A gene perturbation drug relocation method and system based on deep autoencoders

Abstract

This invention provides a gene perturbation drug relocation method and system based on a deep autoencoder, comprising: acquiring and standardizing gene expression profile data of disease and drug-perturbed cell lines; training a deep autoencoder model based on the standardized disease gene expression profile data; screening key gene targets for disease treatment based on the trained deep autoencoder model; and calculating drug importance scores based on the key gene targets for disease treatment and the standardized drug-perturbed cell line gene expression profile data to achieve drug relocation. This invention learns the nonlinear relationships between genes through a deep autoencoder, laying the foundation for screening key disease genes at the gene expression perturbation level; and proposes a novel drug relocation method based on key disease genes and drug-perturbed gene expression data, accelerating the drug development process.

Description

Technical Field

[0001] This invention belongs to the field of drug relocation technology, specifically relating to a gene perturbation drug relocation method and system based on a deep autoencoder. Background Technology

[0002] Drug development is not only time-consuming and costly, but also inherently risky. To overcome these challenges, drug repositioning has emerged as an effective strategy for discovering new indications for approved drugs, offering significant advantages in accelerating the drug development process. Drug repositioning, as a crucial strategy for innovative drug development, can significantly improve development efficiency by identifying new therapeutic indications for marketed or clinical-stage drug candidates. The drug development cycle based on drug repositioning strategies can be shortened to 3-12 years, saving approximately 60%-75% of time compared to traditional drug development processes. To date, this strategy has achieved numerous successes globally: the classic example of sildenafil successfully transforming from a cardiovascular disease treatment into Viagra for erectile dysfunction, and the successful expansion of artemisinin from an antimalarial drug to an anti-tuberculosis treatment, discovered by Chinese scientist Professor Tu Youyou, demonstrating its unique advantage in delaying the evolution of tuberculosis resistance, are both exemplary cases of successful drug repositioning. Since repositioned drugs have already completed key stages such as systematic preclinical research, safety evaluation, and formulation process development, this strategy can significantly reduce R&D risks, shorten the R&D cycle (saving an average of 5-7 years), and substantially reduce R&D investment (saving approximately 40%-60% in costs). Based on existing research data, the risk-benefit ratio of drug repositioning is significantly better than that of traditional drug development models, making it one of the most valuable drug development strategies for clinical application and translation.

[0003] The CMap database, a crucial resource for pharmacogenomics research, has collected over 6100 gene expression profiles of more than 1300 small molecule drugs under different dosage treatments in five human tumor cell lines (MCF7, ssMCF7, HL60, PC3, and SKMEL5). This large-scale dataset provides essential experimental data support for drug retargeting studies based on gene expression perturbations. Studies by Lamb et al. have demonstrated that the CMap database can effectively identify potential disease-related therapeutics. In addition to providing massive gene expression data, CMap innovatively employs a KS (Kolmogorov-Smirnov) statistic combined with nonparametric rank analysis to establish a gene expression pattern matching strategy, providing a reliable computational framework for drug retargeting. Drug relocation studies based on the CMap cell line drug perturbation gene expression profile dataset have made many breakthroughs, but the following challenges still exist: 1) Drug perturbation genes have low overlap with disease-related genes, making it difficult to directly assess the impact of drugs on disease genes from the perspective of overlapping genes; 2) There are complex nonlinear regulatory relationships between genes, and there is currently a lack of appropriate drug relocation models to describe the changes in the expression of other genes after a specific gene is intervened and perturbed. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a gene perturbation drug relocation method and system based on deep autoencoders, thereby accelerating the drug development process.

[0005] A gene perturbation drug relocation method based on a deep autoencoder includes:

[0006] Acquire and standardize disease gene expression profile data and drug-perturbed cell line gene expression profile data;

[0007] A deep autoencoder model was trained based on the standardized disease gene expression profile data.

[0008] Based on a trained deep autoencoder model, key gene targets for disease treatment are screened.

