A virtual cell analysis platform based on cell perturbation data

Abstract

The application discloses a virtual cell analysis platform based on cell disturbance data. The platform comprises three parts: a large language model data preprocessing system, a database and an online analysis platform, and a direction matching algorithm. The large language model data preprocessing system comprises an automatic data information extraction and retrieval tool, an automatic data file download tool, a large language model analysis engine and a gene expression data acquisition and feature extraction tool. The application can automatically preprocess the original data by applying the large language model data preprocessing system to the target scientific problem, and then store the data in the database. The characteristic direction matching algorithm is combined to output the drugs and key targets related to the target data set. The efficiency of target point and drug screening is improved. The application is suitable for large-scale omics data mining and disturbance analysis, accelerates the drug research and development, disease mechanism and translational medicine research process, and reduces the research and development cost.

Description

Technical Field

[0001] This application belongs to the field of bioinformatics technology, specifically relating to a virtual cell analysis platform based on cell perturbation data. Background Technology

[0002] High-throughput sequencing technology can quantitatively detect the expression levels of the entire genome in a single operation, generating massive amounts of omics data. These data systematically reveal global changes in genes during disease occurrence, development, and drug intervention, providing a data foundation for understanding the molecular mechanisms of life activities and for precision medicine.

[0003] With the development of multi-omics technologies, cell perturbation experiments, and computational analysis methods, computational analysis based on cell state characterization and perturbation response characteristics has gradually become an important technical means for drug screening and disease research. By analyzing the transcriptomic changes of cells under different perturbation conditions, cell state and its response characteristics can be characterized in a computational environment, thereby providing support for candidate drug screening, mechanism of action research, and experimental validation.

[0004] Drug screening based on cellular perturbation transcriptome features primarily identifies inverse relationships by comparing the changes in gene expression at the gene expression level between disease characteristics and drug action characteristics. For example, when upregulated genes in disease characteristics show a downregulated trend in drug perturbation features, and downregulated genes in disease characteristics show an upregulated trend in drug perturbation features, the drug can be considered valuable as a candidate for further research. Compared to de novo drug development, screening using approved drugs or clinical-stage compounds helps shorten the development cycle and reduce development costs.

[0005] However, existing drug screening methods based on cell perturbation data still suffer from the following problems: experimental design description formats vary across different data sources; sample grouping information and perturbation conditions often require manual processing, resulting in low data preprocessing efficiency and susceptibility to human error; the cell background and perturbation types differ significantly across datasets, making direct comparisons of candidate drug ranking results insufficient; furthermore, existing methods offer limited support for standardized validation processes for screening results, hindering the formation of a closed-loop analysis process from data mining to experimental validation. To address these issues, the applicant has developed a cell perturbation data mining method based on a large language model and a virtual cell analysis platform. Summary of the Invention

[0006] This application provides a virtual cell analysis platform based on cell perturbation data. The technical solution is as follows: First, a large language model data preprocessing system is applied to perform unified automated data preprocessing on the raw data. Then, the processed cell perturbation data is stored in a database. A feature-directed matching algorithm is then used to output a list of drugs and key targets related to the target dataset. This significantly improves the efficiency of target and drug screening, is suitable for large-scale omics data mining and perturbation analysis, accelerates drug development, disease mechanism and translational medicine research, and reduces R&D costs. The analysis platform comprises three parts: a large language model data preprocessing system, a database and online analysis platform, and a transcriptome-based directional matching algorithm. The large language model data preprocessing system includes: automated data information extraction and retrieval tools, automated data file download tools, a large language model analysis engine, and gene expression data acquisition and feature extraction tools. The following is a detailed description of the functions of each part.

[0007] I. Large Language Model Data Preprocessing System.

[0008] 1. Automated Data Acquisition and Retrieval Tool: A program written in R (version 4.2.2) operates the tool, constructing queries based on drug names and using the unique identifier of a Series to call the Entrez summary interface to retrieve metadata and achieve data retrieval. The search results are presented in a table containing six columns: accession, gpl, title, summary, gdstype, and query_keyword. Specifically: accession is the unique identifier for Series data, starting with "GSE" (GSEID), which allows direct searching for complete data and research information. gpl is the detection platform number, indicating the experimental technology platform used in the study, starting with "GPL". Different platforms correspond to different microarrays or sequencing technologies. title is the official name of the study corresponding to the data, summarizing its core content. summary is a detailed description of the study in the title. gdstype describes the technical methods used in the gene expression detection experiment. query_keyword is the query term used to retrieve the entry.

