A method for screening gene-targeted therapies for hematological diseases

Abstract

This invention relates to the field of gene-targeted therapy drug screening technology, and provides a method for screening gene-targeted therapy drugs for hematological diseases, comprising: Step 1, obtaining blood tumor samples from subjects and constructing patient-derived blood organoids in vitro while preserving the immune microenvironment; Step 2, performing single-cell multi-omics sequencing on the blood organoids to obtain a baseline gene expression matrix, chromatin accessibility matrix, and surface protein abundance matrix, where g is the number of genes and is the number of cells; Step 3, identifying a set of driver candidate target genes based on the baseline multi-omics data. By constructing patient-derived blood organoids while preserving the immune microenvironment, the screening reliability is high; by integrating single-cell multi-omics, CRISPR perturbation, and graph neural networks, targeted drugs are accurately identified; and by establishing a comprehensive scoring system for quantitative ranking, personalized and precise drug recommendations are achieved, effectively improving screening efficiency and clinical value.

Description

Technical Field

[0001] This invention relates to the field of gene-targeted therapy drug screening technology, specifically a method for screening gene-targeted therapy drugs for hematological diseases. Background Technology

[0002] Hematological diseases, especially hematological malignancies, pose a serious threat to human health. Targeted therapy has become the core direction of clinical treatment due to its high specificity and few side effects. Precise and efficient drug screening is a key prerequisite for achieving targeted therapy. The screening results directly determine the treatment effect and patient prognosis, and are of great significance for improving the diagnosis and treatment level of hematological diseases and improving the quality of life of patients.

[0003] Currently, the screening of targeted drugs for hematological diseases mostly uses traditional cell lines or animal models. These models cannot preserve the immune microenvironment of human hematological malignancies, making it difficult to reproduce the actual interactions and signaling pathways between tumor cells, T cells, and stromal cells. This leads to significant model bias and insufficient clinical relevance of the screening results. Furthermore, existing screening methods often rely on single-indicator assessments, lacking a comprehensive analysis of gene targets, network perturbations, and drug affinity. They also lack a scientific and comprehensive scoring system, making it impossible to accurately quantify and rank candidate drugs, failing to match individual patient tumor characteristics, and easily resulting in ineffective treatment. Therefore, there is an urgent need for a gene-targeted therapy screening method for hematological diseases that can address these issues. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for screening gene-targeted therapeutic drugs for hematological diseases, which solves the problems of large biases, single evaluation methods, and inability to achieve individualized and precise screening in existing hematological targeted drug screening models.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for screening gene-targeted therapeutic drugs for hematological diseases, comprising:

[0006] Step 1: Obtain blood tumor samples from the subjects and construct patient-derived blood organoids in vitro while preserving the immune microenvironment;

[0007] Step 2: Perform single-cell multi-omics sequencing on the blood organoids to obtain the baseline gene expression matrix. Chromatin accessibility matrix and surface protein abundance matrix, where g is the number of genes. Cell count;

[0008] Step 3: Based on the baseline multi-omics data, identify the set of driver candidate target genes. ;

[0009] Step 4: Perform multiple gene perturbations on the blood organoids using a combined CRISPR perturbation library, and perform single-cell multi-omics sequencing on the perturbated organoids to obtain the perturbation response gene expression matrix. And the corresponding perturbed chromatin accessibility matrix and surface protein abundance matrix;

[0010] Step 5, according to the above and Constructing gene regulatory networks and the network after disturbance And calculate the network perturbation distance caused by each perturbation combination. ;

[0011] Step Six: Encode the drug molecule structure into a drug molecule graph characterization using a graph isomorphism network, and then connect it to the gene regulatory network. Heterogeneous graph fusion is performed on the topological features, and a trained graph neural network model is used to predict the targeting affinity index of each candidate drug for the set of driver candidate target genes. ;

[0012] Step 7: Combine the network disturbance distance and the target affinity index The gene-targeted therapy score for each candidate drug is calculated using the following formula. :

[0013] ;

[0014] in, The weighting coefficients and , and These represent the minimum and maximum values ​​of the targeting affinity index among all candidate drugs. This represents the maximum network perturbation distance among all candidate drugs. This represents the number of driver candidate target genes. For genes Normalized importance score;

[0015] The network disturbance distance The selection method is as follows: based on the subset of driver candidate target genes predicted by the candidate drug, select the network perturbation distance corresponding to the subset from all perturbation combination network distances obtained in step five. If there is no exact match, select the network perturbation distance corresponding to the perturbation combination with the highest Jaccard similarity to the subset.

