A training method and device for a cancer intervention response prediction model

Abstract

The present invention relates to a training method and device for a cancer intervention response prediction model, comprising: acquiring training sample data; inputting the training sample data into the cancer intervention response prediction model, and processing the training sample data in sequence through an encoding module, a construction extraction module and a prediction module of the cancer intervention response prediction model to obtain corresponding cancer intervention response training results; constructing a target loss function based on the cancer intervention response training results, and training the cancer intervention response prediction model based on the target loss function to obtain a trained cancer intervention response prediction model, thereby improving the prediction accuracy of the model by improving the training method of the cancer intervention response prediction model.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a training method and device for a cancer intervention response prediction model. Background Art

[0002] Cancer is a complex and variable disease that has become a major cause of death that seriously threatens human health. Each cancer patient has unique genomic characteristics, so it is impossible to achieve a treatment that is effective for all types of cancer. Therefore, it is of great clinical significance to use predictive models to predict the response to cancer interventions based on the genomic characteristics of each patient and provide personalized treatment plans.

[0003] As one of the main means of cancer intervention, drug therapy and its effect on cancer response are currently a key issue of concern to researchers. In actual situations, it is very difficult to obtain real clinical data on the response of the same patient to multiple anticancer drugs. Therefore, the data sets used for drug response prediction mainly rely on a large amount of public cancer cell line data, such as the Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC). By utilizing these valuable data, researchers can explore the mechanism of drug response in cancer intervention treatment and develop machine learning methods to predict cancer intervention responses.

[0004] Although there are models that can predict drug response, the existing drug response prediction models at least have the problem of poor model prediction effect. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method and device for training a cancer intervention response prediction model, which solves the technical problem of poor model prediction effect in the prior art.

[0007] (II) Technical solution

[0008] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0009] In a first aspect, an embodiment of the present invention provides a method for training a cancer intervention response prediction model, comprising: obtaining training sample data; wherein the training sample data comprises multi-omics data of a cancer cell line, a drug molecular map, and cell line-drug response data;

[0010] Inputting the training sample data into the cancer intervention response prediction model, and sequentially processing through the encoding module, the construction extraction module and the prediction module of the cancer intervention response prediction model to obtain the corresponding cancer intervention response training results;

[0011] The encoding module includes an adaptive multi-omics encoding module and a drug encoding module. The adaptive multi-omics encoding module is configured to perform feature extraction on multi-omics data to obtain a cell line feature representation under an attribute branch. The drug encoding module is configured to perform feature extraction on a drug molecule graph to obtain a drug feature representation under an attribute branch.

[0012] The construction and extraction module is configured to construct a sensitive association map and a drug resistance association map based on the cell line-drug reaction data, and perform feature extraction on the sensitive association map and the drug resistance association map to obtain feature representations in the sensitive association map and the drug resistance association map under the association branch, and fuse and fully connect the feature representations in the sensitive association map and the drug resistance association map to obtain cell line feature representations and drug feature representations under the association branch;

[0013] The prediction module is configured to fuse and fully connect the cell line feature representation under the attribute branch and the cell line feature representation under the association branch to obtain the final feature representation of the cell line, and also fuse and fully connect the drug feature representation under the attribute branch and the drug feature representation under the association branch to obtain the final feature representation of the drug, and use the linear correlation coefficient and the ScaledSigmoid activation function to process the final feature representation of the cell line and the final feature representation of the drug to obtain the cancer intervention response training result;

[0014] Based on the cancer intervention response training results, a target loss function is constructed, and the cancer intervention response prediction model is trained based on the target loss function to obtain a trained cancer intervention response prediction model.

[0015] In a possible embodiment, the multi-omics data of the cancer cell line includes genomic mutation data, gene expression data and DNA methylation data of the cell line; the response data can be divided into sensitive response data and drug-resistant response data.

[0016] In a possible embodiment, the adaptive multi-omics encoding module is specifically configured to: map the genomic mutation data into a potential representation of a unified dimension to obtain a first potential representation corresponding to the gene mutation data, a second potential representation corresponding to the gene expression data, and a third potential representation corresponding to the DNA methylation data, and concatenate the first potential representation, the second potential representation, and the third potential representation, and obtain a basic multi-omics representation through linear layer transformation and LeakyReLU activation function mapping, and calculate the similarity between cell lines based on the first potential representation, the second potential representation, and the third potential representation through a cosine kernel function, and select K most similar other cell lines for each cell line to construct edges, and finally obtain an adjacency matrix based on omics m, and also concatenate the first potential representation, the second potential representation, and the third potential representation as the initial node attributes of the cell line, and based on the initial node attributes and the normalized adjacency matrix, perform information transfer between cell line nodes through a graph convolutional neural network to obtain an interactively enhanced omics-specific representation, and input the basic multi-omics representation and the interactively enhanced omics-specific representation into a dot product attention network to obtain a cell line feature representation under an attribute branch.

[0017] In one possible embodiment, an extraction module is constructed, which is specifically configured to: use a multi-layer graph convolutional network to perform feature extraction on the sensitive association map and the drug-resistant association map to obtain a first cell line representation and a first drug feature representation from the sensitive association map, and obtain a second cell line representation and a second drug feature representation from the sensitive association map, and fuse the first cell line representation and the second cell line representation to obtain a cell line feature representation under an associated branch, and fuse the first drug feature representation and the second drug feature representation to obtain a drug feature representation under an associated branch.

