Cell transcriptomics-based phenotypic drug molecule discovery method and related assembly

Abstract

The embodiment of the invention relates to the crossing field of generating drug molecules by using a large language model, in particular to a phenotypic drug molecule discovery method, device and equipment based on cell transcriptomics and a computer readable storage medium. The method comprises the following steps: acquiring transcriptome sequencing result data to be analyzed, and extracting a differential expression gene list from the transcriptome sequencing result data; obtaining a differential gene list text description corresponding to the differential expression gene list; inputting the differential gene list text description into a trained molecular generation model to obtain at least one candidate molecular structure representation; the molecular generation model is used for converting the differential gene text description serving as gene expression information into a molecular reasoning result causing the change. The method can break through the dependence of traditional drug research and development on a known compound library, brand new molecules meeting functional requirements are generated, and the chemical space of drug discovery is expanded.

Description

Technical Field

[0001] The embodiments of the present invention relate to the interdisciplinary field of generating drug molecules using large language models, and particularly to a method, apparatus, device, and computer-readable storage medium for phenotypic drug molecule discovery based on cell transcriptomics. Background Technology

[0002] Phenotypic Drug Discovery (PDD) is an important research direction in the field of drug development for complex diseases (such as cancer and neurodegenerative diseases). It does not presuppose specific molecular targets, but directly observes phenotypic changes at the cellular or tissue level to screen or design candidate compounds that can reverse disease-related phenotypes, providing new avenues for the treatment of multifactorial diseases.

[0003] With the development of high-throughput sequencing technology, transcriptome sequencing (RNA-seq) and single-cell transcriptome sequencing (scRNA-seq) have become key supporting technologies for phenotypic drug discovery. These technologies can comprehensively capture the gene expression profile of cells under specific conditions. The transcriptome sequencing results of each cell contain the expression values ​​of approximately 20,000 genes, which can be constructed into a one-dimensional vector. The vectors from a large number of cells are aggregated to form a large-scale gene expression matrix. This high-dimensional expression profile can accurately reflect the state and function of cells, becoming a core tool for characterizing cell phenotypes, constructing disease models, and elucidating drug mechanisms of action.

[0004] Artificial intelligence technologies, especially generative models (such as variational autoencoders, generative adversarial networks, and Transformer-based molecular generators), are maturing in molecular design and drug development, offering new possibilities for phenotype-driven drug discovery by combining transcriptomics data with AI. However, directly combining transcriptomics phenotypic information with molecular generation techniques still faces several challenges. Transcriptome feature-matching retrieval methods, such as CMap and TransiGen, can only search and screen within known molecular sets and cannot design or generate entirely new compounds. While some generative methods (such as VAE methods) possess molecular generation capabilities, their direct reliance on quantitative features of the original transcriptome often leads to batch effects and differences in sequencing platforms, resulting in insufficient model generalization and difficulty in stably reproducing generation results across experimental conditions, thus limiting their application value in real-world drug discovery scenarios.

[0005] Therefore, how to overcome the limitations of fixed chemical libraries while improving the reliability and applicability of phenotypic drug discovery has become an urgent problem for those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide at least one method, apparatus, device, and computer-readable storage medium for phenotypic drug molecule discovery based on cell transcriptomics, which can overcome the limitations of fixed chemical libraries and improve the reliability and applicability of phenotypic drug discovery.

[0007] To address the aforementioned technical problems, at least one embodiment of this application provides a method for phenotypic drug molecule discovery based on cell transcriptomics, comprising: Obtain the transcriptome sequencing results data to be analyzed, and extract a list of differentially expressed genes from the transcriptome sequencing results data; Obtain the text description of the differentially expressed gene list corresponding to the differentially expressed gene list; The textual description of the differentially expressed gene list is input into a trained molecular generation model to obtain at least one candidate molecular structure representation; the molecular generation model is used to convert the textual description of the differentially expressed genes, which serves as gene expression information, into molecular inference results that lead to the change.

[0008] In one embodiment, the phenotypic drug molecule discovery method based on cell transcriptomics further includes: Obtain a molecular generation training set; the molecular generation training set includes multiple sets of first training data, each set of first training data includes a text description of a differentially expressed gene list with corresponding relationships and a candidate molecular structure representation, wherein the text description of the differentially expressed gene list is a text description of a differentially expressed gene list between disease cell model transcriptome data and healthy cell model transcriptome data, and the candidate molecular structure representation is a structural representation of the molecule that causes the change; The molecular generation training set is input into the scientific foundation big model to perform secondary training on the scientific foundation big model for a biological matching task, thereby obtaining the trained molecular generation model.

[0009] In one embodiment, the step of inputting the molecular generation training set into the large-scale scientific model for secondary training of the large-scale scientific model on a biological matching task further includes: The scientific foundation model was retrained for a biological matching task using known compound-transcriptome perturbation pairs as supervisory signals.