[0009] 2. Automated data file download tool: Implemented by R language programming, it is used to obtain the GSEID of Series data and its Uniform Resource Locator (URL) mapping table to obtain Series matrix files. The URL of the series matrix file is composed of FTP address and GSEID, and the data download is achieved by automatically accessing the URL through R script. The downloaded Series matrix file is a standardized data file that integrates all relevant information of a research series into a text file. The Series matrix file is usually in txt format, and its content is divided into two parts: metadata and gene expression matrix, specifically: (1) Metadata part: The lines starting with ! are metadata, which are divided into two levels: A. Series-level information describes the overall information of the entire research project: !Series_title: Research title; !Series_geo_accession: Series GSEID; !Series_status: Data acquisition status and date; !Series_pubmed_id: Related literature; !Series_summary: Research abstract (purpose, methods, results, conclusions); !Series_overall_design: Overview of experimental design; !Series_type: Data type; !Series_contributor: Contributor name; !Series_ sample_id: A list of sample IDs included in this series, starting with GSM; !Series_contact_*: Various information about the contact person (name, email, laboratory, department, institute, address, etc.); !Series_supplementary_file: Download link for supplementary files; !Series_platform_id: The experimental platform ID used, starting with GPL; !Series_platform_taxid: The species taxonomy ID corresponding to the platform; !Series_sample_taxid: The species taxonomy ID of the sample; !Series_relation: Association with other databases; B.Sample-level information describes information for each individual sample: !Sample_title: Sample title; !Sample_geo_accession: Sample ID, starting with GSM; !Sample_status: Sample status; !Sample_submission_date: Sample submission date; !Sample_last_update_date: Sample last update date; !Sample_type: Sample type; !Sample_channel_count: Number of channels; !Sample_source_name_ch1: Sample source; !Sample_organism_ch1: Sample species; !Sample_characteristics_ch1: Sample characteristics (may have multiple entries, such as strain, cell line, cell type, developmental stage, genotype, sex, etc.); !Sample_growth_protocol_ch1: Growth or culture protocol; !Sample_molecule_ch1: Molecular type (e.g., total RNA); !Sample_extract _protocol_ch1: Extraction protocol (how to extract RNA, how to construct the library), !Sample_taxid_ch1: Sample species classification ID, !Sample_description: Sample description, !Sample_data_processing: Data processing steps (including software, process, parameters, etc.), !Sample_platform_id: Experimental platform number used for the sample, starting with GPL, !Sample_contact_*: Various information about the sample contact person, !Sample_data_row_count: Number of rows in the gene expression count matrix, !Sample_instrument_model: Sequencing instrument model, !Sample_library_selection: Library selection method, !Sample_library_source: Library source, !Sample_library_strategy: Library strategy, !Sample_relation: Relationship between the sample and other databases, !Sample_supplementary_file: Supplementary files for the sample. (2) The gene expression matrix section (only column name rows are present in high-throughput sequencing experiments) is located between !series_matrix_table_begin and !series_matrix_table_end: First column: ID_REF - gene or transcript identifier; subsequent columns: expression values ​​for each sample, with the column header being the GSM number.

[0010] 3. Large Language Model Analysis Engine: This engine extracts Series metadata to construct prompt words and calls the DeepSeek Large Language Model API. It is used for automated annotation of experimental design information for each Series and for quality control of annotation results. Automated annotation first extracts the overall description, title, GSEID, sample features, abstract, and experimental design information from the Series metadata, adding original constraints (column name restrictions, delimiters, and row meanings) to construct prompt word text. Then, it sends a JSON-encoded prompt word message body to the Large Language Model API. Finally, the text returned by the Large Language Model is converted into a table, yielding the annotation results. The annotation results include GSEID, cell type, control group sample number, experimental group sample number, experimental perturbation condition type, and experimental perturbation condition. Quality control of annotation results is conducted from three aspects: consistency of sample identifiers, accuracy of group design, and accuracy of drug information. Specifically, the following steps are taken: First, GSEID, cell type, control group sample number, experimental group sample number, experimental perturbation condition type, and experimental perturbation condition column are extracted from the annotation results as data to be controlled. Then, the overall description, title, GSEID, sample characteristics, abstract, and experimental design information from the Series metadata are extracted as reference data for verifying annotation accuracy. Next, the data to be controlled and the reference data are merged with the quality control constraint prompts and submitted to the large language model. The table text returned by the large language model is the quality control result. The quality control result combines the original annotations, quality control results, and necessary experimental information, ultimately including: GSEID, cell type, control group sample number, experimental group sample number, experimental perturbation condition type, experimental perturbation condition, quality control result ("+" indicates pass, "-" indicates fail; three "+"s indicate that the three parts of quality control for sample identifier consistency, group design accuracy, and drug information accuracy have passed), species, sequencing experimental platform number, and gene expression detection experimental type.