[0018] In one possible embodiment, the prediction module is specifically configured to: input the cell line feature representation under the attribute branch and the cell line feature representation under the association branch into the fusion module to obtain the final feature representation of the cell line, and input the drug feature representation under the attribute branch and the drug feature representation under the association branch into the fusion module to obtain the final feature representation of the drug, and use the linear correlation coefficient and the ScaledSigmoid activation function with a scaling coefficient to predict the final feature representation of the cell line and the final feature representation of the drug to obtain the cancer intervention response training result.

[0019] In a possible embodiment, the process of determining the target loss function includes: using a cosine kernel function to respectively calculate the representation similarity between cell lines and the representation similarity between drugs under the attribute branch, association branch and fusion module; based on the representation similarity between cell lines and the representation similarity between drugs, using a mean square error loss to constrain the representation similarity between the attribute branch, association branch and fusion module to obtain a representation distillation loss; based on the cancer intervention response training results, using a mean square error function to calculate the prediction distillation loss between the attribute branch, association branch and fusion module; based on the result data related to the fusion module in the cancer intervention response training results, calculating the prediction task loss; combining the representation distillation loss, the prediction distillation loss and the prediction task loss to obtain the target loss function.

[0020] In a second aspect, an embodiment of the present application provides a training device for a cancer intervention response prediction model, comprising: an acquisition module, for acquiring training sample data; wherein the training sample data includes multi-omics data of cancer cell lines, drug molecular graphs, and cell line-drug reaction data;

[0021] An input processing module is used to input the training sample data into the cancer intervention response prediction model, and process it in sequence through the encoding module, construction extraction module and prediction module of the cancer intervention response prediction model to obtain the corresponding cancer intervention response training results;

[0022] The encoding module includes an adaptive multi-omics encoding module and a drug encoding module. The adaptive multi-omics encoding module is configured to perform feature extraction on multi-omics data to obtain a cell line feature representation under an attribute branch. The drug encoding module is configured to perform feature extraction on a drug molecule graph to obtain a drug feature representation under an attribute branch.

[0023] The construction and extraction module is configured to construct a sensitive association map and a drug resistance association map based on the cell line-drug reaction data, and perform feature extraction on the sensitive association map and the drug resistance association map to obtain feature representations in the sensitive association map and the drug resistance association map under the association branch, and fuse and fully connect the feature representations in the sensitive association map and the drug resistance association map to obtain cell line feature representations and drug feature representations under the association branch;

[0024] The prediction module is configured to fuse and fully connect the cell line feature representation under the attribute branch and the cell line feature representation under the association branch to obtain the final feature representation of the cell line, and also fuse and fully connect the drug feature representation under the attribute branch and the drug feature representation under the association branch to obtain the final feature representation of the drug, and use the linear correlation coefficient and the ScaledSigmoid activation function to process the final feature representation of the cell line and the final feature representation of the drug to obtain the cancer intervention response training result;

[0025] A training module is constructed to construct a target loss function based on the cancer intervention response training results, and the cancer intervention response prediction model is trained based on the target loss function to obtain a trained cancer intervention response prediction model.

[0026] In a possible embodiment, the multi-omics data of the cancer cell line includes genomic mutation data, gene expression data and DNA methylation data of the cell line; the response data can be divided into sensitive response data and drug-resistant response data.

[0027] In a possible embodiment, the adaptive multi-omics encoding module is specifically configured to: map the genomic mutation data into a potential representation of a unified dimension to obtain a first potential representation corresponding to the gene mutation data, a second potential representation corresponding to the gene expression data, and a third potential representation corresponding to the DNA methylation data, and concatenate the first potential representation, the second potential representation, and the third potential representation, and obtain a basic multi-omics representation through linear layer transformation and LeakyReLU activation function mapping, and calculate the similarity between cell lines based on the first potential representation, the second potential representation, and the third potential representation through a cosine kernel function, and select K most similar other cell lines for each cell line to construct edges, and finally obtain an adjacency matrix based on omics m, and also concatenate the first potential representation, the second potential representation, and the third potential representation as the initial node attributes of the cell line, and based on the initial node attributes and the normalized adjacency matrix, perform information transfer between cell line nodes through a graph convolutional neural network to obtain an interactively enhanced omics-specific representation, and input the basic multi-omics representation and the interactively enhanced omics-specific representation into a dot product attention network to obtain a cell line feature representation under an attribute branch.

[0028] In one possible embodiment, an extraction module is constructed, which is specifically configured to: use a multi-layer graph convolutional network to perform feature extraction on the sensitive association map and the drug-resistant association map to obtain a first cell line representation and a first drug feature representation from the sensitive association map, and obtain a second cell line representation and a second drug feature representation from the sensitive association map, and fuse the first cell line representation and the second cell line representation to obtain a cell line feature representation under an associated branch, and fuse the first drug feature representation and the second drug feature representation to obtain a drug feature representation under an associated branch.

[0029] In one possible embodiment, the prediction module is specifically configured to: input the cell line feature representation under the attribute branch and the cell line feature representation under the association branch into the fusion module to obtain the final feature representation of the cell line, and input the drug feature representation under the attribute branch and the drug feature representation under the association branch into the fusion module to obtain the final feature representation of the drug, and use the linear correlation coefficient and the ScaledSigmoid activation function with a scaling coefficient to predict the final feature representation of the cell line and the final feature representation of the drug to obtain the cancer intervention response training result.

