An anti-cancer cell sensitivity prediction method, system, device and medium fusing network relationships of genes

Abstract

A kind of anti-cancer cell sensitivity prediction method, system, equipment and medium of fusion gene network relationship, method includes: the vector of the expression amount of each gene of cell line is as the original gene expression feature of cell line, compressed low-dimensional hidden vector is used as the gene expression feature of cell line by using self-encoder, graph self-encoding is carried out to gene interaction network to obtain gene interaction feature, gene network feature is calculated according to gene expression feature and gene interaction feature, the feature of each atom in drug compound molecule is obtained by establishing drug compound molecular graph, and the predicted anti-cancer cell sensitivity is obtained according to EIGA model;System, equipment and medium are used to realize a kind of anti-cancer cell sensitivity prediction method of fusion gene network relationship;The present application fully considers the interaction relationship between genes and the topological structure in compound molecule by designing new algorithm, realizes the anti-cancer cell sensitivity prediction on cell line and single cell level.

Description

Technical Field

[0001] This invention relates to the field of cell prediction technology, and in particular to a method, system, device and medium for predicting the sensitivity of anti-cancer cells by incorporating gene network relationships. Background Technology

[0002] Predicting the sensitivity of cancer cells or cell lines to specific anticancer drugs is a key challenge in precision medicine. Deep learning algorithms have achieved excellent results in fields such as drug discovery. Compared with other machine learning algorithms, these algorithms have a stronger ability to capture complex nonlinear relationships and are more suitable for drug response modeling tasks. Published related studies have confirmed this and demonstrated the potential of deep learning in this field.

[0003] Molecular Pharmaceutics (Matteo Manica, Ali Oskooei, Jannis Born, Vigneshwari Subramanian, Julio Sáez-Rodríguez, and María Rodríguez Martínez. Toward Explainable Anticancer Compound Sensitivity Prediction via MultimodalAttention-Based Convolutional Encoders. Molecular Pharmaceutics (201916(12),4797-4806)) proposed a convolutional encoder model based on multimodal attention to predict the sensitivity of anticancer compounds. First, a subset of genes is obtained by propagation through a protein-protein interaction network. Second, the original SMILES string is converted into a vector sequence through embedding and input into three parallel convolutional channels and one non-convolutional channel. Each channel has an independent layer to generate gene weights to filter gene expression data. Then, the vector sequence obtained by embedding and the filtered gene expression data are input into four contextual attention layers with gene expression data as context. Finally, the outputs of these 16 layers are concatenated and the IC50 is obtained through a series of hidden layers. In addition, the study on cancer drug sensitivity inference based on gene expression (Chawla, S., Rockstroh, A., Lehman, M. et al. Gene expression based inference of cancer drug sensitivity. Nat Commun 13, 5680 (2022).) proposed the Precily model, which concatenates the processed gene expression features with the molecular features of drug compounds and inputs them into a series of linear layers.

[0004] The drawbacks of existing technologies are: first, the characterization of cancer cell lines usually only considers gene expression levels and ignores the interactions between genes; second, the characterization of anticancer compounds usually uses SMILES strings directly and ignores the topological structure of the compound molecules. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the present invention aims to provide a method for predicting the sensitivity of anticancer cells by integrating gene network relationships. By designing a new algorithm, the method fully considers the interaction relationships between genes and the topological structure in compound molecules, thereby achieving the prediction of anticancer cell sensitivity at the cell line and single cell levels.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting the sensitivity of cancer cells by fusing gene network relationships includes the following steps:

[0008] Step 1: The vector composed of the expression levels of each gene in the cell line is used as the original gene expression feature of the cell line. An autoencoder built with a fully connected layer with an activation function is used to compress the original gene expression feature into a low-dimensional latent vector of 64-1024 dimensions, and this low-dimensional latent vector is used as the gene expression feature of the cell line.