[0011] 4. Gene Expression Data Acquisition and Feature Extraction Tools: Programs written in Python and R were used to automate the download of gene expression profile files. The R packages DESeq2 and limma were used to calculate gene expression level differences. The download of gene expression profile files was automated using the program from the GREIN database (GEO RNA-seq Experiments Interactive Navigator, https: / / www.ilincs.org / apps / grein) and the Maayanlab database (https: / / labs.icahn.mssm.edu / maayanlab / ). The downloaded expression profile files are gene counting matrices where the first column is the gene identifier and the remaining columns are sample numbers (starting with GSM). For datasets without gene expression profiles, the raw sequencing data was obtained and the aforementioned format of gene expression profiles was generated through calculation. Differential analysis calculates the difference in gene expression levels between experimental and control samples. For high-throughput sequencing data, the DESeq2 package in R is used to perform differential calculations. The results are a table containing 7 columns: genes, logFC, baseMean, lfcSE, stat, pvalue, and padj. Among them, genes are gene identifiers, logFC is log2FoldChange, which represents the log2 fold change of the experimental group relative to the control group, baseMean is the average count of each sample after size factor normalization for that gene, reflecting the overall gene expression abundance level, lfcSE is the standard error of log2FoldChange, stat is the test statistic, pvalue is the original p-value generated by the above statistical tests, and padj is the p-value after multiple test correction. For microarray data: The differential expression calculation was performed using the R package limma. The result is a table containing genes, logFC, AveExpr, t, P.Value, adj.P.Val, and B. Where: genes: probe / gene identification and annotation information; logFC: log2 fold change of the experimental group relative to the control group; AveExpr: the average log2 expression level of the gene in all samples / channels during the experiment (overall average level); t: modified t statistic (t value after empirical Bayesian smoothing); P.Value: the original p value corresponding to the modified t above; adj.P.Val: p value after multiple test correction; B: the log-odds ratio of differential expression of the gene.

[0012] II. Database and Online Analysis Platform.

[0013] The platform includes a reference drug perturbation feature library, a drug-gene association library, a drug annotation library, and an online analysis platform. The reference drug perturbation feature library stores transcriptomic changes caused by drugs acting on cells. This library uses a hierarchical data structure, organizing data in the format of "gene × drug perturbation feature," where rows represent unified gene identifiers, columns represent perturbation features formed by different drugs in different cellular backgrounds, and the values ​​in the matrix represent the standardized expression change statistics of the corresponding gene under the drug perturbation condition. The reference drug perturbation feature library is used to match user-submitted query features one by one to complete the directional matching calculation and ranking of candidate drugs. The drug-gene association library stores the association relationships between drug perturbation features and genes. This library uses a relational database structure to store pre-organized drug perturbation feature identifiers and their associated gene information. During analysis, the system compares the input differentially expressed genes for diseases with the associated gene information in the drug-gene association library to obtain a list of associated genes corresponding to each candidate drug and their quantity, which helps to interpret the molecular correspondence between candidate drugs and input differentially expressed features. The drug annotation library stores basic annotation information for candidate drugs. The library stores drug names and their corresponding classifications and descriptions in a structured table format. After candidate drugs are ranked, the system integrates drug annotation information with directional matching scores, displaying and exporting the results in tabular form. The online analysis platform adopts a front-end and back-end collaborative task-based processing architecture. After users submit up-regulation and down-regulation genes through the web interface, the back-end service receives the request and generates a unique task identifier, saving the input data, task status, and result files in the corresponding task directory. Subsequently, an independent background analysis process is started to perform gene preprocessing, reference library retrieval, directional matching scoring, result organization, and annotation integration. After the analysis is completed, structured result files and tabular result files are generated for web display and download. Users can check the task running status through the task result page and view the candidate drug ranking results after the task is completed. For users who provide a contact email address, the platform sends email notifications at the start and end of the task.

[0014] III. Directional matching algorithm based on transcriptome features.

[0015] The platform's relevance calculation employs a directional matching algorithm based on transcriptome features. For datasets that have passed quality control, the platform first extracts sets of genes with increased expression and sets of genes with decreased expression based on the differential expression analysis results between the experimental and control groups. These sets are then used as the upregulated and downregulated gene sets for the query feature, respectively. Subsequently, the query feature is matched one by one with each drug perturbation feature in the reference drug perturbation feature library. The strength of the directional matching between candidate drugs and the query feature is calculated, thereby completing the candidate drug ranking. Detailed procedures are attached. Figure 1As shown. Suppose a certain reference drug perturbation feature contains N genes, and an ordered gene sequence is formed according to the statistical value of each gene in descending order under this perturbation condition. For any gene set S in the query feature, first calculate its weighted normalization factor: Where, r j Indicates gene g j The statistics in the perturbation features of the reference drug. Subsequently, the sequence is scanned position-by-position along the ordered gene sequence, and the cumulative hit and non-hit terms are calculated respectively: , And define the running deviation of the i-th position as: The enrichment score for this gene set is determined by the larger absolute value of the maximum positive or negative deviation during the entire scan process. After calculating enrichment scores for the upregulated and downregulated gene sets respectively, a weighted directional matching score of the reference drug perturbation feature relative to the query feature is further obtained. If the enrichment directions of the upregulated and downregulated gene sets are opposite, the weighted directional matching score is defined as follows: Where U represents the upregulated gene set and D represents the downregulated gene set. If the enrichment directions of both are the same, then the perturbation feature of the reference drug and the query feature do not form an effective directional correspondence, and the weighted directional matching score is recorded as 0. Further, the platform can group and normalize the weighted directional matching scores of each candidate drug according to the cell background and perturbation type in the reference library to improve the comparability of candidate drug ranking results under different experimental backgrounds, and preferably rank candidate drugs based on the normalized results. Further, the platform performs statistical significance evaluation on the weighted directional matching scores of candidate drugs and performs multiple tests to correct the significance results, obtaining the corresponding significance index. Finally, the platform outputs the candidate drug identifier, candidate drug name, weighted directional matching score, significance result, multiple correction result, and the number of upregulated and downregulated genes involved in the calculation; when drug annotation data and drug-gene association data are configured, it also outputs the annotation information of the candidate drug, associated gene information, and the number of associated genes for subsequent drug screening and manual evaluation.