[0030] In a possible embodiment, a module for determining a target loss function is configured to: use a cosine kernel function to respectively calculate the representation similarity between cell lines and the representation similarity between drugs under the attribute branch, association branch, and fusion module; based on the representation similarity between cell lines and the representation similarity between drugs, use a mean square error loss to constrain the representation similarity between the attribute branch, association branch, and fusion module to obtain a representation distillation loss; based on the cancer intervention response training results, use a mean square error function to calculate the prediction distillation loss between the attribute branch, association branch, and fusion module; based on the result data related to the fusion module in the cancer intervention response training results, calculate the prediction task loss; combine the representation distillation loss, the prediction distillation loss, and the prediction task loss to obtain the target loss function.

[0031] In a third aspect, an embodiment of the present application provides a storage medium having a computer program stored thereon, and the computer program is executed by a processor to execute the method described in the first aspect or any optional implementation of the first aspect.

[0032] In a fourth aspect, an embodiment of the present application provides an electronic device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the method described in the first aspect or any optional implementation method of the first aspect is performed.

[0033] In a fifth aspect, the present application provides a computer program product, which, when running on a computer, enables the computer to execute the method in the first aspect or any possible implementation of the first aspect.

[0034] (III) Beneficial effects

[0035] The beneficial effects of the present invention are:

[0036] The present invention discloses a training method and device for a cancer intervention response prediction model, which improves the prediction accuracy of the model by improving the training method of the cancer intervention response prediction model through independent learning of two branches of cell line attributes, drug attributes and cell line-drug association information and interactive fusion guided by distillation.

[0037] In order to make the above-mentioned objectives, features and advantages to be achieved by the embodiments of the present application more obvious and understandable, the following specifically cites preferred embodiments and describes them in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 A flowchart of a method for training a cancer intervention response prediction model provided in an embodiment of the present application is shown;

[0040] Figure 2 A schematic diagram of a cancer intervention response prediction model provided in an embodiment of the present application is shown;

[0041] Figure 3 A specific schematic diagram of a cancer intervention response prediction model provided in an embodiment of the present application is shown;

[0042] Figure 4 A schematic diagram of a training device for a cancer intervention response prediction model provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0043] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0044] At present, the existing cancer intervention response prediction models ignore the effective use of cell line-drug association information, or fuse attribute information and association information through module series connection, resulting in the loss of some original attribute information during the information propagation process. Therefore, the existing drug response prediction has at least the problem of poor prediction effect.

[0045] Based on this, an embodiment of the present invention proposes a training method and device for a cancer intervention response prediction model based on dual relationship distillation, which improves the training method of the cancer intervention response prediction model by performing two-branch independent learning and distillation-guided interactive fusion of cell line attributes, drug attributes and cell line-drug association information, thereby improving the prediction accuracy of the model.

[0046] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0047] See also Figure 1 , Figure 1 FIG. 1 is a flow chart showing a method for training a cancer intervention response prediction model provided in an embodiment of the present application. Figure 1 As shown, the training method can be executed by a training device for a cancer intervention response prediction model, and the specific device of the training device can be set according to actual needs, and the embodiments of the present application are not limited thereto. For example, the training device can be a computer or a server. Specifically, the training method includes:

[0048] Step S110, obtaining training sample data, wherein the training sample data includes multi-omics data of cancer cell lines, drug molecular graphs, and cell line-drug reaction data.

[0049] It should be understood that the data included in the multi-omics data of the cancer cell line, the specific number of drug molecule maps, and the data included in the cell line-drug reaction data can all be set according to actual needs, and the embodiments of the present application are not limited to this.

[0050] Optionally, the multi-omics data of the cancer cell line may include genomic mutation data, gene expression data and DNA methylation data of the cell line obtained from the public databases CCLE and GDSC, wherein the genomic mutation data may include raw gene mutation data.

[0051] Optionally, the cell line-drug reaction data can be divided into sensitive reaction data and drug resistance reaction data. And, the number of cell lines and the number of drugs are N respectively. C and N D , and the number of reaction data is N R .

[0052] Step S120, inputting the training sample data into the cancer intervention response prediction model, and processing them in sequence through the encoding module, construction extraction module and prediction module of the cancer intervention response prediction model to obtain corresponding cancer intervention response training results.

[0053] It should be understood that those skilled in the art may modify the cancer intervention response prediction model to other names according to actual needs, and the embodiments of the present application are not limited thereto.

[0054] For example, the cancer intervention response prediction model may also be called a drug response prediction model, etc.

[0055] It should also be understood that the specific modules and functions included in the encoding module, the specific modules and functions included in the construction and extraction module, and the specific modules and functions included in the prediction module can all be set according to actual needs, and the embodiments of the present application are not limited to this.

[0056] Alternatively, if Figure 2 and Figure 3 As shown, the encoding module may include an adaptive multi-omics encoding module and a drug encoding module, wherein the adaptive multi-omics encoding module is configured to perform feature extraction on the multi-omics data of the cancer cell line to obtain a cell line feature representation under an attribute branch; and the drug encoding module is configured to perform feature extraction on the drug molecule graph to obtain a drug feature representation under an attribute branch.

[0057] It should be noted here that this attribute branch is designed to extract features of the original attribute information of cell lines and drugs, and convert the multi-omics data of cell lines and the molecular structure of drugs into potential representations.

[0058] It should also be understood that the feature extraction process of the adaptive multi-omics encoding module can be set according to actual needs, and the embodiments of the present application are not limited to this.

[0059] Optionally, the original multi-omics data (i.e., genomic mutation data) of the cancer cell line can be first mapped into a potential representation of a unified dimension through an adaptive multi-omics encoding module, as follows:

[0060]

[0061]

[0062]

[0063] In the formula, and represent the original gene mutation data of the ith cell line, the gene expression data of the ith cell line, and the DNA methylation data of the ith cell line, respectively; f t (·) and f e(·) indicates a fully connected network (or fully connected processing); f g (·) represents a one-dimensional convolutional network (or one-dimensional convolution processing); and They respectively represent the potential representation of the original gene mutation data of the i-th cell line, the potential representation of the gene expression data of the i-th cell line, and the potential representation of the DNA methylation data of the i-th cell line.