[0009] Step 2: Perform graph autoencoding on the gene interaction network to obtain gene interaction features;

[0010] Step 3: Calculate the gene network characteristics of the cell line based on the gene expression characteristics of the cell line in Step 1 and the gene interaction characteristics in Step 2;

[0011] Step 4: Treat the atoms of the drug compound molecule as nodes and the chemical bonds as edges to obtain the drug compound molecule diagram. Input the drug compound molecule diagram into the molgnet model to obtain the characteristics of each atom in the drug compound molecule.

[0012] Step 5: Input the gene expression characteristics of the cell line in Step 1, the gene network characteristics of the cell line in Step 3, and the characteristics of each atom in the drug compound molecule in Step 4 into the EIGA model to obtain the predicted anti-cancer cell sensitivity.

[0013] In step 3, the gene network characteristics of the cell line are calculated. The specific process is as follows:

[0014] The set of genes with the highest original gene expression characteristics of each cell line, ranging from 2% to 50%, is taken as the high-expression gene subset of that cell line. The mean value of the gene interaction characteristics of all genes in the high-expression gene subset is taken to obtain the gene network characteristics of the cell line.

[0015] In step 5, the specific steps are as follows:

[0016] Step 5.1: Input the features of each atom in the drug compound molecule and the gene expression features mapped by linear layer transformation and activation function into multi-head attention network A to obtain molecular characterization A; input the features of each atom in the drug compound molecule and the gene network features mapped by linear layer transformation and activation function into multi-head attention network B to obtain molecular characterization B; add molecular characterization A and molecular characterization B to obtain the overall molecular characterization.

[0017] The outputs of the multi-head attention networks A and B are calculated as follows:

[0018]

[0019] Where, d k To characterize the dimension of the vectors representing each atom in the molecule; matrices K and V are obtained by passing a series of vectors representing each atom in the molecule through two independent linear layers; matrix Q is obtained by passing gene expression features or gene network features through a linear layer with an activation function and a linear layer without an activation function.

[0020] Step 5.2: The molecular overall characterization is concatenated with the gene expression features and gene network features after being mapped by linear layer transformation and activation function, and then input into a series of fully connected layers with activation functions to obtain the predicted cancer cell sensitivity.

[0021] A cancer cell sensitivity prediction system that integrates gene network relationships includes:

[0022] Original gene expression feature compression module: An autoencoder built using a fully connected layer with an activation function compresses the original gene expression features to obtain low-dimensional gene expression features;

[0023] Gene interaction relationship encoding module: Performs graph autoencoding on the gene interaction network to obtain gene interaction features containing gene network relationship information;

[0024] Information fusion module: used to fuse gene expression features and gene interaction features to obtain gene network features;

[0025] Drug compound molecule encoding module: used to obtain the characteristics of each atom in the drug compound molecule;

[0026] EIGA prediction module: Used to input gene expression characteristics of cell lines, gene network characteristics of cell lines, and characteristics of each atom in drug compound molecules to obtain predicted cancer cell sensitivity.

[0027] A device for predicting cancer cell sensitivity based on fusion gene network relationships, comprising:

[0028] Memory: A computer program used to store a method for predicting the sensitivity of cancer cells based on fused gene network relationships;

[0029] A processor for executing the computer program to implement a method for predicting the sensitivity of cancer cells by fusing gene network relationships.

[0030] A computer-readable storage medium comprising:

[0031] The computer-readable storage medium stores a computer program that, when executed by a processor, enables a method for predicting the sensitivity of cancer cells by fusing gene network relationships.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. Step 1 of this invention uses an autoencoder to compress the original gene expression features, which can obtain a low-dimensional vector containing gene expression information. It has the characteristics of reducing noise information and reducing dimensionality for easier processing.

[0034] 2. Step 3 of the present invention uses fusion gene expression information and gene interaction information to obtain a vector containing the overall gene interaction information of the cell line, which has the characteristic of containing a richer amount of information in the characterization of the cell line.

[0035] 3. Step 4 of this invention adopts a scheme that treats atoms as vertices and chemical bonds as edges, that is, uses a topological graph to model drug molecules, and uses the Molgnet model to calculate the characteristics of each atom in the drug compound molecule. This allows the topological structure of the compound molecule to be fully considered in the drug sensitivity prediction, and has the characteristic of containing a richer amount of information in the characterization of drugs.

