Cell-targeted drug screening model training method, drug screening method and device

Abstract

The application provides a cell-targeting drug primary screening model training method, a drug primary screening method and equipment. The method comprises the following steps: obtaining molecular fingerprint feature data corresponding to each compound sample with label information to form a corresponding original data set, performing data screening on the original data set by using a distance measurement method, and training a regression model based on a graph neural network to learn the structure topology of each compound sample. The model is trained as a cell-targeting drug primary screening model for predicting whether a compound has potential cell-targeting drug activity. The application can effectively improve the comprehensiveness of molecular representation during the training process of the cell-targeting drug primary screening model, effectively improve the generalization ability of the cell-targeting drug primary screening model obtained by training, effectively improve the applicability of the cell-targeting drug primary screening model obtained by training, and improve the accuracy and reliability of the prediction result of the potential cell-targeting drug activity predicted by the model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for training cell-targeted drug screening models, a drug screening method, and equipment. Background Technology

[0002] Compound screening is a crucial step in drug discovery and development, aiming to identify compounds with therapeutic potential from thousands of potential candidates. To improve the efficiency of initial screening for cell-targeted drugs, molecular machine learning can be employed. Looking back at the development of feature engineering for molecular machine learning, early research relied primarily on traditional hand-designed features. However, due to the limitations of traditional methods in handling the highly diverse structures of molecules, the need for more flexible, data-driven approaches has grown. In recent years, with the rise of big data and large models, the field has evolved from initially relying on hand-extracted and heuristically designed features to using more flexible and expressive data-driven features. In recent years, molecular machine learning methods, employing deep learning and self-supervised learning, have learned data-driven molecular representations from massive amounts of unlabeled molecular data. This approach is currently the mainstream method for initial screening of cell-targeted drugs.

[0003] However, traditional cell-targeted drug screening methods mainly rely on fixed molecular fingerprints, which can lead to the loss of important information related to drug physiological metabolism, especially when dealing with molecules with highly diverse structures. While data-driven molecular characterization has addressed the shortcomings of molecular fingerprints to some extent, these methods typically focus on modeling molecules using a single specific modality, resulting in varying degrees of information loss. Furthermore, traditional cell-targeted drug screening methods, such as self-supervised learning-based molecular characterization methods, lack generalization ability.

[0004] Therefore, there is an urgent need to design a training method for cell-targeted drug screening models that can improve the accuracy of cell-targeted drug screening prediction and the model's generalization ability. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method for training a cell-targeted drug screening model, a drug screening method, and an apparatus to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of this application provides a method for training a cell-targeted drug screening model, comprising:

[0007] The molecular fingerprint feature data corresponding to each compound sample with label information is obtained to form the corresponding original dataset, wherein the label information is used to indicate whether the corresponding compound sample has potential cell-targeting drug activity;

[0008] The distance metric is used to filter the compound samples in the original dataset to obtain the corresponding target dataset, and a training set is divided from the target dataset.

[0009] A graph neural network-based regression model is trained based on the training set to learn the structural topological information of each compound sample. This graph neural network-based regression model is then trained to become a cell-targeted drug screening model for predicting whether a compound has potential cell-targeted drug activity based on its molecular fingerprint feature data.

[0010] In some embodiments of this application, before obtaining the molecular fingerprint feature data corresponding to each compound sample with tagged information to form the corresponding original dataset, the method further includes:

[0011] Obtain raw information data of each candidate compound used for target cell activation or inhibition, along with their respective tag information;

[0012] The original information data of each candidate compound is converted into SMILES strings;

[0013] For candidate compounds whose SMILES strings are successfully converted and whose generated SMILES strings are unique, the SMILES strings of the candidate compounds are used as compound samples.

[0014] Candidate compounds for which SMILES string conversion failed were deleted;

[0015] For candidate compounds whose SMILES strings are successfully converted and whose generated SMILES strings are not unique, one of the various SMILES strings of the candidate compound is selected as the compound sample.

[0016] In some embodiments of this application, the target cells include: tumor cells, immune cells, cardiomyocytes, neuronal cells, endocrine cells, hepatocytes, stem cells, or respiratory epithelial cells;

[0017] The immune cells include: T cells, B cells, NK cells, myeloid-derived cells, or antigen-presenting cells.

[0018] In some embodiments of this application, obtaining the molecular fingerprint feature data corresponding to each compound sample with tagged information to form the corresponding original dataset includes:

[0019] Based on a pre-defined latent feature space representation method, molecular fingerprint feature data corresponding to each compound sample with label information is obtained.

[0020] Generate an original dataset containing the molecular fingerprint feature data and the tag information corresponding to each of the compound samples;

[0021] The latent feature space representation method includes at least one of the following: ECFP2 molecular fingerprint representation method, ECFP4 molecular fingerprint representation method, ECFP6 molecular fingerprint representation method, and MACCS Key molecular fingerprint representation method.

[0022] In some embodiments of this application, the step of using a distance metric to filter the compound samples in the original dataset to obtain a corresponding target dataset, and then dividing the training set from the target dataset, includes:

[0023] Molecular fingerprint feature data of multiple compound samples are randomly selected from the original dataset to serve as reference anchors respectively;

[0024] Based on the number of reference anchor points, a first hyperparameter for representing the radius threshold of the feature space hypersphere, and a second hyperparameter for representing the label difference threshold, a distance metric is used to cluster each of the compound samples in the original dataset to filter out heterogeneous samples in the original dataset, thereby obtaining the corresponding target dataset.

[0025] The target dataset is divided into a training set and a test set;

[0026] The distance measurement methods include: Hamming distance measurement method or Euclidean distance measurement method.

[0027] In some embodiments of this application, training a graph neural network-based regression model based on the training set, so that the graph neural network-based regression model learns the structural topological information of each compound sample, and then training the graph neural network-based regression model into a cell-targeted drug screening model for predicting whether a compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound, includes:

[0028] The graph neural network-based regression model is iteratively trained based on the training set and preset training hyperparameters, so that the graph neural network-based regression model can represent nodes for atoms, edges for chemical bonds, and global information containing the structural topology information of the compound sample as learnable vectors. During the iterative training process, the potential cell-targeted drug activity prediction results obtained by the graph neural network-based regression model each time are filtered according to the preset prediction threshold range.

[0029] The trained graph neural network-based regression model is tested based on the test set. If the graph neural network-based regression model passes the model test, it is used as a preliminary screening model for cell-targeted drugs to predict whether a compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound.

