for cell-drug interactions.

[0056] 2-5 A fully connected multilayer perceptron is used as a classifier to assess the correlation between cell line pharmacogenomics information and drug response at the cell line level, based on the results derived in 2-4. We use this as input to a classifier to predict drug sensitivity at the cell line level, with the loss function being:

[0057]

[0058] In formula (2), Predict probabilities for the classifier. This is the true label for the cell line.

[0059] In sections 2-5, the model uses a linear layer with 128 input dimensions and 2 output dimensions as a classifier.

[0060] 3) Domain adaptation based on cross-attention:

[0061] 3-1 Input data consists of n target domains Unlabeled samples, where d and s represent drug structure data and single-cell gene expression data, respectively. We process the dataset using methods 1-4 and then input it into the model.

[0062] 3-2 Using the same methods as steps 2-2, 2-3, and 2-4 in step 2), obtain the l-dimensional vector of single-cell drug interaction. The loss function for the single-cell encoder is:

[0063]

[0064] In formula (3), n is the number of cells. This is single-cell gene expression data, s′ i This is a dataset of single-cell gene expression data with added noise. For noise distribution, E θ For encoding networks, D ψFor the decoding network, a binomial distribution is chosen here.

[0065] 3-3 For domain adaptation, a domain discriminator helps all encoders extract domain-invariant cell-drug interaction knowledge. The drug encoder, cell encoder, and cell-drug interaction encoder are trained adversarially as model feature extractors and the domain discriminator: the training objective of the domain discriminator is to maximize the applicability of the input features a i The feature extractor must be correctly categorized into the source and target domains. The feature extractor extracts features that the domain discriminator cannot correctly identify if feature a is incorrect. i The domain from which they originate creates an adversarial relationship, and the loss function is:

[0066]

[0067] In formula (4), d i It is data x i The true tag of the domain, d i =1 indicates the source domain, d i =0 indicates the target domain; p(d i ∣D(x i )) is the predicted probability given by the domain discriminator;

[0068] 3-4 The feature extractor, classifier, and domain discriminator in steps 2) and 3) are jointly trained and updated using adversarial learning; the total loss training objective is given by formula (5):

[0069]

[0070] 4) Single-cell drug sensitivity prediction:

[0071] 4-1 Test Dataset Preparation: The required dataset for the test is: the structure data of the drug to be tested, and the standardized single-cell gene expression dataset from step 1.

[0072] 4-2 After obtaining the well-trained feature extractor and classifier assembly, the test dataset is input into the model to predict the sensitivity information of single cells to drugs.

[0073] In the network model:

[0074] Supervised learning: Building a model to predict drug response label classification in cell lines;

[0075] Unsupervised learning: Transferring drug cell knowledge at the cell line level and using the model to predict drug response labels for single cells.

[0076] In this embodiment, the GDSC database cell line gene expression data includes 825 cell line gene expression data and 2933 gene features; the single cell gene expression data includes 66 single cell gene expression data and 2933 gene features; the GDSC database cell line drug sensitivity data includes 56125 data points on the responses of 237 drugs to 825 cell lines, including cell line name, drug name, and z-score value; and the drug structure data corresponding to the drugs in the GDSC database from the ZINC database includes SMILES descriptors corresponding to 237 drugs.

[0077] An eight-layer graph attention neural network (GNN) was used as the drug encoder, and an eight-layer cross-attention neural network (CNA) module was used to extract the knowledge features between the drug and cells. The feature extractor in the model encoded the data into 128-dimensional features. Adagrad was used as the optimizer during training, with a learning rate of 0.001 to update the encoder. Subsequently, a graph attention neural network trained by adversarial learning, a denoising autoencoder, a cross-attention mechanism, and a classifier were used as the test model. 80% of the data was used as the training set, and 20% was used for validation. During the model testing phase, the model was input with SMILES data of the drug Gefitinib and gene expression data from 66 head and neck cancer single cells in the target domain, and was trained for a total of 100 epochs.

[0078] The results of predicting the sensitivity of head and neck cancer single-cell JUH006 to the drug Gefitinib are as follows: Figure 2 As shown.