[0016] In summary, this application has the following beneficial effects: This application provides a virtual cell analysis platform based on cell perturbation data. By optimizing prompt word design, integrating real-world omics data, and implementing strict quality control, it achieves efficient and automated mining and annotation of omics data and accurately calculates changes in cell gene expression caused by drugs. Specifically, it is reflected in the following technical features: 1. Construct prompt words based on Series metadata and call a large language model to automatically annotate experimental design information; 2. Quality control of annotation results is performed based on the consistency of sample identifiers, the accuracy of grouping design, and the accuracy of drug information; 3. The upregulated and downregulated gene sets formed by differential expression are directionally matched with the perturbation features of the reference drug one by one; 4. Calculate a weighted directional matching score when the enrichment directions of the upregulated gene set and the downregulated gene set are opposite; 5. The scoring results were grouped and normalized according to cell background and perturbation type, and further statistical significance assessment and multiple test correction were performed; 6. Furthermore, this application also supplements the verification process based on organoid models, namely, forming a closed-loop verification path through organoid model construction, transcriptome or single-cell omics sequencing, platform screening, and organoid validity verification; This application addresses the need for automated annotation and quality control of cellular perturbation data from heterogeneous sources with inconsistent experimental designs. It enables the ranking of candidate drugs with comparable and statistically evaluable capabilities based on the perturbation transcriptome characteristics of reference drugs across different cellular backgrounds and perturbation types, while simultaneously forming a validation loop using organoid models. This significantly improves data annotation efficiency and accuracy, substantially reduces manual costs and random errors, and is applicable to large-scale omics data mining and drug screening. It accelerates drug development, scientific hypothesis generation, and therapeutic research, thereby promoting the rapid development of biomedical disciplines and the healthcare industry. Attached Figure Description

[0017] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 A flowchart of a virtual cell analysis platform based on cell perturbation data.

[0019] Figure 2 The flowchart illustrates the analysis and verification of the therapeutic effect of imatinib mesylate on gastric signet ring cell carcinoma using a virtual cell analysis platform based on cellular perturbation data and combined with gastric signet ring cell carcinoma organoids.

[0020] Figure 3 Volcano plot of gene expression characteristic of gastric signet ring cell carcinoma, with a threshold of |fold change (log2)| > 1 and a corrected P-value < 0.05.

[0021] Figure 4 Image showing the results of drug sensitivity cell viability assay (CellTiter-Glo, CTG) for gastric signet ring cell carcinoma organoids from Patient 1.

[0022] Figure 5Image showing the results of drug sensitivity cell viability assay (CellTiter-Glo, CTG) for gastric signet ring cell carcinoma organoids from Patient 2.

[0023] Figure 6 Image showing the results of drug sensitivity cell viability assay (CellTiter-Glo, CTG) for gastric signet ring cell carcinoma organoids from Patient 3.

[0024] Figure 7 Image showing the results of drug sensitivity cell viability assay (CellTiter-Glo, CTG) for gastric signet ring cell carcinoma organoids from Patient 4.

[0025] Figure 8 Image showing the results of drug sensitivity cell viability assay (CellTiter-Glo, CTG) for gastric signet ring cell carcinoma organoids from patient 5.

[0026] Figure 9 Image showing the results of drug sensitivity cell viability assay (CellTiter-Glo, CTG) for gastric signet ring cell carcinoma organoids from patient 6. Detailed Implementation

[0027] To better understand this application, the following description, in conjunction with embodiments and accompanying drawings, further illustrates the application. It should be noted that the specific embodiments described herein are merely for illustrative and explanatory purposes and are not intended to limit the scope of this application. Unless otherwise specified in the embodiments, conventional conditions or conditions recommended by the manufacturer are followed. Reagents or instruments used without a specified manufacturer are all commercially available conventional products.

[0028] Example 1: A virtual cell analysis platform based on cell perturbation data.