[0016] Step 8: Sort the candidate drugs according to the gene-targeted therapy score, and output the N drugs with the highest scores as the screening results.

[0017] Preferably, the patient-derived blood organoids that preserve the immune microenvironment in step one are constructed as follows: primary hematologic tumor cells, autologous T lymphocytes, and bone marrow stromal cells are selected from the subject's bone marrow or peripheral blood and co-encapsulated in a 5:2:3 ratio in a methacrylamide gelatin hydrogel containing fibronectin. The cells are then subjected to three-dimensional perfusion culture in IMDM complete medium supplemented with a combination of cytokines, including IL-2 (50 IU / mL), IL-7 (20 ng / mL), IL-15 (20 ng / mL), FLT3 ligand (50 ng / mL), and anti-CD3 / CD28 bispecific activating microbeads. Half of the medium is replaced daily during the culture process, and fresh cytokines are added.

[0018] Preferably, in step three, the set of driver candidate target genes is screened by integrating the differential expression significance, the enrichment degree of chromatin accessibility peak in the promoter region, and the outlier values ​​of surface protein abundance in the baseline gene expression matrix, and then using a random forest importance ranking algorithm.

[0019] Preferably, the CRISPR perturbation library combination in step four employs a split-pool CRISPR screening strategy.

[0020] Preferably, the gene regulatory network in step five Limited to the set of driver candidate target genes The local network consisting of its first-order neighbor genes, with the number of nodes controlled within 100-300, describes the network perturbation distance. The following approximate calculation method is adopted:

[0021] ;

[0022] in, For a local network node set, and These are the eigenvalue vectors of the Laplacian matrix of the local network. These are the balancing parameters.

[0023] Preferably, the graph neural network model in step six is ​​a heterogeneous graph attention network, which includes:

[0024] Drug molecule coding units are used to extract molecular fingerprints of drug atom-bond diagrams using graph isomorphic networks;

[0025] Gene network coding units are used to aggregate the features of nodes and neighbors in the gene regulation network using graph attention networks;

[0026] The cross-modal attention fusion unit is used to interactively fuse drug molecule characterization with node characterization of candidate target genes through a multi-head attention mechanism, and output a targeting affinity index. .

[0027] Preferably, the graph neural network model employs a two-stage training strategy of "pre-training-fine-tuning":

[0028] Pre-training phase: Self-supervised graph contrastive learning is performed using a large-scale general drug-protein interaction database to maximize the mutual information of positive sample drug-target pairs while minimizing the mutual information of negative sample pairs as the training objective.

[0029] Fine-tuning phase: Constructing a weakly supervised signaling system using the perturbation response data of the blood organoids. The definition is as follows: if the cosine similarity between the transcriptome perturbation spectrum after drug treatment and the transcriptome perturbation spectrum after gene knockout exceeds a preset threshold, then the drug-gene pair is marked as a positive sample; otherwise, it is marked as a negative sample.

[0030] Preferably, the training loss function in the fine-tuning phase for:

[0031] ;

[0032] in, Labeled as weak supervision This is a drug structure similarity matrix. Gene functional similarity matrix The regularization coefficient is . It is the Frobenius norm. Sample confidence weights .

[0033] Preferably, in step seven, the normalized importance score is... Obtained through the following methods:

[0034] First calculate the genes Mutual information between single-cell level expression and whole transcriptome expression:

[0035] ;

[0036] in, Indicates gene Discretized representation vector, This represents the expression matrix of the remaining genes.

[0037] Preferably, the CRISPR screening strategy of the splitting pool specifically involves the first round of transducing each aliquot of the blood organoids with a single gRNA lentiviral library, followed by selection with puromycin and mixing; the second round of transduction with an orthogonal single gRNA lentiviral library; and obtaining a dual-gene perturbation cell population by fluorescent dual-labeling sorting.