[0009] Based on key gene targets for disease treatment and standardized drug-perturbed cell line gene expression profiles, drug importance scores are calculated to achieve drug repositioning.

[0010] Preferably, the deep autoencoder model employs a stacked autoencoder constructed from fully connected layers, the autoencoder comprising an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer; wherein a scaling exponential linear unit activation function is introduced between the input layer and the third hidden layer; and the output layer employs a linear function.

[0011] Preferably, the method for screening key gene targets for the treatment of the disease includes:

[0012] Statistical differences in gene expression profiles between the pre-defined disease sample group and the normal sample group were tested to obtain the p-value, which represents the significance of the expression difference for each gene.

[0013] Based on gene expression profile data from disease and normal sample groups, the fold change in expression of each gene was obtained. And calculate the logarithm of the fold change in expression for each gene. ;

[0014] Based on the p-value, a significance index for expression differences, and the fold change in gene expression, disease-related genes and normal genes are identified, and the identification results are obtained.

[0015] Based on the aforementioned judgment results and the logarithm of the fold change in expression for each gene. Construct a gene perturbation scoring function;

[0016] Based on the gene perturbation scoring function, the importance score of disease-related genes is calculated;

[0017] Based on the importance scores of disease-related genes and a trained deep autoencoder model, key gene targets for disease treatment are obtained.

[0018] Preferably, the gene perturbation scoring function is expressed as follows:

[0019] ,

[0020] in, Indicates that the gene is affected After regulation, give the gene Rate the impact it brings; and They represent genes respectively Before / after regulation, disease sample genes The p-value, which indicates the significance of gene expression differences compared to normal samples; and They represent genes respectively Before / after regulation The expression value; Representative gene i, express The logarithm of the fold change in gene expression.

[0021] Preferably, the method for calculating the drug importance score includes: calculating the drug importance score for the therapeutic effect of the drug based on the fold change in gene expression caused by the drug, the fold change in expression of disease-related genes, and key gene targets for disease treatment.

[0022] The present invention also provides a gene perturbation drug relocation system based on a deep autoencoder for implementing the method, comprising:

[0023] The data acquisition module is used to acquire and standardize disease gene expression profile data and drug-perturbed cell line gene expression profile data;

[0024] The model building module is used to train a deep autoencoder model based on the standardized disease gene expression profile data.

[0025] The target screening module is used to screen key gene targets for disease treatment based on a trained deep autoencoder model.

[0026] The drug scoring module is used to calculate drug importance scores based on key gene targets for disease treatment and standardized drug-perturbed cell line gene expression profile data, thereby enabling drug repositioning.

[0027] Preferably, in the model building module, the deep autoencoder model adopts a stacked autoencoder constructed from fully connected layers. The autoencoder includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer. A scaling exponential linear unit activation function is introduced between the input layer and the third hidden layer. The output layer adopts a linear function.

[0028] Preferably, the target screening module includes:

[0029] The indicator calculation unit is used to perform statistical difference tests on the gene expression profile data of the preset disease sample group and normal sample group to obtain the p-value of the expression difference significance index of each gene.

[0030] The logarithmic calculation unit is used to obtain the fold change in expression of each gene based on gene expression profile data from disease and normal sample groups. And calculate the logarithm of the fold change in expression for each gene. ;

[0031] Based on the expression abnormality judgment unit, the disease-related genes and normal genes are judged based on the p-value of the expression difference significance index and the fold change of gene expression, and the judgment result is obtained.

[0032] The scoring function construction unit is used to base the judgment result on the logarithm of the fold change in expression of each gene. Construct a gene perturbation scoring function;

[0033] The importance score calculation unit is used to calculate the importance score of disease-related genes based on the gene perturbation scoring function; the calculation formula for the importance score of disease-related genes is as follows:

[0034] ,

[0035] in, This indicates that by regulating genes The combined impact on the other 4730 genes, i.e., the importance score of disease-related genes, Indicates that the gene is affected disturbance The change score;

[0036] The target screening unit is used to obtain key gene targets for disease treatment based on the importance scores of disease-related genes and a trained deep autoencoder model.

