Drug screening analysis methods, media and equipment based on single-cell transcriptome profiles

Abstract

The present application discloses a drug screening analysis method, medium and equipment based on single-cell transcriptome maps, which relates to the field of single-cell transcriptome data analysis. Constructing a single-cell pan-tumor transcriptome map containing cancer tissues of various cancer types, corresponding paracancerous tissues and corresponding tissues from normal human bodies is helpful to discover common drug targets across different cancer types; through re-clustering and annotation, it is possible to accurately distinguish different cell types, especially the differences between malignant cells and normal cells, which is helpful to discover drugs for specific cell types; based on the single-cell pan-tumor transcriptome map, combined with the pre-processed cell perturbation data matrix, the deep learning model is trained to achieve accurate prediction of cell response under drug treatment, which can quickly identify pan-tumor drugs and tumor-specific drugs, discover new indications for marketed drugs and candidate drugs, evaluate the side effects of drugs, and discover the potential mechanism of action of drugs.

Description

Technical Field

[0001] The present application relates to the field of single-cell transcriptome data analysis, and in particular to a drug screening analysis method, medium and equipment based on single-cell transcriptome maps. Background Art

[0002] In the field of cancer research, the heterogeneity of tumors and their interaction with the microenvironment are key to understanding cancer occurrence, development, and clinical treatment effects. Tumors are not just composed of a single population of cancer cells, but a complex ecosystem composed of multiple cell types and their extracellular matrix. This heterogeneity plays a vital role in tumor progression, treatment response, and drug resistance.

[0003] However, current large-scale screening methods for anti-tumor drugs still have significant defects. On the one hand, most of these methods rely on traditional transcriptome sequencing data and cannot fully capture the heterogeneity within the tumor and its interaction with the microenvironment; on the other hand, many studies are still limited to a single cancer type, ignoring the common features contained in pan-tumor data. These limitations not only affect the accuracy and effectiveness of drug screening, but also limit the discovery and clinical application of new treatment strategies. Summary of the invention

[0004] In order to solve at least one of the technical problems mentioned in the background technology, the purpose of the present application is to provide a drug screening analysis method, medium and equipment based on single-cell transcriptome mapping, which can realize large-scale screening of anti-tumor drugs at the single-cell level and analyze the potential mechanism of action of the drugs and their side effects.

[0005] To achieve the above objectives, this application provides the following technical solutions:

[0006] In a first aspect, the present application provides a drug screening analysis method based on a single cell transcriptome profile, comprising the following steps:

[0007] Constructing a single-cell pan-tumor transcriptome map, wherein the single-cell pan-tumor transcriptome map includes cancer tissues of various cancer types, corresponding paracancerous tissues, and single-cell transcriptome sequencing count matrices from corresponding tissues of normal human bodies, as well as annotation information matrices of single cells;

[0008] The annotation information matrix of the single cell is obtained by grouping cells and annotating them based on cell type-specific genes based on the single-cell transcriptome sequencing count matrix; the annotation information matrix of the single cell contains information on the tissue origin, cell type and corresponding cell lineage of each cell in the single-cell pan-tumor transcriptome map;

[0009] Acquiring cell perturbation data and screening the cell perturbation data to obtain a pre-processed cell perturbation data matrix, wherein the pre-processed cell perturbation data matrix includes a plurality of significantly differentially expressed genes and corresponding drug treatment information;

[0010] Constructing a deep learning model and training the deep learning model using the single-cell pan-tumor transcriptome map and the pre-processed cell perturbation data matrix;

[0011] Use the trained deep learning model to evaluate the impact of each drug treatment on each cell, and search for pan-tumor drugs and tumor-specific drugs, as well as potential side effects of drugs.

[0012] Furthermore, the step of obtaining the annotation information matrix of the single cell includes:

[0013] Obtaining single-cell transcriptome sequencing count matrices of cancer tissues of various cancer types, corresponding paracancerous tissues, and corresponding tissues from normal human bodies, grouping each sample, and using lineage-specific genes to initially group cells and annotate and divide them into multiple cell lineages; the multiple cell lineages include epithelial cell lineages;

[0014] Identify and annotate malignant cells in the epithelial cell lineage in the initial clustering;

[0015] The cell types belonging to the same cell lineage in all samples are grouped together, and the cells derived from the same cell lineage are regrouped. Based on cell type-specific genes, the cell types are finely annotated at the cell lineage level to obtain an annotation information matrix for single cells, where the cell lineage includes malignant cells, epithelial cells, stromal cells, endothelial cells, myeloid cells, and lymphoid cells.

[0016] Furthermore, the step of identifying and annotating malignant cells in the epithelial cell lineage in the initial grouping includes:

[0017] checking whether the gene expression level of a specific cancer marker in the epithelial cells is greater than a first threshold;

[0018] Using a copy number variation analysis tool to calculate whether the copy number variation score of each epithelial cell is greater than a second threshold;

[0019] If the epithelial cells meet the gene expression level of a specific cancer marker greater than a first threshold and / or the copy number variation score greater than a second threshold, they are annotated as malignant cells.

[0020] Furthermore, the step of acquiring the cell disturbance data and screening the cell disturbance data to obtain a pre-processed cell disturbance data matrix includes:

[0021] Acquire cell perturbation data from a public database or experimental data, wherein the cell perturbation data includes single-cell transcriptome changes before and after drug treatment;

[0022] According to preset filtering conditions, high-confidence cell perturbation data entries are screened out from the cell perturbation data;

[0023] Each cell perturbation data entry is sorted, and the significantly differentially expressed genes ranked before the preset threshold are selected to generate the preprocessed cell perturbation data matrix.