[0030] In some embodiments of this application, the graph neural network-based regression model includes: a GCN model and / or an MPNN model.

[0031] Another aspect of this application provides a method for primary screening of cell-targeted drugs, comprising:

[0032] Obtain molecular fingerprint feature data of the target compound;

[0033] The molecular fingerprint feature data of the target compound is input into a preset cell-targeted drug screening model, so that the cell-targeted drug screening model outputs a potential cell-targeted drug activity prediction result indicating whether the target compound has potential cell-targeted drug activity. The cell-targeted drug screening model is pre-trained based on the cell-targeted drug screening model training method.

[0034] In some embodiments of this application, the cell-targeted drug screening method further includes:

[0035] Output the potential cell-targeting drug activity prediction result corresponding to the target compound, so that when or after determining that the target compound has potential cell-targeting drug activity based on the potential cell-targeting drug activity prediction result, in vitro cell experiments can be performed on the target compound.

[0036] A third aspect of this application provides a training device for a cell-targeted drug screening model, comprising:

[0037] The feature acquisition module is used to acquire the molecular fingerprint feature data corresponding to each compound sample with label information to form the corresponding original dataset, wherein the label information is used to indicate whether the corresponding compound sample has potential cell-targeting drug activity;

[0038] The data filtering module is used to filter the compound samples in the original dataset using a distance metric to obtain the corresponding target dataset and to divide the training set from the target dataset.

[0039] The model training module is used to train a graph neural network-based regression model based on the training set, so that the graph neural network-based regression model learns the structural topological information of each compound sample, and then trains the graph neural network-based regression model into a cell-targeted drug screening model for predicting whether a compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound.

[0040] A fourth aspect of this application provides a cell-targeted drug screening device, comprising:

[0041] The data acquisition module is used to acquire the molecular fingerprint feature data of the target compound;

[0042] The model prediction module is used to input the molecular fingerprint feature data of the target compound into a preset cell-targeted drug screening model, so that the cell-targeted drug screening model outputs a potential cell-targeted drug activity prediction result indicating whether the target compound has potential cell-targeted drug activity. The cell-targeted drug screening model is pre-trained based on the cell-targeted drug screening model training method.

[0043] A fifth aspect of this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the cell-targeted drug screening model training method, and / or to implement the cell-targeted drug screening method.

[0044] A sixth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the cell-targeted drug screening model training method and / or the cell-targeted drug screening method.

[0045] The seventh aspect of this application provides a computer program product comprising a computer program that, when executed by a processor, implements the cell-targeted drug screening model training method, and / or implements the cell-targeted drug screening method.

[0046] The cell-targeted drug screening model training method provided in this application obtains molecular fingerprint feature data corresponding to each compound sample with labeled information to form a corresponding original dataset. The labeled information indicates whether the corresponding compound sample has potential cell-targeted drug activity. A distance metric is used to filter the data of each compound sample in the original dataset to obtain a corresponding target dataset, and a training set is partitioned from the target dataset. A graph neural network-based regression model is trained based on the training set to learn the structural topological information of each compound sample. This graph neural network-based regression model is then trained into a cell-targeted drug screening model used to predict whether a compound has potential cell-targeted drug activity based on its molecular fingerprint feature data. This study employs a graph neural network-based regression model, utilizing both molecular fingerprints (focusing on functional group information) and topological information (focusing on compound structure). By combining information from these two molecular data structures, it can retain as many molecular features as possible that are helpful for the initial drug screening model, thus effectively improving the comprehensiveness of molecular characterization during the training process of the cell-targeted drug screening model. Furthermore, a novel and efficient data filtering strategy is introduced to remove potentially interfering compound samples from the original dataset. This overcomes the influence of samples with similar characteristics but different label values ​​on the model, effectively improving the generalization ability of the trained cell-targeted drug screening model. Consequently, it enhances the applicability of the trained cell-targeted drug screening model and improves the accuracy and reliability of the model's predictions of potential cell-targeted drug activity.

[0047] Additional advantages, objectives, and features of this application will be set forth in part in the description which follows, and will in part become apparent to those skilled in the art upon review of the following description, or may be learned by practice of the application. The objectives and other advantages of this application can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0048] Those skilled in the art will understand that the purposes and advantages that can be achieved with this application are not limited to those specifically described above, and that the above and other purposes that this application can achieve will be more clearly understood from the following detailed description. Attached Figure Description

[0049] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, do not constitute a limitation thereof. The components in the drawings are not drawn to scale but are merely for illustrating the principles of this application. For ease of illustration and description of certain parts of this application, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to this application. In the drawings:

[0050] Figures 1(a) to 1(d) The diagrams, in order, illustrate how the latent feature spaces extracted by ECFP2, MACCS Key, MolFormer, and ImageMol are mapped to a two-dimensional space using t-SNE technology.

[0051] Figure 2 This is a schematic diagram of the first process of training a cell-targeted drug screening model in one embodiment of this application.

[0052] Figure 3 This is a schematic diagram of the second process of the cell-targeted drug screening model training method in one embodiment of this application.

[0053] Figure 4(a) is a schematic diagram of the effect of the number of reference anchor points K=5, the hyperparameters α1=25 and α2=15 on the screening model in an example of this application.

[0054] Figure 4(b) is a schematic diagram of the effect of the number of reference anchor points K=5, the hyperparameters α1=25 and α2=40 on the screening model in an example of this application.

[0055] Figure 4(c) is a schematic diagram of the effect of the number of reference anchor points K=5, the hyperparameters α1=70 and α2=40 on the screening model in an example of this application.

[0056] Figure 5 This is a flowchart illustrating a cell-targeted drug screening method in one embodiment of this application.

[0057] Figure 6 This is a schematic diagram illustrating the specific algorithm flow of the data filtering strategy in the application example of this application.

[0058] Figure 7 This is a schematic diagram illustrating the execution flow of the cell-targeted drug screening model training and validation method in the application example of this application.

[0059] Figure 8 This is a schematic diagram of the comparison table between model verification and performance evaluation in the application examples of this application. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and their descriptions are used to explain this application, but are not intended to limit it.

[0061] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the structures and / or processing steps closely related to the solution according to this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0062] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0063] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0064] In the following description, embodiments of the present application will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0065] Drug screening involves the efficient screening of large-scale molecular libraries to identify candidate molecules suitable for treating specific diseases or pathophysiological processes. The specific process includes assessing the physiological properties of various compounds, including efficacy, pharmacokinetics, and safety. For example, TH17 cells are an important type of T cell and a crucial component of the immune system. Drug screening can be used to discover compounds that target Tregs or TH17 cells, enabling the exploration of therapeutic drugs for various immune-related diseases.