[0037] Preferably, in the drug scoring module, the formula for calculating the drug importance score is as follows:

[0038] ,

[0039] in, This refers to the drug importance score, where j is the number of drug-perturbed genes. This refers to the fold change in gene expression caused by a drug. This refers to the fold change in the expression of disease-related genes. The score represents the importance of the disease-related gene m.

[0040] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a gene perturbation drug relocation method and system based on deep autoencoders. To systematically and dynamically evaluate the impact of drug intervention on disease genes, firstly, a nonlinear relationship model between genes is constructed using deep autoencoders based on gene expression data; then, by up- and down-regulating a certain input gene using this model, the expression changes of other genes are observed, and the regulatory efficacy index of the gene in the treatment of a certain disease, i.e., gene importance, is calculated; finally, by integrating changes in gene expression caused by drug perturbation, changes in disease gene expression, and gene importance index based on deep autoencoder perturbation, a drug relocation score is defined. This invention learns the nonlinear relationships between genes through deep autoencoders, laying the foundation for screening key disease genes at the level of gene expression perturbation; based on key disease genes and drug perturbation gene expression data, a novel drug relocation method is proposed, accelerating the drug development process. Signaling pathway and survival analyses show that the key genes screened by this invention do indeed participate in multiple signaling pathways closely related to disease occurrence and are significantly correlated with patient prognosis. Case studies and comparisons with other methods all demonstrate that our proposed method has very high accuracy. Attached Figure Description

[0041] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of the gene perturbation drug relocation method based on a deep autoencoder according to an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the structure of the depth autoencoder model in an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] The following is an explanation of some of the terms used in this invention:

[0047] KEGG Human Signaling Pathways: KEGG (Kyoto Encyclopedia of Genes and Genomes) is a comprehensive database that integrates genomic, chemical, and systemic functional information. The KEGG Pathway database specifically includes various biological pathway maps, including human signaling pathways.

[0048] Example 1

[0049] like Figure 1 As shown, a gene perturbation drug relocation method based on a deep autoencoder includes:

[0050] S1: Acquire and standardize disease gene expression profile data and drug-perturbed cell line gene expression profile data.

[0051] In this embodiment, gene expression profiles for various diseases (including disease groups and normal groups) were obtained from the public databases GEO and TCGA, and gene expression data of cell lines before and after drug treatment were obtained from the LINCS database. To avoid batch effects, the gene expression data were standardized using Z-scores.

[0052] (1)

[0053] in, This represents the expression value of a specific gene. The sample mean representing gene expression levels. The standard deviation is denoted as .

[0054] S2: Train a deep autoencoder model based on standardized disease gene expression profile data. For example... Figure 2 As shown. A further implementation method is that the deep autoencoder model adopts a stacked autoencoder constructed from fully connected layers. The autoencoder includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer; wherein, a scaling exponential linear unit activation function is introduced between the input layer and the third hidden layer; and the output layer adopts a linear function.

[0055] Specifically, a deep autoencoder model is trained based on gene expression profile data to infer nonlinear expression relationships between genes. In this invention, a stacked autoencoder is constructed using fully connected layers. The autoencoder structure consists of five layers: an input layer (4731 dimensions, corresponding to genes in the KEGG human signaling pathway), hidden layer 1 (the first hidden layer, 1500 dimensions, SELU activation), hidden layer 2 (the bottleneck layer, the second hidden layer, 1000 dimensions, SELU activation), hidden layer 3 (the third hidden layer, 1500 dimensions), and an output layer (4731 dimensions). The dimension of each layer in the neural network refers to the number of neurons. The specific explanation is as follows:

[0056] 1) The genes for the input and output layers are taken from the KEGG human signaling pathway. Genes appearing in the KEGG signaling pathway have complex regulatory relationships. To ensure the interpretability of the model, we construct a deep autoencoder based only on genes appearing in the signaling pathway to explore the nonlinear relationships between genes.