[0010] In one embodiment, obtaining the text description of the differentially expressed gene list includes: The differentially expressed gene list is input into a trained gene and cell type description generation model to obtain the differentially expressed gene text description.

[0011] In one embodiment, the phenotypic drug molecule discovery method based on cell transcriptomics further includes: Obtain a text conversion training set, which includes multiple sets of second training data, each set of second training data including a biological name and a text description corresponding to the biological name; The text conversion training set is input into the scientific foundation big model to perform secondary training on the scientific foundation big model for biological description, resulting in the trained gene and cell type description generation model.

[0012] In one embodiment, after obtaining the candidate molecule sequence representation, the method further includes: The text description of the differentially expressed gene list and its corresponding at least one candidate molecular structure representation are input into a trained molecular-phenotype matching model to obtain a matching score between the differentially expressed gene text description and each candidate molecular structure representation. At least one of the candidate molecular structure representations is sorted in descending order based on the matching score to obtain candidate molecular sequence representations.

[0013] In one embodiment, the phenotypic drug molecule discovery method based on cell transcriptomics further includes: Obtain a molecular-phenotype matching training set, which includes multiple sets of third training data. Each set of third training data includes a text description of a differential gene list and its corresponding multiple molecular structure representations, as well as a matching score between the text description of the differential gene list and each corresponding molecular representation, wherein the matching score is between 0 and 1. The molecular-phenotype matching training set is input into a preset molecular-phenotype matching model for scoring ability training to obtain the molecular-phenotype matching model.

[0014] At least one embodiment of this application also provides a phenotypic drug molecule discovery device based on cell transcriptomics, comprising: The data acquisition module is used to acquire the transcriptome sequencing results data to be analyzed and extract a list of differentially expressed genes from the transcriptome sequencing results data; The gene description acquisition module is used to acquire the text description of the differentially expressed gene list corresponding to the differentially expressed gene list. A molecular structure generation module is used to input the textual description of the differentially expressed gene list into a trained molecular generation model to obtain at least one candidate molecular structure representation; the molecular generation model is used to convert the textual description of the differentially expressed genes, which serves as gene expression information, into molecular inference results that lead to the change.

[0015] At least one embodiment of this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method for phenotypic drug molecule discovery based on cell transcriptomics.

[0016] At least one embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for phenotypic drug molecule discovery based on cell transcriptomics.

[0017] The embodiments of this application provide a phenotypic drug molecule discovery method based on cell transcriptomics. This method acquires transcriptome sequencing data to be analyzed, extracts a list of differentially expressed genes, and directly focuses on the core differences in gene expression between disease and healthy states. It abandons the limitation of traditional target-guided drug discovery that pre-sets a single target, and can cover the complex pathological features of diseases with multi-gene and multi-pathway synergistic disorders. It provides a basis for subsequent molecular design that reflects the essence of disease phenotype and avoids the limitation of compound efficacy caused by target selection bias. After converting the list of differentially expressed genes into a text description of the differential gene list, it is input into a trained molecular generation model to achieve a precise mapping between phenotypic information and molecular structure. The text conversion solves the pain point that traditional numerical gene expression matrices are difficult to be directly processed by language-based models, and can fully activate the semantic understanding ability of molecular generation models of gene function and pathway association. This method breaks through the traditional technical framework of retrieving known molecules based on a fixed compound library. It can generate novel molecular structures that do not appear in the training data, significantly expanding the space of medicinal chemistry and enhancing the innovation of drug discovery. At the same time, the process from extracting differentially expressed genes from transcriptome data to generating candidate molecules is coherent, without the need for repeated connection of multiple models, reducing the accumulation of errors in intermediate steps. Furthermore, the text input is adapted to the rapid processing characteristics of AI models, which can shorten the cycle from disease phenotype analysis to molecular design, reduce the time and data cost of drug development, and lay an efficient foundation for subsequent candidate molecule screening and experimental validation. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0019] Figure 1 This is a flowchart of a phenotypic drug molecule discovery method based on cell transcriptomics provided in one embodiment of this application; Figure 2 This is a schematic diagram of a phenotypic drug molecule discovery device based on cell transcriptomics provided in one embodiment of this application. Figure 3 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0021] This invention proposes a phenotypic drug molecule discovery method based on cell transcriptomics and a phenotypic drug generation method based on cell transcriptomics. The implementation details of the phenotypic drug molecule discovery method based on cell transcriptomics in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0022] Example 1: The specific process of the phenotypic drug molecule discovery and phenotypic drug generation method based on cell transcriptomics in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Obtain the transcriptome sequencing results data to be analyzed, and extract the list of differentially expressed genes from the transcriptome sequencing results data.