[0036] 4. Step 5 of this invention employs a scheme that uses gene interaction information captured by graph autoencoder and fused with gene expression information to predict drug sensitivity. This allows for a full consideration of the interaction mechanisms between genes in drug sensitivity prediction, rather than just using the expression levels of the genes themselves. This approach offers more comprehensive consideration, more efficient data utilization, higher prediction accuracy, stronger generalization ability in predicting unseen cell lines and drugs, and stronger generalization ability when using single-cell sequencing data for prediction.

[0037] 5. Step 5 of this invention utilizes the attention mechanism to aggregate molecular features, which enables the characterization of drug molecules to integrate the features of drug molecules and cell lines, resulting in more comprehensive consideration and improved accuracy and generalization ability in drug sensitivity prediction.

[0038] This invention, through the design of a novel algorithm that fully considers the interactions between genes and the topological structure of compound molecules, enables the prediction of cancer cell sensitivity at the cell line and single-cell levels. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention.

[0040] Figure 2 This is a diagram of the EIGA model framework of the present invention.

[0041] Figure 3 This is a system block diagram of the present invention. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings.

[0043] like Figure 1 As shown, a method for predicting the sensitivity of cancer cells based on fusion gene network relationships specifically includes the following steps:

[0044] Step 1: The vector of expression levels of the 6163 genes with the largest expression variations across cell lines (Jia, P., Hu, R., Pei, G. et al. Deep generative neural network for accurate drug response imputation. Nat Commun 12, 1740 (2021).) was used as the original gene expression features of the cell lines. Specifically, for cell lines, gene expression levels were RMA (robust multi-chip average) normalized gene expression data from the GDSC (Cancer Drug Sensitivity Genomics) database or TPM (transcriptome sequencing per million transcripts) normalized RNA-seq (transcriptome sequencing) gene expression data from the CCLE (Encyclopedia of Cancer Cell Lines); for single cells, gene expression levels were CPM (counts per million) normalized scRNA-seq (single-cell transcriptome sequencing) gene expression data. Due to gene omissions in the corresponding data, the original gene expression features based on RMA-normalized gene expression data have only 5757 dimensions; the original gene expression features based on TPM-normalized RNA-seq gene expression data and CPM-normalized scRNA-seq gene expression data have only 5863 dimensions. The specific preprocessing methods for the required data are as follows: For cell line TPM-normalized RNA-seq expression profiles, gene IDs were mapped to gene names using the annotation files provided in the dataset. For genes corresponding to multiple IDs, the average of the expression data corresponding to multiple IDs was taken. Then, the TPM-normalized matrix was incremented by one and the natural logarithm was taken. For CPM-normalized single-cell transcriptome sequencing expression profiles, these were obtained from the Broad Institute's single-cell portal (https: / / singlecell.broadinstitute.org / single_cell / study / SCP542 / pan-cancer-cell-line-heterogeneity#study-download) and the comprehensive gene expression website GEO (accession number GSE157220). The CPM-normalized matrix was incremented by one and the natural logarithm was taken. Then, single-cell expression profiles from the same cell line were averaged. The fully connected layer in the autoencoder has a dimension of [n, 2048, 1024, 512, 1024, 2048, n], where n is the dimension of the original gene expression features. The activation function of the fully connected layer is ReLU. The output of the fully connected layer with an output dimension of 512 is taken as the gene expression features of the cell line.

[0045] Step 2: Merge the four gene interaction networks and perform graph autoencoding on the merged network to obtain gene interaction features of 12,938 genes (Forster, DT, Li, SC, Yashiroda, Y. et al. BIONIC: biological network integration using convolutions. Nat Methods 19, 1250–1261 (2022).), where the gene interaction features of each gene are 512-dimensional.