[0064] Subsequently, the potential representations of each omics can be concatenated and transformed through a linear layer and mapped with a LeakyReLU activation function to obtain the basic multi-omics representation, as follows:

[0065]

[0066] In the formula, represents the basic multi-omics representation of the ith cell line, and the superscript b is only configured to represent the basic multi-omics, not a variable; || represents the splicing operation; W b and b b represents the learnable parameters in the linear layer.

[0067] Subsequently, based on the potential representation of each omics, the similarity between cell lines can be calculated by the cosine kernel function k(), and the K most similar other cell lines are selected for each cell line to construct edges, and finally the adjacency matrix A based on omics m is obtained. m Among them, the specific formula for similarity calculation is as follows:

[0068]

[0069] In the formula, represents the similarity between the i-th cell line and the j-th cell line based on omics m; m∈{g,t,e} represents different types of omics data; T represents the transpose operation of a vector or matrix.

[0070] And, the normalized adjacency matrix calculation formula of omics m is as follows:

[0071]

[0072] In the formula, represents the normalized adjacency matrix of omics m; represents the diagonal matrix of group m; I represents the identity matrix.

[0073] Subsequently, the potential representations of each group can be concatenated as the initial node attributes of the cell line, as follows:

[0074]

[0075]

[0076] In the formula, H m,(0) Represents the initial node attributes of the cell line.

[0077] And, through the graph convolutional neural network (GCN), information is transmitted between cell line nodes to obtain the interactively enhanced omics-specific representation. The GCN may contain L layers in total. The following takes the l+1th layer of the L layer of GCN as an example (where l+1 is less than or equal to L, and l and L are both positive integers). The specific formula is as follows:

[0078]

[0079] In the formula, H m,(l+1) represents the output of the l+1th layer for omics m; represents the normalized adjacency matrix of omics m; H m,(l) represents the output of the lth layer for omics m; represents the linear transformation matrix for the l-th layer of omics m.

[0080] And, the output of the Lth layer of GCN (i.e. the output of the last layer of GCN) is used as the omics-specific representation of omics m You can get and The details are as follows:

[0081] H m =H m,(L) ;

[0082] Where m∈{g,t,e} represents different types of omics data.

[0083] Finally, the basic multi-omics representation and the omics-specific representation obtained above can be input into the dot product attention network to obtain the cell line feature representation under the attribute branch. The specific formula is as follows:

[0084]

[0085] In the formula, At represents the attribute branch; C represents the cell line; H At,C N C The collection of cell line feature representations under each attribute branch; Represents the feature representation of the i-th cell line under the attribute branch, and i is less than or equal to N C A positive integer.

[0086] It should also be understood that the feature extraction process of the drug coding module can also be set according to actual needs, and the embodiments of the present application are not limited to this.

[0087] Optionally, the drug encoding module may take a drug molecule graph as input, where the atoms of the drug molecules are nodes and the chemical bonds are edges. Also, a set of drug molecule graphs may be defined as in and Represents the j-th drug molecule graph The feature matrix and adjacency matrix of N D Indicates the total number of drugs. In addition, the drug encoding module can include a multi-layer graph convolutional network and a global maximum pooling layer to extract features from the drug molecule graph. The specific process is as follows:

[0088]

[0089]

[0090] In the formula, Represents all node features of the molecular graph of the jth drug obtained after passing through the multi-layer graph convolutional network GCN; It represents the feature representation of the j-th drug under the attribute branch; GMP is the global maximum pooling operation.

[0091] Finally, we can get the representation N D The set of drug feature representations H At,D ,Right now And 0<j≤N D .

[0092] Furthermore, the construction and extraction module is configured to construct a sensitive association map and a drug-resistant association map based on the cell line-drug reaction data, and perform feature extraction on the sensitive association map and the drug-resistant association map to obtain feature representations in the sensitive association map and feature representations in the drug-resistant association map under the association branch, and fuse and fully connect the feature representations in the sensitive association map and the feature representations in the drug-resistant association map to obtain cell line feature representations and drug feature representations under the association branch.

[0093] It should be noted here that the association branch aims to extract corresponding cell line feature representation and drug feature representation from the cell line-drug reaction data.

[0094] Specifically, a sensitive association map is constructed based on the cell line-drug response data. Drug resistance association diagram

[0095] Among them, X s and A s Represent the characteristic matrix and adjacency matrix of the sensitive association graph respectively, and A s ∈(N C +N D )×(N C +ND ) and Ar∈(N C +N D )×(N C +N D ), which is defined as follows:

[0096]

[0097]

[0098]

[0099]

[0100] In the formula, R s and R r Represent sensitive association matrix and drug resistance association matrix respectively; R represents known cell line-drug reaction data; represents the Hadamard product operation; M represents the indicator matrix, if M ij is equal to 1, indicating that cell line c i and drugs j There is a known relationship between them, otherwise it does not exist; and, X r and A r The feature matrix and adjacency matrix of the drug resistance association graph are represented respectively, and the cell lines and drugs are used as nodes in the graph, and the feature matrix is ​​initialized in the same way as Where X C and X D Trainable randomly initialized feature representations corresponding to cell lines and drugs, respectively.