[0023] The occurrence and development of diseases are often accompanied by abnormal gene expression, such as oncogene activation and tumor suppressor gene silencing. The system receives transcriptome sequencing results for diseases requiring drug analysis. The types of transcriptome sequencing results can include conventional transcriptome sequencing (RNA-seq), single-cell transcriptome sequencing (scRNA-seq), and high-throughput sequencing data related to the LINCSL1000 project. This embodiment does not limit the specific type of transcriptome sequencing.

[0024] Extracting a list of differentially expressed genes from transcriptome sequencing data reveals that these genes directly reflect disease states at the molecular level. For example, the unlimited proliferation of tumors may be associated with the upregulation of cell cycle-related genes, while immunodeficiency may be associated with the downregulation of immune-related genes. By extracting these genes, abstract disease phenotypes can be transformed into quantifiable and analyzable molecular features, unbiasedly covering expression changes across the entire genome and avoiding the limitations of pre-defined targets. This approach is particularly suitable for polygenic diseases with complex etiologies. Furthermore, differential gene extraction allows for focusing on core disease phenotypic features, screening key genes that truly reflect disease states from high-dimensional gene expression profiles, avoiding interference from irrelevant genes, and providing precise phenotypic targets for subsequent molecular design.

[0025] Raw transcriptome sequencing results are susceptible to interference from multiple non-biological factors: On the one hand, low-quality reads (short sequence fragments) and background signals below the gene expression threshold can introduce data noise during sequencing. If used directly for comparison, this noise may be mistakenly identified as gene expression differences. On the other hand, in actual experiments, disease samples and control samples are often tested in multiple batches (e.g., sequencing is completed in 2-3 batches due to large sample sizes). Differences in instrument status, reagent purity, and operating conditions between different batches can lead to batch effects. That is, the same gene may also show differences in expression levels in different batches of normal samples. Such non-disease-related differences may interfere with the true comparison results between disease samples and control samples. To eliminate interference from non-biological factors and ensure that subsequent comparative analyses focus on gene expression changes caused by the disease itself, before extracting the list of differentially expressed genes from the transcriptome sequencing data, data preprocessing and raw signal correction can be performed on the disease transcriptome sequencing results (such as removing low-quality sequencing reads, filtering genes with extremely low expression levels, and correcting signal bias caused by base mismatches). This yields standardized transcriptome sequencing results with noise eliminated from the basic data. Subsequently, batch correction algorithms (such as ComBat, SVA, etc.) are used to eliminate non-biological differences caused by different sample testing batches (such as sequencing at different times, different testing instruments, and different reagent batches), ultimately obtaining transcriptome data that truly reflects the biological differences in gene expression. Of course, other data optimization processes or methods can be used, and this embodiment does not limit this approach.

[0026] Step 102: Obtain the text description of the differentially expressed gene list corresponding to the differentially expressed gene list.

[0027] The list of differentially expressed genes obtained through bioinformatics analysis is transformed into a text format that conforms to natural language logic and includes gene expression change characteristics and potential biological associations. For example, by combining information such as gene expression fold change and functional annotation, text descriptions can be generated in target disease samples, such as TP53 gene expression, which is related to apoptosis regulation, being upregulated by 2.5 times, and MYC gene expression, which is related to cell proliferation, being upregulated by 3.1 times; while PTEN gene expression, which is related to tumor suppression, is downregulated by 2.8 times, and BRCA1 gene expression, which is related to DNA repair, is downregulated by 2.3 times. This transforms the original list containing only gene names and the direction of change into an information carrier rich in biological semantics that AI models can understand.

[0028] By converting the list of differentially expressed genes into a textual description of the differentially expressed gene list, the technical bottleneck of traditional numerical / structured gene data being difficult for language-based AI models to process directly is overcome. Through textual conversion, not only can the data format be adapted to the model's input requirements, but the association between gene expression changes and biological functions (such as apoptosis and DNA repair) can also be presented. This allows the model to accurately identify the disease pathology mechanisms behind differentially expressed genes (such as the association between TP53 upregulation and PTEN downregulation and malignant cell proliferation), providing phenotypic guidance with clear biological significance for subsequent molecular generation.

[0029] Step 103: Input the text description of the differential gene list into the trained molecular generation model to obtain at least one candidate molecular structure representation; the molecular generation model is used to convert the text description of differential genes as gene expression information into molecular inference results that lead to the change.

[0030] The differential gene text description molecule generation submodule, obtained after processing by the gene and cell type description generation module, further adds an in vitro functional verification step after the candidate molecules are output. This step further verifies the function of the candidate analysis sequences to determine their actual effect. A specific process and logic are as follows: The output candidate molecules are subjected to in vitro functional verification processing to obtain changes in differentially expressed genes in cells before and after processing; the inverse correlation coefficient between the differentially expressed genes induced by the candidate molecules and the differentially expressed genes of the target disease is calculated based on these changes; if the inverse correlation coefficient reaches a threshold, the candidate molecules are determined to reverse the target disease phenotype; if the inverse correlation coefficient does not reach the threshold, the candidate molecules and their corresponding inverse correlation coefficients are fed back as incremental training data to the scientific language model.