[0046] Step 3: Take the set of the 256 genes with the highest original gene expression characteristic values ​​of each cell line as the high-expression gene subset of that cell line. Take the mean of the gene interaction characteristics of all genes in the high-expression gene subset to obtain the gene network characteristics of the cell line. If there are genes in the high-expression gene subset that do not have gene interaction characteristics, these genes are ignored in the calculation.

[0047] Step 4: Treat the atoms of the drug compound molecule as nodes and the chemical bonds as edges to obtain the molecular graph of the drug compound molecule. Specifically, the atomic and chemical bond information of the drug compound molecule is obtained using the rdkit library in Python, with the molecule's SMILES string as the index. For the SMILES string of the molecule, the PubChemCID of the drug compound molecule is obtained through GDSC and then retrieved using the pubchempy library in Python. In addition, the PubChemCID or SMILES string of some molecules is manually retrieved from the PubChem database and the LINCS database. Subsequently, the molecular diagram was input into molgnet (Pengyong Li, Jun Wang, Yixuan Qiao, Hao Chen, Yihuan Yu, Xiaojun Yao, Peng Gao, Guotong Xie, Sen Song, An effective self-supervised framework for learning expressive molecular global representations to drug discovery, Briefings in Bioinformatics, Volume 22, Issue 6, November 2021, bbab109) to obtain the features of each atom in the drug compound molecule, where each atom has 768 dimensions.

[0048] Step 5: Input the gene expression characteristics of the cell line, the gene network characteristics of the cell line, and the characteristics of each atom in the drug compound molecule into the EIGA model to obtain the predicted anti-cancer cell sensitivity.

[0049] like Figure 2 As shown, the EIGA model uses gene expression features and gene network features together as a characterization of cell lines. It uses molecular graphs to characterize drug compound molecules and inputs them into molgnet to obtain the features of each atom in the molecule. Furthermore, it applies gene expression features and gene network features to the atom features through attention networks to obtain the overall molecular characterization. Finally, it fuses gene expression features, gene network features, and the overall molecular characterization and inputs them into a series of fully connected layers with ReLU activation functions to predict LN IC50 (half-inhibition concentration of natural logarithmic transformation).

[0050] The detailed processing procedure for the EIGA model is as follows:

[0051] The features of each atom in the drug compound molecule and the gene expression features mapped by the linear layer transformation and the ReLU activation function are input into a multi-head attention network A to obtain molecular representation E; the features of each atom in the drug compound molecule and the gene network features mapped by the linear layer transformation and the ReLU activation function are input into a multi-head attention network B to obtain molecular representation I; molecular representation E and molecular representation I are added to obtain the overall molecular representation.

[0052] The output of an attention network is calculated as follows:

[0053]

[0054] Where, d k To characterize the dimension of the vectors representing each atom in the molecule, i.e., 768 dimensions; the matrix formed by a series of vectors representing each atom in the molecule is passed through two independent linear layers to obtain matrices K and V; the gene expression features or gene network features are passed through a linear layer with the activation function ReLU and a linear layer without the activation function to obtain matrix Q.

[0055] The molecular whole characterization is concatenated with gene expression features and gene network features that have been mapped by linear layer transformation and ReLU activation function, and then input into a series of fully connected layers with ReLU activation function to obtain the predicted cancer cell sensitivity.

[0056] From the GDSC database, we can obtain the LN IC50 (half-inhibitory concentration of natural logarithmic transformation) of cell line drug pairs as a sensitivity value for anti-cancer cells for model training and testing.

[0057] First, the EIGA model was tested using a relaxed partitioning strategy, which allows for overlap in cell lines or drugs between the training, validation, and test sets (but excludes overlapping cell line-drug pairs). Five-fold cross-validation was used, and the average of the predictions from each fold was taken as the final prediction. For comparison, the same training, validation, and test sets were used as those used during the training of the MCA model under the relaxed partitioning strategy. The gene expression features in this dataset were RMA-normalized basal expression profiles. Due to the missing SMILES string, there were 15910 cell line-drug pairs in the test set. The decision variable was the normalized LN IC50. The test results are shown in Table 1.