[0024] Furthermore, the step of constructing a deep learning model and training the deep learning model using the single-cell pan-tumor transcriptome map and the pre-processed cell perturbation data matrix includes:

[0025] Constructing a deep learning model, wherein the deep learning model adopts a variational autoencoder structure;

[0026] The single-cell pan-tumor transcriptome map is divided into training data and validation data, which are respectively integrated with the preprocessed cell perturbation data matrix to obtain a training set and a validation set. Based on the training set and the validation set, the deep learning model is trained until the loss function converges to obtain a trained deep model; the training set and the validation set are both binary matrices containing the perturbation change characteristics of a single cell.

[0027] Furthermore, the step of using the trained deep learning model to evaluate the effect of each drug treatment on each cell includes:

[0028] The trained deep learning model is used to calculate the potential score and potential direction of each cell in response to a specific drug treatment; the potential score represents the intensity of the cell's response to the specific drug treatment; the potential direction represents the change trend of the cell in response to the specific drug treatment;

[0029] Based on the potential scores and potential directions, the therapeutic effects and potential side effects of different drugs on specific cancer types are evaluated, and drug screening is performed.

[0030] Furthermore, it also includes:

[0031] Based on the potential score and potential direction of each cell in response to a specific drug treatment, a Bayesian test was used to calculate the effect of drug treatment on each cell type, generating an enrichment score matrix of cell types and drug treatments, with each row representing a cell type and each column representing a drug treatment.

[0032] In a second aspect, an embodiment of the present application provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned drug screening and analysis method based on single-cell transcriptome mapping.

[0033] In a third aspect, an embodiment of the present application provides a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned drug screening and analysis method based on single-cell transcriptome mapping when executing the computer program.

[0034] Compared with the prior art, the beneficial effects of the present application are: constructing a single-cell pan-tumor transcriptome map containing cancer tissues of multiple cancer types, corresponding adjacent cancer tissues and corresponding tissues from normal human bodies, which is helpful to discover common drug targets across different cancer types; through re-clustering and annotation based on cell type-specific genes, it is possible to accurately distinguish different cell types, especially the differences between malignant cells and normal cells, thereby improving the accuracy and specificity of drug screening and helping to discover drugs targeting specific cell types; based on the single-cell pan-tumor transcriptome map, combined with the pre-treated cell perturbation data matrix, a deep learning model is trained to achieve accurate prediction of cell response under drug treatment, which can quickly identify pan-tumor drugs and tumor-specific drugs, discover new indications for marketed drugs and candidate drugs, evaluate the side effects of drugs, and discover the potential mechanism of action of drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flowchart of a drug screening analysis method based on a single-cell transcriptome profile provided in Example 1 of the present application;

[0036] Figure 2 A UMAP projection diagram of training data predicted based on a pre-trained deep learning model provided in Example 1 of the present application;

[0037] Figure 3 A UMAP projection map of the entire single-cell pan-tumor transcriptome map predicted based on a pre-trained deep learning model provided in Example 1 of the present application;

[0038] Figure 4 The number of repetitions of the compounds corresponding to the 10 most significant drug treatments in each cell type in a single-cell pan-tumor transcriptome map provided in Example 2 of the present application;

[0039] Figure 5 The number of repetitions of the 10 most significant drug treatments in each cell type in a stromal cell and endothelial cell lineage provided in Example 2 of the present application;

[0040] Figure 6 The latent space score corresponding to a tumor-associated fibroblast in drug treatment LJP008_SKL_24H:G19_DOWN and MOA001_U2OS_24H:P04_DOWN provided in Example 2 of the present application;

[0041] Figure 7 The latent space scores corresponding to all cells in a single-cell pan-tumor transcriptome map in drug treatment CPC006_PC3_24H:J14_DOWN and CPC006_HT29_6H:D19_UP provided in Example 2 of the present application;

[0042] Figure 8 The corresponding latent space scores of tumor-associated fibroblasts and tumor-associated endothelial cells in drug treatment CPC006_PC3_24H:J14_DOWN and CPC006_HT29_6H:D19_UP provided in Example 2 of the present application;

[0043] Fig. 9 A potential spatial score of a drug treatment enriched in lung-derived malignant cells in the entire epithelial cell lineage provided in Example 3 of the present application;

[0044] Fig.10 The latent space scores corresponding to all cells in a single-cell pan-tumor transcriptome map under drug treatment LPROT004_YAPC_6H:BRD-A14634327:1_DOWN and PBIOA018_A549_24H:M02_DOWN provided in Example 3 of the present application;

[0045] Fig.11 The latent space scores corresponding to cells derived from the lung in drug treatment LPROT004_YAPC_6H:BRD-A14634327:1_DOWN and PBIOA018_A549_24H:M02_DOWN provided in Example 3 of the present application;

[0046] Fig.12 The cell viability of an A549 cell line provided in Example 3 of the present application after being treated with compounds Volasertib, Prednisolone and Fostamatinib for different time periods;

[0047] Fig.13 The absolute value and ranking of the gene score corresponding to drug treatment PBIOA018_A549_24H:M02_DOWN in a decoder provided in Example 4 of the present application;