[0101] Subsequently, a multi-layer graph convolutional network can be used to extract features from the sensitive association graph and the drug resistance association graph. Taking the sensitive association graph as an example, the specific formula is as follows:

[0102]

[0103] In the formula, H S,(l+1) Represents the output of the sensitive association graph S in the l+1th layer of the graph convolutional network; represents the angle matrix; H S,(l) Represents the output of the sensitive association graph S in the l-th layer of the graph convolutional network; Represents the linear transformation matrix of the l+1th layer of the graph convolutional network.

[0104] It should be noted here that the above formula is represented by H because other variables in this formula, such as A and D, are in the form of matrices, so the representation here is also for all cell lines and drug sets.

[0105] And, after L layers of graph convolutional networks, the cell line representation H that can be obtained from the sensitive association graph S,C and drug characteristics indicate H S,D , and is defined as follows:

[0106]

[0107]

[0108] In the formula, represents the feature representation of the i-th cell line in the sensitive association map, and is obtained by concatenating the results of the L-layer GCN output in the sensitive association map, that is,

[0109] Similarly, the calculation process of obtaining the characteristic representation of the jth drug in the sensitive association map is consistent with the above process. And, for the drug resistance association map, similar operations can be used to obtain the cell line representation H in the drug resistance association map R,C and drugs expressed as H R,D ,as well as And i is less than or equal to N C positive integers, and And j is less than or equal to N D A positive integer.

[0110] Thirdly, the feature representations captured in the sensitive association graph and the drug-resistant association graph can be fused to obtain the cell line feature representation H under the association branch. As,C and drug characteristics indicate H As,D ,as well as And i is less than or equal to N C positive integers, and And j is less than or equal to N D The specific process is as follows:

[0111]

[0112]

[0113] In the formula, represents the cell line feature representation of the i-th cell line; and They are respectively represented as the learnable parameters in the corresponding fully connected layer.

[0114] Furthermore, the prediction module is configured to fuse and fully connect the cell line feature representation under the attribute branch and the cell line feature representation under the association branch to obtain the final feature representation of the cell line, and also to fuse and fully connect the drug feature representation under the attribute branch and the drug feature representation under the association branch to obtain the final feature representation of the drug, and to process the final feature representation of the cell line and the final feature representation of the drug using a linear correlation coefficient and a ScaledSigmoid activation function to obtain a cancer intervention response training result.

[0115] Specifically, the two cell line feature representations under the attribute branch and the two drug feature representations under the association branch are respectively input into the fusion module to obtain the final cell line feature representation H Fu,C and the final characteristics of the drug represent H Fu,D ,as well as And i is less than or equal to N C positive integers, and And j is less than or equal to N D The specific process is as follows:

[0116]

[0117]

[0118] In the formula, and They are respectively represented as the learnable parameters in the corresponding fully connected layer.

[0119] Subsequently, the linear correlation coefficient and the ScaledSigmoid activation function with a scaling factor can be used to predict the cell line c i and drugs j Reaction between The specific process is as follows:

[0120]

[0121]

[0122]

[0123] Where B∈{At,As,Fu} represents the attribute branch At, the association branch As and the final fusion module Fu; γ represents the set scaling factor; ρ(·) represents the calculation function of the linear correlation coefficient; u i Represents input variables The average value of u j Represents input variables The average value of .

[0124] Step S130, constructing a target loss function based on the cancer intervention response training result, and training the cancer intervention response prediction model based on the target loss function to obtain a trained cancer intervention response prediction model.

[0125] It should be noted here that after the cancer intervention response prediction model is trained, the input variables can be directly input into the trained model to obtain the drug response prediction results, without having to repeat steps S110 to S130 during each use.

[0126] It should be understood that the specific function of the target loss function can be set according to actual needs, and the embodiments of the present application are not limited to this.

[0127] Optionally, the dual relation distillation mechanism includes representation distillation loss functions L between cell lines / drugs under different branches rd and predicted reaction distillation of cell line-drug under different branches pd , the specific process is as follows:

[0128] First, the cosine kernel function k() can be used to calculate the representation similarity S between the cell lines under each branch and fusion module (i.e., the part marked with the Fu superscript above) B,C and the similarity between drugs S B,D , where C represents the cell, D represents B∈{At,As,Fu} represents the attribute branch At, the association branch AS and the final fusion module Fu. The specific process is as follows:

[0129]

[0130]

[0131] 0<i≤N C ;

[0132] 0<p≤N C ;

[0133] 0<j≤N D ;

[0134] 0<q≤N D .

[0135] In the formula, represents the representation similarity between the i-th cell line and the p-th cell line; B∈{At,As,Fu} represents the attribute branch At, the association branch AS and the final fusion module Fu; represents the representation similarity between the j-th drug and the q-th drug.

[0136] Secondly, the mean square error loss is used to constrain the representation similarity between the two branches and the final fusion module, taking the attribute branch as an example:

[0137]

[0138] In the formula, c i represents the i-th cell line; C represents the cell line collection; c p represents the pth cell line; d j represents the jth drug; D represents the drug set; d q represents the qth drug.

[0139] Similarly, a similar method can be used to obtain the distillation loss between the associated branches and the final fusion module The specific formula is as follows:

[0140]

[0141] Furthermore, the overall distillation loss is expressed as:

[0142] And then the mean square error function is used to calculate the prediction distillation loss L between the two branches and the fusion module pd , the specific formula is as follows:

[0143]

[0144] And then calculate the loss function L of the prediction task ce , the specific formula is as follows:

[0145]

[0146] Where M represents the indicator matrix. If M ij is equal to 1, indicating that cell line c i and drugs j There is a known relationship between them, otherwise it does not exist; R ij Indicates cell line c i and drugs j If R ij If R ij If it is equal to 0, it means drug resistance.