[0031] First, in vitro functional validation is performed on the selected high-matching candidate molecules. For example, the candidate molecules are added to the cell model corresponding to the target disease, and the transcriptome changes of the cells before and after treatment are detected to obtain data on the differential changes in gene expression in the cells after the candidate molecules are applied. Second, the inverse correlation coefficient between the differentially expressed genes induced by the candidate molecules and the differentially expressed genes of the target disease is calculated. That is, if the genes upregulated in the disease are downregulated after molecular treatment, and the genes downregulated in the disease are upregulated after treatment, the higher the inverse correlation coefficient, the stronger the effect of the molecule on reversing the abnormal genes of the disease, thus quantitatively evaluating the actual ability of the candidate molecules to correct the disease phenotype. Finally, the results of the inverse correlation coefficient are handled in two ways: if the coefficient reaches a preset threshold, it means that the candidate molecule can effectively reverse the phenotype of the target disease and can be identified as a potential drug candidate; if the coefficient does not reach the threshold, the sequence information of the candidate molecule and its corresponding inverse correlation coefficient are used as incremental training data and fed back into the scientific big language model to optimize the model's subsequent molecular generation ability. By employing in vitro functional validation and quantitative evaluation of inverse correlation coefficients, it is ensured that the ultimately identified candidate molecules not only theoretically match the disease but also possess the ability to reverse the disease phenotype at the actual cellular level. This significantly reduces the risk of failure in subsequent in vivo experiments and enhances the practical effectiveness of the candidate molecules. Of course, the above in vitro validation step can be omitted and can be included according to actual usage needs; this embodiment does not impose any limitations on this.

[0032] Based on the above introduction, the phenotypic drug molecule discovery method based on cell transcriptomics provided in this embodiment inputs the text description of the differentially expressed gene list into the molecular generation model. This module aims to reverse the pathological phenotype corresponding to the differentially expressed gene spectrum of the target disease. Based on the mapping relationship between the differentially expressed gene text description and molecular structure learned during the pre-training process, it generates a molecular structure representation form that conforms to chemical laws (such as the SMILES format) by prediction, and finally outputs multiple sets of candidate molecular sequence representations with potential activity.

[0033] Traditional transcriptomics-based drug discovery methods (such as Connectivity Maps) can only search for matching molecules in a known compound library, failing to expand the chemical space. This step, however, uses a molecular generation model to predict molecular sequences de novo, generating entirely new molecules not found in the training data. This overcomes the limitations of a fixed chemical library and significantly enhances the innovation of drug discovery. Furthermore, the generation process is based on differentially expressed gene profiles. The pre-trained model understands the molecular structures required to regulate the disordered expression of specific differentially expressed gene combinations. The generated candidate molecules can directly target the pathological essence of the disease, avoiding the aforementioned limitations. Based on this, the phenotypic drug molecule discovery method based on cell transcriptomics provided in this embodiment obtains the transcriptomics sequencing results data to be analyzed and extracts a list of differentially expressed genes. It directly focuses on the core differences in gene expression between diseased and healthy states, abandoning the limitations of traditional target-guided drug discovery that presupposes a single target. It can cover the complex pathological features of diseases involving multiple genes and multiple pathway dysregulation, providing a basis for subsequent molecular design that reflects the essence of the disease phenotype and avoiding the limitation of compound efficacy caused by target selection bias. After converting the list of differentially expressed genes into a text description of the differentially expressed gene list, it is input into a trained molecular generation model to achieve a precise mapping between phenotypic information and molecular structure. The text conversion solves the pain point that traditional numerical gene expression matrices are difficult to be directly processed by language-based AI models, and can fully activate the semantic understanding ability of molecular generation models of gene function and pathway association. This method breaks through the traditional technical framework of retrieving known molecules based on a fixed compound library. It can generate novel molecular structures that do not appear in the training data, significantly expanding the space of medicinal chemistry and enhancing the innovation of drug discovery. At the same time, the process from extracting differentially expressed genes from transcriptome data to generating candidate molecules is coherent, without the need for repeated connection of multiple models, reducing the accumulation of errors in intermediate steps. Furthermore, the text input is adapted to the rapid processing characteristics of AI models, which can shorten the cycle from disease phenotype analysis to molecular design, reduce the time and data cost of drug development, and lay an efficient foundation for subsequent candidate molecule screening and experimental validation.

[0034] This embodiment proposes a molecular generation model training process, which specifically includes the following steps: Step 104: Obtain the molecular generation training set.