[0048] Fig.14 The latent space scores corresponding to all cells in a single-cell pan-tumor transcriptome map and malignant cells derived from the liver in drug treatment CPC006_A549_24H:BRD-K56343971-001-02-3:10_UP provided in Example 5 of the present application;

[0049] Fig.15 The latent space scores corresponding to all cells in a single-cell pan-tumor transcriptome map in drug-treated ASG003_MCF7_48H:E07_UP and the 16 cell types with the highest scores provided in Example 5 of the present application;

[0050] Fig.16 The latent space scores corresponding to all cells in a single-cell pan-tumor transcriptome map in drug treatment LJP005_SKBR3_24H:F19_DOWN and the 16 cell types with the highest scores provided in Example 5 of the present application;

[0051] Fig.17 The corresponding latent space scores of tumor-associated fibroblasts and malignant cells derived from the liver under drug treatment ASG003_MCF7_48H:E07_UP and LJP005_SKBR3_24H:F19_DOWN provided in Example 5 of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0053] In order to solve the above problems existing in the prior art, the embodiments of the present application disclose a drug screening analysis method, medium and equipment based on single-cell transcriptome mapping, which are described in detail below.

[0054] Embodiment 1:

[0055] like Figure 1 As shown, Figure 1 A flowchart of a drug screening analysis method based on a single cell transcriptome profile is provided in Example 1 of the present application. The method comprises:

[0056] A single-cell pan-tumor transcriptome map was constructed, which included cancer tissues of various cancer types, corresponding adjacent cancer tissues, and single-cell transcriptome sequencing count matrices and annotation information matrices of single cells from corresponding tissues of normal humans.

[0057] The annotation information matrix of a single cell is obtained by grouping cells and annotating them based on cell type-specific genes based on the single-cell transcriptome sequencing count matrix; the annotation information matrix of a single cell contains the information of the tissue origin, cell type and corresponding cell lineage of each cell in the single-cell pan-tumor transcriptome map. This step can be subdivided into the following steps:

[0058] Single-cell transcriptome sequencing count matrices of cancer tissues of various cancer types, corresponding adjacent cancer tissues, and corresponding tissues from normal humans were obtained, and each sample was clustered. Lineage-specific genes were used to initially cluster the cells and annotate them into five cell lineages, including epithelial cells, stromal cells, endothelial cells, myeloid cells, and lymphoid cells.

[0059] Seurat software can be used to merge, normalize, logarithmize, reduce dimension and cluster the single-cell transcriptome sequencing data of each patient's cancer and adjacent adjacent samples to obtain cell clusters for a single patient. Based on the differentially expressed genes of each cell cluster, these cell clusters are divided into five major cell lineages. Among them, the epithelial cell lineage mainly expresses EPCAM and KRT, the endothelial cell lineage mainly expresses ENG, VWF and PLVAP, the stromal cell lineage mainly expresses DCN and COL1A1, the lymphoid cell lineage mainly expresses CD3D, CD8A, FOXP3, MS4A1, IGKC and JCHAIN, and the myeloid cell lineage mainly expresses CD1C, IL1B, C1QA, S100A8 and TPSAB1.

[0060] Identify and annotate malignant cells in the epithelial lineage in the initial cohort.

[0061] For the epithelial cell lineage, check whether the gene expression level of specific cancer markers in epithelial cells is greater than a first threshold; use the copy number variation analysis tool to calculate whether the copy number variation score of each epithelial cell is greater than a second threshold; if the epithelial cell meets the requirement that the gene expression level of the specific cancer marker is greater than the first threshold and / or the copy number variation score is greater than the second threshold, it is annotated as a malignant cell.

[0062] Three methods are used simultaneously to jointly determine whether epithelial cells are malignant cells. First, for each sample's epithelial cell lineage, each cell population is checked to see whether it highly expresses genes that are significantly highly expressed in malignant tissues relative to normal tissues in the corresponding cancer. For example, VDAC1 is highly expressed in malignant tissues in bladder cancer, RNF43 is highly expressed in colorectal cancer, GPC3 is highly expressed in hepatocellular carcinoma, CD151 is highly expressed in intrahepatic cholangiocarcinoma, CEACAM6 is highly expressed in lung adenocarcinoma, and LAMC2 is highly expressed in pancreatic cancer. Second, using epithelial cells in normal human samples corresponding to cancer tissues in the Human Cell Landscape (HCL) as a reference, the inferCNV software is used to infer the copy number variation scores of all epithelial cells in each patient. Potential malignant cells should have higher inferCNV scores compared to normal epithelial cells derived from HCL and epithelial cells derived from adjacent tissues. Third, epithelial cells derived from cancer tissues that have both high expression of malignant tissue gene markers and high inferCNV scores are considered potential malignant cells. Dimensionality reduction is performed on potential malignant cells from different cancer types, and malignant cells from different samples should form independent clusters.

[0063] The cell types belonging to the same cell lineage in all samples were grouped together, and the cells derived from the same cell lineage were regrouped. Based on cell type-specific genes, cell type annotation was performed at the cell lineage level to obtain an annotation information matrix for single cells. The cell lineages included malignant cells, epithelial cells, stromal cells, endothelial cells, myeloid cells, and lymphoid cells.

[0064] Through re-clustering and annotation based on cell type-specific genes, different cell types can be accurately distinguished, especially the differences between malignant cells and normal cells, which improves the accuracy and specificity of drug screening and helps to discover drugs targeting specific cell types.

