A cancer drug response prediction method based on deep transfer learning

Abstract

The application provides a cancer drug response prediction method based on deep transfer learning, and belongs to the technical field of cancer diagnosis. Firstly, according to the gene expression profiles of cell lines and single cells, differential genes between response cells and non-response cells are screened to reduce the scale of input features of a deep learning model; then, an autoencoder is used as a feature extractor to further learn the low-dimensional representation of the cell line and single cell expression matrix; then, a heterogeneous graph of the cell line and the single cell is constructed, and a multi-head graph attention network is used for semantic propagation between samples to better realize transfer learning; finally, a multilayer perceptron is used as a predictor to output the drug response prediction result of the cell; the application can be applied to the fields of cancer diagnosis, treatment and individualized drug treatment, and provides a reference for selecting a more effective treatment scheme in clinic.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cancer diagnosis, and particularly relates to a cancer drug response prediction method based on deep transfer learning. BACKGROUND

[0002] There is heterogeneity between cancer cells in different states, which means that individual cells are not completely the same in response to cancer drugs. Predicting drug response at the single-cell level can guide clinical drug use to eliminate tumors and residues as much as possible and avoid cancer recurrence. At present, the number of published single-cell omics data is limited, which is difficult to use to train high-performance single-cell drug response prediction models.

[0003] However, the cell line data provided by databases such as GDSC and CCLE can be used as an effective supplementary resource. By using a deep transfer learning method, the relationship between gene expression and drug response contained in the cell line data can be transferred to single cells, providing strong support for drug response prediction at the single-cell level.

[0004] The existing drug response prediction method from cell lines to single cells does not consider that irrelevant gene features will introduce noise and increase the complexity of the model, and does not perform semantic propagation between cell lines and single cells, resulting in the need to improve the prediction performance after transfer. SUMMARY

[0005] To solve the above problems, the application provides a cancer drug response prediction method based on deep transfer learning, which uses labeled cancer cell line data to predict the response of single cells of patients to cancer drugs.

[0006] The application is implemented by the following technical solutions:

[0007] A cancer drug response prediction method based on deep transfer learning:

[0008] The method specifically comprises the following steps:

[0009] Step 1, gene screening; according to the gene expression profiles of cell lines and single cells, difference genes between responsive cells and non-responsive cells are screened out;

[0010] Step 2, feature extraction is performed on the genes screened in step 1; an autoencoder is used as a feature extractor to learn the low-dimensional representation of the cell line and single cell expression matrix;

[0011] Step 3, based on the low-dimensional features in step 2, a heterogeneous graph of cell lines and single cells is constructed, and semantic propagation between samples is performed by using a multi-head graph attention network;

[0012] Step 4, after feature extraction in step 2 and semantic propagation in step 3, use multi-layer perceptron as predictor to output drug response prediction results of cells;

[0013] Step 5, on the basis of steps 2, 3, and 4, design multiple domain discriminators at different levels to align the feature distribution of cell lines and single cells at different levels;

[0014] Step 6, use the real labels of cell lines and pseudo labels of single cells to supervise the update of the whole model by aligning the centroids of the same type of cells;

[0015] Step 7, after completing the training of steps 1 to 6, input new single cell data to perform drug response prediction.

[0016] Further, in step 1,

[0017] Gene screening only retains genes that can show differences between responsive and non-responsive cells; since in this migration problem, the drug response labels of cell lines are known, while the labels of single cells are unknown, different gene screening methods need to be adopted for the two types of data;

[0018] For cell line data, use R package DESeq2 to identify differentially expressed genes between the two types of samples; select genes with adjusted Wald test p-value less than 0.01 and differential expression fold change |log2foldchange| greater than or equal to 1;

[0019] For single cell data, use unsupervised method to screen genes; first randomly select 30 genes from all genes as sample features, use K-Means clustering algorithm to cluster the samples into two clusters, repeat 100,000 times, calculate the cluster evaluation index profile coefficient of each round; then count the frequency of each gene appearing in the rounds with profile coefficient greater than the third quartile; finally, select the genes with the highest frequency, the number of which is equal to the number of genes selected from cell line data.