[0035] First, a molecular generation training set is constructed, comprising multiple sets of initial training data. Each set includes a textual description of a differentially expressed gene list and a candidate molecular structure representation, with corresponding relationships. The differentially expressed gene list is a textual description of the differentially expressed genes between the transcriptome data of the disease cell model and the transcriptome data of the healthy cell model. The candidate molecular structure representation is the structural representation of the molecule that causes the change. It should be understood that this design of the differentially expressed gene textual description-molecular structure representation correspondence in the training set allows the model to directly learn the regulatory rules of molecules on transcriptome phenotypes, ensuring that the generated candidate molecules can target and reverse abnormal expression of disease-related genes, thus addressing the weakness of traditional gene generation models in terms of weak correlation between molecules and disease phenotypes.

[0036] Step 105: Input the molecular generation training set into the scientific foundation big model to perform secondary training on the scientific foundation big model for biological matching tasks, and obtain the trained molecular generation model.

[0037] The molecular generation training set is input into a large-scale scientific model, and the model is then retrained through a biological matching task (i.e., learning the mapping rules between differential gene text descriptions and molecular structure representations). The result is a trained model that can be directly used for subsequent drug molecule generation. This retraining based on the large-scale scientific model inherits the text understanding and sequence generation capabilities of the basic model, while the biological matching task enhances its professional understanding of gene function and molecular mechanisms of action. Compared to models trained from scratch, this approach more efficiently grasps the phenotype-molecule mapping logic, shortening the model development cycle. Furthermore, the training process focuses on the core association of molecules causing differential changes, ensuring that the candidate molecules generated by the model are not only structurally sound but also possess the potential to reverse disease phenotypes. This provides higher-quality candidates for subsequent experimental validation, thereby improving the accuracy and efficiency of drug discovery.

[0038] In the secondary training of the scientific foundation model for the biological matching task, in order to further address the problem that the supervision signal may be insufficient or deviate from the real biological mechanism if the training set is simply generated by molecules, known compound-transcriptome perturbation pairs can be used as supervision signals to perform secondary training of the scientific foundation model for the biological matching task.

[0039] Specifically, when inputting the molecular generation training set into the large-scale scientific model for secondary training, it does not rely solely on the correspondence between the textual descriptions of differentially expressed genes and the representations of candidate molecular structures in the training set. Instead, it introduces known compound-transcriptome perturbation pairs as supervisory signals for training. That is, it uses real data showing that specific perturbations occur in the cell transcriptome after treating cells with known compounds (such as a compound downregulating the expression of disease-related upregulated genes and upregulating the expression of downregulated genes) to guide the model in learning the precise mapping relationship between the textual descriptions of differentially expressed genes (reflecting transcriptome changes) and the molecular structure representations (corresponding compounds). This ensures that the correlation between the molecular structures generated by the model and the transcriptome phenotypic regulatory effects conforms to real biological laws.

[0040] The known supervisory role of compound-transcriptome perturbation pairs can effectively constrain the training direction of the model, preventing the model from generating invalid molecules with reasonable structures but no actual transcriptome regulatory activity. This ensures that the generated candidate molecules not only conform to chemical laws but also have the potential to reverse disease-related transcriptome disorders, significantly reducing the cost of subsequent experimental validation. In addition, this supervisory signal can enhance the model's learning of the correlation between molecular structural features and transcriptome perturbation effects (such as the ease with which certain molecular structures regulate the expression of genes related to a certain pathway), helping the model to form a more accurate structure-function understanding. When faced with new differential gene text descriptions, it can generate molecular structures that meet regulatory requirements more quickly and accurately, further improving the efficiency and success rate of drug discovery.

[0041] It should be noted that the scientific foundation model refers to a basic model that uses a general-purpose large language model (such as the Transformer architecture model) as its underlying framework, and is pre-trained with massive amounts of scientific data (covering literature, database information, experimental data, etc. in disciplines such as life sciences, chemistry, and medicine). It possesses the potential for scientific semantic understanding, professional knowledge integration, and specific scientific task processing. It is not built from scratch for a single drug discovery scenario, but rather, based on general language understanding capabilities, it incorporates professional scientific knowledge and can be adapted to the specific needs of phenotypic drug molecule discovery through subsequent targeted secondary training (such as biological description generation, molecular structure prediction, etc.). This embodiment does not limit the specific model type of the scientific foundation model; for example, the NatureLM model, a Transformer-based autoregressive generative model, or a cross-modal large language model incorporating graph neural networks (GNNs) can be used.

[0042] Example 3: In the above embodiments, there are no limitations on the method of generating the differential gene list text description. In order to ensure the accuracy, standardization and biological depth of the differential gene text description, step 102 obtains the differential gene list text description corresponding to the differentially expressed genes. Specifically, the differentially expressed gene list can be input into a trained gene and cell type description generation model to obtain the differential gene text description.