[0065] The cell perturbation data are obtained and screened to obtain a pre-processed cell perturbation data matrix, which contains a plurality of significantly differentially expressed genes and corresponding drug treatment information.

[0066] Download the latest version of cell perturbation data from the Connectivity Map database. The Level 5 dataset contains 8 billion cell perturbation data entries, covering gene expression profiles of more than 240 human cell lines treated with 39,320 small molecules at different concentrations and for different times.

[0067] After filtering out low-confidence gene expression profile entries and screening by strict standards, a total of 8642 high-confidence gene expression profile entries were obtained.

[0068] The filtering criteria for high confidence entries of cell perturbation data are as follows:

[0069] The cell lines and treatment conditions corresponding to the gene expression profile entries are complete, there is no null value, is_ncs_exemplar=1;

[0070] With high-quality technical and functional replications, is_hiq=1, qc_pass=1, median_recall_rank_spearman<=5, median_recall_rank_wtcs_50<=5;

[0071] The treatment conditions corresponding to the gene expression profile entries had at least three replicates, nsample>=3;

[0072] The entries have significantly differentially expressed genes with high transcriptional activity scores, cc_q75>=0.2, tas>=0.2.

[0073] The significantly differentially expressed genes in the high-confidence gene expression profile entries were extracted and saved in gmt format, and a total of 17,284 treatments with significantly differentially expressed genes (cell lines and treatment conditions corresponding to the gene expression profile entries) were obtained, namely the pre-processed cell perturbation data matrix.

[0074] The preprocessed cell perturbation data matrix generation process is as follows:

[0075] Based on the high-confidence cell perturbation data entries, the gene expression data corresponding to the high-confidence entries were extracted from the gene expression spectrum; the expression data of each entry was sorted, and the top 10% significantly highly expressed genes or significantly underexpressed genes were selected, and the corresponding entries and the corresponding differentially expressed genes were saved in gmt format. The preprocessed cell perturbation data matrix is ​​separated by tabs, where the rows represent the differentially expressed gene sets corresponding to each entry, and contains three parts. The first column is the name of the high-confidence entry, the second column is the description of the high-confidence entry, that is, the cell line and treatment conditions corresponding to the entry, and the remaining columns are the significantly differentially expressed genes corresponding to the entry.

[0076] A deep learning model was constructed and trained using the single-cell pan-tumor transcriptome map and the pre-processed cell perturbation data matrix. The trained deep learning model was used to evaluate the effect of each drug treatment on each cell, and to search for pan-tumor drugs and tumor-specific drugs, as well as potential side effects of drugs.

[0077] The single-cell pan-tumor transcriptome map was integrated with the pre-processed cell perturbation data matrix to obtain a binary matrix of the perturbation change characteristics of a single cell for deep learning model input.

[0078] Among them, the binary matrix generation process of the perturbation change characteristics of a single cell is as follows:

[0079] The single-cell pan-tumor transcriptome profiles for model training were collected and saved using the AnnData data structure in Scanpy. The single-cell pan-tumor transcriptome profiles were integrated with the pre-processed cell perturbation data matrix, and a binary matrix was used to indicate whether the significantly differentially expressed genes of each treatment were expressed in a single cell; in the binary matrix, each column represents a treatment, and each row represents the gene expression characteristics in the treatment; the binary matrix was saved in the .uns of the AnnData data structure, which is an ordered dictionary in dict format. The single-cell count matrix was then standardized, logarithmized, and highly variable genes were screened, and treatments with expression characteristics of less than 12 genes in the binary matrix were deleted, and the processed AnnData data results were used for model input.

[0080] A deep learning model for drug perturbation prediction at the single-cell level was established to capture the response of individual cells to pharmacological treatment. The model adopted a variational autoencoder structure to merge individual cells from different conditions, used a nonlinear encoder to achieve flexibility, and used a linear decoder to achieve interpretability.

[0081] The specific structure of the deep learning model is as follows:

[0082] The model adopts a variational autoencoder structure. The encoder contains 4 hidden layers, each with 512; the decoder has the same number of hidden layers but in reverse order; in the bottleneck layer, the latent space dimension is equal to the number of entries in the cell perturbation data matrix, and the weight of each latent dimension to the output layer is programmed according to the binary matrix representing the change characteristics of the single cell perturbation to ensure that the latent entry can only contribute to the genes in the characteristic entry. The received gene expression and the research label of each cell are trained to encode the latent variables representing the cell perturbation entries. Considering that the number of entries in the perturbation data is very large and some cells may have redundant perturbation features, the group lasso regularization layer in the latent space uses an attention-like mechanism to enrich the perturbation entries with rich feature information for each cell. The cell perturbation entries processed by group lasso and gene-level sparse regularization are input to the linear decoder. The cell perturbation matrix is ​​used to program the neural network architecture by connecting the model parameters of the decoder to learn the specific cell perturbations for each latent dimension.

[0083] Based on the pre-trained end-to-end model, the impact of each drug treatment on each cell is evaluated, pan-tumor drugs and tumor-specific drugs are searched, and the potential targets of the drugs and the potential side effects of the drugs are analyzed explanatoryally based on the model.

[0084] In prediction, the latent representation of the prediction set and the model training set is set to the same latent space, and the batch effect between them is corrected. The latent score and latent direction of each cell are calculated based on the model latent space. The latent score represents the response intensity of the cell to a specific drug treatment, and the latent direction represents the change trend of the cell to a specific drug treatment. Based on the latent score and latent direction, the therapeutic effects and potential side effects of different drugs on specific cancer types are evaluated, and drug screening is performed.