[0066] As an important branch of molecular machine learning, current mainstream drug screening methods generally employ a feature extractor and classifier model, or a feature extractor and regressor model. Therefore, they mainly involve two steps: feature engineering and building and training the classifier. The most fundamental challenges of feature engineering are determining the appropriate data structure for molecular modeling and how to embed the molecule into a suitable latent space. In earlier molecular machine learning research, molecular representation methods were typically based on manually extracted features—a type of feature extraction method that uses fixed-length bit vectors to represent the molecular feature structure.

[0067] With the development of graph neural networks and their related models and optimization techniques, graph-based molecular representation has become one of the mainstream methods in molecular feature engineering. Graph neural networks have a natural and close relationship with molecular structure; for example, molecular structure can be represented using a topological graph data structure, where atoms are represented as nodes and chemical bonds as edges. Besides topological graphs, there are two other representative graph-based molecular representation methods: one is a geometric graph that can represent Euclidean space metrics, and the other is a knowledge graph that can mine the relationships between molecules. In addition to the above representation methods, some molecular representation learning methods study structural string representations, and some works attempt to connect the fields of computer vision and molecular representation learning tasks, re-examining molecular representation learning tasks from the perspective of natural RGB images.

[0068] Here are three of the most representative drug screening methods in recent years: MolCLR, MolFormer, and ImageMol.

[0069] Among them, the self-supervised learning-based molecular representation learning framework MolCLR models each molecular structure with a topological graph and pre-trains a graph neural network using a contrastive learning-based self-supervised learning framework to extract molecular representation features. It is entirely based on the general self-supervised learning framework SimCLR, where the encoder network f(·) uses graph neural networks (GNNs) and the projection head g(·) uses multilayer perceptrons (MLPs). Benefiting from pre-training on massive datasets of tens of millions of data points, MolCLR not only achieves state-of-the-art performance on several challenging benchmark datasets but has also been shown to embed molecules into representation features that can distinguish chemical properties. In contrast, the self-supervised learning-based molecular representation learning framework MolFormer treats molecular data as strings of SMILES (Simplified Molecular Input Line Entry System), employing the masked language model most widely used in Natural Language Understanding (NLP) to achieve self-supervised learning tasks, and also demonstrates considerable performance. The final work employs ImageMol, a self-supervised learning-based molecular representation learning framework. This work combines molecular modeling with the common RGB images of computer vision, leveraging pre-training methods in computer vision, such as image reconstruction and jigsaw puzzle tasks, for molecular representation learning. Overall, existing implementations mainly focus on molecular representation learning and molecular property prediction, improving model performance through methods such as self-supervised learning, especially when dealing with molecules with challenging structures. Among these, SMILES is a string representation for molecular structures. It is a chemical description language that uses ASCII characters to represent molecular structures, allowing computers to easily process and interpret molecular information.

[0070] Therefore, it can be seen that existing methods for training cell-targeted drug screening models mainly have the following problems:

[0071] (1) Limitations of molecular characterization techniques: Traditional molecular characterization studies mainly rely on fixed molecular fingerprints. However, this method may lose important information related to drug physiological metabolism, especially when dealing with molecules with highly diverse structures. Although data-driven molecular characterization has solved the shortcomings of molecular fingerprints to some extent, these methods usually focus on modeling molecules with a single specific modality, resulting in varying degrees of information loss.

[0072] For example, directly using topological graphs for model training only considers the topological relationships within the molecular structure, neglecting the structural and spatial information of functional groups; while SMILES includes topological and functional group structural information, it does not cover spatial information; and methods based on general RGB images may ignore prior knowledge in chemistry, such as the chirality of molecules. In studies of cell promotion or inhibition levels, the lack of molecular information can critically affect drug model screening and the discovery of new therapeutic targets. This is because a comprehensive understanding of molecular properties is essential for the precise regulation of cellular responses.

[0073] (2) Generalization issues in specific domains: The effectiveness of existing methods remains limited to public datasets and specific datasets. For example, the three self-supervised molecular representation methods mentioned above were only comprehensively evaluated on some tasks of the MoleculeNet benchmark dataset, indirectly indicating that these methods suffer from poor adaptability and out-of-domain issues. Among them, ImageMol is the method closest to real-world application value. This work not only evaluated their models on the MoleculeNet benchmark but also conducted a very detailed demonstration on molecular inhibitors against SARS coronavirus, proving that the method can explore potential 3C-like protease inhibitors for treating COVID-19. However, although existing methods such as ImageMol perform well on molecular compound datasets with the same distribution as the training set, their generalization ability is insufficient on data in unknown domains that do not conform to the independent and identically distributed assumption. Figures 1(a) to 1(d) As shown, by mapping the latent feature spaces extracted from Extended-Connectivity FingerPrints (ECFP), Molecular Access System Keys (MACCSKey), MolFormer, and ImageMol to a two-dimensional space using t-SNE technology, researchers found that none of these four methods showed significant effectiveness in distinguishing between inhibitors and promoters. Among them, MACCS Key is a 166-bit fingerprint used to describe molecular structure.

[0074] Based on this, in order to solve the above-mentioned technical problems of existing methods, the embodiments of this application provide a cell-targeted drug screening model training method, a cell-targeted drug screening model training device for executing the cell-targeted drug screening model training method, a cell-targeted drug screening method, a cell-targeted drug screening device for executing the cell-targeted drug screening method, a computer device, a computer-readable storage medium, and a computer program product, which can improve the accuracy of cell-targeted drug screening prediction and the model generalization ability.

[0075] The following examples will provide a detailed description.

[0076] Based on this, embodiments of this application provide a method for training a cell-targeted drug screening model that can be implemented by a cell-targeted drug screening model training device, see [link to relevant documentation]. Figure 2 The training method for the cell-targeted drug screening model specifically includes the following:

[0077] Step 100: Obtain the molecular fingerprint feature data corresponding to each compound sample with label information to form the corresponding original dataset, wherein the label information is used to indicate whether the corresponding compound sample has potential cell-targeting drug activity.