[0020] Further, in step 2,

[0021] Use autoencoder (AE) as feature extractor to learn low-dimensional representations of cell line expression matrix X c and single cell expression matrix X s respectively;

[0022] The structure of autoencoder consists of encoder and decoder; the encoder is used to map the expression matrix to a low-dimensional subspace, represented as follows:

[0023] X′ c =E c (X c )

[0024] X′ s = E s (X s )

[0025] where E c and E s represent the encoder for cell lines and single cells, respectively, which are composed of 3 fully connected linear layers; the decoder is used to reconstruct a matrix similar to the initial expression matrix from the low-dimensional representation, denoted as follows:

[0026] X″ c = D c (X′ c )

[0027] X″ s = D s (X′ s )

[0028] where D c and D s represent the decoder for cell lines and single cells, respectively, which are symmetric to the structure of the encoder;

[0029] The feature extractor is optimized by minimizing the mean square error (MSE) between the input matrix and the reconstructed matrix, and the loss function is denoted as:

[0030]

[0031] Further, in step 3,

[0032] A multi-head graph attention network is used to propagate semantics between cell lines and single cell samples; first, a heterogeneous graph is constructed, where each node represents a cell line or a single cell sample; the adjacency matrix is calculated by a Gaussian kernel function, representing the similarity between cells:

[0033]

[0034] where δ is a hyperparameter with a value of 40; x' i and x' j are the low-dimensional representations of cells i and j obtained by the feature extractor, respectively;

[0035] In order to reduce the size and computational scale of the heterogeneous graph, only the top 40 nearest neighbors of each node are retained; that is, each sample retains 20 nearest neighbors of intra-domain samples and 20 nearest neighbors of cross-domain samples;

[0036] Then, a graph attention network is used to learn the relationship between nodes in the heterogeneous graph, aggregate node features, and perform representation learning; specifically, the attention coefficient of cell i and its neighbor j is calculated by the attention mechanism:

[0037]

[0038] where W is the weight matrix of linear transformation, a is the weight vector of attention, denotes the neighbor set of cell i; the updated feature representation of cell i is denoted as:

[0039]

[0040] Finally, multi-head attention is adopted to stabilize the learning process, i.e., the above-mentioned node feature update based on attention mechanism is independently performed K times, and then the features updated K times are connected to obtain the feature representation of cell i:

[0041]

[0042] where, denotes the attention coefficient calculated by the kth attention mechanism, W k denotes the weight matrix of the kth linear transformation;

[0043] Through step 3, the complex relationship between nodes in the heterogeneous network is captured, and cross-domain semantic migration between cell lines and single-cell samples is achieved.

[0044] Further, in step 4,

[0045] A fully connected multi-layer perceptron (MLP) is used as a predictor to estimate the drug response of cells, which is denoted as:

[0046] Y = P(H)

[0047] where P denotes the predictor shared by cell lines and single cells, containing 3 hidden layers; H denotes the embedding representation learned by the multi-head graph attention network for cell lines and single cells; the parameters of the predictor are optimized by using the classification loss (i.e., cross-entropy) between the real and predicted cell line drug response labels, and the loss function is denoted as:

[0048]

[0049] where, denotes the real drug response label of the cell line, Y c denotes the predicted result of the cell line output by the predictor, denotes the set of cell line samples, denotes the real label of cell line i, y i denotes the predicted value of cell line i.

[0050] Further, in step 5,

[0051] Three different level domain discriminators are designed in the model; the first domain discriminator is set after the feature extractor, which is used to distinguish whether the low-dimensional representation learned by the encoder is from the cell line or the single cell; while the encoder E c and E s The domain discriminator needs to be deceived so that it cannot effectively make a judgment; the loss of the domain discriminator is calculated using cross-entropy:

[0052]

[0053] where D1 represents the first domain discriminator, which consists of 4 fully connected linear layers; x' represents the set of single cell samples, x' i represents the encoder E c output low-dimensional representation of cell line i, x' j represents the encoder E s output low-dimensional representation of single cell j;

