A method and device for screening drug activity

Abstract

The present invention provides a method and device for screening drug activity, which relates to the field of drug analysis technology, including obtaining first information, second information and third information; performing feature extraction processing according to the first information to obtain a molecular descriptor; performing feature extraction processing according to the second information to obtain cell mechanism feature data; performing fusion processing according to the molecular descriptor and the cell mechanism feature data to obtain a fusion feature set; performing modeling processing according to the fusion feature set, and obtaining an activity prediction model by simulating the interaction between different compounds and compound combinations in a cell environment; performing spectrum analysis and component identification processing according to the molecular descriptor and the third information to obtain drug component information; inputting the drug component information into the activity prediction model for activity screening processing to obtain an overall activity prediction result. The present invention further improves the accuracy of the drug activity prediction model by simulating the interaction between different compound combinations in a cell environment.

Description

Technical Field

[0001] The present invention relates to the technical field of drug analysis, and in particular to a method and device for screening drug activity. Background Art

[0002] In modern drug development, drug activity screening is a crucial step, and its main goal is to screen compounds with potential biological activity through efficient and accurate methods. Current technologies mainly rely on high-throughput screening (HTS) methods, which are able to test thousands of compounds simultaneously to find effective drug molecules.

[0003] However, although HTS technology has been widely used in drug screening, it still has some significant problems: first, the HTS method usually requires a lot of experimental resources and time, resulting in high costs; in addition, the screening results of the HTS method often have high false positive and false negative rates, requiring a large number of subsequent verification experiments, further increasing the burden of time and cost.

[0004] Based on the above-mentioned shortcomings of the prior art, a method and device for screening drug activity is urgently needed. Summary of the invention

[0005] The purpose of the present invention is to provide a method and device for screening drug activity to improve the above problems. In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] In one aspect, the present application provides a method for screening drug activity, comprising:

[0007] Acquiring first information, second information and third information, wherein the first information includes drug-related compounds and their biological activity data, the second information includes data on the mechanism of action of the compounds at the cellular level, and the third information includes nuclear magnetic resonance spectroscopy data of the drug to be screened;

[0008] Performing feature extraction processing according to the first information to obtain a molecular descriptor by extracting global and local structural features of the compound and identifying functional groups associated with biological activity;

[0009] Performing feature extraction processing according to the second information, obtaining cell mechanism feature data by identifying changes in cell morphology, protein distribution and signal pathway activation status under the action of the compound;

[0010] Performing fusion processing according to the molecular descriptor and the cell mechanism feature data, and constructing a fusion feature set by combining feature selection and weighted average strategy;

[0011] Performing modeling processing according to the fusion feature set, and obtaining an activity prediction model by simulating the interactions of different compounds and compound combinations in a cell environment;

[0012] Performing spectrum analysis and component identification processing according to the molecular descriptor and the third information to obtain drug component information, wherein the drug component information includes drug components and their structural information;

[0013] The drug component information is input into the activity prediction model for activity screening to obtain an overall activity prediction result.

[0014] On the other hand, the present application also provides a drug activity screening device, comprising:

[0015] an acquisition module, configured to acquire first information, second information and third information, wherein the first information includes data on drug-related compounds and their biological activities, the second information includes data on the action mechanism of the compounds at the cellular level, and the third information includes nuclear magnetic resonance spectroscopy data of the drug to be screened;

[0016] An extraction module, used for performing feature extraction processing according to the first information, obtaining a molecular descriptor by extracting global and local structural features of the compound and identifying functional groups related to biological activity;

[0017] an identification module, configured to perform feature extraction processing according to the second information, and obtain cell mechanism feature data by identifying changes in cell morphology, protein distribution, and activation status of signal pathways under the action of the compound;

[0018] A fusion module, used for performing fusion processing according to the molecular descriptor and the cell mechanism feature data, and obtaining a fusion feature set by combining feature selection and weighted average strategy;

[0019] A construction module is used to perform modeling processing according to the fusion feature set, and obtain an activity prediction model by simulating the interactions of different compounds and compound combinations in a cell environment;

[0020] An analysis module, used for performing spectrum analysis and component identification processing according to the molecular descriptor and the third information to obtain drug component information, wherein the drug component information includes drug components and their structural information;

[0021] The prediction module is used to input the drug component information into the activity prediction model to perform activity screening processing to obtain an overall activity prediction result.