[0085] The difference between two cell types is identified by Bayesian test, and an enrichment score matrix is ​​generated, where each row represents a cell type and each column represents a drug treatment. For each cell type, the larger the score of the cell perturbation entry, the more sensitive each cell type is to the drug treatment corresponding to the entry. The absolute value of the decoder weight of the gene in each entry is extracted and sorted to measure the genetic contribution of each cell perturbation entry; the higher the absolute value of the gene weight, the greater the impact of the gene on the drug corresponding to the entry, indicating the relative importance of the gene in the cell perturbation entry.

[0086] The following is an actual example to illustrate the preprocessing of single-cell maps and the pre-training of deep learning models.

[0087] First, we collected single-cell transcriptome sequencing data from cancer, adjacent tissues, and healthy normal tissues, performed cell type analysis based on a single patient, and identified malignant cells. Then, we regrouped and annotated the single-cell data of all patient samples by cell lineage to obtain cell type annotation information at the cell lineage level. A total of 388,646 cells were collected, and 86 cell types were annotated, which can be divided into six cell lineages: malignant cells, epithelial cells, stromal cells, endothelial cells, myeloid cells, and lymphoid cells.

[0088] Secondly, 90% of the cells (346,129 cells) in the single-cell atlas were randomly selected as training data, and the remaining 10% of the cells (42,517 cells) were used as validation data. Both the training data and validation data sets contain single cells from cancer, adjacent to cancer, and healthy normal tissues.

[0089] The original single-cell expression matrix of the training data was normalized and logarithmically transformed in Scanpy to find highly variable genes, and then integrated with the pre-processed cell perturbation data matrix to generate a binary matrix containing the characteristics of single-cell perturbation changes, and entries with less than 12 genes were filtered out for model input;

[0090] A series of pre-training was performed to optimize the model's hyperparameters, including the multiplication of the KL divergence part of the loss, the group lasso regularization coefficient, and the number and size of hidden layers. The model parameters were finally determined to be alpha_kl=0.001, alpha=0.95, hidden layer=[512, 512, 512, 512].

[0091] like Figure 2 As shown, Figure 2 The UMAP projection diagram of the training data predicted by the pre-trained deep learning model is shown. After training, features can be successfully extracted from the validation data and mapped to the training data. Cells of different types, sources, lineages, and tissues in the prediction data are divided into different clusters, and clusters from the same cell lineage tend to cluster together.

[0092] like Figure 3 As shown, Figure 3 The UMAP projection map of the entire single-cell pan-tumor transcriptome map predicted by the pre-trained deep learning model is shown. The pre-trained model is used to predict the entire single-cell pan-tumor transcriptome map. The cells of different cell types in the validation data and the training data form independent clusters, and the cell clusters from the same cell lineage are very close in the UMAP projection map. This shows that the model has successfully learned the unique and common features of cells from different tissues and different sources, especially the differential features of tumor cells and normal cells, and can successfully map these cells to various cell types.

[0093] The single-cell pan-tumor transcriptome profile was randomly divided into 10 parts, and any 9 of the 10 parts were used as training data and 1 part was used as test data for 10-fold cross-validation analysis. The 10-fold cross-validation results showed that feature extraction and cell clustering were highly reproducible, indicating the robustness of the constructed deep learning model.

[0094] Embodiment 2:

[0095] Based on Example 1, the trained deep learning model can be used to screen candidate drugs targeting pan-tumor cells.

[0096] Based on the trained deep learning model, the potential score and potential direction of each cell are calculated, and the effect of drug treatment on each cell type is calculated through Bayesian test to generate an enrichment score matrix of cell types and drug treatments, with each row representing a cell type and each column representing a drug treatment.

[0097] The 10 most significant drug treatments in each cell type are collected, and based on the potential scores corresponding to these drug treatments in the model, different cell lineages in the single-cell atlas, as well as different cell types in each cell lineage, can be distinguished.

[0098] like Figure 4 As shown, Figure 4 The number of repetitions of the compounds corresponding to the 10 most significant drug treatments in each cell type in the single-cell pan-tumor transcriptome map is shown. Drug treatment is a collection of important features extracted from cell perturbation data, reflecting the effect of compound treatment on cells. The 10 most significant drug treatments in each cell type were further counted. Some drug treatments appeared multiple times, and the small molecule compounds corresponding to these treatments have been extensively tested in clinical trials or laboratory experiments in multiple cancer types, and some of these compounds are marketed drugs approved by the US Food and Drug Administration (FDA).

[0099] For example, Azacitidine is approved for the treatment of myelodysplastic syndrome and acute myeloid leukemia, and related clinical trials are ongoing in glioma (clinical trial number NCT03666559) and pancreatic cancer (clinical trial number NCT03264404). Tosedostat is an aminopeptidase inhibitor that has shown efficacy in clinical trials for the treatment of acute myeloid leukemia and various solid tumors.

[0100] like Figure 5 As shown, Figure 5 The number of repetitions of the top 10 most significant drug treatments in each cell type in the stromal cell and endothelial cell lineages is shown. In the stromal cell lineage and endothelial cell lineage, the more significant drug treatments are repeated in multiple cell types. In the tumor-associated fibroblast type, there is a large overlap in the top 10 most significant drug treatments in each cell type, but there is almost no overlap with the significant drug treatments in normal fibroblasts and smooth muscle cells; similar results are observed in the endothelial cell lineage.