[0054] The second domain discriminator is set after the multi-head graph attention network, which is used to distinguish whether the feature representation after semantic propagation belongs to the cell line or the single cell; the domain discriminator distinguishes the data features of the cell line and the single cell, guiding the graph attention network to learn consistent cross-domain feature representation while achieving semantic propagation; the loss function of the second domain discriminator is represented as:

[0055]

[0056] where D2 represents the second domain discriminator, which consists of 3 fully connected linear layers;

[0057] The third domain discriminator is set between the feature extractor and the multi-head graph attention network, which is used to judge whether the output of the feature extractor and the output of the multi-head graph attention network belong to the same class; in this way, the feature distribution of the cell line and the single cell is further aligned to the unified distribution learned from the graph attention network; the loss function of the third domain discriminator is represented as:

[0058]

[0059] where D3 represents the third domain discriminator, which first uses a linear layer to reduce the output of the feature extractor to the same dimension as the output of the graph attention network, and then obtains the discrimination result through 4 fully connected linear layers.

[0060] Further, in step 6,

[0061] After learning by the model, the centroids of the same type of cells belonging to different domains are close, so the loss function is defined as:

[0062]

[0063] where Y s represents the predicted result of single cells output by the predictor, i.e., pseudo labels, and represents a function for calculating the Euclidean distance, and respectively represent the centroids of sensitive and drug-resistant cell lines, and respectively represent the centroids of single cells with pseudo labels of sensitive and drug-resistant.

[0064] A cancer drug response prediction system based on deep transfer learning:

[0065] The prediction system comprises a training module and a prediction module;

[0066] The training module first performs gene screening; according to the gene expression profiles of cell lines and single cells, differential genes between responsive cells and non-responsive cells are screened out;

[0067] Then, the screened genes are subjected to feature extraction; an autoencoder is used as a feature extractor to learn the low-dimensional representation of the cell line and single cell expression matrix;

[0068] On the basis of the low-dimensional features, a heterogeneous graph of cell lines and single cells is constructed, and semantic propagation between samples is performed by using a multi-head graph attention network; then, a multilayer perceptron is used as a predictor to output the drug response prediction result of the cell;

[0069] A plurality of domain discriminators of different levels are designed to align the feature distributions of cell lines and single cells at different levels; finally, the centroids of the same type of cells are aligned to supervise the update of the entire model;

[0070] The prediction module inputs new single cell data after the training phase to perform drug response prediction.

[0071] An electronic device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the above method when executing the computer program.

[0072] A computer-readable storage medium for storing computer instructions, the computer instructions being executed by a processor to implement the steps of the above method.

[0073] Advantages of the present application

[0074] Compared with the traditional machine learning method, the deep learning model is adopted to automatically extract features, which reduces the dependence on professional knowledge and improves the prediction ability of the model.

[0075] Through the migration learning technology, the existing cell line data is effectively utilized, the problem of insufficient single cell data amount is solved, and the generalization ability of the model is improved.

[0076] By constructing a heterogeneous graph and applying a graph attention network, more effective information fusion and knowledge migration between different data sources (cell lines and single cell data) are achieved.

[0077] The domain discriminator can help the model learn a feature representation with similar distribution between the source domain (cell line) and the target domain (single cell) through adversarial training, thereby improving the performance of cross-domain prediction of the model.

[0078] Through feature extraction and dimension reduction, the calculation amount is reduced while the prediction accuracy is maintained or improved.

[0079] The present application can be applied to the fields of cancer diagnosis, treatment and individualized drug treatment, and provides a reference for selecting a more effective treatment plan in the clinic. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 The method flowchart of the present application.

[0081] Figure 2 The performance (performance index AUC) of single cell drug response prediction of the present application.

[0082] Figure 3 The performance (performance index F1 score) of single cell drug response prediction of the present application. DETAILED DESCRIPTION

[0083] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0084] The experimental methods used in the following embodiments are conventional methods unless otherwise specified. The materials, reagents, methods and instruments used are conventional materials, reagents, methods and instruments in the art unless otherwise specified, and can be obtained by commercial channels by those skilled in the art.

[0085] In combination Figures 1 to 3 The present application proposes a cancer drug response prediction method based on deep migration learning, which specifically comprises the following steps:

[0086] Step 1, gene screening. First, according to the gene expression profiles of cell lines and single cells, the differential genes between responsive cells and non-responsive cells are screened, reducing the scale of input features of the deep learning model, reducing the complexity of the model and the introduced noise.