[0029] This platform automatically acquires cellular perturbation data for analysis, uses a large language model to automate data annotation and quality control, and calculates and analyzes changes in whole-genome expression profiles caused by gene expression perturbation factors, including drugs, as a basis for perturbation screening. This platform significantly improves the efficiency of target and drug screening, saving costs in target discovery and new drug development. (See attached image) Figure 1As shown, the technical solution of this application is as follows: First, a large language model data preprocessing system is applied to perform unified automated data preprocessing on the original data. Then, the processed perturbation data is stored in a database, and a feature-directed matching algorithm is used to output a list of drugs and key targets related to the target dataset. The analysis platform includes three parts: a large language model data preprocessing system, a database and online analysis platform, and a transcriptome-based directional matching algorithm. The large language model data preprocessing system includes: automated data information extraction and retrieval tools, automated data file download tools, a large language model analysis engine, and gene expression data acquisition and feature extraction tools. The following is a detailed description of the functions of each part.

[0030] I. Large Language Model Data Preprocessing System.

[0031] 1. Automated Data Extraction and Retrieval Tool: A program written in R (version 4.2.2) operates the tool, constructing queries based on drug names and using the unique identifier of a Series to call the Entrez summary interface to retrieve metadata and achieve data retrieval. The search results are presented in a table containing six columns: accession, gpl, title, summary, gdstype, and query_keyword. Specifically: accession is the unique identifier for Series data, starting with "GSE" (GSEID), which allows direct searching of complete data and research information in the database. gpl is the detection platform number, indicating the experimental technology platform used in the study, starting with "GPL". Different platforms correspond to different microarrays or sequencing technologies. title is the official name of the study corresponding to the data, summarizing its core content. summary is a detailed description of the study in the title. gdstype describes the technical methods used in the gene expression detection experiment. query_keyword is the query term used to retrieve the entry.

[0032] 2. Automated data file download tool: Implemented by R language programming, it is used to obtain the GSEID of Series data and its Uniform Resource Locator (URL) mapping table, and access the URL to download Series matrix files. The URL of the Series matrix file is composed of FTP address and GSEID, and the data download is achieved by automatically accessing the URL through R script. The downloaded Series matrix file is a standardized data file that integrates all relevant information of a research series into a text file. The Series matrix file is usually in txt format, and its content is divided into two parts: metadata and gene expression matrix, specifically: (1) Metadata part: The lines starting with ! are metadata, which are divided into two levels: A. Series-level information describes the overall information of the entire research project: !Series_title: Research title; !Series_geo_accession: Series GSEID; !Series_status: Data acquisition status and date; !Series_pubmed_id: Related literature; !Series_summary: Research abstract (purpose, methods, results, conclusions); !Series_overall_design: Overview of experimental design; !Series_type: Data type; !Series_contributor: Contributor name; !Series_ sample_id: A list of sample IDs included in this series, starting with GSM; !Series_contact_*: Various information about the contact person (name, email, laboratory, department, institute, address, etc.); !Series_supplementary_file: Download link for supplementary files; !Series_platform_id: The experimental platform ID used, starting with GPL; !Series_platform_taxid: The species taxonomy ID corresponding to the platform; !Series_sample_taxid: The species taxonomy ID of the sample; !Series_relation: Association with other databases; B.Sample-level information describes information for each individual sample: !Sample_title: Sample title; !Sample_geo_accession: Sample ID, starting with GSM; !Sample_status: Sample status; !Sample_submission_date: Sample submission date; !Sample_last_update_date: Sample last update date; !Sample_type: Sample type; !Sample_channel_count: Number of channels; !Sample_source_name_ch1: Sample source; !Sample_organism_ch1: Sample species; !Sample_characteristics_ch1: Sample characteristics (may have multiple entries, such as strain, cell line, cell type, developmental stage, genotype, sex, etc.); !Sample_growth_protocol_ch1: Growth or culture protocol; !Sample_molecule_ch1: Molecular type (e.g., total RNA); !Sample_extract _protocol_ch1: Extraction protocol (how to extract RNA, how to construct the library), !Sample_taxid_ch1: Sample species classification ID, !Sample_description: Sample description, !Sample_data_processing: Data processing steps (including software, process, parameters, etc.), !Sample_platform_id: Experimental platform number used for the sample, starting with GPL, !Sample_contact_*: Various information about the sample contact person, !Sample_data_row_count: Number of rows in the gene expression count matrix, !Sample_instrument_model: Sequencing instrument model, !Sample_library_selection: Library selection method, !Sample_library_source: Library source, !Sample_library_strategy: Library strategy, !Sample_relation: Relationship between the sample and other databases, !Sample_supplementary_file: Supplementary files for the sample. (2) The gene expression matrix section (only column name rows are present in high-throughput sequencing experiments) is located between !series_matrix_table_begin and !series_matrix_table_end: First column: ID_REF - gene or transcript identifier; subsequent columns: expression values ​​for each sample, with the column header being the GSM number.