[0101] Drug treatments LJP008_SKL_24H:G19_DOWN and MOA001_U2OS_24H:P04_DOWN were recurrently found in the significant drug treatments in the stromal cell lineage, corresponding to the compounds Azacitidine and Parbenddazole, respectively. Both drug treatments were significantly enriched in tumor-associated fibroblasts, but not in the significant drug treatments in other stromal cell types and cell types of non-stromal cell lineages.

[0102] like Figure 6 As shown, Figure 6The corresponding latent space scores of tumor-associated fibroblasts in drug treatments LJP008_SKL_24H:G19_DOWN and MOA001_U2OS_24H:P04_DOWN are shown. Using the latent space scores corresponding to these two drug treatments in the model, tumor-associated fibroblasts and other cell types can be distinguished. This further shows that Azacitidine and Parbenddazole are sensitive to these tumor-associated fibroblast types, which come from multiple tissues, indicating that Azacitidine and Parbenddazole have the ability to target pan-tumor cells. In addition to FDA-approved indications, Azacitidine is also in clinical trials for multiple cancer types. Parbenddazole is an FDA-approved drug for the treatment of parasitic infections in animals and is currently undergoing clinical trials as a candidate drug for multiple tumor types, including acute myeloid leukemia, pancreatic cancer, and head and neck squamous cell carcinoma.

[0103] also, Figure 5 Drug treatments CPC006_PC3_24H:J14_DOWN and CPC006_HT29_6H:D19_UP were also found to be repetitive among the significant drug treatments in both stromal and endothelial cell lineages.

[0104] like Figure 7 As shown, Figure 7 The corresponding latent space scores of all cells in the single-cell pan-tumor transcriptome profile in drug treatments CPC006_PC3_24H:J14_DOWN and CPC006_HT29_6H:D19_UP are shown. It can be found that these two drug treatments are significantly enriched in tumor-associated fibroblasts and tumor-associated endothelial cells.

[0105] like Figure 8 As shown, Figure 8 The corresponding latent space scores of tumor-associated fibroblasts and tumor-associated endothelial cells in drug treatments CPC006_PC3_24H:J14_DOWN and CPC006_HT29_6H:D19_UP are shown. The latent space scores corresponding to these two drug treatments in the model can distinguish tumor-associated fibroblasts and tumor-associated endothelial cells from normal stromal cells, normal endothelial cells, and cell types of other lineages.

[0106] The compound corresponding to drug treatment CPC006_PC3_24H:J14_DOWN, erastin, is a ferroptosis inducer, which has been reported to have anticancer activity and can induce iron-dependent cell death of cancer cells in solid cancer and blood cancer. This shows that the candidate drug erastin screened by the Shennong framework has the ability to target pan-tumor cells.

[0107] Embodiment three:

[0108] Based on Example 1, the trained deep learning model can also be used to screen candidate drugs that target specific tumor cells.

[0109] Unlike tumor-associated fibroblasts, which are derived from multiple tissues, malignant cells are tissue-specific. There was less overlap in significant drug responses among cell types of the epithelial lineage, and malignant cells derived from different tissues had significantly different drug responses, but some drug responses were repeated in malignant cells derived from the same tissue.

[0110] like Fig. 9 As shown, Fig. 9 The latent spatial scores of drug treatments enriched in lung-derived malignant cells across the epithelial cell lineage are shown. Some drug treatments were significantly enriched in lung-derived malignant cells but not in normal lung epithelial cells.

[0111] like Fig.10 As shown, Fig.10 The corresponding latent space scores of all cells in the single-cell pan-tumor transcriptome profile in drug treatments LPROT004_YAPC_6H:BRD-A14634327:1_DOWN and PBIOA018_A549_24H:M02_DOWN are shown. The compounds corresponding to these drug treatments include FDA-approved drugs, candidate drugs, and laboratory compounds, most of which have been reported to inhibit tumor development, especially lung cancer. For example, drug treatments LPROT004_YAPC_6H:BRD-A14634327:1_DOWN and PBIOA018_A549_24H:M02_DOWN are enriched in malignant cells derived from the lung, especially alveolar type II epithelial cells (C2), bronchiolar exocrine cells (C13), and ciliated cells (C19).

[0112] like Fig.11 As shown, Fig.11The corresponding latent space scores of cells derived from the lung in drug treatments LPROT004_YAPC_6H:BRD-A14634327:1_DOWN and PBIOA018_A549_24H:M02_DOWN are shown. The latent space scores corresponding to these two drug treatments in the model can distinguish malignant cells from lungs from normal lung epithelial cells. In normal lung epithelial cells and other types of epithelial cells, the enrichment scores of these two drug treatments are low, indicating that the compounds corresponding to these two treatments are insensitive to these cell types and have the ability to specifically target malignant cells.

[0113] Drug treatment PBIOA018_A549_24H:M02_DOWN corresponding compound Volasertib is a candidate drug for acute myeloid leukemia and other solid tumors, and has been reported to enter multiple phases of clinical trials, including non-small cell lung cancer. However, the overall anti-tumor activity in clinical trials was not high, and the experiment in lung cancer was terminated.