[0147] Finally, the representation distillation loss, prediction distillation loss, and prediction task loss can be combined as the final target loss function. The specific formula is as follows:

[0148] L=L ce +λ rd L rd +λ pd L pd ;

[0149] In the formula, λ rd and λ pd represents the hyperparameters that balance different loss functions.

[0150] Therefore, with the help of the above-mentioned technical solution, the present application improves the prediction accuracy of the model by improving the training method of the cancer intervention response prediction model through independent learning of two branches of cell line attributes, drug attributes and cell line-drug association information and interactive fusion guided by distillation.

[0151] It should be understood that the above-mentioned training method of the cancer intervention response prediction model is only exemplary, and those skilled in the art can make various modifications based on the above-mentioned method, and the modified scheme also falls within the protection scope of this application.

[0152] It should be pointed out that although the various steps in the above embodiments are described in a specific order, those skilled in the art can understand that in order to achieve the effects of the present invention, different steps do not have to be performed in such an order. They can be performed simultaneously (in parallel) or in other orders. These changes are within the scope of protection of the present invention.

[0153] See also Figure 4 , Figure 4 The structural block diagram of a training device 400 for a cancer intervention response prediction model provided in an embodiment of the present application is shown. It should be understood that the training device 400 is capable of executing each step in the above method embodiment. The specific functions of the training device 400 can be found in the description above. To avoid repetition, the detailed description is appropriately omitted here. The training device 400 includes at least one software function module that can be stored in a memory in the form of software or firmware or solidified in the operating system (OS) of the training device 400. Specifically, the training device 400 includes:

[0154] An acquisition module 410 is used to acquire training sample data; wherein the training sample data includes multi-omics data of cancer cell lines, drug molecular graphs, and cell line-drug reaction data;

[0155] An input processing module 420 is used to input the training sample data into the cancer intervention response prediction model, and process it in sequence through the encoding module, the construction extraction module and the prediction module of the cancer intervention response prediction model to obtain the corresponding cancer intervention response training results;

[0156] The encoding module includes an adaptive multi-omics encoding module and a drug encoding module. The adaptive multi-omics encoding module is configured to perform feature extraction on multi-omics data to obtain a cell line feature representation under an attribute branch. The drug encoding module is configured to perform feature extraction on a drug molecule graph to obtain a drug feature representation under an attribute branch.

[0157] The construction and extraction module is configured to construct a sensitive association map and a drug resistance association map based on the cell line-drug reaction data, and perform feature extraction on the sensitive association map and the drug resistance association map to obtain feature representations in the sensitive association map and the drug resistance association map under the association branch, and fuse and fully connect the feature representations in the sensitive association map and the drug resistance association map to obtain cell line feature representations and drug feature representations under the association branch;

[0158] The prediction module is configured to fuse and fully connect the cell line feature representation under the attribute branch and the cell line feature representation under the association branch to obtain the final feature representation of the cell line, and also fuse and fully connect the drug feature representation under the attribute branch and the drug feature representation under the association branch to obtain the final feature representation of the drug, and use the linear correlation coefficient and the ScaledSigmoid activation function to process the final feature representation of the cell line and the final feature representation of the drug to obtain the cancer intervention response training result;

[0159] A training module 430 is constructed to construct a target loss function based on the cancer intervention response training results, and to train the cancer intervention response prediction model based on the target loss function to obtain a trained cancer intervention response prediction model.

[0160] In a possible embodiment, the multi-omics data of the cancer cell line includes genomic mutation data, gene expression data and DNA methylation data of the cell line; the response data can be divided into sensitive response data and drug-resistant response data.

[0161] In a possible embodiment, the adaptive multi-omics encoding module is specifically configured to: map the genomic mutation data into a potential representation of a unified dimension to obtain a first potential representation corresponding to the gene mutation data, a second potential representation corresponding to the gene expression data, and a third potential representation corresponding to the DNA methylation data, and concatenate the first potential representation, the second potential representation, and the third potential representation, and obtain a basic multi-omics representation through linear layer transformation and LeakyReLU activation function mapping, and calculate the similarity between cell lines based on the first potential representation, the second potential representation, and the third potential representation through a cosine kernel function, and select K most similar other cell lines for each cell line to construct edges, and finally obtain an adjacency matrix based on omics m, and also concatenate the first potential representation, the second potential representation, and the third potential representation as the initial node attributes of the cell line, and based on the initial node attributes and the normalized adjacency matrix, perform information transfer between cell line nodes through a graph convolutional neural network to obtain an interactively enhanced omics-specific representation, and input the basic multi-omics representation and the interactively enhanced omics-specific representation into a dot product attention network to obtain a cell line feature representation under an attribute branch.

[0162] In one possible embodiment, an extraction module is constructed, which is specifically configured to: use a multi-layer graph convolutional network to perform feature extraction on the sensitive association map and the drug-resistant association map to obtain a first cell line representation and a first drug feature representation from the sensitive association map, and obtain a second cell line representation and a second drug feature representation from the sensitive association map, and fuse the first cell line representation and the second cell line representation to obtain a cell line feature representation under an associated branch, and fuse the first drug feature representation and the second drug feature representation to obtain a drug feature representation under an associated branch.