[0033] 3. Large Language Model Analysis Engine: This engine extracts Series metadata to construct prompt words and calls the DeepSeek Large Language Model API. It is used for automated annotation of experimental design information for each Series and for quality control of annotation results. Automated annotation first extracts the overall description, title, GSEID, sample features, abstract, and experimental design information from the Series metadata, adding original constraints (column name restrictions, delimiters, and row meanings) to construct prompt word text. Then, it sends a JSON-encoded prompt word message body to the Large Language Model API. Finally, the text returned by the Large Language Model is converted into a table, yielding the annotation results. The annotation results include GSEID, cell type, control group sample number, experimental group sample number, experimental perturbation condition type, and experimental perturbation condition. Quality control of annotation results is conducted from three aspects: consistency of sample identifiers, accuracy of group design, and accuracy of drug information. Specifically, the following steps are taken: First, GSEID, cell type, control group sample number, experimental group sample number, experimental perturbation condition type, and experimental perturbation condition column are extracted from the annotation results as data to be controlled. Then, the overall description, title, GSEID, sample characteristics, abstract, and experimental design information from the Series metadata are extracted as reference data for verifying annotation accuracy. Next, the data to be controlled and the reference data are merged with the quality control constraint prompts and submitted to the large language model. The table text returned by the large language model is the quality control result. The final quality control result combines the original annotations, quality control results, and necessary experimental information, ultimately including: GSEID, cell type, control group sample number, experimental group sample number, experimental perturbation condition type, experimental perturbation condition, quality control result ("+" indicates pass, "-" indicates fail; three "+"s indicate that the three parts of quality control for sample identifier consistency, group design accuracy, and drug information accuracy have passed), species, sequencing experimental platform number, and gene expression detection experimental type.

[0034] 4. Gene Expression Data Acquisition and Feature Extraction Tools: Programs written in Python and R were used to automate the download of gene expression profile files. The R packages DESeq2 and limma were used to calculate gene expression level differences. The download of gene expression profile files was automated using the program from the GREIN database (GEO RNA-seq Experiments Interactive Navigator, https: / / www.ilincs.org / apps / grein) and the Maayanlab database (https: / / labs.icahn.mssm.edu / maayanlab / ). The downloaded expression profile files are gene counting matrices where the first column is the gene identifier and the remaining columns are sample numbers (starting with GSM). For datasets without gene expression profiles, the raw sequencing data was obtained and the aforementioned format of gene expression profiles was generated through calculation. Differential analysis calculates the difference in gene expression levels between the experimental and control groups. For high-throughput sequencing data, the DESeq2 package in R is used to perform differential calculations. The results are a table with 7 columns: genes, logFC, baseMean, lfcSE, stat, pvalue, and padj. Among them, genes are gene identifiers, logFC is log2FoldChange, which represents the log2 fold change of the experimental group relative to the control group, baseMean is the average count of each sample after size factor normalization for that gene, reflecting the overall gene expression abundance level, lfcSE is the standard error of log2FoldChange, stat is the test statistic, pvalue is the original p-value generated by the above statistical tests, and padj is the p-value after multiple test correction. For microarray data: The differential expression calculation was performed using the R package limma. The result is a table containing genes, logFC, AveExpr, t, P.Value, adj.P.Val, and B. Where: genes: probe / gene identification and annotation information; logFC: log2 fold change of the experimental group relative to the control group; AveExpr: the average log2 expression level of the gene in all samples / channels during the experiment (overall average level); t: modified t statistic (t value after empirical Bayesian smoothing); P.Value: the original p value corresponding to the modified t above; adj.P.Val: p value after multiple test correction; B: the log-odds ratio of differential expression of the gene.

[0035] II. Database and Online Analysis Platform The platform includes a reference drug perturbation feature library, a drug-gene association library, a drug annotation library, and an online analysis platform. The reference drug perturbation feature library stores transcriptomic changes caused by drugs acting on cells. This library uses a hierarchical data structure, organizing data in the format of "gene × drug perturbation feature," where rows represent unified gene identifiers, columns represent perturbation features formed by different drugs in different cellular backgrounds, and the values ​​in the matrix represent the standardized expression change statistics of the corresponding gene under the drug perturbation condition. The reference drug perturbation feature library is used to match user-submitted query features one by one to complete the directional matching calculation and ranking of candidate drugs. The drug-gene association library stores the association relationships between drug perturbation features and genes. This library uses a relational database structure to store pre-organized drug perturbation feature identifiers and their associated gene information. During analysis, the system compares the input differentially expressed genes with the associated gene information in the drug-gene association library to obtain a list of associated genes corresponding to each candidate drug and their quantity, which helps to interpret the molecular correspondence between candidate drugs and input differentially expressed features. The drug annotation library stores basic annotation information for candidate drugs. The library uses a structured table format to store drug names and their corresponding classifications and descriptions. After candidate drugs are ranked, the system integrates the drug annotation information with the directional matching score results to form a result table that can be directly displayed and exported. The online analysis platform adopts a front-end and back-end collaborative task-based processing architecture. After users submit up-regulation and down-regulation genes through the web interface, the back-end service receives the request and generates a unique task identifier, saving the input data, task status, and result files in the corresponding task directory; then, an independent background analysis process is started to perform gene preprocessing, reference library retrieval, directional matching scoring, result organization, and annotation integration; after the analysis is completed, structured result files and tabular result files are generated for web display and download. Users can check the task running status through the task result page and view the candidate drug ranking results after the task is completed; for users who fill in their contact email, the platform sends corresponding notification information when the task starts and ends.