[0043] Specifically, once a list of differentially expressed genes is obtained (such as structured data containing upregulated genes, downregulated genes, and their fold changes in expression), it is input into a specially trained gene and cell type description generation model. Based on learned gene function knowledge, pathway association rules, and text generation capabilities, the model automatically outputs differential gene text descriptions that conform to natural language logic and are rich in biological semantics (e.g., integrating gene function annotations to generate descriptions such as gene A, which is involved in cell cycle regulation, being upregulated by 2.3 times, while gene B, which inhibits cell proliferation, is downregulated by 1.8 times in the target disease).

[0044] Model-based transformation ensures the consistency and reproducibility of text descriptions, avoiding subsequent model input biases caused by differences in transformation styles among different operators, and improving the stability of the entire drug discovery process. In addition, specially trained models can deeply integrate professional knowledge such as gene function and cell type characteristics. The generated text not only contains information on gene expression changes, but also reveals the underlying biological significance (such as the possibility that gene E downregulation may lead to enhanced immune escape). This enables subsequent molecular generation models to accurately capture the core mechanisms of disease phenotypes and improve the targeting of candidate molecules. Moreover, automated transformation significantly improves processing efficiency, especially when dealing with massive differential gene data generated by single-cell transcriptomes. The model can quickly complete the transformation, avoiding the time-consuming and error-prone manual processing, providing efficient support for high-throughput drug screening. It also lays a unified text input foundation for cross-experimental and cross-platform transcriptome data integration and analysis, further enhancing the generalization ability of the method.

[0045] In this embodiment, the training process for the gene and cell type description generation model is not limited. To ensure that the gene and cell type description generation model has professional and accurate biological semantic conversion capabilities, this embodiment further proposes a training method, which includes: obtaining a text conversion training set, which includes multiple sets of second training data, each set of second training data including biological names and corresponding text descriptions; inputting the text conversion training set into the scientific foundation big model to perform secondary training on the scientific foundation big model for biological description, thereby obtaining the trained gene and cell type description generation model.

[0046] First, a text conversion training set is constructed, which includes multiple sets of secondary training data. Each set of secondary training data includes a biological name (e.g., gene name TP53, cell type name hepatocellular carcinoma) and its corresponding text description (e.g., TP53: a tumor suppressor gene that regulates apoptosis and genome stability; mutations in this gene easily lead to malignant cell proliferation; hepatocellular carcinoma: a malignant tumor cell with high metabolic activity and epithelial-mesenchymal transition characteristics). This training set is then input into a large-scale scientific model. Through secondary training targeting the biological descriptions (i.e., learning the mapping rules between biological names and text descriptions), the model masters the expression logic of professional knowledge such as gene function and cell characteristics, ultimately resulting in a gene and cell type description generation model capable of accurately generating differential gene text descriptions.

[0047] While large-scale scientific models possess general text generation capabilities, they lack specialized knowledge in the biological field. Directly using them for translation may lead to errors in gene function descriptions and inaccuracies in cellular feature representations. Specialized training on text translation training sets allows the model to systematically learn biological nomenclature conventions, functional annotation logic, and professional terminology, avoiding semantic distortion caused by non-specialized translations. Furthermore, limiting the training data structure to biological names and text descriptions ensures the model focuses on the core task of transforming abstract names into concrete descriptions, rather than generalized text generation. This improves the specificity and accuracy of the translation, providing high-quality input for subsequent molecular generative models.

[0048] This embodiment only uses the above training process as an example for introduction. The training methods of other models can be referred to the introduction of this embodiment, and will not be repeated here.

[0049] Example 4: While the trained scientific language model can generate logically consistent molecules based on patterns, some candidate molecules, despite having compliant structures, may exhibit low correlation with core disease genes. To address this deficiency, this embodiment proposes further evaluation and screening of candidate molecules. Specifically, after generating candidate molecule sequence representations, the following steps can be performed: Step 106: Input the text description of the differentially expressed gene list and its corresponding at least one candidate molecular structure representation into the trained molecular-phenotype matching model to obtain the matching score between the differentially expressed gene text description and each candidate molecular structure representation.

[0050] After generating candidate molecules, two key pieces of information are input into the trained molecular-phenotype matching model: one is the differential gene description corresponding to the target disease, that is, the abnormal gene expression pattern unique to the disease, such as natural language text of gene A being upregulated by 3 times and gene B being downregulated by 2 times; the other is the sequence representation of the candidate molecules generated by the previous scientific big language model for the disease, such as the string SMILES.

[0051] The molecular-phenotype matching model evaluates each candidate molecule based on differential gene descriptions and molecular fitness judgment logic learned during training, and outputs a quantitative matching score. The higher the score, the better the candidate molecule matches the need to correct the gene abnormalities of the target disease, and theoretically, it has a better potential therapeutic effect.