[0078] In one or more embodiments of this application, the tag information is used to indicate whether the corresponding compound sample has potential cell-targeting drug activity. Specifically, it can be indicated by the expression level of cytokines obtained by biological experiments of the compound corresponding to the compound sample. For example, if the expression level of cytokines of a compound is equal to or greater than a preset value, it is determined to have potential cell-targeting drug activity; otherwise, it is determined not to have potential cell-targeting drug activity.

[0079] It is understood that the compounds corresponding to the compound samples are pre-acquired compounds that are highly correlated with the activation or inhibition of the target cells.

[0080] Step 200: Use a distance metric to filter the data of each compound sample in the original dataset to obtain the corresponding target dataset, and divide the training set from the target dataset.

[0081] In step 200, the data filtering strategy using distance metrics can be used to solve the problem of difficulty in distinguishing between promoters and inhibitors in datasets such as Treg, and can keep the label differences of samples with similar characteristics within an acceptable range, so as to screen out heterogeneous samples in the original dataset.

[0082] Step 300: Train a graph neural network-based regression model based on the training set, so that the graph neural network-based regression model learns the structural topological information of each compound sample, and then trains the graph neural network-based regression model into a cell-targeted drug screening model for predicting whether a compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound.

[0083] In this application, the graph neural network-based regression model is used to model molecules. Specifically, the graph neural network uses atoms in the molecule as nodes and chemical bonds as edges to model the molecule using a graph data structure. Specifically, the graph neural network-based regression model represents nodes, edges, and global information as learnable vectors, while the structural topology information of the compound is a part of the global information.

[0084] As described above, the cell-targeted drug screening model training method provided in this application adopts a regression model based on graph neural networks. It utilizes two molecular data structures: molecular fingerprints focusing on molecular functional groups and topological information focusing on compound structures. By comprehensively using information from these two molecular data structures, it can retain as many molecular features as possible that may be helpful to the drug screening model, thereby effectively improving the comprehensiveness of molecular characterization during the training process. Furthermore, by introducing a novel and efficient data screening strategy to remove potentially interfering compound samples from the original dataset, it overcomes the influence of samples with similar characteristics but different label values ​​on the model. This effectively improves the generalization ability of the trained cell-targeted drug screening model, thereby enhancing its applicability and improving the accuracy and reliability of the model's predictions of potential cell-targeted drug activity.

[0085] To further improve the effectiveness and reliability of training cell-targeted drug screening models, a cell-targeted drug screening model training method is provided in the embodiments of this application, see [link to relevant documentation]. Figure 3 The training method for the cell-targeted drug screening model specifically includes the following steps prior to step 100:

[0086] Step 010: Obtain raw information data of each candidate compound used for target cell activation or inhibition, as well as their corresponding tag information.

[0087] Specifically, the process begins by collecting raw information data on candidate compounds highly correlated with the activation or inhibition of target cells. This raw information data may include compound names and identification numbers, which can be encoded using CAS (Chemical Abstracts Service Registry Numbers). Next, the expression levels of cytokines for each candidate compound are collected through biological experiments. Based on the cytokine expression levels of each candidate compound, a tag is generated to indicate whether the candidate compound possesses potential cell-targeting drug activity. CAS codes are standardized identification codes, a type of digital identifier used to uniquely identify chemical substances.

[0088] Step 020: Perform SMILES string conversion on the original information data of each candidate compound; for candidate compounds whose SMILES string conversion is successful and whose generated SMILES string is unique, use the SMILES string of the candidate compound as a compound sample; delete candidate compounds whose SMILES string conversion is unsuccessful; for candidate compounds whose SMILES string conversion is successful but whose generated SMILES string is not unique, select one of the various SMILES strings of the candidate compound as a compound sample.

[0089] In other words, the raw information data represented in CAS encoding is converted into the corresponding SMILES string through the API interface of the PubChem database for the chemical modules. Two special cases may occur during this conversion process:

[0090] (1) If the CAS code of a compound does not have a corresponding SMILES string, it will not be used;

[0091] (2) The CAS code of a compound corresponds to multiple SMILES strings. The first SMILES string of the PubChem database query result can be used as the corresponding conversion result by default.

[0092] The drug screening principle corresponding to the cell-targeted drug screening model training method provided in this application embodiment is based on actual screening data. Therefore, although only Treg or Th17 cell types are used as examples in the embodiment, in practice any type of cell can be used.

[0093] Therefore, to further improve the applicability and diversity of cell-targeted drug screening, in the cell-targeted drug screening model training method provided in this application embodiment, the target cells can be various types of cells. For example, the target cells can be selected from, but are not limited to, important cell types such as tumor cells, immune cells, cardiomyocytes, neurons, endocrine cells, hepatocytes, stem cells, or respiratory epithelial cells. Among them, the immune cells can be selected from, but are not limited to, T cells, B cells, NK cells, myeloid-derived cells, or antigen-presenting cells (such as macrophages or dendritic cells). These cells are all quite important cell types for drug targeting.

[0094] To further improve the comprehensiveness of molecular characterization during the training process of cell-targeted drug screening models, a cell-targeted drug screening model training method is provided in this application embodiment, see [link to relevant documentation]. Figure 3 Step 100 in the cell-targeted drug screening model training method specifically includes the following:

[0095] Step 110: Based on the preset latent feature space representation method, obtain the molecular fingerprint feature data corresponding to each compound sample with label information.

[0096] Step 120: Generate an original dataset containing the molecular fingerprint feature data and the label information corresponding to each of the compound samples; wherein the latent feature space representation method includes at least one of the following: ECFP2 molecular fingerprint representation method, ECFP4 molecular fingerprint representation method, ECFP6 molecular fingerprint representation method and MACCS Key molecular fingerprint representation method.

[0097] It is understood that the ECFP2 molecular fingerprint representation method can be preferentially selected as the latent feature space representation method in the embodiments of this application. On this basis, the radius range of ECFP can be expanded, that is, the ECFP4 or ECFP6 method can be used to extract features from the compound, or MACCS Key or other features can be used as the representation method of the compound.

[0098] To further improve the generalization ability of cell-targeted drug screening models, a cell-targeted drug screening model training method is provided in this application embodiment, see [link to relevant documentation]. Figure 3 Step 200 in the cell-targeted drug screening model training method specifically includes the following:

[0099] Step 210: Randomly select molecular fingerprint feature data of multiple compound samples from the original dataset to serve as reference anchors respectively.