[0058] Table 1 Comparison of model performance under the relaxed partitioning strategy

[0059]

[0060] As can be seen from Table 1, the EIGA model performs significantly better than the MCA model.

[0061] Secondly, the EIGA model was tested using a strict partitioning strategy, meaning that there was no overlap between cell lines or drugs in the training, validation, and test sets. 25-fold cross-validation was used. For comparison with the MCA model, the same training and test sets were used. The gene expression features in this dataset were RMA-normalized basal expression profiles, and the test set contained 1517 cell line-drug pairs. The decision variable was the normalized LN IC50. The test results are shown in Table 2.

[0062] Table 2 Comparison of model performance under strict partitioning strategy

[0063]

[0064] As can be seen from Table 2, the EIGA model performs significantly better than the MCA model.

[0065] Furthermore, the EIGA model was tested using a strict cell line-only partitioning strategy, meaning there was no overlap between the training, validation, and test sets, although the drugs overlapped. The gene expression characteristics in this dataset were TPM-normalized RNA-seq profiles. Five-fold cross-validation was used, and the average of the predictions from each fold was taken as the final prediction. For comparison with the Precily model, the same training and test sets were used. The decision variable was LN IC50. The test results are shown in Table 3.

[0066] Table 3 compares the model performance under the strict cell line division strategy only.

[0067]

[0068] As can be seen from Table 3, the EIGA model performs better than the Precily model.

[0069] Finally, the EIGA model was tested based on single-cell expression profiles using a strict cell line-specific partitioning strategy. The gene expression features in the training and validation sets were TPM-normalized RNA-seq profiles, while those in the test set were CPM-normalized scRNA-seq profiles. Since the counting method for CPM data is independent of gene length, CPM and TPM data can be treated equally. Five-fold cross-validation was used, and the average of the predictions from each fold was taken as the final prediction. For comparison with the Precily model, the same training and test sets were used. The test set contained 17279 cell-drug pairs. The decision variable was LN IC50. The test results are shown in Table 4.

[0070] Table 4 shows the model performance on single-cell sequencing data under a strict cell line segmentation strategy only.

[0071]

[0072] As can be seen from Table 4, the EIGA model performs better than the Precily model.

[0073] In summary, the EIGA model achieves highly accurate predictions under a relaxed partitioning strategy, and also outperforms previous models under strict partitioning strategies and strict partitioning strategies targeting only cell lines. Furthermore, the model demonstrates good predictive performance on single-cell sequencing data.

[0074] like Figure 3 As shown, a cancer cell sensitivity prediction system that integrates gene network relationships specifically includes:

[0075] Original gene expression feature compression module: An autoencoder built using a fully connected layer with an activation function compresses the original gene expression features to obtain low-dimensional gene expression features, which are used in step 1 of a method for predicting the sensitivity of anti-cancer cells by fusing gene network relationships;

[0076] Gene interaction relationship encoding module: Perform graph autoencoding on gene interaction network to obtain gene interaction features containing gene network relationship information, which are used in step 2 of a method for predicting the sensitivity of anti-cancer cells by incorporating gene network relationships;

[0077] Information fusion module: Gene network features are obtained from fused gene expression features and gene interaction features, which are used in step 3 of a method for predicting the sensitivity of anti-cancer cells based on fused gene network relationships;

[0078] Drug compound molecule encoding module: obtains the characteristics of each atom in the drug compound molecule, which is used in step 4 of an anti-cancer cell sensitivity prediction method that integrates gene network relationships;

[0079] EIGA prediction module: Input the gene expression characteristics of the cell line, the gene network characteristics of the cell line, and the characteristics of each atom in the drug compound molecule to obtain the predicted anti-cancer cell sensitivity, which is used in step 5 of an anti-cancer cell sensitivity prediction method that integrates gene network relationships.

[0080] A device for predicting cancer cell sensitivity based on fusion gene network relationships, comprising:

[0081] Memory: Used to store computer programs;

[0082] A processor for executing the computer program to implement the method for predicting the sensitivity of cancer cells based on fused gene network relationships.