[0038] This invention provides a method for screening gene-targeted therapeutic drugs for hematological diseases. It has the following beneficial effects:

[0039] 1. This invention uses patient-derived blood organoids while preserving the immune microenvironment, including tumor cells, T cells, and stromal cells, to realistically reproduce the cell interactions and signaling pathways of hematologic malignancies in the human body. Compared with traditional cell lines or animal models, it significantly reduces model bias and greatly improves the clinical relevance and reliability of drug screening results.

[0040] 2. This invention precisely identifies driver target genes through single-cell multi-omics, combines CRISPR perturbation to quantify gene network response, and then uses heterogeneous graph neural networks to predict drug targeting affinity; thus achieving precise evaluation of the entire chain from gene targets, network perturbations, and drug affinity, avoiding bias from a single indicator and improving the hit efficiency of targeted drugs.

[0041] 3. This invention establishes a comprehensive scoring system that integrates target affinity, network perturbation distance, and gene importance, enabling scientific quantitative ranking of candidate drugs. The output Top-N drugs directly match the individual tumor characteristics of patients, providing personalized, precise, and feasible targeted drug treatment plans for hematological diseases and reducing the risk of ineffective drug use. Attached Figure Description

[0042] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0043] 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.

[0044] Example:

[0045] Please see the appendix Figure 1 This invention provides a method for screening gene-targeted therapeutic drugs for hematological diseases, comprising:

[0046] Blood tumor samples were obtained from the subjects, and patient-derived blood organoids with preserved immune microenvironment were constructed in vitro. In step one, patient-derived blood organoids with preserved immune microenvironment were constructed as follows: primary hematologic tumor cells, autologous T lymphocytes, and bone marrow stromal cells were selected from the subjects' bone marrow or peripheral blood and encapsulated in a 5:2:3 ratio in a methacrylamide gel hydrogel containing fibronectin. The cells were then subjected to three-dimensional perfusion culture in IMDM complete medium supplemented with a combination of cytokines, including IL-2 (50 IU / mL), IL-7 (20 ng / mL), IL-15 (20 ng / mL), FLT3 ligand (50 ng / mL), and anti-CD3 / CD28 bispecific activating microbeads. Half of the medium was changed and fresh cytokines were added daily during the culture process.

[0047] Specifically, primary hematologic tumor cells, autologous T lymphocytes, and bone marrow stromal cells were isolated from the subjects' bone marrow or peripheral blood and co-encapsulated in a 5:2:3 ratio in a methacrylamide gelatin hydrogel containing fibronectin. The mixture was then subjected to three-dimensional perfusion culture in IMDM complete medium supplemented with a combination of cytokines. The cytokine combination included IL-2 (50 IU / mL), IL-7 (20 ng / mL), IL-15 (20 ng / mL), FLT3 ligand (50 ng / mL), and anti-CD3 / CD28 bispecific activating microbeads. Half of the medium was replaced daily during the culture process, and fresh cytokines were added.

[0048] Please see the appendix Figure 1 Single-cell multi-omics sequencing was performed on blood organoids to obtain baseline gene expression matrices. Chromatin accessibility matrix and surface protein abundance matrix, where g is the number of genes. Cell count;

[0049] Specifically, a single-cell sequencing platform was used to perform single-cell RNA sequencing, chromatin open region sequencing (ATAC-seq), and surface protein flow cytometry mass spectrometry on the constructed blood organoids. Through data quality control, alignment, and normalization, a baseline gene expression matrix was finally obtained. (g represents the number of genes, c represents the number of cells), chromatin accessibility matrix and surface protein abundance matrix, used for subsequent target gene identification.

[0050] Please see the appendix Figure 1 Based on baseline multi-omics data, a set of driver candidate target genes was identified. In step one, the patient-derived blood organoids that preserve the immune microenvironment are constructed as follows: primary hematologic tumor cells, autologous T lymphocytes, and bone marrow stromal cells are selected from the bone marrow or peripheral blood of the subjects and co-encapsulated in a 5:2:3 ratio in a methacrylamide gel containing fibronectin. They are then subjected to three-dimensional perfusion culture in IMDM complete medium supplemented with a combination of cytokines, including IL-2 (50 IU / mL), IL-7 (20 ng / mL), IL-15 (20 ng / mL), FLT3 ligand (50 ng / mL), and anti-CD3 / CD28 bispecific activating microbeads. Half of the medium is changed and fresh cytokines are added daily during the culture process.