[0022] The beneficial effects of the present invention are:

[0023] The present invention uses algorithms such as molecular graph convolutional networks and self-attention mechanisms to conduct in-depth analysis of the global and local structural features of compounds, significantly improving the efficiency and accuracy of feature extraction; by simulating the interactions of different compound combinations in the cellular environment, the accuracy of the drug activity prediction model is further improved.

[0024] Other features and advantages of the present invention will be set forth in the following description, and in part will become apparent from the description, or may be understood by practicing embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0026] Figure 1 A schematic diagram of the process for screening drug activity described in an embodiment of the present invention;

[0027] Figure 2 is a flow chart of the molecular graph convolutional network described in an embodiment of the present invention;

[0028] Figure 3 is a schematic diagram of a multi-layer perceptron described in an embodiment of the present invention;

[0029] Figure 4 Schematic diagram of the structure of the drug activity screening device described in an embodiment of the present invention.

[0030] Labels in the figure: 1. Acquisition module; 2. Extraction module; 3. Recognition module; 4. Fusion module; 5. Construction module; 6. Parsing module; 7. Prediction module. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0032] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0033] Embodiment 1:

[0034] like Figure 1 As shown, this embodiment provides a method for screening drug activity, which includes steps S100 to S700:

[0035] Step S100, obtaining first information, second information and third information;

[0036] Wherein, the first information includes drug-related compounds and their biological activity data, the second information includes data on the mechanism of action of the compounds at the cellular level, and the third information includes nuclear magnetic resonance spectroscopy data of the drug to be screened;

[0037] For example, the drug-related compounds and their biological activity data in the first information may come from chemical databases such as PubChem, ChEMBL or special experimental results, which can provide the molecular structure, physicochemical properties and activity performance of the compounds under different biological experimental conditions; the action mechanism data of the compounds at the cellular level in the second information may come from the results of cell experiments or high-throughput screening (HTS) technology, which includes the effects of the compounds on cell morphology, cell cycle, protein expression and signal pathway activation, and these data can be obtained through high-content screening (HCS) technology or other biological imaging technologies; the nuclear magnetic resonance spectroscopy (NMR) data of the drugs to be screened in the third information can provide detailed structural information of the compounds, and analyze their chemical shifts and molecular environments by identifying the characteristic peaks generated by the compounds in the magnetic field;

[0038] Therefore, through the first information, the second information and the third information, not only the comprehensiveness and diversity of the data are ensured, but also a solid foundation is provided for subsequent feature extraction and model construction, thereby significantly improving the accuracy and reliability of activity prediction.

[0039] Step S200, performing feature extraction processing according to the first information, extracting global and local structural features of the compound and identifying functional groups related to biological activity to obtain molecular descriptors;

[0040] Preferably, this step comprehensively and accurately extracts the global and local structural features and functional groups of the compound by combining deep learning algorithms and chemical informatics methods. Molecular descriptors are numerical values ​​or symbols used to represent the properties of compounds. They are single data points or a set of data that reflect the chemical structure and possible biological activity of the molecule. They include not only the global properties of the molecule, such as molecular weight, logarithmic P value (LogP), polar surface area, etc., but also local properties, such as the type, quantity and spatial arrangement relationship of atoms or functional groups.

[0041] Step S300, performing feature extraction processing according to the second information, and obtaining cell mechanism feature data by identifying changes in cell morphology, protein distribution and signal pathway activation status under the action of the compound;

[0042] This step can comprehensively analyze the mechanism of action of compounds on cells from multiple levels and angles, and generate cell mechanism feature data. These feature data can not only provide more detailed cell response information, but also provide rich basic data for subsequent model construction and activity prediction, significantly improving the accuracy and credibility of the model. The mechanism data in the second information are the original experimental results, including data from biological imaging and biochemical analysis. In order to accurately use them for the establishment of activity prediction models, these data need to be further processed for feature extraction. This processing step includes systematically extracting key information from aspects such as cell morphology changes, protein distribution, and signal pathway activation. For example, image processing technology is used to quantify specific changes in cell morphology, immunofluorescence staining technology is used to analyze protein localization and concentration changes, and biochemical methods are used to evaluate the activation state of signal pathways. The data obtained after processing are not only in a unified format, but also effectively reflect the biological effects of the compounds.

[0043] Step S400: performing fusion processing based on the molecular descriptors and the cell mechanism feature data, and constructing a fusion feature set by combining feature selection and weighted average strategy.

[0044] Step S500, performing modeling processing according to the fusion feature set, and obtaining an activity prediction model by simulating the interactions of different compounds and compound combinations in a cell environment;

[0045] The activity prediction model generated in this step is used to simulate the interactions of different compound combinations, which helps to reveal their biological effects and improve the accuracy and reliability of predictions.