[0100] Step 220: Based on the number of the reference anchor points, a first hyperparameter for representing the radius threshold of the feature space hypersphere and a second hyperparameter for representing the label difference threshold, cluster the compound samples in the original dataset using a distance metric to filter out heterogeneous samples in the original dataset, thereby obtaining the corresponding target dataset. The distance metric includes Hamming distance or Euclidean distance.

[0101] Step 230: Divide the target dataset into a training set and a test set.

[0102] The steps 210 and 220 are illustrated using the Hamming distance metric as an example:

[0103] To address the issue of inseparable promoters and inhibitors in datasets such as Treg, this application proposes a data filtering strategy based on ECFP2 features. This strategy employs an efficient approximate distance metric (Hamming distance) to keep the label differences between samples with similar features within an acceptable range, thereby filtering out heterogeneous samples in the dataset.

[0104] Where K represents the number of baseline anchor points randomly selected from the original dataset, and α1 and α2 represent the radius threshold and label difference threshold of the hypersphere in the feature space, respectively. Figures 4(a) to 4(c) The visualizations show the effects of different values ​​of α1 and α2 in two-dimensional space when K=5. The larger α1 is, the larger the neighborhood considered by each anchor point; the larger α2 is, the higher the tolerance of each anchor point for label differences in samples. This design concept can be understood as clustering samples into K clusters based on anchor points, with the radius of the annulus of each cluster being α1.

[0105] To further improve the comprehensiveness of molecular characterization during the training process of cell-targeted drug screening models, a cell-targeted drug screening model training method is provided in this application embodiment, see [link to relevant documentation]. Figure 3 Step 300 in the cell-targeted drug screening model training method specifically includes the following:

[0106] Step 310: Iteratively train the graph neural network-based regression model based on the training set and preset training hyperparameters, so that the graph neural network-based regression model can represent nodes representing atoms, edges representing chemical bonds, and global information containing the structural topology information of the compound sample as learnable vectors. During the iterative training process, the potential cell-targeted drug activity prediction results obtained by the graph neural network-based regression model each time are screened according to the preset prediction threshold range.

[0107] Step 320: Test the trained graph neural network-based regression model based on the test set. If the graph neural network-based regression model passes the model test, then the graph neural network-based regression model is used as a preliminary screening model for cell-targeted drugs to predict whether the compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound.

[0108] The regression model based on graph neural networks includes: Graph Convolutional Networks (GCN) model and / or Message Passing Neural Networks (MPNN) model.

[0109] Specifically, on samples after data screening, this application implemented two graph neural network-based regression models—GCN and MPNN—based on the DeepChem library. These models aim to perform further topological analysis on compounds screened by the ECFP functional group to achieve drug screening. Both models used default network structures, and the remaining training hyperparameters were set as follows: batch size of 16, learning rate of 0.001, and number of training epochs of 100. To prevent abnormal prediction scores from the model output, during the iterative training phase, the cytokine expression levels displayed in the regression model's predictions were pruned, restricting the prediction results to [y...]. min ,y max Within the range of ], where y min and y max These are the minimum and maximum values ​​of the label information in the original dataset, respectively.

[0110] Based on the cell-targeted drug screening model training method provided in the above embodiments, this application also provides an embodiment of a cell-targeted drug screening method, see [link to embodiment]. Figure 5 The cell-targeted drug screening method specifically includes the following:

[0111] Step 400: Obtain the molecular fingerprint feature data of the target compound.

[0112] Step 500: Input the molecular fingerprint feature data of the target compound into a preset cell-targeted drug screening model, so that the cell-targeted drug screening model outputs a potential cell-targeted drug activity prediction result indicating whether the target compound has potential cell-targeted drug activity, wherein the cell-targeted drug screening model is pre-trained based on the cell-targeted drug screening model training method.

[0113] It should be noted that the cell-targeted drug screening model training method mentioned in step 500 can be implemented with reference to the aforementioned embodiment of the cell-targeted drug screening model training method, and will not be repeated here.

[0114] As can be seen from the above description, the cell-targeted drug screening method provided in this application embodiment can effectively improve the comprehensiveness of molecular characterization during the training process of the cell-targeted drug screening model, and can effectively improve the generalization ability of the trained cell-targeted drug screening model, thereby improving the accuracy and reliability of the model's prediction of potential cell-targeted drug activity.

[0115] To further improve the timeliness and reliability of subsequent processing in the initial screening of cell-targeted drugs, the cell-targeted drug initial screening method provided in this application is described in [reference needed]. Figure 5 The cell-targeted drug screening method, after step 500, further includes the following:

[0116] Step 600: Output the potential cell-targeting drug activity prediction result corresponding to the target compound, so as to conduct in vitro cell experiments on the target compound when or after determining that the target compound has potential cell-targeting drug activity based on the potential cell-targeting drug activity prediction result.

[0117] Specifically, the cell-targeted drug screening device outputs the potential cell-targeted drug activity prediction results corresponding to the target compound, so that researchers can conduct in vitro cell experiments on the target compound when or after determining that the target compound has potential cell-targeted drug activity based on the potential cell-targeted drug activity prediction results.

[0118] In other words, after initially screening compounds with potential drug activity, the next step is to conduct more in-depth evaluations to verify the safety and efficacy of these compounds. This stage of evaluation includes more complex and detailed in vitro cell experiments, and may even involve the use of animal models. These experiments are designed to simulate the behavior of compounds in vivo, providing more comprehensive information about their pharmacological effects, toxicity, and metabolic properties.

[0119] The purpose of these experiments is to ensure that compounds are not only theoretically effective but also safe and effective in real biological systems. If compounds demonstrate good safety and expected efficacy in these further evaluations, they can be selected as candidates for biotesting. Biotesting is a crucial step in the drug development process, involving further testing of compounds in biological systems to validate their pharmacological effects and safety, and laying the foundation for subsequent clinical trials. This stage is a critical transition from laboratory to clinical application, ensuring that only the most promising compounds are further developed.