[0052] Step 107: Sort at least one candidate molecular structure representation in descending order based on matching score to obtain candidate molecular sequence representation.

[0053] All candidate molecules are sorted in descending order of matching score, resulting in an ordered sequence of candidate molecules. This sorting result (represented by the candidate molecule sequence) visually presents the priority of each candidate molecule's matching degree with the disease phenotype, facilitating the selection of molecules with high matching scores for validation in subsequent experimental screening, reducing the cost of ineffective experiments, and improving drug discovery efficiency.

[0054] In this embodiment, the training method for the molecular-phenotype matching model is not limited. Optionally, a training method for the molecular-phenotype matching model may include the following steps: Step 108: Obtain the molecular-phenotype matching training set. The molecular-phenotype matching training set includes multiple sets of third training data. Each set of third training data includes a text description of the differentially expressed gene list and its corresponding multiple molecular structure representations, as well as a matching score between the text description of the differentially expressed gene list and each corresponding molecular representation. The matching score is between 0 and 1.

[0055] The training set consists of multiple sets of structured data, each containing three core elements: first, a textual description of a list of differentially expressed genes for a specific disease (e.g., gene expression characteristics of EGFR upregulation by 3-fold and TP53 downregulation by 2-fold in lung cancer cells); second, multiple molecular structural representations related to the textual description (e.g., SMILES sequences of different compounds that can affect the expression of the aforementioned genes); and third, a matching score between each molecular structural representation and the textual description of the differentially expressed genes—this score ranges from 0 to 1, where 1 indicates the highest degree of matching between the molecule and the phenotype (e.g., the molecule can perfectly reverse the abnormal expression of the differentially expressed genes), and 0 indicates the lowest degree of matching. The score directly reflects the potential ability of the molecule to regulate the corresponding disease phenotype.

[0056] Step 109: Input the molecular-phenotype matching training set into the preset molecular-phenotype matching model for scoring ability training to obtain the molecular-phenotype matching model.

[0057] First, a pre-defined basic model architecture is constructed (usually a deep neural network or a cross-modal model that integrates text and chemical features). Then, a molecular-phenotype matching training set containing text descriptions of differentially expressed genes, corresponding to multiple molecular structure representations, and matching scores between the two in the 0-1 interval is input into the model. Through iterative learning, the model learns the quantitative rules of the correlation strength between the text descriptions of differentially expressed genes and the molecular structure representations.

[0058] During training, the model continuously adjusts its parameters to minimize the difference between the predicted matching score and the actual matching score in the training set, ultimately forming a molecular-phenotypic matching model that can stably output reasonable matching scores. The core capability of this model is that it can autonomously assess the degree of matching between newly input differentially expressed gene text descriptions and candidate molecule structural representations, and output a score between 0 and 1. This provides a quantitative basis for prioritizing candidate molecules, thereby supporting a subsequent efficient drug screening process.

[0059] Example 5: This embodiment relates to a phenotypic drug molecule discovery device based on cell transcriptomics. A schematic diagram of this phenotypic drug molecule discovery device based on cell transcriptomics can be seen as follows: Figure 2 As shown, it includes a data acquisition module 201, a gene description acquisition module 202, and a molecular structure generation module 203.

[0060] Among them, the data acquisition module 201 is used to acquire the transcriptome sequencing results data to be analyzed and extract the list of differentially expressed genes from the transcriptome sequencing results data; Gene description acquisition module 202 is used to acquire the text description of the differentially expressed gene list corresponding to the differentially expressed gene list; The molecular structure generation module 203 is used to input the text description of the differential gene list into the trained molecular generation model to obtain at least one candidate molecular structure representation; the molecular generation model is used to convert the text description of the differential genes as gene expression information into molecular inference results that lead to the change.

[0061] In the phenotypic drug discovery device based on cell transcriptomics provided in this embodiment, the data acquisition module extracts a list of differentially expressed genes to accurately capture the core gene expression differences between disease and healthy states, providing a phenotypic basis reflecting the essence of the disease for subsequent analysis and avoiding the limitations of traditional target pre-setting. The gene description acquisition module converts the gene list into text descriptions, solving the problem of adapting numerical gene data to the model, while incorporating biological semantics, allowing the model to understand gene functional associations and pathological mechanisms. The molecular structure generation module, using a trained molecular generation model, directly generates candidate molecules from the text descriptions of differentially expressed genes, breaking through the limitations of fixed compound libraries to expand the chemical space, and ensuring the targeted association between the generated molecules and the disease phenotype. The three modules work together to achieve an efficient closed loop of phenotypic analysis, semantic conversion, and molecular generation, improving the accuracy, innovation, and efficiency of drug discovery, and adapting to the research needs of complex diseases with multi-gene regulatory characteristics.