[0051] Specifically, the study integrates three core indicators from the baseline gene expression matrix: significant differential gene expression, enrichment of chromatin accessibility peaks in gene promoter regions, and outlier values ​​of surface protein abundance. After standardizing the data for these three indicators, the data are input into a random forest model. A model importance ranking algorithm is then used to screen genes highly correlated with the occurrence, proliferation, and immune escape of hematological malignancies, forming a set of driver candidate target genes. .

[0052] Please see the appendix Figure 1 We used a combined CRISPR perturbation library to perform multiple gene perturbations on blood organoids, and then performed single-cell multi-omics sequencing on the perturbated organoids to obtain the perturbation response gene expression matrix. The perturbation chromatin accessibility matrix and surface protein abundance matrix were obtained. In step four, the CRISPR perturbation library was combined using a split-pool CRISPR screening strategy. Specifically, the split-pool CRISPR screening strategy involved first transducing blood organoids into equal portions using a single gRNA lentiviral library, then mixing them after screening with puromycin. In the second round, the orthogonal single gRNA lentiviral library was used for secondary transduction, and dual-gene perturbation cell populations were obtained by fluorescent double labeling sorting.

[0053] Specifically, a split-pool CRISPR screening strategy was used to construct a combined perturbation library. In the first round, single gRNA lentiviral libraries were used to transduce equal fractions of blood organoids. After removing untransduced cells by puromycin screening, the fractions were mixed. In the second round, orthogonal single gRNA lentiviral libraries were used for secondary transduction, and dual-gene perturbation cell populations were obtained by fluorescent double-labeling sorting. Single-cell multi-omics sequencing was performed on the perturbation blood organoids, and data analysis yielded the perturbation response gene expression matrix Y, as well as the corresponding perturbation chromatin accessibility matrix and surface protein abundance matrix.

[0054] Please see the appendix Figure 1 ,according to and Constructing gene regulatory networks and the network after disturbance And calculate the network perturbation distance caused by each perturbation combination. In step five, the gene regulatory network Limited to a set of driver candidate target genes The local network consists of its first-order neighbor genes, with the number of nodes controlled within 100-300 and the network perturbation distance. The following approximate calculation method is adopted:

[0055] ;

[0056] in, For a local network node set, and These are the eigenvalue vectors of the Laplacian matrix of the local network. For balance parameters;

[0057] Specifically, a baseline gene regulatory network is constructed based on the baseline gene expression matrix X. Limited to a set of driver candidate target genes A local network consisting of its first-order neighbor genes, with the number of nodes controlled within 100-300; based on the perturbation response matrix. Constructing a perturbation-based gene regulatory network Using approximate formulas Calculate the network perturbation distance, where For a local network node set, and The eigenvectors of the network's Laplacian matrix. These are the balancing parameters.

[0058] Please see the appendix Figure 1 The drug molecule structure is encoded into a drug molecule graph representation using a graph isomorphism network, and then compared with gene regulatory networks. Heterogeneous graph fusion was performed using topological features, and a trained graph neural network model was used to predict the targeting affinity index of each candidate drug for the set of driver candidate target genes. In step six, the graph neural network model is a heterogeneous graph attention network, which includes:

[0059] Drug molecule coding units are used to extract molecular fingerprints of drug atom-bond diagrams using graph isomorphic networks;

[0060] Gene network coding units are used to aggregate the features of nodes and neighbors in the gene regulation network using graph attention networks;

[0061] The cross-modal attention fusion unit is used to interactively fuse drug molecule characterization with node characterization of candidate target genes through a multi-head attention mechanism, and output a targeting affinity index. The graph neural network model employs a two-stage training strategy of "pre-training-fine-tuning":

[0062] Pre-training phase: Self-supervised graph contrastive learning is performed using a large-scale general drug-protein interaction database to maximize the mutual information of positive sample drug-target pairs while minimizing the mutual information of negative sample pairs as the training objective.