[0087] Gene expression profile data usually contains about 20,000 genes, but not all genes play a role in cell drug response prediction. Taking all genes as input features of the prediction model will introduce unnecessary noise on the one hand, and increase the scale and complexity of the model on the other hand.

[0088] Therefore, the present application first screens the genes, and only keeps the genes that can show the difference between responsive cells and non-responsive cells. Since in this migration problem, the drug response labels of cell lines are known, and the labels of single cells are unknown, different gene screening methods need to be taken for the two types of data.

[0089] For cell line data, the present application uses R package DESeq2 to identify differentially expressed genes between the two types of samples. Genes with adjusted Wald test p-value less than 0.01 and differential expression fold change |log2foldchange| greater than or equal to 1 are screened.

[0090] For single cell data, the present application uses an unsupervised method to screen genes. First, 30 genes are randomly selected from all genes as sample features, and the K-Means clustering algorithm is used to cluster the samples into two clusters, repeated 100,000 times, and the clustering evaluation index silhouette coefficient of each round is calculated. Then, the frequency of each gene appearing in the round whose silhouette coefficient is greater than the third quartile is counted. Finally, the genes with the highest frequency are screened, and the number is equal to the number of genes screened from the cell line data.

[0091] Step 2, feature extraction. Use autoencoder as feature extractor to further learn the low-dimensional representation of cell line and single cell expression matrix, which helps to more accurately capture the key information in the data, while removing noise and redundancy. This helps to simplify the complexity of the model and improve learning efficiency.

[0092] Use autoencoder (AE) as feature extractor to learn the low-dimensional representation of cell line expression matrix X c and single cell expression matrix X s respectively. The structure of the autoencoder consists of two parts: encoder and decoder. The encoder is used to map the expression matrix to a low-dimensional subspace, represented as follows:

[0093] X′ c =E c (S c )

[0094] X′s = E s (X s )

[0095] where E c and E s represent the encoder for cell lines and single cells, respectively, which are composed of 3 fully connected linear layers. The decoder is used to reconstruct a matrix similar to the initial expression matrix from the low-dimensional representation, denoted as follows:

[0096] X″ c = D c (X′ c )

[0097] X″ s = D s (X′ s )

[0098] where D c and D s represent the decoder for cell lines and single cells, respectively, which are symmetric to the structure of the encoder.

[0099] The feature extractor is optimized by minimizing the mean square error (MSE) between the input matrix and the reconstructed matrix, and the loss function is represented as:

[0100]

[0101] Step 3: Based on the low-dimensional features, construct a heterogeneous graph for cell lines and single cells. The nodes in the heterogeneous graph represent cell samples, and the edges represent the similarity or relationship between cells. In this way, the interaction and similarity information between cells can be better utilized. Use the multi-head graph attention network to perform semantic propagation between samples, that is, transfer information between cell lines and single cell samples, so that the model can learn cross-domain and consistent feature representations. This helps improve the model's prediction ability for single cell drug response. Improve the transfer learning ability between cross-domain samples and better achieve transfer learning.

[0102] Use the multi-head graph attention network to perform semantic propagation between cell lines and single cell samples. First, construct a heterogeneous graph, where each node represents a cell line or single cell sample. The adjacency matrix is calculated by the Gaussian kernel function, representing the similarity relationship between cells:

[0103]

[0104] where δ is a hyperparameter with a value of 40; x' i and x' jrespectively, are the low-dimensional representations of cell i and cell j after the feature extractor. Furthermore, to reduce the size and computational scale of the heterogeneous graph, the present application only keeps the top 40 nearest neighbors of each node. At the beginning of training, inevitably, the low-dimensional representations of cell samples in the same domain are more similar, resulting in that the nearest neighbors of each sample are mostly samples in the same domain, which will not be conducive to the semantic propagation between cross-domain samples. Therefore, the present application forces a sample to keep 20 nearest neighbors of samples in the same domain and 20 nearest neighbors of cross-domain samples. For example, the nearest neighbors of a single cell sample include 20 single cell samples and 20 cell line samples.