[0163] In one possible embodiment, the prediction module is specifically configured to: input the cell line feature representation under the attribute branch and the cell line feature representation under the association branch into the fusion module to obtain the final feature representation of the cell line, and input the drug feature representation under the attribute branch and the drug feature representation under the association branch into the fusion module to obtain the final feature representation of the drug, and use the linear correlation coefficient and the ScaledSigmoid activation function with a scaling coefficient to predict the final feature representation of the cell line and the final feature representation of the drug to obtain the cancer intervention response training result.

[0164] In a possible embodiment, a module for determining a target loss function (not shown) is configured to: use a cosine kernel function to respectively calculate the representation similarity between cell lines and the representation similarity between drugs under the attribute branch, association branch and fusion module; based on the representation similarity between cell lines and the representation similarity between drugs, use a mean square error loss to constrain the representation similarity between the attribute branch, association branch and fusion module to obtain a representation distillation loss; based on the cancer intervention response training results, use a mean square error function to calculate the prediction distillation loss between the attribute branch, association branch and fusion module; based on the result data related to the fusion module in the cancer intervention response training results, calculate the prediction task loss; combine the representation distillation loss, the prediction distillation loss and the prediction task loss to obtain the target loss function.

[0165] Since the device described in the above embodiment of the present invention is a device used to implement the method of the above embodiment of the present invention, based on the method described in the above embodiment of the present invention, a person skilled in the art can understand the specific structure and deformation of the device, so it is not described here. All devices used in the method of the above embodiment of the present invention belong to the scope of protection of the present invention.

[0166] The present application provides a storage medium having a computer program stored thereon. The computer program is executed by a processor to execute the method described in the embodiment.

[0167] The present application also provides a computer program product, which, when executed on a computer, enables the computer to execute the method described in the method embodiment.

[0168] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions.

[0170] It should be noted that in the claims, any reference numerals placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In the claims enumerating several means, several of these means may be embodied by the same hardware. The use of the words first, second, third, etc., is for convenience of expression only and does not indicate any order. These words may be understood as part of the component name.

[0171] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0172] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments after knowing the basic creative concept. Therefore, the claims should be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present invention.

[0173] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention should also include these modifications and variations.

Claims

1. A method for training a cancer intervention response prediction model, characterized in that: include: Acquire training sample data; wherein the training sample data includes multi-omics data of cancer cell lines, drug molecular maps, and cell line-drug reaction data; the multi-omics data of cancer cell lines includes genome mutation data, gene expression data, and DNA methylation data of cell lines; Inputting the training sample data into a cancer intervention response prediction model, and sequentially processing the training sample data through an encoding module, a construction extraction module, and a prediction module of the cancer intervention response prediction model to obtain corresponding cancer intervention response training results; The encoding module includes an adaptive multi-omics encoding module and a drug encoding module, wherein the adaptive multi-omics encoding module is configured to perform feature extraction on the multi-omics data to obtain a cell line feature representation under an attribute branch, and the drug encoding module is configured to perform feature extraction on the drug molecule graph to obtain a drug feature representation under an attribute branch; The construction and extraction module is configured to construct a sensitive association map and a drug resistance association map based on the cell line-drug reaction data, and perform feature extraction on the sensitive association map and the drug resistance association map to obtain a feature representation in the sensitive association map and a feature representation in the drug resistance association map under an association branch, and fuse and fully connect the feature representation in the sensitive association map and the feature representation in the drug resistance association map to obtain a cell line feature representation and a drug feature representation under an association branch; The prediction module is configured to fuse and fully connect the cell line feature representation under the attribute branch and the cell line feature representation under the association branch to obtain a final cell line feature representation, and also fuse and fully connect the drug feature representation under the attribute branch and the drug feature representation under the association branch to obtain a final drug feature representation, and process the cell line final feature representation and the drug final feature representation using a linear correlation coefficient and a ScaledSigmoid activation function to obtain the cancer intervention response training result; Based on the cancer intervention response training results, constructing a target loss function, and training the cancer intervention response prediction model based on the target loss function to obtain a trained cancer intervention response prediction model; The construction and extraction module is specifically configured to: use a multi-layer graph convolutional network to perform feature extraction on the sensitive association map and the drug resistance association map, so as to obtain a first cell line representation and a first drug feature representation from the sensitive association map, and a second cell line representation and a second drug feature representation obtained from the sensitive association map, and fuse the first cell line representation and the second cell line representation to obtain a cell line feature representation under the association branch, and fuse the first drug feature representation and the second drug feature representation to obtain a drug feature representation under the association branch; The process of determining the target loss function includes: The cosine kernel function is used to respectively calculate the representation similarity between the attribute branch, the association branch and the cell lines under the fusion module and the representation similarity between the drugs; Based on the representation similarity between the cell lines and the representation similarity between the drugs, a mean square error loss is used to constrain the representation similarity between the attribute branch, the association branch, and the fusion module to obtain a representation distillation loss; Based on the cancer intervention response training results, a mean square error function is used to calculate the prediction distillation loss between the attribute branch, the association branch and the fusion module; Calculating the prediction task loss based on the result data related to the fusion module in the cancer intervention response training result; The representation distillation loss, the prediction distillation loss, and the prediction task loss are combined to obtain the target loss function.

2. The training method according to claim 1, characterized in that: The reaction data can be divided into sensitive reaction data and resistant reaction data.