[0079] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A single-cell drug sensitivity prediction method integrating cross-attention and domain adaptation, characterized by: First, a predictive model is built to predict the drug response label classification of cell lines; then, cell-drug interaction knowledge at the cell line level is transferred, and the predictive model is used to predict the drug response labels of single cells. The specific steps include: 1) Data preprocessing: The 1-1 dataset includes a cell line gene expression dataset, a single cell gene expression dataset, a cell line drug sensitivity dataset, and a drug structure dataset corresponding to the drugs. 1-2 Binarize the z-score values ​​in the cell line drug sensitivity dataset to 0 and 1 to obtain the binarized cell line drug sensitivity dataset and define it as dataset A; Both the cell line gene expression dataset and the single cell gene expression dataset contain the same genes, and the two datasets are standardized in the same way. The standardized cell line gene expression dataset includes data with rows named as gene names and columns named as cell line names. The standardized single cell gene expression dataset includes data with rows named as gene names and columns named as single cell names. The drug structure dataset includes: a column of drug name data and a column of drug structure data; Then, extract the column data of single cell names from the standardized single-cell gene expression dataset, and add a column named "single cell name" to dataset A; repeatedly add the single cell name data to the column of single cell names in dataset A to obtain a new dataset A; the new dataset A includes: a column of cell line name data, a column of drug name data, a column of single cell name data, and a column of binarized cell line drug sensitivity data; among which, the binarized cell line drug sensitivity data is used as label data; 1-3 Divide the new dataset A into training dataset A train and validation dataset A val ; Construct training dataset A train and validation dataset A val There are four datasets in total, including the source domain and the target domain. The source domain dataset is the training dataset. Training dataset for the target domain Validation dataset of the source domain and the validation dataset for the target domain 2) Predictive models are used to predict drug sensitivity in cell lines: The input data for the 2-1 prediction model includes n source domains. There are labeled samples, where d and c represent drug structure data and cell line gene expression data, respectively, and y represents the sensitivity label of the drug to the cell line. Input the dataset obtained in step 1) into the prediction model; 2-2 In the prediction model, a drug encoder is used to transform the drug structure data into an l-dimensional vector. The drug structure data of each row in the drug structure dataset is used to form a molecular graph, and the molecular graph data is used as the input of the drug encoder. 2-3 In the prediction model, a cell encoder is used to map cell line gene expression data to an l-dimensional latent vector. 2-4 In the prediction model, an encoder for cell-drug interactions is used, taking the processing results from 2-2 and 2-3 as input, to obtain an l-dimensional vector of cell line-drug interactions. 2-5 In the prediction model, a classifier is used to convert the results obtained in 2-4. As input to the classifier, it predicts drug sensitivity at the cell line level; Drug encoders, cell encoders, and encoders for cell-drug interactions serve as model feature extractors; The drug encoder is a graph attention neural network; the cell encoder is a denoising autoencoder; the encoder for cell-drug interaction is a multi-head cross-attention network module; and the classifier is a fully connected multilayer perceptron. 3) Domain adaptation based on cross-attention: The input data for the 3-1 prediction model includes n target domains. Unlabeled samples, where d and s represent drug structure data and single-cell gene expression data, respectively; input the dataset obtained in step 1) into the prediction model; 3-2 In the prediction model: First, a drug encoder is used to transform the drug structure data into an l-dimensional vector. Next, a cell encoder is used to map single-cell gene expression data to an l-dimensional latent vector. Finally, an encoder for cell-drug interactions is used, taking the results of the first two steps as input, to obtain an l-dimensional vector of single-cell-drug interactions. 3-3 For domain adaptation, a domain discriminator helps the feature extractor extract domain-invariant cell-drug interaction knowledge. The feature extractor and the domain discriminator are trained adversarially. The training objective of the domain discriminator is to maximize the accuracy of the input feature a. i The feature extractor must be correctly categorized into the source domain and the target domain. The feature extractor extracts features that the domain discriminator cannot correctly identify when feature a is not correctly identified. i The domain from which they come forms an adversarial relationship; 3-4. Adversarial learning is used to jointly train and update the feature extractor, classifier, and domain discriminator in steps 2) and 3). 4) Single-cell drug sensitivity prediction: 4-1 Test Dataset Preparation: The dataset required for the test is: the structure data of the drug to be tested, and the standardized single-cell gene expression dataset from step 1). 4-2 After obtaining the well-trained feature extractor and classifier assembly, the test dataset is input into the prediction model to predict the sensitivity information of single cells to the drug.