[0105] Next, the present application uses a graph attention network to learn the relationship between nodes in the heterogeneous graph, aggregate and learn the representation of node features. Specifically, the attention coefficient of cell i and its neighbor j is calculated by the attention mechanism:

[0106]

[0107] where W is the weight matrix of linear transformation, and a is the weight vector of attention, represents the neighbor set of cell i. The updated feature representation of cell i is:

[0108]

[0109] Finally, the present application uses multi-head attention to stabilize the learning process, i.e. independently performing the above node feature update based on the attention mechanism K times, and then connecting the K times updated features to obtain the feature representation of cell i:

[0110]

[0111] where, represents the attention coefficient calculated by the kth attention mechanism, and W k represents the weight matrix of the kth linear transformation.

[0112] Through the above steps, the complex relationship between nodes in the heterogeneous network can be effectively captured, and the cross-domain semantic migration between cell lines and single cell samples can be realized.

[0113] Step 4, after feature extraction and semantic propagation, a multi-layer perceptron is used as a predictor to output the drug response prediction result of the cell, which combines the learned low-dimensional features and cross-domain features to estimate the response of the cell to the drug, and predict the sensitivity or drug resistance of the single cell to the specific drug.

[0114] The present application uses a fully connected multi-layer perceptron (MLP) as a predictor to estimate the drug response of the cell, which is represented as follows:

[0115] Y = P(H)

[0116] where P represents the predictor shared by cell lines and single cells, containing 3 hidden layers; H represents the embedding representation learned by cell lines and single cells through multi-head graph attention network. The parameters of the predictor are optimized by using the classification loss (i.e., cross-entropy) between the real and predicted cell line drug response labels, and the loss function is represented as:

[0117]

[0118] where, represents the real drug response label of the cell line, Y C represents the predicted result of the cell line output by the predictor, represents a set of cell line samples, represents the real label of the cell line i, y i represents the predicted value of the cell line i.

[0119] Step 5, domain discriminator, in order to further narrow the difference between the feature distribution of cell lines and single cells, a plurality of different levels of domain discriminators are designed, through adversarial learning, prompting the feature extractor and the graph attention network to generate more consistent feature representation, improving the generalization ability of the model.

[0120] In order to better realize the feature alignment of cell lines and single cells, the application additionally designs 3 different levels of domain discriminators in the model. The first domain discriminator is set after the feature extractor, which is used to distinguish whether the low-dimensional representation learned by the encoder is from the cell line or the single cell. While the encoder E c and E s It is necessary to deceive the domain discriminator so that it cannot effectively make a judgment. The loss of the domain discriminator is calculated using cross-entropy:

[0121]

[0122] where D1 represents the first domain discriminator, which is composed of 4 fully connected linear layers; represents a set of single cell samples, x' i represents the low-dimensional representation of the cell line i output by the encoder E c represents the low-dimensional representation of the cell line i output by the encoder E j represents the low-dimensional representation of the single cell j output by the encoder E s represents the low-dimensional representation of the single cell j output by the encoder E

[0123] The second domain discriminator is set after the multi-head graph attention network, and its task is to distinguish whether the feature representation after semantic propagation belongs to the cell line or the single cell. The domain discriminator distinguishes the data features of the cell line and the single cell, guiding the graph attention network to learn consistent cross-domain feature representation while realizing semantic propagation. The loss function of the second domain discriminator is represented as:

[0124]

[0125] where D2 denotes the second domain discriminator, which consists of 3 fully connected linear layers.

[0126] The third domain discriminator is set between the feature extractor and the multi-head graph attention network, which is used to determine whether the output of the feature extractor and the output of the multi-head graph attention network belong to the same class. In this way, the feature distribution of the cell lines and the single cells can be further aligned to the unified distribution learned from the graph attention network. The loss function of the third domain discriminator is represented as:

[0127]

[0128] where D3 denotes the third domain discriminator, which first uses a linear layer to reduce the output of the feature extractor to the same dimension as the output of the graph attention network, and then obtains the discrimination result through 4 fully connected linear layers.