[0046] Step S600: performing spectrum analysis and component identification processing according to the molecular descriptor and the third information to obtain drug component information, wherein the drug component information includes drug component names and structure information thereof;

[0047] It can be understood that this step comprehensively and accurately identifies and confirms the drug components and their structures by combining spectral analysis and chemical database matching.

[0048] Step S700: input drug component information into an activity prediction model for activity screening to obtain drug activity prediction results.

[0049] Furthermore, the step S200 includes steps S210 to S240:

[0050] Step S210: performing global structural feature extraction processing according to the first information, by converting the molecular structure of the compound into a graph, and using a molecular graph convolutional network to perform a convolution operation on the graph, extracting local information of each node and aggregating it within the entire graph, thereby obtaining a global structural feature representation of the compound;

[0051] Specifically, the molecular structure of the compound needs to be converted into the form of a graph first. Each compound molecule can be represented as a graph, in which nodes represent atoms and edges represent bonds between atoms. Take aspirin as an example. The chemical formula of aspirin is C9H8O4, and its structure contains a benzene ring, a carboxylic acid group, and an ester group. When converting aspirin into a molecular graph, each atom is represented as a node in the graph, and carbon, hydrogen, and oxygen atoms are distinguished by different colors. The chemical bonds between atoms are represented as edges between nodes, such as the carbon atoms in the benzene ring are connected by alternating single bonds and double bonds, while the double bonds in the carboxylic acid group connect carbon and oxygen. Furthermore, each chemical bond in aspirin is not only a line connecting two nodes, but the properties of the chemical bond (such as the type of bond and the length of the bond) are also encoded in the graph. In addition, specific functional groups, such as carboxylic acids and ester groups, are also emphasized in the graph through different subgraphs. The graph representation method can retain the topological structure of the compound, making it suitable for subsequent graph convolution operations. Next, as Figure 2 As shown in Figure 1, a molecular graph convolutional network is used to perform convolution operations on the graph. By performing convolution operations on the graph, the molecular graph convolutional network can capture the information of each node and its neighboring nodes, thereby extracting local structural features.

[0052] Specifically, in each layer of convolution operation, each node aggregates the information of its neighboring nodes. This process is performed layer by layer, so that the features of the node contain not only its own information, but also the structural information of its neighborhood. Through the convolution operation of the multi-layer molecular graph convolutional network, feature aggregation is finally performed across the entire graph. After multi-layer aggregation, the information of each node forms a global feature representation that can reflect the structural characteristics of the entire molecule. This global structural feature representation contains the complex relationships and interactions between the various parts of the molecule, and is a comprehensive description of the overall structure of the compound.

[0053] Step S220, performing local structural feature extraction processing according to the global structural features, by identifying the interaction between groups, using the self-attention mechanism to calculate the mutual dependence between substructures, and obtaining local structural features, wherein the groups include one or more of aromatic rings, amino groups and hydroxyl groups;

[0054] Specifically, first, key substructures are separated from the global structural features. The key substructures include functional groups such as aromatic rings, amino groups and hydroxyl groups, which play an important role in the biological activity of the compound. Therefore, accurately identifying these substructures is an important part of feature extraction. Then the self-attention mechanism is used to calculate the interdependencies between these substructures. The self-attention mechanism is a deep learning technique that was originally applied to natural language processing, but is also effective in molecular feature extraction. Through the self-attention mechanism, the model can calculate the interrelationships and dependencies between different substructures within the compound. The self-attention mechanism measures the strength of the interaction between them by assigning weights to each pair of substructures. In this way, the model can identify which interactions between substructures have an important impact on the overall properties and biological activity of the compound.

[0055] For example, the self-attention mechanism can generate an attention matrix by calculating the similarity and distance between substructures. Each element in this matrix represents the degree of attention of a substructure to other substructures, thereby quantifying the mutual dependence between them. This method can effectively capture complex structural features, such as the conjugation effect of aromatic rings, the hydrogen bonding effect of amino groups, and the polar properties of hydroxyl groups. Ultimately, the local structural features can be obtained through the above method, reflecting the interactions and dependencies between substructures within the compound, forming a comprehensive local feature representation, accurately identifying and quantifying the interactions between important substructures within the compound, and generating high-quality local structural features, which not only improves the accuracy and depth of feature extraction, but also provides a rich information basis for subsequent activity prediction.