[0083] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or any conventional processor. The processor is the control center of the device for predicting the sensitivity of cancer cells by realizing fusion gene network relationships, and it connects various parts of the device using various interfaces and lines.

[0084] When the processor executes the computer program, it implements the steps of the above-mentioned method for predicting the sensitivity of anticancer cells by fusing gene network relationships. For example, it uses the vector composed of the expression levels of each gene in the cell line as the original gene expression features of the cell line, and uses an autoencoder built with a fully connected layer with an activation function to compress the original gene expression features into a low-dimensional latent vector of 64-1024 dimensions, and uses this low-dimensional latent vector as the gene expression features of the cell line; it performs graph autoencoding on the gene interaction network to obtain gene interaction features; it calculates gene network features based on gene expression features and gene interaction features; it establishes a molecular graph of drug compounds to obtain the features of each atom in the drug compound molecule; it obtains the predicted sensitivity of anticancer cells based on the EIGA model; and thus realizes the method for predicting the sensitivity of anticancer cells by fusing gene network relationships.

[0085] Alternatively, when the processor executes the computer program, it implements the functions of each module in the above system, such as: a raw gene expression feature compression module: using an autoencoder built with a fully connected layer with an activation function to compress raw gene expression features to obtain low-dimensional gene expression features; a gene interaction relationship encoding module: performing graph autoencoder on the gene interaction network to obtain gene interaction features containing gene network relationship information; an information fusion module: used to fuse gene expression features and gene interaction features to obtain gene network features; a drug compound molecule encoding module: used to obtain the features of each atom in the drug compound molecule; and an EIGA prediction module: used to input the gene expression features of the cell line, the gene network features of the cell line, and the features of each atom in the drug compound molecule to obtain the predicted anti-cancer cell sensitivity.

[0086] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a preset function, the instruction segments describing the execution process of the computer program in the device for predicting the sensitivity of cancer cells to fusion gene network relationships. For example, the computer program can be divided into a raw gene expression feature compression module; a gene interaction relationship encoding module; an information fusion module; a drug compound molecule encoding module; and an EIGA prediction module. The specific functions of each module are as follows: Raw gene expression feature compression module: using an autoencoder built with a fully connected layer with an activation function to compress raw gene expression features to obtain low-dimensional gene expression features; Gene interaction relationship encoding module: performing graph autoencoder on the gene interaction network to obtain gene interaction features containing gene network relationship information; Information fusion module: used to fuse gene expression features and gene interaction features to obtain gene network features; Drug compound molecule encoding module: used to obtain the features of each atom in the drug compound molecule; EIGA prediction module: used to input the gene expression features of the cell line, the gene network features of the cell line, and the features of each atom in the drug compound molecule to obtain the predicted anti-cancer cell sensitivity.

[0087] The device for predicting the sensitivity of cancer cells by realizing fusion gene network relationships can be a desktop computer, laptop, handheld computer, or cloud server, etc. The device may include, but is not limited to, processors and memory. Those skilled in the art will understand that the above is an example of a device for predicting the sensitivity of cancer cells by realizing fusion gene network relationships, and does not constitute a limitation on such a device. It may include more components than described above, or combine certain components, or use different components. For example, the device may also include input / output devices, network access devices, buses, etc.

[0088] The memory can be used to store the computer program and / or modules, and the processor implements various functions of the device for predicting the sensitivity of anti-cancer cells by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory.

[0089] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function (such as sound playback or image playback). The data storage area may store data created based on the use of the phone (such as audio data or a phonebook). Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMediaCards (SMC), Secure Digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0090] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for predicting the sensitivity of cancer cells based on fused gene network relationships.

[0091] If the integrated module / unit of the system for predicting the sensitivity of anti-cancer cells by fusing gene network relationships is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0092] This invention implements all or part of the process in the above-described method for predicting the sensitivity of cancer cells based on fused gene network relationships. It can also be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program implements the steps of the above-described method for predicting the sensitivity of cancer cells based on fused gene network relationships. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0093] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0094] It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable storage medium does not include electrical carrier signals and telecommunication signals.