[0120] To further illustrate the effectiveness of the aforementioned cell-targeted drug screening model training method and cell-targeted drug screening method, this application also provides a method for training and validating a cell-targeted drug screening model, using immune cells as an example for illustration. (See [link to relevant documentation]). Figure 7 The training and validation method specifically includes the following:

[0121] Step 1: Compound Database and Data Preprocessing

[0122] This application example collected 1400 candidate compounds (compound name, number, etc.) highly correlated with immune cell activation or inhibition, and collected the expression levels of cytokines measured in biological experiments (label information corresponding to the compounds), constructing a dataset for cell-targeted drug screening. First, the raw data represented by CAS encoding was converted into corresponding SMILES strings through the PubChem API interface. During this conversion process, two special cases may occur: (1) the CAS encoding of the compound does not have a corresponding SMILES string; (2) the CAS encoding of the compound corresponds to multiple SMILES strings. In the Treg dataset, 26 compounds exhibited case (1), and these compounds were not used in subsequent experiments. For compounds exhibiting case (2), this application example defaulted to using the first SMILES string of the PubChem database query result as their corresponding conversion result. The preprocessed compounds are represented as SMILES strings. Finally, the ECFP2 fingerprints corresponding to these compounds were calculated as inputs to the subsequent data screening model.

[0123] Step 2: Data Filtering Strategy

[0124] To address the inseparability of promoters and inhibitors in the Treg dataset, this application proposes a data filtering strategy based on ECFP2 features. This strategy employs an efficient approximate distance metric (Hamming distance) to keep the label differences between samples with similar features within an acceptable range, thereby filtering out heterogeneous samples in the dataset. The specific algorithm flow of the data filtering strategy is as follows: Figure 6 As shown, the algorithm takes the original dataset, the number of baseline anchors K, and two threshold-related hyperparameters α1 and α2 as input, and outputs the filtered dataset. K represents the number of baseline anchors randomly selected from the original dataset, and α1 and α2 represent the radius threshold and label difference threshold of the feature space hypersphere, respectively.

[0125] Furthermore, there are multiple ways to set the hyperparameters α1 and α2. In one example of this application, an adaptive dataset setting method can be used, where the threshold for the hypersphere radius and the threshold for label difference tolerance are obtained by calculating quantiles based on the distribution of the overall dataset. Alternatively, these two hyperparameters can be pre-set empirically.

[0126] It should be noted that in the data filtering strategy proposed in the application examples, the latent feature space representation method (ECFP2 in this application example) and the distance metric function (Hamming distance is used in this application example) are interchangeable. For example, the radius of ECFP can be expanded, i.e., ECFP4 or ECFP6 can be used to extract features from compounds, or MACCS Key or other features can be used as the representation method for compounds. Regarding distance metrics, Hamming distance is the most common method for measuring distance between binary vectors, but in practice, Euclidean distance can also play a similar role.

[0127] Step 3: Implement and deploy the regression model

[0128] On the sample after data screening, four regression models were designed and validated based on the DeepChem library, including the traditional machine learning models SVR and Ridge, as well as two regression models based on graph neural networks: GCN and MPNN. The aim was to perform further topological analysis on compounds screened by ECFP functional groups to achieve drug screening tasks.

[0129] exist Figure 7 In the diagram, the MPNN model is omitted from the display, but in actual execution, step 3 trains the four models mentioned above, including the MPNN model. The network structures of all four models use default settings, and the remaining training hyperparameters are set as follows: batch size of 16, learning rate of 0.001, and number of training epochs of 100. To prevent the model from outputting abnormal prediction scores, the prediction scores of the regression model are manually clipped during the inference phase, i.e., limited to the range [y_min, y_max], where y_min and y_max are the minimum and maximum values ​​of the labeled values ​​in the original dataset, respectively.

[0130] In this application, the four models (SVR, Ridge, GCN, and MPNN) mentioned in the application examples are all independent and parallel models, with no correlation during deployment. Furthermore, the reason for proposing the use of multiple models for training and validation is to demonstrate the effectiveness of the data filtering method proposed in the previous step. In summary, the training process can be single or multiple. The training and evaluation processes of each model are completed independently, with the same execution logic and flow. Their inputs are datasets filtered by the data filtering strategy, which are further divided into training and test sets. The final output is a trained model and its corresponding evaluation results on the test set.

[0131] Furthermore, the application examples in this application can also train simpler linear regression models, such as linear regression, LASSO regression, and ridge regression, or other machine learning models, such as support vector regression, random forests, and multilayer perceptrons, all of which can be used for similar tasks. Even representative works such as MolFormer can fine-tune models on filtered datasets to achieve the same results. However, the effectiveness of these methods still needs to be evaluated. For example, while linear models can fit regression tasks, these methods lose information about the topological structure of compounds.

[0132] Step 4: Model Validation and Performance Evaluation

[0133] If SVR and Ridge modeling methods are used, the corresponding topological information is ignored when modeling molecules (because the feature extraction method for these two methods is still ECFP). Consequently, compared with GCN and MPNN graph neural network modeling methods, the experimental results of traditional machine learning models SVR and Ridge are usually worse.

[0134] To verify the above, all experimental results in the application examples of this application employ five-fold cross-validation. In addition to considering the mean absolute error (MAE) and root mean square error (RMSE) used on the MoleculeNet benchmark dataset, this application example also introduces two additional evaluation metrics for regression tasks: R² and Spearman correlation coefficient. These additional evaluation metrics help to more comprehensively assess the model's performance on drug screening tasks. In particular, the Spearman correlation coefficient considers only the ranking of variables, rather than their actual values. It measures the monotonic relationship between two variables and is unaffected by outliers. Therefore, it provides a more robust evaluation when dealing with molecular attribute data that may contain extreme values.

[0135] Because this application uses five-fold cross-validation, the above steps will be repeated five times based on different data partitions when facing the same input data (the filtered dataset), such as... Figure 8 The mean and standard deviation of each indicator for each model in the model validation and performance evaluation comparison table shown are also calculated in this way.

[0136] In the model validation and performance evaluation comparison table, the first two columns are hyperparameters related to the data filtering strategy. α1 has two options: 25 and 70; α2 has three options: 15, 30, and 40. " / " indicates that no data filtering strategy was used, and the original dataset was directly used for subsequent model training and validation (blank control group, used to illustrate the effectiveness of the data filtering strategy proposed in this application). The third column shows the number of samples in the filtered dataset based on different hyperparameters. Without the data filtering strategy, the total number of samples in the dataset is 1449; the number of samples in the filtered dataset is much smaller than the original dataset size. The fourth column lists the four models used, and columns 5-8 show the corresponding metrics for independent training and testing of the four models under the current data filtering strategy's hyperparameter settings.

[0137] "#samples" indicates the number of samples, "Model" indicates the model, and "MAE", "RMSE", "R2" and "Spearman's corr." are different regression model evaluation metrics. MAE (Mean Absolute Error) refers to the mean absolute error, RMSE (Root Mean Square Error) refers to the root mean square error, R2 (R-Square) refers to the goodness of fit, and "Spearman's corr." refers to the Spearman correlation coefficient.