[0056] Step S230, performing functional group identification processing according to the local structural features, identifying specific functional groups related to biological activity by analyzing and matching the local structural features, and comparing the local structural features with the patterns of known active groups to confirm the existence of specific functional groups in the molecule and their relative positions, and obtaining functional group features;

[0057] Specifically, the local structural features are first analyzed to identify the key functional groups in the compound. This step usually uses chemical informatics methods to match the local structural features with a database of known functional groups. Through this matching process, functional groups in the compound, such as hydroxyl, amino, carboxyl, etc., can be quickly located and identified. These groups play a key role in the biological activity of the compound.

[0058] Preferably, the parsing and matching process is achieved by using fingerprinting and substructure search algorithms. Fingerprinting converts molecules into feature vectors, generates molecular fingerprints through hashing algorithms, and matches known active groups in the database. Substructure search algorithms directly search for known active group patterns in the local structure of the compound through graph matching technology. Next, the identified functional groups are compared with the patterns of known active groups. This step confirms the presence and relative positions of these functional groups in the molecule by comparing the spatial arrangement and chemical environment of local structural features with known active groups. For example, by three-dimensional structure comparison, the spatial arrangement of functional groups can be determined to ensure that they are in the correct position and conformation required for biological activity, thus, by comparison and confirmation, and using known structure-activity relationships, the functional groups can be accurately located. Through the above steps, functional group features are finally obtained, which not only contain the chemical composition and spatial structure information of the functional groups, but also reflect their relevance to biological activity.

[0059] Step S240: performing feature fusion processing according to the global structural feature representation, the local structural feature and the functional group feature, and adjusting the contribution ratio of each feature through a weighted average strategy to construct a molecular descriptor set.

[0060] Therefore, this step can fully consider the contribution of each feature when fusing features through a weighted average strategy, ensuring that the final set of molecular descriptors contains both global structural information and the details of local structures and functional groups, so as to more accurately reflect the overall characteristics of the compound through this multi-level feature fusion method and improve the prediction accuracy of the model.

[0061] Further, step S300 includes steps S310 to S340:

[0062] Step S310, performing cell morphology change analysis and processing according to the second information, identifying and extracting cell morphology change data to obtain cell morphology characteristics by processing cell image data under the action of the compound;

[0063] Specifically, first, cell image data before and after compound treatment is obtained, and the cell image data is usually collected by a high content screening (HCS) system or a fluorescence microscope. The image data is preprocessed, including denoising, correction, and standardization, to ensure data quality and consistency. Further, the preprocessed cell images are analyzed using image processing and computer vision techniques. Convolutional neural networks automatically learn features in images and extract morphological features of cells through multi-layer convolution and pooling operations. The morphological features include cell size, shape, nuclear morphology, distance and distribution between cells, etc., which are important indicators for evaluating cell responses.

[0064] In addition, the cells are separated from the background using image segmentation technology to further analyze the morphological changes of individual cells. The image segmentation method includes a Canny algorithm based on edge detection, a segmentation method based on regional growth, and a U-Net model based on deep learning, etc. After being processed by the above image segmentation method, the morphological characteristics of individual cells can be accurately identified and extracted. Furthermore, by calculating and comparing the changes in cell morphology before and after treatment, the specific effects of the compound on cell morphology can be identified, such as cell expansion, contraction, morphological variation, and changes in cell spacing, which not only reflect the physical changes of the cell, but also indirectly reveal the biological activity and mechanism of action of the compound, providing important information support for drug screening and evaluation.

[0065] Step S320, performing protein distribution feature extraction processing according to the cell morphology characteristics, by processing the protein staining image data, identifying and extracting the distribution characteristics of the protein in the cell, including protein concentration, distribution pattern and localization characteristics, to obtain the protein distribution characteristics;

[0066] Specifically, this step uses image segmentation and analysis technology to separate the cell area from the background in the protein staining image and extract the distribution information of the protein. Furthermore, a convolutional neural network is also used to automatically identify and extract the distribution characteristics of the protein in the cell, including the concentration, distribution pattern and localization characteristics of the protein.

[0067] Thus, through multi-layer convolution operations, the complex distribution pattern of proteins in cells can be captured, including the specific distribution of proteins in the nucleus, cytoplasm and cell membrane, as well as the relative concentration of proteins in different cell structures. In addition, image analysis software such as CellProfiler is used to automatically analyze protein staining images to extract the spatial distribution characteristics and quantitative information of proteins, including the average concentration of proteins, localization patterns in cells (such as nuclear localization, membrane localization, etc.), and localization characteristics of proteins in specific subcellular structures, etc., and finally the protein distribution characteristics are obtained. The protein distribution characteristics can not only reveal the effects of compounds on protein expression and localization, but also reflect the specific mechanisms of cellular responses, which can provide important information for understanding the dynamic changes of proteins under the action of compounds.