[0057] 2) A Scaled Exponential Linear Unit (SELU) activation function is introduced between the input layer and hidden layer 3. Its self-normalization property can alleviate the gradient vanishing problem. The mathematical form is:

[0058] (2)

[0059] Where λ = 1.0507 and α = 1.6733 are the default values ​​for the SELU function.

[0060] 3) The output layer (4,731 dimensions) uses a linear function and does not use an activation function.

[0061] 4) The loss function uses the mean absolute error (MAE):

[0062] (3)

[0063] in This indicates the number of samples, and 4731 indicates the number of genes. Indicates the first test set In the nth sample The true expression value of the gene (or the gene in the input layer) (the true value of each gene) Indicates the first test set In the nth sample The expression value of the nth gene predicted by the deep autoencoder (or the nth gene in the output layer) (Predicted values ​​for each gene). A smaller mean absolute error indicates a better fit of the deep autoencoder to the decoded genes.

[0064] During the training of this encoder, we use Mini-batch Stochastic Gradient Descent (Mini-batch SGD) as a widely used optimization algorithm in deep learning. The training set batch size is set to 100, and the training is conducted for 500 rounds. The model parameters are updated iteratively to minimize the objective function.

[0065] S3: Based on a trained deep autoencoder model, screen key gene targets for disease treatment.

[0066] Deep autoencoder models trained on S2 can simulate the mutual regulation and influence of gene expression. For example, the expression of a certain gene node in the input layer can be upregulated or downregulated, and the effect on other genes (i.e., changes in the expression of genes in the output layer) can be observed. In order to identify important target genes that have a positive effect on disease treatment, we comprehensively consider disease-related genes and the changes in the expression of other genes after being regulated by a certain gene.

[0067] A further implementation method includes screening key gene targets for disease treatment, comprising:

[0068] S31: Perform statistical difference tests on the gene expression profile data of the preset disease sample group and normal sample group to obtain the p-value of the expression difference significance index of each gene.

[0069] S32: Based on gene expression profile data from the disease sample group and the normal sample group, obtain the fold change in expression of each gene. And calculate the logarithm of the fold change in expression for each gene. :

[0070] (4)

[0071] in Indicates genes in disease sample groups The expression mean, Indicates genes in the normal sample group The mean expression value is obtained by taking the logarithm of the expression fold, and the positive or negative value of the logarithm corresponds to the up- or down-regulation of gene expression.

[0072] S33: Based on the p-value, a significant indicator of expression difference, and the fold change in gene expression. This involves identifying disease-related genes from normal genes and obtaining the results; specifically, This indicates significantly differential gene expression (or abnormal gene expression, i.e., disease-related genes), while the opposite indicates normal genes.

[0073] S34: Based on the aforementioned judgment results and the logarithm of the fold change in expression of each gene. This invention constructs a gene perturbation scoring function. From the perspective of disease treatment, the goal of treatment is to induce abnormally expressed genes to return to normal, that is, to reduce the expression of abnormally upregulated genes and increase the expression of abnormally downregulated genes. Based on this idea, this invention studies the positive effects of regulating a certain gene on changes in other genes. The specific gene perturbation scoring function expression is as follows:

[0074] (5)

[0075] in, Indicates that the gene is affected After regulation, give the gene Rate the impact it brings; and They represent genes respectively Before / after regulation, disease sample genes The p-value, which indicates the significance of gene expression differences compared to normal samples; and They represent genes respectively Before / after regulation The expression value; Representative gene i, express The logarithm of the fold change in gene expression. Specifically, if a gene... If differential expression is upregulated before regulation and downregulated after regulation, or differential expression is downregulated before regulation and upregulated after regulation, then 1 point is awarded; if the gene Before and after regulation, genes If all remain in normal condition, score 0; if the regulatory gene... Before, genes It is a normal gene, a regulatory gene. After that, genes If it becomes an abnormal gene, -1 point will be awarded.