3. The training method according to claim 2, characterized in that: The adaptive multi-omics encoding module is specifically configured to: map the genomic mutation data into a potential representation of a unified dimension to obtain a first potential representation corresponding to the gene mutation data, a second potential representation corresponding to the gene expression data, and a third potential representation corresponding to the DNA methylation data, and concatenate the first potential representation, the second potential representation, and the third potential representation, and obtain a basic multi-omics representation through linear layer transformation and LeakyReLU activation function mapping, and calculate the similarity between cell lines through a cosine kernel function based on the first potential representation, the second potential representation, and the third potential representation, and select K most similar other cell lines for each cell line to construct edges, and finally obtain an adjacency matrix based on omics m, and also concatenate the first potential representation, the second potential representation, and the third potential representation as the initial node attributes of the cell line, and based on the initial node attributes and the normalized adjacency matrix, perform information transfer between cell line nodes through a graph convolutional neural network to obtain an interactively enhanced omics-specific representation, and input the basic multi-omics representation and the interactively enhanced omics-specific representation into a dot product attention network to obtain a cell line feature representation under the attribute branch.

4. The training method according to claim 3, characterized in that: The prediction module is specifically configured to: input the cell line feature representation under the attribute branch and the cell line feature representation under the association branch into a fusion module to obtain the final feature representation of the cell line, and input the drug feature representation under the attribute branch and the drug feature representation under the association branch into a fusion module to obtain the final feature representation of the drug, and use a linear correlation coefficient and a ScaledSigmoid activation function with a scaling coefficient to perform prediction processing on the final feature representation of the cell line and the final feature representation of the drug to obtain the cancer intervention response training result.

5. A training device for a cancer intervention response prediction model, characterized in that: include: An acquisition module, used to acquire training sample data; wherein the training sample data includes multi-omics data of cancer cell lines, drug molecular maps and cell line-drug reaction data; the multi-omics data of cancer cell lines includes genome mutation data, gene expression data and DNA methylation data of cell lines; An input processing module, used to input the training sample data into the cancer intervention response prediction model, and process the training sample data in sequence through the encoding module, the construction extraction module and the prediction module of the cancer intervention response prediction model to obtain corresponding cancer intervention response training results; The encoding module includes an adaptive multi-omics encoding module and a drug encoding module, wherein the adaptive multi-omics encoding module is configured to perform feature extraction on the multi-omics data to obtain a cell line feature representation under an attribute branch, and the drug encoding module is configured to perform feature extraction on the drug molecule graph to obtain a drug feature representation under an attribute branch; The construction and extraction module is configured to construct a sensitive association map and a drug resistance association map based on the cell line-drug reaction data, and perform feature extraction on the sensitive association map and the drug resistance association map to obtain a feature representation in the sensitive association map and a feature representation in the drug resistance association map under an association branch, and fuse and fully connect the feature representation in the sensitive association map and the feature representation in the drug resistance association map to obtain a cell line feature representation and a drug feature representation under an association branch; The prediction module is configured to fuse and fully connect the cell line feature representation under the attribute branch and the cell line feature representation under the association branch to obtain a final cell line feature representation, and also fuse and fully connect the drug feature representation under the attribute branch and the drug feature representation under the association branch to obtain a final drug feature representation, and process the cell line final feature representation and the drug final feature representation using a linear correlation coefficient and a ScaledSigmoid activation function to obtain the cancer intervention response training result; Constructing a training module, which is used to construct a target loss function based on the cancer intervention response training result, and to train the cancer intervention response prediction model based on the target loss function to obtain a trained cancer intervention response prediction model; The construction and extraction module is specifically configured to: use a multi-layer graph convolutional network to perform feature extraction on the sensitive association map and the drug-resistant association map to obtain a first cell line representation and a first drug feature representation from the sensitive association map, and a second cell line representation and a second drug feature representation obtained from the sensitive association map, and fuse the first cell line representation and the second cell line representation to obtain a cell line feature representation under the association branch, and fuse the first drug feature representation and the second drug feature representation to obtain a drug feature representation under the association branch; the target loss function determination module is configured to: use the cosine kernel function to respectively calculate the attribute branch, the association branch and the The representation similarity between the cell lines and the representation similarity between the drugs under the fusion module; based on the representation similarity between the cell lines and the representation similarity between the drugs, the mean square error loss is used to constrain the representation similarity between the attribute branch, the association branch and the fusion module to obtain the representation distillation loss; based on the cancer intervention response training results, the mean square error function is used to calculate the prediction distillation loss between the attribute branch, the association branch and the fusion module; based on the result data related to the fusion module in the cancer intervention response training results, the prediction task loss is calculated; the representation distillation loss, the prediction distillation loss and the prediction task loss are combined to obtain the target loss function.

6. The training device according to claim 5, characterized in that The reaction data can be divided into sensitive reaction data and resistant reaction data.

7. The training device according to claim 6, characterized in that The adaptive multi-omics encoding module is specifically configured to: map the genomic mutation data into a potential representation of a unified dimension to obtain a first potential representation corresponding to the gene mutation data, a second potential representation corresponding to the gene expression data, and a third potential representation corresponding to the DNA methylation data, and concatenate the first potential representation, the second potential representation, and the third potential representation, and obtain a basic multi-omics representation through linear layer transformation and LeakyReLU activation function mapping, and calculate the similarity between cell lines through a cosine kernel function based on the first potential representation, the second potential representation, and the third potential representation, and select K most similar other cell lines for each cell line to construct edges, and finally obtain an adjacency matrix based on omics m, and also concatenate the first potential representation, the second potential representation, and the third potential representation as the initial node attributes of the cell line, and based on the initial node attributes and the normalized adjacency matrix, perform information transfer between cell line nodes through a graph convolutional neural network to obtain an interactively enhanced omics-specific representation, and input the basic multi-omics representation and the interactively enhanced omics-specific representation into a dot product attention network to obtain a cell line feature representation under the attribute branch.

Images