[0036] III. Directional matching algorithm based on transcriptome features.

[0037] The platform's relevance calculation employs a directional matching algorithm based on transcriptome features. For datasets that have passed quality control, the platform first extracts sets of genes with increased expression and sets of genes with decreased expression based on the differential expression analysis results between the experimental and control groups. These sets are then used as the upregulated and downregulated gene sets for the query feature, respectively. Subsequently, the query feature is matched one by one with each drug perturbation feature in the reference drug perturbation feature library. The strength of the directional matching between candidate drugs and the query feature is calculated, thereby completing the candidate drug ranking. Detailed procedures are attached. Figure 1As shown. Suppose a certain reference drug perturbation feature contains N genes, and an ordered gene sequence is formed according to the statistical value of each gene in descending order under this perturbation condition. For any gene set S in the query feature, first calculate its weighted normalization factor: Where, r j Indicates gene g j The statistics in the perturbation features of the reference drug. Subsequently, the sequence is scanned position-by-position along the ordered gene sequence, and the cumulative hit and non-hit terms are calculated respectively: , And define the running deviation of the i-th position as: The enrichment score for this gene set is determined by the larger absolute value of the maximum positive or negative deviation during the entire scan process. After calculating enrichment scores for the upregulated and downregulated gene sets respectively, a weighted directional matching score of the reference drug perturbation feature relative to the query feature is further obtained. If the enrichment directions of the upregulated and downregulated gene sets are opposite, the weighted directional matching score is defined as follows: Where U represents the upregulated gene set and D represents the downregulated gene set. If the enrichment directions of the two are the same, the perturbation feature of the reference drug and the query feature do not form an effective directional correspondence, and the weighted directional matching score is recorded as 0. Further, the platform can group and normalize the weighted directional matching scores of each candidate drug according to the cell background and perturbation type in the reference library to improve the comparability of candidate drug ranking results under different experimental backgrounds, and preferably rank candidate drugs based on the normalized results. Further, the platform performs statistical significance evaluation on the weighted directional matching scores of candidate drugs and performs multiple tests to correct the significance results, obtaining the corresponding significance index. Finally, the platform outputs the candidate drug identifier, candidate drug name, weighted directional matching score, significance result, multiple correction result, and the number of upregulated and downregulated genes involved in the calculation; when drug annotation data and drug-gene association data are configured, it also outputs the annotation information of the candidate drug, associated gene information, and the number of associated genes for subsequent drug screening and manual evaluation.

[0038] In summary, the virtual cell analysis platform based on cell perturbation data can perform the following functions: (1) automatically collect, retrieve, standardize, and extract features of cell perturbation data based on a large language model, and automatically update and store it in the platform database; (2) provide an open interface and a user-friendly analysis interface to realize the analysis functions of data upload and result download; (3) realize the rapid discovery of drugs and targets based on cell perturbation data, shorten the research and development cycle, and reduce research and development costs.

[0039] Example 2: Using the aforementioned virtual cell analysis platform based on cell perturbation data, combined with organoids from 6 patients with gastric signet ring cell carcinoma, the therapeutic effect of imatinib mesylate on gastric signet ring cell carcinoma was discovered.