[0076] S25: Calculate the importance score of disease-related genes based on the gene perturbation scoring function;

[0077] S26: Based on the importance scores of disease-related genes and a trained deep autoencoder model, key gene targets for disease treatment are obtained.

[0078] Specifically, for disease treatment, each differentially expressed gene is a potential therapeutic target. However, the benefits, i.e., the therapeutic effect, brought about by targeting different genes are certainly different. This invention represents the therapeutic importance score of a gene based on the comprehensive impact of a gene before and after regulation on other genes. The specific calculation method is shown in formula (6):

[0079] (6)

[0080] in This indicates that by regulating genes The combined impact on the other 4730 genes, i.e., the importance score of disease-related genes, Indicates that the gene is affected disturbance The change score (the calculation method is shown in Equation 5).

[0081] S4: Based on key gene targets for disease treatment, calculate drug importance scores to achieve drug repositioning.

[0082] A further implementation method involves calculating the drug importance score by: calculating the drug importance score for the therapeutic effect of the drug based on the fold change in drug-perturbed gene expression, the fold change in disease-related gene expression, and key gene targets for disease treatment. The drug-perturbed gene expression data used in this embodiment is based on cell line experiments.

[0083] Specifically, to identify disease-related drugs, after determining the key disease targets, rules were designed to calculate drug importance scores. The perturbation of gene expression by drugs in different cell lines was obtained from the LINCS dataset, and then the drug importance score for the therapeutic effect on the disease was calculated. The formula is as follows:

[0084] (7)

[0085] in This refers to the drug importance score; a higher score indicates that the drug has therapeutic potential for the disease, and j represents the number of genes that the drug perturbs. This refers to the fold change in gene variation caused by a drug. This refers to the fold change in disease-related genes. This represents the importance score of the disease-related gene m, obtained by perturbing the disease sample genes and performing correlation analysis. The above is the rule for calculating drug importance scores in this invention. Combined with the importance score of disease-related genes, the importance score of disease-related drugs is obtained, thereby achieving the purpose of drug repositioning.

[0086] Example 2

[0087] This invention also provides a gene perturbation drug relocation system based on a deep autoencoder, which includes a method for implementing the method, comprising:

[0088] The data acquisition module is used to acquire and standardize disease gene expression profile data and drug-perturbed cell line gene expression profile data;

[0089] The model building module is used to train a deep autoencoder model based on standardized disease gene expression profile data.

[0090] The target screening module is used to screen key gene targets for disease treatment based on a trained deep autoencoder model.

[0091] The drug scoring module is used to calculate drug importance scores based on key gene targets for disease treatment and standardized drug-perturbed cell line gene expression profile data, thereby enabling drug repositioning.

[0092] A further implementation method is that, in the model building module, the deep autoencoder model adopts a stacked autoencoder constructed from fully connected layers. The autoencoder includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer. A scaling exponential linear unit activation function is introduced between the input layer and the third hidden layer. The output layer adopts a linear function.

[0093] A further implementation method includes a target screening module comprising:

[0094] The indicator calculation unit is used to perform statistical difference tests on the gene expression profile data of the preset disease sample group and normal sample group to obtain the p-value of the expression difference significance index of each gene.

[0095] The logarithmic calculation unit is used to obtain the fold change in expression of each gene based on gene expression profile data from disease and normal sample groups. And calculate the logarithm of the fold change in expression for each gene. ;

[0096] Based on the expression abnormality judgment unit, and based on the p-value (significance index of expression difference) and the fold change in gene expression, To determine whether a disease-related gene is a normal gene and to obtain the determination result;

[0097] The scoring function building block is used to construct the score based on the judgment result and the logarithm of the fold change in expression of each gene. Construct a gene perturbation scoring function;

[0098] The importance score calculation unit is used to calculate the importance score of disease-related genes based on the gene perturbation scoring function; the calculation formula for the importance score of disease-related genes is as follows:

[0099] ,

[0100] in, This indicates that by regulating genes The combined impact on the other 4730 genes, i.e., the importance score of disease-related genes, Indicates that the gene is affected disturbance The change score;

[0101] The target screening unit is used to obtain key gene targets for disease treatment based on the importance scores of disease-related genes and a trained deep autoencoder model.