[0129] Using multiple domain discriminators can learn feature representations with similar distributions between cell lines and single cells at multiple levels, thereby improving the generalization ability of the model in single-cell drug response prediction.

[0130] Although the main role of the domain discriminator is to assist feature learning, the discrimination result of the feature can be used to evaluate and adjust the model to ensure that the features generated by the feature extractor and the graph attention network are consistent between different domains.

[0131] Step 6, centroid alignment. Finally, the real labels of the cell lines and the pseudo labels of the single cells given by the predictor are used to supervise the update of the entire model. That is, through the loss function of centroid alignment, the model update is guided so that the centroids of the cells of the same type (sensitive or resistant) in different domains (cell lines and single cells) are close to each other in the feature space. This step helps to improve the prediction accuracy of the model.

[0132] Specifically, the present application hopes that after the learning of the model, the centroids of the same type of cells belonging to different domains are close, so the loss function is defined as:

[0133]

[0134] where Y s represents the prediction result of the single cells output by the predictor, i.e. the pseudo label, and Φ represents the function for calculating the Euclidean distance, and represent the centroids of the sensitive and resistant cell lines, respectively, and represent the centroids of the single cells with pseudo labels of sensitive and resistant, respectively.

[0135] Step 7, steps 1 to 6 are the training stage of the model, after the training is completed, the trained deep transfer learning model can be used to predict the drug response of new unknown single cell data.

[0136] Embodiments:

[0137] The cell line data comes from two public cell line omics databases: CCLE (https: / / sites.broadinstitute.org / ccle) and GDSC (https: / / www.cancerrxgene.org / ). The six sets of single cell data are downloaded from the GEO database (https: / / www.ncbi.nlm.nih.gov / geo / ), and the specific information is shown in the following table.

[0138] Table 1 Cell line data control table

[0139]

[0140]

[0141] Select the cell line domain of the same drug as the single cell data, and apply the method proposed in the application to predict the drug response of the single cell sample.

[0142] The performance of the application is tested on six single cell data sets, and compared with the existing deep transfer learning method scDEAL for single cell drug response prediction.

[0143] The results are shown in Figure 2 , 3 The prediction performance index AUC of the method of the application on all data sets is more than 0.9, and the average AUC is 0.961, which is much better than the 0.898 of the scDEAL method. Figure 1

[0144] The F1 score of the application exceeds the scDEAL method on half of the data sets, and the average F1 score is 0.920, which is also better than the scDEAL method.

[0145] An electronic device comprising a memory and a processor, the memory storing a computer program, the processor implementing the steps of the above method when executing the computer program.

[0146] A computer readable storage medium for storing computer instructions, the computer instructions being executed by a processor to implement the steps of the above method.

[0147] ​The memory in the embodiments of the application can be a volatile memory or a nonvolatile memory, or can include both volatile and nonvolatile memory. Where the nonvolatile memory is a read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically EPROM (EEPROM), or flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example, and not limitation, many forms of RAM are available, for example, static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM). Note that the memory described herein is intended to include, among others, these and any other memory suitable for storing the data associated with the methods described herein.

[0148] In the above embodiments, all or part of the methods can be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the methods can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center through a wired manner such as a coaxial cable, an optical fiber, a digital subscriber line (DSL), or a wireless manner such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that includes one or more available media sets. The available medium can be a magnetic medium such as a floppy disk, a hard disk, a magnetic tape, an optical medium such as a digital video disc (DVD), or a semiconductor medium such as a solid state disc (SSD), etc.

[0149] In the implementation process, each step of the above method can be completed by the integrated logic circuit of hardware in the processor or the instruction in the form of software. The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware processor execution or executed by a combination of hardware and software modules in the processor. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, or other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0150] It should be noted that the processor in the embodiments of the present application can be an integrated circuit chip with signal processing capability. In the implementation process, each step of the method embodiments can be completed by the integrated logic circuit of hardware in the processor or the instructions in the form of software. The processor mentioned above can be a general processor, a digital signal processor DSP, an application specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The disclosed methods, steps and logic block diagrams in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as hardware code processor execution completion, or executed by hardware and software module combination in the code processor. The software module can be located in the random access memory, the flash memory, the read only memory, the programmable read only memory or the electrically erasable programmable memory, the register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory, and combines the hardware to complete the steps of the above method.