[0114] Fig. 9 Other compounds with significant drug treatments include Atorvastatin, Exemestane, Prednisolone, and Fostamatinib. Atorvastatin belongs to a class of statins that are FDA-approved for the treatment of high blood cholesterol and are currently being studied in combination with chemotherapy drugs for the prevention and treatment of cancer, including non-small cell lung cancer. Exemestane is a hormone therapy used to treat breast cancer that has been reported to have antiproliferative effects in lung cancer cells and is being tested in preclinical and clinical trials in non-small cell lung cancer and small cell lung cancer (NCT02666105).

[0115] like Fig.12 As shown, Fig.12 The cell viability of the A549 cell line after being treated with the compounds Volasertib, Prednisolone and Fostamatinib at different times is shown. The results of the drug screening experiment showed that the cell viability was measured after the A549 cell line was treated with different compounds at different concentrations for 24 and 48 hours. The cells treated with Prednisolone and Fostamatinib showed very low cell viability, similar to the treatment results of Volasertib. The anti-lung cancer activity of Prednisolone and Fostamatinib was confirmed, especially Fostamatinib, and the results of the cell screening experiment showed that it had a significantly low cell viability value. These results show that the Shennong framework has the ability to screen subsequent drugs targeting specific tumor cells, and can also be used to discover new indications for marketed drugs.

[0116] Embodiment 4:

[0117] Based on Example 1, the trained deep learning model can also be used to screen and analyze the mechanism of action of targeted tumor-specific drugs.

[0118] like Fig.13 As shown, Fig.13 The absolute value and ranking of the gene scores corresponding to the drug treatment PBIOA018_A549_24H:M02_DOWN in the decoder are shown. The weights of the genes in the drug treatment PBIOA018_A549_24H:M02_DOWN are extracted from the model decoder and sorted by absolute value. This treatment corresponds to Volasertib. It has been reported in the literature that Volasertib can lead to downregulation of BCL2L1 expression, disrupt the interaction between PLK1 and NEK7, or displace HMGB2 from mitotic chromosomes, thereby affecting mitotic arrest and apoptosis of cancer cells. These genes have extremely high weights in the decoders corresponding to drug treatments, which shows that the Shennong framework has the ability to parse the mechanism of action of targeted tumor-specific drugs.

[0119] Embodiment five:

[0120] The trained deep learning models can also be used to discover side effects of candidate drugs.

[0121] Figure 7 In the experiment, CPC006_PC3_24H:J14_DOWN was enriched in normal endothelial cells after drug treatment, indicating that normal endothelial cells were sensitive to the compound corresponding to this treatment, erastin. CPC006_PC3_24H:J14_DOWN was enriched in tumor-associated stromal cells, tumor-associated endothelial cells, and normal endothelial cells at the same time, indicating that erastin can target normal endothelial cells in addition to pan-tumor cells, and has side effects. It has been reported in the literature that erastin, which has anti-tumor activity, has certain toxicity to healthy tissues, which is consistent with the prediction results of the deep learning model.

[0122] In addition, among drug treatments targeting specific tumor cells, some drug treatments were enriched in both malignant cells derived from cancer tissues and epithelial cells derived from normal tissues, indicating that the compounds corresponding to these drug treatments simultaneously targeted tumor-specific cells and normal epithelial cells, and had certain side effects on healthy tissues.

[0123] like Fig.14 As shown, Fig.14The corresponding latent space scores for all cells in the single-cell pan-tumor transcriptome profile and malignant cells from the liver in the drug treatment CPC006_A549_24H:BRD-K56343971-001-02-3:10_UP are shown. The drug treatment CPC006_A549_24H:BRD-K56343971-001-02-3:10_UP has significant enrichment scores in malignant cells from the liver (C9, C11, and C17), acinar cells from the normal pancreas (C4), and exocrine cells (C8). The corresponding compound to this drug treatment, vemurafenib, is a BRAF inhibitor used for targeted therapy of Eldheim-Chester disease and melanoma, and has been reported to induce pancreatitis in patients.

[0124] Figure 5 In the data, drug treatments LJP005_SKBR3_24H:F19_DOWN and ASG003_MCF7_48H:E07_UP were duplicated among the significant drug treatment entries in the stromal cell lineage and endothelial cell lineage.

[0125] like Figure 15-17 As shown, Fig.15 The corresponding latent space scores of all cells in the single-cell pan-tumor transcriptome profile in drug-treated ASG003_MCF7_48H:E07_UP and the 16 cell types with the highest scores are shown; Fig.16 The corresponding latent space scores of all cells in the single-cell pan-tumor transcriptome profile in drug-treated LJP005_SKBR3_24H:F19_DOWN and the 16 cell types with the highest scores are shown; Fig.17 The corresponding latent space scores of tumor-associated fibroblasts and malignant cells from the liver in drug treatments ASG003_MCF7_48H:E07_UP and LJP005_SKBR3_24H:F19_DOWN are shown. In addition to enrichment in tumor-associated stromal cells and malignant cells, these two drug treatments are also enriched in normal stromal cells and normal epithelial cells, including fibroblasts (C0), hepatocytes (C21), ductal cells (C18), and epithelial cells (C20) derived from normal tissues.

[0126] This indicates that the compounds corresponding to these two drug treatments have significant effects on epithelial cells and stromal cells derived from normal liver and pancreas. The compound lopinavir corresponding to the drug treatment ASG003_MCF7_48H:E07_UP has been reported to cause liver damage. The compound GSK-690693 corresponding to the drug treatment LJP005_SKBR3_24H:F19_DOWN has been reported to have significant anti-tumor proliferation ability, but because it can cause hyperglycemia-related side effects and limited anti-tumor activity, its related clinical trials have been withdrawn or terminated (NCT00666081 and NCT00493818). These results show that the deep learning model constructed in the embodiment of the present application has the ability to discover side effects of candidate drugs.