[0138] Therefore, the application examples in this application aim to overcome the limitations of existing molecular characterization techniques, particularly in drug screening models and the discovery of new therapeutic targets. It primarily addresses the following two core issues:

[0139] ① Limitations of single-molecule characterization

[0140] Traditional molecular characterization methods typically rely on fixed molecular fingerprints or specific modalities, such as topological maps or SMILES encoding. These methods often fail to fully capture the complexity of molecules, especially the structural and spatial information of functional groups. To address this issue, this application employs a fusion method that considers both molecular functional group information and topological maps of compound structures. By integrating information from these two data structures, this method can more comprehensively preserve molecular features beneficial to drug screening models.

[0141] ② The problem of domain generalization

[0142] Existing molecular characterization methods may perform well on specific datasets, but they often underperform when dealing with data in unknown domains, lacking generalization ability. To address this challenge, this application proposes a novel data screening strategy. Samples are clustered into K clusters based on random anchor points. By adjusting the neighborhood radius of each cluster and the tolerance of each anchor point for label differences, samples are cleaned, potentially interfering compound samples are removed from heterogeneous datasets, reducing the negative impact of samples with similar characteristics but different label values ​​on the model. Subsequently, a graph neural network is used to model, train, and cross-validate the screened data, forming an effective initial drug screening strategy.

[0143] This application proposes an automated drug screening process. The first stage employs a latent feature space representation method and distance metric to filter noisy data. The second stage utilizes a different molecular modeling method, considering multiple molecular data modalities, to achieve a drug screening method for the immunology field. This method can efficiently discover potential drugs, rapidly and effectively identifying potential drugs from thousands of possible candidate molecules. This helps reduce R&D costs and improve the efficiency of the entire drug development process.

[0144] Regarding the limitations of single-molecule characterization: This application example considers both molecular fingerprints, which focus on information about molecular functional groups, and topological maps, which focus on compound structures. By combining the information from both molecular data structures, it retains as many molecular features as possible that may be helpful for drug screening models.

[0145] Regarding the issue of domain generalization: This application introduces a novel and efficient data screening strategy. This strategy focuses on removing potentially interfering compound samples from heterogeneous datasets, overcoming the influence of samples with similar characteristics but different label values ​​on the model. Then, a graph neural network is used to model, train, and cross-validate the screened data, thus forming an effective initial drug screening strategy.

[0146] In other words, this application proposes a molecular representation learning method that considers multiple molecular modalities. The proposed data filtering strategy effectively removes noisy samples from heterogeneous datasets and, to some extent, addresses the adaptability problem of graph neural networks (GCNs) and MPNNs in molecular machine learning, enabling the model to perform generalizable inference on molecular compound datasets with unstable labels. In summary, the basic principle of this application is to improve the accuracy and generalization ability of drug screening models by combining multimodal molecular data structures and efficient data filtering methods, thus opening new avenues for discovering novel therapeutic targets.

[0147] The application examples in this application simultaneously consider two molecular data structures: molecular fingerprints, which focus on information about molecular functional groups, and topological graphs, which focus on compound structures. By combining information from these two molecular data structures, the complexity and diversity of molecules can be better captured. Compared with existing molecular modeling methods that only consider a single modality, the application examples in this application can uncover key features related to physiological metabolism in compounds to a greater extent.

[0148] Data filtering strategies address domain adaptability issues: This application example improves the robustness, generalization, and domain adaptability of graph neural network models through data filtering strategies, thereby avoiding the problem of significant performance degradation outside the target domain. Simultaneously, this application example can effectively process and analyze small sample datasets for targeted drug screening and effectively ignore noisy samples in the dataset, improving the model's generalization ability.

[0149] From a software perspective, this application also provides a cell-targeted drug screening model training device for performing all or part of the cell-targeted drug screening model training method, wherein the cell-targeted drug screening model training device specifically includes the following:

[0150] The feature acquisition module is used to acquire the molecular fingerprint feature data corresponding to each compound sample with label information to form the corresponding original dataset, wherein the label information is used to indicate whether the corresponding compound sample has potential cell-targeting drug activity;

[0151] The data filtering module is used to filter the compound samples in the original dataset using a distance metric to obtain the corresponding target dataset and to divide the training set from the target dataset.

[0152] The model training module is used to train a graph neural network-based regression model based on the training set, so that the graph neural network-based regression model learns the structural topological information of each compound sample, and then trains the graph neural network-based regression model into a cell-targeted drug screening model for predicting whether a compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound.

[0153] The embodiments of the cell-targeted drug screening model training device provided in this application can be used to execute the processing flow of the cell-targeted drug screening model training method in the above embodiments. Its functions will not be repeated here, but can be referred to the detailed description of the cell-targeted drug screening model training method embodiments above.

[0154] The cell-targeted drug screening model training device can perform the training of the cell-targeted drug screening model either on a server or on a client device. The choice depends on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations in this regard. If all operations are performed on the client device, the client device may further include a processor for the specific processing of the cell-targeted drug screening model training.

[0155] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0156] The server and the client device can communicate using any suitable network protocol, including those not yet developed as of the date of this application. Such network protocols may include, for example, TCP / IP, UDP / IP, HTTP, HTTPS, etc. Furthermore, such network protocols may also include RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer Protocol) protocols used on top of the aforementioned protocols.

[0157] As can be seen from the above description, the cell-targeted drug screening model training device provided in this application embodiment can effectively improve the comprehensiveness of molecular characterization during the training process of the cell-targeted drug screening model, and can effectively improve the generalization ability of the trained cell-targeted drug screening model, thereby effectively improving the applicability of the trained cell-targeted drug screening model, and can improve the accuracy and reliability of the model's prediction of potential cell-targeted drug activity.

[0158] From a software perspective, this application also provides a cell-targeted drug screening device for performing all or part of the cell-targeted drug screening method, wherein the cell-targeted drug screening device specifically includes the following:

[0159] The data acquisition module is used to acquire the molecular fingerprint feature data of the target compound;

[0160] The model prediction module is used to input the molecular fingerprint feature data of the target compound into a preset cell-targeted drug screening model, so that the cell-targeted drug screening model outputs a potential cell-targeted drug activity prediction result indicating whether the target compound has potential cell-targeted drug activity. The cell-targeted drug screening model is pre-trained based on the cell-targeted drug screening model training method.