2. The single-cell drug sensitivity prediction method integrating cross-attention and domain adaptation according to claim 1, characterized in that: In step 1), the method for binarizing the z-score value is as follows: based on the z-score values ​​in the cell line drug sensitivity dataset, z-score values ​​greater than 0 are converted to 1, and z-score values ​​less than or equal to 0 are converted to 0.

3. The single-cell drug sensitivity prediction method integrating cross-attention and domain adaptation according to claim 1, characterized in that: In step 1), 80% of dataset A is used as training dataset A. train 20% as validation dataset A val .

4. The single-cell drug sensitivity prediction method integrating cross-attention and domain adaptation according to claim 1, characterized in that: Step 1) also includes 1-4: For the training dataset of the source domain The construction process includes: The standardized cell line gene expression profile dataset, drug structure dataset, and training dataset A are used to... train Read it in; The drug structure dataset is traversed row by row. For each row of drug structure data, a molecular graph is generated using the RKDit tool, where nodes represent atoms and edges represent chemical bonds. At the end of the traversal, a dictionary of drug names is generated with the bond values ​​as the values ​​of the molecular graph data. Then, iterate through the row-by-row dataset A train The process obtains the cell line name, drug name, and label data for each row; extracts the corresponding gene feature data from the standardized cell line gene expression dataset based on the cell line name, and extracts the corresponding molecular map data from the dictionary based on the drug name; then saves the gene feature data, molecular map data, and label data; and saves the data to a file after the traversal is complete. For the validation dataset of the source domain Construction process and training dataset of the source domain The construction approach is the same, and dataset A is used. train Replace with dataset A val ; For the target domain training dataset Replace the standardized cell line gene expression dataset with the standardized single cell gene expression dataset; For the target domain validation dataset Replace the standardized cell line gene expression dataset with the standardized single-cell gene expression dataset, and set dataset A to... train Replace with dataset A val .

5. The single-cell drug sensitivity prediction method integrating cross-attention and domain adaptation according to claim 1, characterized in that: Drug SMILES descriptors are used as drug structure data.

6. The single-cell drug sensitivity prediction method integrating cross-attention and domain adaptation according to claim 1, characterized in that: In step 2), a 2-layer, 8-head graph attention neural network is used as the drug encoder; a 2-layer, 8-head cross-attention network model is used as the encoder for cell-drug interaction; and a linear layer with 128 input dimensions and 2 output dimensions is used as the classifier.

7. The single-cell drug sensitivity prediction method integrating cross-attention and domain adaptation according to claim 1, characterized in that: In steps 2) 2-3, a denoising autoencoder is used as a cell encoder to encode the cell line gene expression data into an l-dimensional latent vector. The loss function is: In formula (1), n ​​is the number of cells. For cell line gene expression data, c′ i Gene expression data of cell lines after noise has been added. For noise distribution, E θ For encoding networks, D ψ For the decoding network, a binomial distribution is chosen here. In steps 2) 2-5, a fully connected multilayer perceptron acts as a classifier to assess the correlation between cell line pharmacogenomics information and drug response at the cell line level, based on the results derived in steps 2-4. Using this as input to a classifier to predict drug sensitivity at the cell line level, the loss function is: In formula (2), Predict probabilities for the classifier. This is a true label for cell lines; In step 3) 3-2, a denoising autoencoder is used as a cell encoder to encode single-cell gene expression data into an l-dimensional latent vector. The loss function is: In formula (3), n is the number of cells. This is single-cell gene expression data, s′ i This is a dataset of single-cell gene expression data with added noise. For noise distribution, E θ For encoding networks, D ψ For the decoding network, a binomial distribution is chosen here. In step 3) 3-3, the feature extractor and domain discriminator are trained adversarially, with the loss function being: In formula (4), d i It is data x i The true tag of the domain, d i =1 indicates the source domain, d i =0 indicates the target domain; p(d i ∣D(x i )) is the predicted probability given by the domain discriminator; In steps 3) 3-4, adversarial learning is used to jointly train and update the feature extractor, classifier, and domain discriminator from steps 2) and 3); the total loss training objective is given by formula (5):

Images