[0127] Embodiment six:

[0128] An embodiment of the present application also provides a computer storage medium on which a computer program is stored, and when the program is executed by a processor, the above-mentioned drug screening analysis method based on single-cell transcriptome mapping is implemented.

[0129] Embodiment seven:

[0130] An embodiment of the present application also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned drug screening analysis method based on single-cell transcriptome mapping when executing the computer program.

[0131] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the spirit or essential features of the present application. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present application.

Claims

1. A drug screening analysis method based on single-cell transcriptome profiles, characterized in that: The following steps are involved: Constructing a single-cell pan-tumor transcriptome map, wherein the single-cell pan-tumor transcriptome map includes a single-cell transcriptome sequencing count matrix and a single-cell annotation information matrix of multiple cancer tissues, corresponding adjacent cancer tissues, and corresponding normal human tissues; The annotation information matrix of the single cell is obtained by grouping cells and annotating them based on cell type-specific genes based on the single-cell transcriptome sequencing count matrix; the annotation information matrix of the single cell contains information on the tissue origin, cell type and corresponding cell lineage of each cell in the single-cell pan-tumor transcriptome map; Acquiring cell perturbation data and screening the cell perturbation data to obtain a pre-processed cell perturbation data matrix, wherein the pre-processed cell perturbation data matrix includes a plurality of significantly differentially expressed genes and corresponding drug treatment information; Constructing a deep learning model and training the deep learning model using the single-cell pan-tumor transcriptome map and the pre-processed cell perturbation data matrix; Use the trained deep learning model to evaluate the impact of each drug treatment on each cell, and to find pan-tumor drugs and tumor-specific drugs, as well as potential side effects of drugs; The step of obtaining the annotation information matrix of the single cell comprises: Obtaining single-cell transcriptome sequencing count matrices of cancer tissues of various cancer types, corresponding paracancerous tissues, and corresponding tissues from normal human bodies, grouping each sample, and using lineage-specific genes to initially group cells and annotate and divide them into multiple cell lineages; the multiple cell lineages include epithelial cell lineages; Identify and annotate malignant cells in the epithelial cell lineage in the initial clustering; The cell types belonging to the same cell lineage in all samples are grouped together, the cells from the same cell lineage are regrouped, and the cell types are finely annotated at the cell lineage level based on cell type-specific genes to obtain an annotation information matrix for a single cell, wherein the cell lineage includes malignant cells, epithelial cells, stromal cells, endothelial cells, myeloid cells, and lymphoid cells; The step of identifying and annotating malignant cells in the epithelial cell lineage in the primary grouping comprises: checking whether the gene expression level of a specific cancer marker in the epithelial cells is greater than a first threshold; Using a copy number variation analysis tool to calculate whether the copy number variation score of each epithelial cell is greater than a second threshold; If the epithelial cells meet the gene expression level of a specific cancer marker greater than a first threshold and / or the copy number variation score greater than a second threshold, they are annotated as malignant cells.

2. The drug screening analysis method based on single-cell transcriptome profile according to claim 1, characterized in that: The step of acquiring the cell disturbance data and screening the cell disturbance data to obtain a pre-processed cell disturbance data matrix comprises: Acquire cell perturbation data from a public database or experimental data, wherein the cell perturbation data includes single-cell transcriptome changes before and after drug treatment; According to the preset filtering conditions, cell perturbation data entries meeting the preset filtering conditions are screened out from the cell perturbation data; Each cell perturbation data entry is sorted, and the significantly differentially expressed genes ranked before the preset threshold are selected to generate the preprocessed cell perturbation data matrix.

3. The drug screening analysis method based on single-cell transcriptome profile according to claim 1, characterized in that: The step of constructing a deep learning model and training the deep learning model using a single-cell pan-tumor transcriptome map and a pre-processed cell perturbation data matrix comprises: Constructing a deep learning model, wherein the deep learning model adopts a variational autoencoder structure; The single-cell pan-tumor transcriptome map is divided into training data and validation data, which are respectively integrated with the preprocessed cell perturbation data matrix to obtain a training set and a validation set. Based on the training set and the validation set, the deep learning model is trained until the loss function converges to obtain a trained deep model; the training set and the validation set are both binary matrices containing the perturbation change characteristics of a single cell.

4. The drug screening analysis method based on single-cell transcriptome profile according to claim 1, characterized in that: The step of using the trained deep learning model to evaluate the effect of each drug treatment on each cell includes: The trained deep learning model is used to calculate the potential score and potential direction of each cell in response to a specific drug treatment; the potential score represents the intensity of the cell's response to the specific drug treatment; the potential direction represents the change trend of the cell in response to the specific drug treatment; Based on the potential scores and potential directions, the therapeutic effects and potential side effects of different drugs on specific cancer types are evaluated, and drug screening is performed.

5. The drug screening analysis method based on single-cell transcriptome profile according to claim 4, characterized in that: Also includes: Based on the potential score and potential direction of each cell in response to a specific drug treatment, a Bayesian test was used to calculate the effect of drug treatment on each cell type, generating an enrichment score matrix of cell types and drug treatments, with each row representing a cell type and each column representing a drug treatment.

6. A computer storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

7. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.

Images