[0102] A further implementation method involves using the following formula to calculate the drug importance score in the drug scoring module:

[0103] ,

[0104] in, This refers to the drug importance score, where j is the number of drug-perturbed genes. This refers to the fold change in gene expression caused by a drug. This refers to the fold change in the expression of disease-related genes. The score represents the importance of the disease-related gene m.

[0105] Example 3

[0106] This embodiment provides a specific example analysis process:

[0107] (1) Data Preparation: Based on the Cancer Genome Atlas (TCGA) and GEO data platforms, this invention collected transcriptome sequencing data for three types of solid tumors, including breast cancer, lung cancer, and colorectal cancer. Breast cancer data can be further divided into four subtypes: Basal, Her2, LumA, and LumB. The gene expression profile data for the three cancers are compiled and obtained, as shown in Table 1:

[0108] surface

[0109]

[0110] (2) Data preprocessing: The obtained gene expression data were Z-score standardized according to S1 to facilitate the integration and analysis of cross-platform data.

[0111] (3) Training of deep autoencoder models for gene expression: Deep autoencoder models for gene expression were trained for three different types of cancer. Five-fold cross-validation was used during training, and the mean error loss (MAE) loss function was adopted. The model was optimized using mini-batch stochastic gradient descent (Mini-batch SGD, batch size 100). The best prediction accuracy was 0.156 on the training set and 0.256 on the test set.

[0112] (4) Screening of key targets: Based on the S3 calculation process, key therapeutic targets for breast cancer, lung cancer, and colorectal cancer were screened. The signaling pathway enrichment analysis of the top 50 key genes showed that these genes were significantly enriched in cancer-related signaling pathways, including the cell cycle, P53 signaling pathway, and pathways in cancer. This indicates that the key genes screened in this invention do indeed have important biological significance.

[0113] (5) Drug Repositioning Prediction: Based on drug-perturbed gene expression profile data and gene expression profile data of three types of cancer, and using a drug repositioning importance scoring model, we performed predictive analysis on potential drugs for the three types of cancer. Different cell lines were ranked according to their drug therapeutic potential scores, and the proportion of drugs ranked in the top n (n=10, 20, 30, 40, 50) with known therapeutic associations with the target disease was statistically analyzed to quantify the accuracy of drug repositioning. Experimental results in the MCF7 cell line (human breast cancer cell line) for various breast cancer subtypes showed that when the ranking threshold was set to the top 10, the drug repositioning accuracy was as high as 100% for the Basal and Her2 subtypes; when the ranking threshold was set to the top 20, the drug repositioning accuracy also reached 90%, verifying the effectiveness of the model in the field of drug repositioning. Similarly, in the analysis of colorectal cancer and lung cancer samples, most of the top-ranked drugs could be found to have relevant biological evidence. Table 2 shows the proportion of cancer-related drugs among the top n drugs (MCF7); Table 3 shows the proportion of cancer-related drugs among the top n drugs in different cell lines for colorectal cancer samples; Table 4 shows the proportion of cancer-related drugs among the top n drugs in different cell lines for lung cancer samples; Table 5 shows the data compared with other methods (bold indicates best performance).

[0114] Table 2

[0115]

[0116] Table 3

[0117]

[0118] Table 4

[0119]

[0120] Table 5

[0121]