[0063] Fine-tuning phase: Constructing a weakly supervised signal using perturbation response data from blood organoids. It is defined as follows: if the cosine similarity between the transcriptome perturbation spectrum after drug treatment and the transcriptome perturbation spectrum after gene knockout exceeds a preset threshold, then the drug-gene pair is marked as a positive sample; otherwise, it is marked as a negative sample. The training loss function during the fine-tuning phase... for:

[0064] ;

[0065] in, Labeled as weak supervision This is a drug structure similarity matrix. Gene functional similarity matrix The regularization coefficient is . It is the Frobenius norm. Sample confidence weights ;

[0066] Specifically, the candidate drug molecule structure is input into a graph isomorphism network, and atomic-bond graph features are extracted and encoded into a drug molecule graph representation; the baseline gene regulatory network is then used... A topological feature input graph attention network is used to aggregate gene node and neighbor features. A cross-modal attention fusion unit is constructed through a multi-head attention mechanism to interactively fuse drug molecule representations with candidate target gene node representations. A heterogeneous graph attention network model using a two-stage training strategy of "pre-training-fine-tuning" is employed to predict the targeting affinity index of each candidate drug for the driving candidate target gene set. .

[0067] Please see the appendix Figure 1 Combined with network disturbance distance and target affinity index The gene-targeted therapy score for each candidate drug is calculated using the following formula. :

[0068] ;

[0069] in, The weighting coefficients and , and These represent the minimum and maximum values ​​of the targeting affinity index among all candidate drugs. This represents the maximum network perturbation distance among all candidate drugs. This represents the number of driver candidate target genes. For genes Normalized importance score;

[0070] Network Distance The selection method is as follows: Based on the subset of driver candidate target genes predicted by the candidate drug, select the network perturbation distance corresponding to the subset from all perturbation combinations obtained in step five. If there is no exact match, select the network perturbation distance corresponding to the perturbation combination with the highest Jaccard similarity to the subset. The importance score is normalized in step seven. Obtained through the following methods:

[0071] First calculate the genes Mutual information between single-cell level expression and whole transcriptome expression:

[0072] ;

[0073] in, Indicates gene Discretized representation vector, Represents the expression matrix of the remaining genes;

[0074] Specifically, according to the formula Calculate the score; where The weighting coefficients and , and To target the maximum value of the affinity index, The maximum network perturbation distance is given by m, where m is the number of target genes. Gene t-normalized importance score (calculated from gene expression level and whole transcriptome mutual information); network perturbation distance D is selected based on the principle of matching candidate drug target gene subsets or the highest Jaccard similarity.

[0075] Please see the appendix Figure 1 Candidate drugs are ranked according to gene-targeted therapy scores, and the N drugs with the highest scores are output as the screening results.

[0076] Specifically, the gene-targeted therapy scores of all candidate drugs are compiled and sorted in descending order of scores. Taking into account clinical drug needs, drug safety and accessibility, the number of drugs to be screened is set to N (usually 5-10). The top N candidate drugs after sorting are output as the preferred drugs for precision targeted therapy of hematological diseases.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for screening gene-targeted therapeutic drugs for hematological diseases, characterized in that, include: Step 1: Obtain blood tumor samples from the subjects and construct patient-derived blood organoids in vitro while preserving the immune microenvironment; Step 2: Perform single-cell multi-omics sequencing on the blood organoids to obtain the baseline gene expression matrix. Chromatin accessibility matrix and surface protein abundance matrix, where g is the number of genes. Cell count; Step 3: Based on the baseline multi-omics data, identify the set of driver candidate target genes. ; Step 4: Perform multiple gene perturbations on the blood organoids using a combined CRISPR perturbation library, and perform single-cell multi-omics sequencing on the perturbated organoids to obtain the perturbation response gene expression matrix. And the corresponding perturbed chromatin accessibility matrix and surface protein abundance matrix; Step 5, according to the above and Constructing gene regulatory networks and the network after disturbance And calculate the network perturbation distance caused by each perturbation combination. ; Step Six: Encode the drug molecule structure into a drug molecule graph characterization using a graph isomorphism network, and then connect it to the gene regulatory network. Heterogeneous graph fusion is performed on the topological features, and a trained graph neural network model is used to predict the targeting affinity index of each candidate drug for the set of driver candidate target genes. ; Step 7: Combine the network disturbance distance and the target affinity index The gene-targeted therapy score for each candidate drug is calculated using the following formula. : ; in, The weighting coefficients and , and These represent the minimum and maximum values ​​of the targeting affinity index among all candidate drugs. This represents the maximum network perturbation distance among all candidate drugs. This represents the number of driver candidate target genes. For genes Normalized importance score; The network disturbance distance The selection method is as follows: based on the subset of driver candidate target genes predicted by the candidate drug, select the network perturbation distance corresponding to the subset from all perturbation combination network distances obtained in step five. If there is no exact match, select the network perturbation distance corresponding to the perturbation combination with the highest Jaccard similarity to the subset. Step 8: Sort the candidate drugs according to the gene-targeted therapy score, and output the N drugs with the highest scores as the screening results.

2. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 1, characterized in that, In step one, the patient-derived blood organoids that preserve the immune microenvironment are constructed as follows: primary hematologic tumor cells, autologous T lymphocytes, and bone marrow stromal cells are selected from the subject's bone marrow or peripheral blood and co-encapsulated in a 5:2:3 ratio in a methacrylamide gelatin hydrogel containing fibronectin. The cells are then subjected to three-dimensional perfusion culture in IMDM complete medium supplemented with a combination of cytokines, including IL-2 (50 IU / mL), IL-7 (20 ng / mL), IL-15 (20 ng / mL), FLT3 ligand (50 ng / mL), and anti-CD3 / CD28 bispecific activating microbeads. During the culture process, half of the medium is replaced daily and fresh cytokines are added.

3. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 1, characterized in that, In step three, the set of driving candidate target genes is screened by integrating the differential expression significance, the enrichment of chromatin accessibility peaks in the promoter region, and surface protein abundance outliers in the baseline gene expression matrix, and then using a random forest importance ranking algorithm.

4. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 1, characterized in that, In step four, the CRISPR perturbation library is combined using a split-pool CRISPR screening strategy.

5. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 1, characterized in that, gene regulatory network in step five Limited to the set of driver candidate target genes The local network consisting of its first-order neighbor genes, with the number of nodes controlled within 100-300, describes the network perturbation distance. The following approximate calculation method is adopted: ; in, For a local network node set, and These are the eigenvalue vectors of the Laplacian matrix of the local network. For balancing parameters.

6. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 1, characterized in that, The graph neural network model in step six is ​​a heterogeneous graph attention network, which includes: Drug molecule coding units are used to extract molecular fingerprints of drug atom-bond diagrams using graph isomorphic networks; Gene network coding units are used to aggregate the features of nodes and neighbors in the gene regulation network using graph attention networks; The cross-modal attention fusion unit is used to interactively fuse drug molecule characterization with node characterization of candidate target genes through a multi-head attention mechanism, and output a targeting affinity index. .

7. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 6, characterized in that, The graph neural network model adopts a two-stage training strategy of "pre-training-fine-tuning": Pre-training phase: Self-supervised graph contrastive learning is performed using a large-scale general drug-protein interaction database to maximize the mutual information of positive sample drug-target pairs while minimizing the mutual information of negative sample pairs as the training objective. Fine-tuning phase: Constructing a weakly supervised signaling system using the perturbation response data of the blood organoids. The definition is as follows: if the cosine similarity between the transcriptome perturbation spectrum after drug treatment and the transcriptome perturbation spectrum after gene knockout exceeds a preset threshold, then the drug-gene pair is marked as a positive sample; otherwise, it is marked as a negative sample.

8. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 7, characterized in that, The training loss function in the fine-tuning phase for: ; in, Labeled as weak supervision This is a drug structure similarity matrix. Gene functional similarity matrix The regularization coefficient is . It is the Frobenius norm. Sample confidence weights .

9. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 1, characterized in that, The normalized importance score in step seven Obtained through the following methods: First calculate the genes Mutual information between single-cell level expression and whole transcriptome expression: ; in, Indicates gene Discretized representation vector, This represents the expression matrix of the remaining genes.

10. The method for screening gene-targeted therapeutic drugs for hematological diseases according to claim 4, characterized in that, The CRISPR screening strategy for the split pool specifically involves the first round of transducing equal portions of the blood organoids using a single gRNA lentiviral library, followed by selection with puromycin and then mixing. The second round involves secondary transduction using an orthogonal single gRNA lentiviral library, followed by sorting using fluorescent dual labeling to obtain a dual-gene perturbation cell population.

Images