[0151] The above describes in detail the cancer drug response prediction method based on deep transfer learning proposed by the present application, and the principle and implementation mode of the present application are described. The above embodiment is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the above description should not be understood as the limitation of the present application.

Claims

1. A cancer drug response prediction method based on deep transfer learning, characterized in that: The method specifically comprises the following steps: Step 1, gene screening; According to the gene expression profile of cell lines and single cells, the differential genes between response cells and non-response cells are screened out; Step 2, feature extraction is performed on the genes screened in step 1; Use autoencoder as feature extractor to learn the low-dimensional representation of cell line and single cell expression matrix; Step 3, on the basis of low-dimensional features in step 2, construct the heterogeneous graph of cell lines and single cells, and use multi-head graph attention network for semantic propagation between samples; Step 4, after feature extraction in step 2 and semantic propagation in step 3, use multilayer perceptron as predictor to output the drug response prediction result of the cell; Use fully connected multilayer perceptron (MLP) as predictor to estimate the drug response of the cell, which is represented as follows: wherein, represents a predictor shared by cell lines and single cells, containing 3 hidden layers; represents the embedding representation learned by the multi-head graph attention network for cell lines and single cells; the parameters of the predictor are optimized using the classification loss, i.e. cross-entropy, between the true and predicted cell line drug response labels, and the loss function is represented as: wherein, represents a true drug response label of a cell line, represents a predicted result of a cell line output by a predictor, represents a set of cell line samples, represents a cell line a true label of a cell line, represents a predicted value of a cell line a predicted value of a cell line. Step 5, on the basis of steps 2, 3 and 4, design multiple domain discriminators at different levels to align the feature distribution of cell lines and single cells at different levels; Three different level discriminators are designed in the model; the first discriminator is set after the feature extractor, which is used to distinguish whether the low-dimensional representation learned by the encoder is from the cell line or the single cell; while the encoder and The domain discriminator needs to be deceived so that it cannot effectively make a judgment; the loss of the domain discriminator is calculated by cross-entropy: wherein, denotes the first domain discriminator, consisting of 4 fully connected linear layers; denotes a collection of single-cell samples, denotes an encoder output cell lines low-dimensional representation of, denotes an encoder output single cells low-dimensional representation of; The second domain discriminator is set after the multi-head graph attention network to distinguish whether the feature representation after semantic propagation belongs to the cell line or the single cell; The domain discriminator distinguishes the data features of the cell line and the single cell, and guides the graph attention network to learn consistent cross-domain feature representation while realizing semantic propagation; The loss function of the second domain discriminator is represented as: wherein, denotes the second domain discriminator, consisting of 3 fully connected linear layers; The third domain discriminator is set between the feature extractor and the multi-head graph attention network to determine whether the output of the feature extractor and the output of the multi-head graph attention network belong to the same class; In this way, the feature distribution of the cell line and the single cell is further aligned to the unified distribution learned from the graph attention network; The loss function of the third domain discriminator is represented as: wherein, denotes the third domain discriminator, which first reduces the output of the feature extractor to the same dimension as the output of the graph attention network using a linear layer, and then obtains the discrimination result through 4 fully connected linear layers. Step 6, use the real label of the cell line and the pseudo label of the single cell to supervise the update of the whole model by aligning the centroid of the same class of cells; After learning the model, the centroids of the same class of cells belonging to different domains are close, so the loss function is defined as: wherein, represents the predicted result of a single cell output by the predictor, i.e., a pseudo label, represents a function for calculating the Euclidean distance, and respectively represent the centroids of sensitive and drug-resistant cell lines, and respectively represent the centroids of single cells with pseudo labels of sensitive and drug-resistant; Step 7, after the training of steps 1 to 6 is completed, input new single cell data to perform drug response prediction.