[0095] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

Claims

1. A method for predicting anticancer cell sensitivity by fusing gene network relationships, characterized by, The method comprises the following steps: Step 1: taking the vector composed of the expression amounts of genes of the cell lines as the original gene expression characteristics of the cell lines, using the autoencoder built by the fully connected layer with an activation function to compress the original gene expression characteristics into a low-dimensional hidden vector of 64-1024 dimensions, and taking the low-dimensional hidden vector as the gene expression characteristics of the cell lines; Step 2: graph autoencoding the gene interaction network to obtain gene interaction characteristics; Step 3: calculating the gene network characteristics of the cell lines according to the gene expression characteristics of the cell lines in step 1 and the gene interaction characteristics in step 2; The specific process is as follows: taking 2%-50% of the genes with the highest original gene expression characteristic values of each cell line as a high-expression gene subset of the cell line, and taking the mean value of the gene interaction characteristics of all genes in the high-expression gene subset to obtain the gene network characteristics of the cell line; Step 4: taking the atoms of the drug compound molecules as nodes and the chemical bonds as edges to obtain a drug compound molecule graph, inputting the drug compound molecule graph into a molgnet model to obtain the characteristics of each atom in the drug compound molecule; Step 5: inputting the gene expression characteristics of the cell lines in step 1, the gene network characteristics of the cell lines in step 3 and the characteristics of each atom in the drug compound molecule in step 4 into an EIGA model to obtain the predicted anticancer cell sensitivity; The specific steps are as follows: Step 5.1: inputting the characteristics of each atom in the drug compound molecule and the gene expression characteristics mapped through a linear layer transformation and an activation function into a multi-head attention network A to obtain a molecular representation E; inputting the characteristics of each atom in the drug compound molecule and the gene network characteristics mapped through a linear layer transformation and an activation function into a multi-head attention network B to obtain a molecular representation I, and adding the molecular representation E and the molecular representation I to obtain a molecular overall representation; The output of the multi-head attention network A and B is calculated as follows: wherein d k is the dimension of the vector representing each atom in the molecule; the matrix formed by the series of vectors representing each atom in the molecule is obtained by two independent linear layers to obtain the matrix K, V; the gene expression features or gene network features are obtained by a linear layer with an activation function and a linear layer without an activation function to obtain the matrix Q; Step 5.2: inputting the molecular overall representation and the gene expression characteristics and gene network characteristics obtained by concatenating the gene expression characteristics and gene network characteristics mapped through a linear layer transformation and an activation function into a series of fully connected layers with an activation function to obtain the predicted anticancer cell sensitivity.

2. A system for predicting anticancer cell sensitivity by fusing gene network relationships, comprising: It comprises: An original gene expression characteristic compression module: an autoencoder built by a fully connected layer with an activation function is used to compress the original gene expression characteristics to obtain low-dimensional gene expression characteristics; A gene interaction relationship encoding module: graph autoencoding the gene interaction network to obtain gene interaction characteristics containing gene network relationship information; An information fusion module: used for fusing gene expression characteristics and gene interaction characteristics to obtain gene network characteristics; A drug compound molecule encoding module: used for obtaining the characteristics of each atom in the drug compound molecule; An EIGA prediction module: used for inputting the gene expression characteristics of the cell lines, the gene network characteristics of the cell lines and the characteristics of each atom in the drug compound molecule to obtain the predicted anticancer cell sensitivity.

3. An apparatus for implementing anticancer cell sensitivity prediction using fusion gene network relationship, comprising: It comprises: A memory: used for storing a computer program for implementing the method for predicting the anticancer cell sensitivity by fusing gene network relationships according to claim 1. A processor for implementing the method of predicting sensitivity of anticancer cells to a fusion gene network relationship according to claim 1 when the computer program is executed.

4. A computer-readable storage medium, characterized in that, Comprise: The computer readable storage medium stores a computer program, and the computer program can implement the method of predicting sensitivity of anticancer cells to a fusion gene network relationship according to claim 1 when the processor is executed.

Images