[0062] It should be noted that the contents of the phenotypic drug molecule discovery device based on cell transcriptomics provided in this embodiment can be referred to in conjunction with the phenotypic drug molecule discovery method based on cell transcriptomics provided in the above embodiments, and the repeated parts will not be described again in this embodiment.

[0063] Furthermore, it is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.

[0064] Example 6: Another embodiment of this application relates to an electronic device, such as... Figure 3 As shown, it includes: at least one processor 301; and a memory 302 communicatively connected to at least one processor 301; wherein the memory 302 stores instructions executable by at least one processor 301, the instructions being executed by at least one processor 301 to enable at least one processor 301 to perform the steps of the phenotypic drug molecule discovery method based on cell transcriptomics in the above embodiments.

[0065] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0066] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0067] Example 7: Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0068] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0069] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A phenotypic drug molecule discovery method based on cell transcriptomics, characterized in that, include: Obtain the transcriptome sequencing results data to be analyzed, and extract a list of differentially expressed genes from the transcriptome sequencing results data; Obtain the text description of the differentially expressed gene list corresponding to the differentially expressed gene list; The textual description of the differentially expressed gene list is input into a trained molecular generation model to obtain at least one candidate molecular structure representation; the molecular generation model is used to convert the textual description of the differentially expressed genes, which serves as gene expression information, into molecular inference results that lead to the change.

2. The phenotypic drug molecule discovery method based on cell transcriptomics according to claim 1, characterized in that, Also includes: Obtain the molecular generation training set; The molecular generation training set includes multiple sets of first training data. Each set of first training data includes a text description of a differentially expressed gene list and a candidate molecular structure representation with corresponding relationships. The text description of the differentially expressed gene list is a text description of a differentially expressed gene list between disease cell model transcriptome data and healthy cell model transcriptome data. The candidate molecular structure representation is a structural representation of the molecule that causes the change. The molecular generation training set is input into the scientific foundation big model to perform secondary training on the scientific foundation big model for a biological matching task, thereby obtaining the trained molecular generation model.

3. The phenotypic drug molecule discovery method based on cell transcriptomics according to claim 2, characterized in that, The step of inputting the molecular generation training set into the scientific foundation model for secondary training of the scientific foundation model on a biological matching task further includes: The scientific foundation model was retrained for a biological matching task using known compound-transcriptome perturbation pairs as supervisory signals.

4. The phenotypic drug molecule discovery method based on cell transcriptomics according to claim 1, characterized in that, The text description of obtaining the differentially expressed genes corresponding to the differentially expressed genes includes: The differentially expressed gene list is input into a trained gene and cell type description generation model to obtain the differentially expressed gene text description.

5. The phenotypic drug molecule discovery method based on cell transcriptomics according to claim 4, characterized in that, Also includes: Obtain a text conversion training set, which includes multiple sets of second training data, each set of second training data including a biological name and a text description corresponding to the biological name; The text conversion training set is input into the scientific foundation big model to perform secondary training on the scientific foundation big model for biological description, resulting in the trained gene and cell type description generation model.

6. The phenotypic drug molecule discovery method based on cell transcriptomics according to claim 1, characterized in that, After obtaining the candidate molecule sequence representation, the process further includes: The text description of the differentially expressed gene list and its corresponding at least one candidate molecular structure representation are input into a trained molecular-phenotype matching model to obtain a matching score between the differentially expressed gene text description and each candidate molecular structure representation. At least one of the candidate molecular structure representations is sorted in descending order based on the matching score to obtain candidate molecular sequence representations.

7. The phenotypic drug molecule discovery method based on cell transcriptomics according to claim 6, characterized in that, Also includes: Obtain a molecular-phenotype matching training set, which includes multiple sets of third training data. Each set of third training data includes a text description of a differential gene list and its corresponding multiple molecular structure representations, as well as a matching score between the text description of the differential gene list and each corresponding molecular representation, wherein the matching score is between 0 and 1. The molecular-phenotype matching training set is input into a preset molecular-phenotype matching model for scoring ability training to obtain the molecular-phenotype matching model.

8. A phenotypic drug molecule discovery device based on cell transcriptomics, characterized in that, include: The data acquisition module is used to acquire the transcriptome sequencing results data to be analyzed and extract a list of differentially expressed genes from the transcriptome sequencing results data; The gene description acquisition module is used to acquire the text description of the differentially expressed gene list corresponding to the differentially expressed gene list. A molecular structure generation module is used to input the textual description of the differentially expressed gene list into a trained molecular generation model to obtain at least one candidate molecular structure representation; the molecular generation model is used to convert the textual description of the differentially expressed genes, which serves as gene expression information, into molecular inference results that lead to the change.

9. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the phenotypic drug molecule discovery method based on cell transcriptomics as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the phenotypic drug molecule discovery method based on cell transcriptomics as described in any one of claims 1 to 7.

Images