2. The prediction method of claim 1, wherein: In step 1, Gene screening only retains genes that can show the difference between response cells and non-response cells; Since the drug response label of the cell line is known in this transfer problem, and the label of the single cell is unknown, different gene screening methods need to be adopted for the two types of data; For cell line data, use R package DESeq2 to identify differentially expressed genes between the two samples; Screen out genes with adjusted Wald test p-value less than 0.01 and differential expression fold change |log2foldchange| greater than or equal to 1; For single-cell data, an unsupervised method is used to screen genes; first, 30 genes are randomly selected from all genes as sample features, and the sample is clustered into two clusters using the K-Means clustering algorithm, repeated 100,000 times, and the silhouette coefficient of each round is calculated; then the frequency of each gene in the round whose silhouette coefficient is greater than the third quartile is counted; finally, the genes with the highest frequency are screened, and the number is equal to the number of genes screened from cell line data.

3. The method of claim 2, wherein: In step 2, Using autoencoders (AEs) as feature extractors, low-dimensional representations of cell line expression matrices and single-cell expression matrices are learned, respectively; The structure of the autoencoder is composed of an encoder and a decoder; the encoder is used to map the expression matrix to a low-dimensional subspace, represented as follows: where, and denote the encoder for cell lines and single cells, respectively, which both consist of 3 fully connected linear layers; the decoder is used to reconstruct a matrix similar to the initial expression matrix from the low-dimensional representation, denoted as follows: wherein, and respectively represent the decoders for cell lines and single cells, which are symmetric to the structure of the encoders; The feature extractor is optimized by minimizing the mean square error (MSE) between the input matrix and the reconstructed matrix, and the loss function is represented as: 。 4. The prediction method of claim 3, wherein: In step 3, A multi-head graph attention network is used to propagate semantics between cell lines and single-cell samples; first, a heterogeneous graph is constructed, where each node represents a cell line or single-cell sample; the adjacency matrix is calculated by a Gaussian kernel function, representing the similarity between cells: wherein, is a hyper-parameter and takes the value 40; and are the low-dimensional representations of the cells and cells obtained by the feature extractor, respectively. In order to reduce the size and computational scale of the heterogeneous graph, only the top 40 nearest neighbors of each node are retained; that is, each sample retains 20 nearest neighbors of intra-domain samples and 20 nearest neighbors of cross-domain samples; Then the graph attention network is used to learn the relationship between nodes in the heterogeneous graph, aggregate and represent the node features; specifically, the attention coefficient of the cell and its neighbors is calculated by the attention mechanism: wherein, is a weight matrix of the linear transformation, is a weight vector of the attention, denotes a neighborhood set of cells ; the updated feature representation of a cell is Finally, multi-head attention is adopted to stabilize the learning process, i.e., the above-mentioned node feature update based on attention mechanism is independently performed Then, the updated features are connected to obtain the feature representation of the cell Then, the updated features are connected to obtain the feature representation of the cell ​ wherein, denotes the attention coefficient computed by the j-th denotes the attention coefficient computed by the j-th denotes the weight matrix of the j-th linear transformation;​ Through step 3, the complex relationship between nodes in the heterogeneous network is captured, and cross-domain semantic migration between cell lines and single-cell samples is achieved.

5. A prediction system for performing the deep transfer learning-based cancer drug response prediction method of any one of claims 1 to 4, characterized in that: The prediction system comprises a training module and a prediction module; The training module first performs gene screening; according to the gene expression profiles of cell lines and single cells, the differential genes between responsive cells and non-responsive cells are screened out; Then the screened genes are subjected to feature extraction; an autoencoder is used as a feature extractor to learn the low-dimensional representation of cell line and single-cell expression matrices; On the basis of low-dimensional features, a heterogeneous graph of cell lines and single cells is constructed, and a multi-head graph attention network is used for semantic propagation between samples; then a multi-layer perceptron is used as a predictor to output the drug response prediction result of the cell; A plurality of domain discriminators of different levels are designed to align the feature distributions of cell lines and single cells at different levels; finally, the centroids of the same type of cells are aligned to supervise the update of the entire model; The prediction module inputs new single-cell data after the training phase to perform drug response prediction. 6.An electronic device comprising a memory and a processor, the memory storing a computer program, wherein, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 4.

7. A computer readable storage medium for storing computer instructions, characterized in that, The computer instructions are executed by the processor to implement the steps of the method of any one of claims 1 to 4.

Images