[0040] The incidence of gastric signet ring cell carcinoma is continuously increasing, accounting for 8% to 30% of gastric cancer subtypes. Currently, the treatment of gastric signet ring cell carcinoma still faces severe challenges. For a long time, there have been limited targeted therapy options for gastric signet ring cell carcinoma; apart from apatinib, no other targeted drugs with clear indications for gastric signet ring cell carcinoma have been included in first-line treatment regimens. Since 2010, personalized medication guidance based on gene mutation detection has been widely used in tumor treatment, but the analysis results for patients with gastric signet ring cell carcinoma are often unsatisfactory. Organoids are multicellular three-dimensional structures cultured in vitro from adult stem cells, embryonic stem cells, or induced pluripotent stem cells from primary tissues. They can simulate the structure and function of primary tumors and are good models for tumor research and drug efficacy evaluation. Therefore, using the cell perturbation data mining method and virtual cell analysis platform based on the large language model described in this application, combined with organoids from 6 patients with gastric signet ring cell carcinoma, we screened and verified the therapeutic effect of imatinib mesylate on gastric signet ring cell carcinoma. The analysis and verification process is attached. Figure 2 As shown, we first performed transcriptome sequencing and organoid model construction on surgical specimens of 6 gastric signet ring cell carcinomas to obtain the characteristic expression profiles of gastric signet ring cell carcinomas (see attached). Figure 3 The cell perturbation data mining method based on the large language model and the virtual cell analysis platform were used to analyze potential therapeutic drugs. Imatinib mesylate was selected as the lead drug. Finally, the lead drug was placed into the established gastric signet ring cell carcinoma organoid model at the set concentration for co-culture. Drug sensitivity cell viability was detected and the half maximal inhibitory concentration (IC50) was calculated. The results showed that the half-maximal inhibitory concentrations (IC50) of gastric signet ring cell carcinoma organoids from the six patients were 69.06 μM, 63.04 μM, 30.16 μM, 67.52 μM, 29.42 μM, and 33.88 μM, respectively, in the efficacy curves of the positive control drug apatinib. Imatinib mesylate showed good efficacy curves with IC50 values ​​of 49.43 μM, 44.01 μM, 18.51 μM, 31.78 μM, 16.61 μM, and 13.17 μM, respectively, all significantly lower than the IC50 of the positive control drug apatinib. This demonstrates that imatinib mesylate has a killing effect on gastric signet ring cell carcinoma organoids. Detailed results are attached. Figure 4-9 .

[0041] In summary, drugs or targets screened using the aforementioned virtual cell analysis platform based on cell perturbation data can be experimentally validated in organoid models. The specific steps are as follows: 1) Construct an organoid model of the target scientific question and perform transcriptomic or single-cell omics sequencing on it; 2) Input the differential expression features obtained from sequencing into the aforementioned virtual cell analysis platform based on cell perturbation data to screen lead drugs or candidate targets; 3) Validate the effectiveness of the screened lead drugs or candidate targets in the organoid model.

Claims

1. A virtual cell analysis platform based on cell perturbation data, characterized in that, include: The system includes a large language model data preprocessing system, a reference drug perturbation feature library, a drug-gene association library, a drug annotation library, a directional matching analysis module, and an online analysis platform. The large language model data preprocessing system is used to automatically retrieve, download, design and label experimental data, and perform quality control on cell perturbation data, and generate preprocessed data for differential expression analysis. The reference drug perturbation feature library is used to store transcriptome changes caused by drugs acting on cells and to match them one by one with the query features. The directional matching analysis module is used to form an upregulated gene set and a downregulated gene set based on the differential expression results between the experimental group and the control group, and to perform directional matching calculations between the upregulated gene set and the downregulated gene set and each drug perturbation feature in the reference drug perturbation feature library to obtain the candidate drug ranking results. The drug-gene association library and drug annotation library are used to associate and integrate the ranking results of candidate drugs; The online analysis platform is used to receive query data submitted by users, call the large language model data preprocessing system and the directional matching analysis module to perform analysis, and output candidate drug results.

2. The virtual cell analysis platform based on cell perturbation data according to claim 1, characterized in that, The large language model data preprocessing system includes: an automated data information extraction and retrieval tool, an automated data file download tool, a large language model analysis engine, and a gene expression data acquisition and feature extraction tool; The automated data extraction and retrieval tool is used to construct a query based on the drug name or perturbation conditions and retrieve cell perturbation data. The automated data file download tool is used to download Series matrix files and corresponding expression spectrum files; The large language model analysis engine is used to construct prompt words based on Series metadata, automatically perform structured annotation of experimental design information, and perform quality control on the annotation results based on the consistency of sample identifiers, the accuracy of group design, and the accuracy of drug information. The gene expression data acquisition and feature extraction tool is used to acquire gene expression profile files and perform differential expression analysis on experimental group samples and control group samples to extract upregulated gene sets and downregulated gene sets.

3. The virtual cell analysis platform based on cell perturbation data according to claim 1, characterized in that, The directional matching analysis module is used for: The enrichment of the upregulated and downregulated gene sets in each drug perturbation feature was scored; When the enrichment directions of the upregulated gene set and the downregulated gene set are opposite, calculate the weighted directional matching score of the drug perturbation feature relative to the query feature; When the enrichment directions of the upregulated gene set and the downregulated gene set are the same, the weighted directional matching score is recorded as 0; The weighted directional matching scores were grouped and normalized according to the cell background and perturbation type to which the reference drug perturbation characteristics belonged, and the normalized results were subjected to statistical significance evaluation and multiple test correction.

4. The virtual cell analysis platform based on cell perturbation data according to claim 1, characterized in that, The selected drugs and targets are experimentally validated in organoid models. The specific steps are as follows: 1) Construct an organoid model of the target scientific question and perform transcriptome or single-cell omics sequencing on it; 2) Input the differential expression features obtained from sequencing into a virtual cell analysis platform based on cell perturbation data to screen lead drugs or candidate targets; 3) Validate the effectiveness of the screened lead drugs or candidate targets in the organoid model.

Images