[0161] As can be seen from the above description, the cell-targeted drug screening device provided in this application embodiment can improve the accuracy and reliability of the predicted activity results of potential cell-targeted drugs by the model.

[0162] This application also provides an electronic device, which may include a processor, a memory, a receiver, and a transmitter. The processor is used to execute the cell-targeted drug screening model training method and / or cell-targeted drug screening method mentioned in the above embodiments. The processor and memory can be connected via a bus or other means, taking a bus connection as an example. The receiver can be connected to the processor and memory via wired or wireless means.

[0163] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0164] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the cell-targeted drug screening model training method and / or cell-targeted drug screening method in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the cell-targeted drug screening model training method and / or cell-targeted drug screening method in the above method embodiments.

[0165] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0166] The one or more modules are stored in the memory, and when executed by the processor, they execute the cell-targeted drug screening model training method and / or cell-targeted drug screening method in the embodiment.

[0167] In some embodiments of this application, the user equipment may include a processor, a memory, and a transceiver unit. The transceiver unit may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected via a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0168] As one implementation method, the functions of the receiver and transmitter in this application can be implemented by transceiver circuits or dedicated transceiver chips, and the processor can be implemented by dedicated processing chips, processing circuits or general-purpose chips.

[0169] As another implementation approach, the server provided in this application embodiment can be implemented using a general-purpose computer. That is, the program code implementing the processor, receiver, and transmitter functions is stored in memory, and the general-purpose processor implements the processor, receiver, and transmitter functions by executing the code in memory.

[0170] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned cell-targeted drug screening model training method and / or cell-targeted drug screening method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0171] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the cell-targeted drug screening model training method and / or the cell-targeted drug screening method.

[0172] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave.

[0173] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0174] In this application, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0175] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to the embodiments of this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for training a cell-targeted drug screening model, characterized in that, include: Obtain raw information data of each candidate compound used for target cell activation or inhibition, along with their respective tag information; The original information data of each candidate compound is converted into SMILES strings; For candidate compounds whose SMILES strings are successfully converted and whose generated SMILES strings are unique, the SMILES strings of the candidate compounds are used as compound samples. Candidate compounds for which SMILES string conversion failed were deleted; For candidate compounds whose SMILES strings are successfully converted and whose generated SMILES strings are not unique, one of the various SMILES strings of the candidate compound is selected as the compound sample. The molecular fingerprint feature data corresponding to each compound sample with label information is obtained to form the corresponding original dataset, wherein the label information is used to indicate whether the corresponding compound sample has potential cell-targeting drug activity; The distance metric is used to filter the compound samples in the original dataset to obtain the corresponding target dataset, and a training set is divided from the target dataset. A graph neural network-based regression model is trained based on the training set to learn the structural topological information of each compound sample. The graph neural network-based regression model is then trained to become a cell-targeted drug screening model for predicting whether a compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound. The step of using a distance metric to filter the compound samples in the original dataset to obtain a corresponding target dataset, and then dividing the target dataset into a training set, includes: Molecular fingerprint feature data of multiple compound samples are randomly selected from the original dataset to serve as reference anchors respectively; Based on the number of reference anchor points, a first hyperparameter for representing the radius threshold of the feature space hypersphere, and a second hyperparameter for representing the label difference threshold, a distance metric is used to cluster each of the compound samples in the original dataset to filter out heterogeneous samples in the original dataset, thereby obtaining the corresponding target dataset. The target dataset is divided into a training set and a test set; The distance measurement methods include: Hamming distance measurement method or Euclidean distance measurement method.

2. The method for training a cell-targeted drug screening model according to claim 1, characterized in that, The target cells include: tumor cells, immune cells, cardiomyocytes, neuronal cells, endocrine cells, hepatocytes, stem cells, or respiratory epithelial cells; The immune cells include: T cells, B cells, NK cells, myeloid-derived cells, or antigen-presenting cells.

3. The method for training a cell-targeted drug screening model according to claim 1, characterized in that, The process of obtaining molecular fingerprint feature data corresponding to each compound sample with labeled information to form the corresponding original dataset includes: Based on a pre-defined latent feature space representation method, molecular fingerprint feature data corresponding to each compound sample with label information is obtained. Generate an original dataset containing the molecular fingerprint feature data and the tag information corresponding to each of the compound samples; The latent feature space representation method includes at least one of the following: ECFP2 molecular fingerprint representation method, ECFP4 molecular fingerprint representation method, ECFP6 molecular fingerprint representation method, and MACCS Key molecular fingerprint representation method.

4. The method for training a cell-targeted drug screening model according to claim 1, characterized in that, The step of training a graph neural network-based regression model based on the training set, so that the graph neural network-based regression model learns the structural topological information of each compound sample, and then trains the graph neural network-based regression model into a cell-targeted drug screening model for predicting whether a compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound, includes: The graph neural network-based regression model is iteratively trained based on the training set and preset training hyperparameters, so that the graph neural network-based regression model can represent nodes for atoms, edges for chemical bonds, and global information containing the structural topology information of the compound sample as learnable vectors. During the iterative training process, the potential cell-targeted drug activity prediction results obtained by the graph neural network-based regression model each time are filtered according to the preset prediction threshold range. The trained graph neural network-based regression model is tested based on the test set. If the graph neural network-based regression model passes the model test, it is used as a preliminary screening model for cell-targeted drugs to predict whether a compound has potential cell-targeted drug activity based on the molecular fingerprint feature data of the compound.

5. The method for training a cell-targeted drug screening model according to any one of claims 1 to 4, characterized in that, The regression models based on graph neural networks include: GCN models and / or MPNN models.

6. A method for primary screening of cell-targeted drugs, characterized in that, include: Obtain molecular fingerprint feature data of the target compound; The molecular fingerprint feature data of the target compound is input into a preset cell-targeted drug screening model, so that the cell-targeted drug screening model outputs a potential cell-targeted drug activity prediction result indicating whether the target compound has potential cell-targeted drug activity, wherein the cell-targeted drug screening model is trained in advance based on the cell-targeted drug screening model training method according to any one of claims 1 to 5.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the cell-targeted drug screening model training method as described in any one of claims 1 to 5, and / or implements the cell-targeted drug screening method as described in claim 6.

8. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the cell-targeted drug screening model training method as described in any one of claims 1 to 5, and / or implements the cell-targeted drug screening method as described in claim 6.

Images