[0122] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for drug repositioning of gene perturbation based on deep autoencoder, characterized in that, The method comprises the following steps: obtaining and standardizing disease gene expression profile data and drug perturbed cell line gene expression profile data; training a deep autoencoder model based on the standardized disease gene expression profile data; screening disease treatment key gene targets based on the trained deep autoencoder model; calculating drug importance scores based on the disease treatment key gene targets and the standardized drug perturbed cell line gene expression profile data to realize drug repositioning; The method for screening the disease treatment key gene targets comprises the following steps: performing statistical difference test on gene expression profile data of a preset disease sample group and a normal sample group to obtain an expression difference significance index p value of each gene; Based on the gene expression profile data of the disease sample group and the normal sample group, the expression change fold of each gene is obtained , and the logarithm of the expression change fold of each gene is calculated ; based on the expression difference significance index p value and the gene expression change fold judging disease-related genes and normal genes to obtain a judgment result; based on the result of the judging and the log of the change fold of each gene expression , constructing a gene perturbation scoring function; calculating disease-related gene importance scores based on the gene perturbation scoring function; obtaining disease treatment key gene targets based on the disease-related gene importance scores and the trained deep autoencoder model; The method for calculating the drug importance scores comprises the following steps: calculating drug importance scores of the drugs for disease treatment based on drug perturbed gene expression change fold, disease-related gene expression change fold and disease treatment key gene targets.

2. The method of claim 1, wherein, The deep autoencoder model adopts a stacked autoencoder constructed by full connection layers, and the autoencoder comprises an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer; wherein a scaled exponential linear unit activation function is introduced between the input layer and the third hidden layer; and the output layer adopts a linear function.

3. The method of claim 1, wherein, The gene perturbation scoring function expression is as follows: ， wherein, represents the gene regulated by the gene , and the impact score of the gene ; and respectively represent the gene regulated by the gene , and the expression difference significance indicator p value of the gene i of the disease sample compared with the gene i of the normal sample before / after the gene ; and respectively represent the expression value of the gene before / after the gene is regulated; represents the gene i, represents the logarithm of the gene expression change multiple of the gene i.

4. A deep autoencoder-based genetic perturbation drug repositioning system for implementing the method of any one of claims 1-3, characterized in that, The method comprises the following steps: a data acquisition module for acquiring and standardizing disease gene expression profile data and drug perturbed cell line gene expression profile data; a model construction module for training a deep autoencoder model based on the standardized disease gene expression profile data; a target screening module for screening disease treatment key gene targets based on the trained deep autoencoder model; a drug scoring module for calculating drug importance scores based on the disease treatment key gene targets and the standardized drug perturbed cell line gene expression profile data to realize drug repositioning.

5. The system of claim 4, wherein, In the model construction module, the deep autoencoder model adopts a stacked autoencoder constructed by full connection layers, and the autoencoder comprises an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer; wherein a scaled exponential linear unit activation function is introduced between the input layer and the third hidden layer; and the output layer adopts a linear function.

6. The system of claim 4, wherein, The target screening module comprises the following steps: an index calculation unit for performing statistical difference test on gene expression profile data of a preset disease sample group and a normal sample group to obtain an expression difference significance index p value of each gene; A logarithm calculation unit is configured to obtain the expression change fold of each gene based on the gene expression profile data of the disease sample group and the normal sample group and calculate the logarithm of the expression change fold of each gene ; a based on expression abnormality judgment unit for judging disease-related genes and normal genes based on the expression difference significance index p value and gene expression change fold to obtain a judgment result; The scoring function construction unit is configured to construct a gene perturbation scoring function based on the judgment result and the logarithm of the expression change fold of each gene. , construct a gene perturbation scoring function; an importance score calculation unit for calculating disease-related gene importance scores based on the gene perturbation scoring function; and the calculation formula of the disease-related gene importance scores is as follows: ， wherein, represents the perturbation of the gene The combined effect on the other 4730 genes, i.e. the disease-related gene importance score, represents the change score of the perturbation of the gene by the gene . The target screening unit is used for obtaining disease treatment key gene targets based on disease related gene importance scores and a trained deep auto-encoder model.

7. The system of claim 4, wherein, In the drug scoring module, the calculation formula of the drug importance score is as follows: ， wherein, denotes the drug importance score, j is the number of drug perturbed genes, denotes the drug perturbed gene expression change fold, denotes the disease associated gene expression change fold, denotes the disease associated gene m importance score.

Images