[0068] Step S330, performing signal pathway activation feature extraction processing according to the protein distribution characteristics, analyzing the phosphorylated protein data and using the pathway analysis algorithm to identify and extract the activation state of the cell signal pathway under the action of the compound, and obtaining the signal pathway activation feature, which includes the activation level of the key node protein and the activation order of the signal pathway;

[0069] Since phosphorylation is usually an important regulatory mechanism for signal transduction, and phosphorylated protein data can be obtained through high-throughput technologies such as mass spectrometry or phosphorylated protein arrays to provide information about the dynamic changes of proteins under phosphorylation states, pathway analysis algorithms are then used to identify the activation state of the signaling pathway.

[0070] In this embodiment, preferably, a dynamic Bayesian network is used to model and infer causal relationships in time series data. The dynamic Bayesian network can process the temporal changes in phosphorylated protein data, identify the activation levels of key node proteins and the activation order of signal pathways, and thus, a dynamic model of intracellular signal pathways can be constructed through the dynamic Bayesian network to capture changes in signal transduction after compound treatment. The specific operation includes the following steps:

[0071] First, the dynamic Bayesian network was used to analyze the data to identify the key node proteins in the signal pathway and their interaction relationships. Then, the activation state of each node protein was determined by analyzing the changes in the phosphorylation level of the protein. Finally, the activation order of the signal pathway was inferred based on the model, revealing how compounds affect cell function by regulating signal transduction pathways.

[0072] Signaling pathway activation features include the activation level of key node proteins and the activation order of signaling pathways. For example, in some cases, compounds may promote cell proliferation by activating the MAPK / ERK pathway, or induce cell apoptosis by inhibiting the PI3K / Akt pathway, which not only provides a deep understanding of the mechanism of action of the compound, but also provides an important reference for subsequent activity prediction and drug development.

[0073] Step S340: Perform principal component analysis based on cell morphology characteristics, protein distribution characteristics and pathway activation characteristics to construct cell mechanism characteristic data.

[0074] The principal component analysis is a dimensionality reduction technique that projects high-dimensional data into a low-dimensional space to extract the most representative principal components in the data, thereby reducing the complexity and redundancy of the data to retain the most important information.

[0075] Embodiment 2:

[0076] The difference between this embodiment and embodiment 1 is that step S400 includes steps S410 to S440:

[0077] Step S410, performing feature selection processing according to the molecular descriptors and the cell mechanism feature data, recursively training the model by using a recursive feature elimination algorithm and removing the features that contribute the least to the prediction, to obtain an optimized feature set;

[0078] It is understood that this step integrates the molecular descriptors and cell mechanism feature data into a complete feature matrix. Then, a basic model is trained with all the features in the feature matrix, preferably using a support vector machine or a linear regression model. After the model training is completed, the contribution of each feature to the model prediction result is calculated.

[0079] Furthermore, the calculation of the contribution can be represented by the coefficient or weight of the feature, such as the regression coefficient in a linear model or the feature weight in a support vector machine. Then, remove the feature with the smallest contribution, that is, the feature with the smallest impact on the prediction accuracy of the model. Next, retrain the model using the remaining features, calculate the contribution of the feature again, and continue to remove the feature with the smallest contribution. Repeat the above steps until a predetermined number of features remain, which are the optimized feature set.

[0080] Step S420: perform evaluation processing based on the optimized feature set, and obtain a weighted feature set by calculating the importance weight of each feature to the model prediction result. Thus, by quantifying the contribution of each feature to the model prediction performance, the feature weights can be reasonably allocated in subsequent steps to improve the prediction ability of the model.

[0081] Step S430, performing feature fusion processing according to the weighted feature set, by jointly training the molecular descriptors and the cell mechanism features, and combining the correlation and complementarity between the features, to obtain preliminary fusion features;

[0082] Specifically, the molecular descriptors and cell mechanism features are first standardized using the weight information in the weighted feature set to ensure that the scales of features from different sources are consistent. Next, a multimodal learning method is used to jointly train the molecular descriptors and cell mechanism features. Multimodal learning can process data from different sources, and through the joint training model, the correlation and complementarity between features can be fully explored.

[0083] Preferably, in the training process, this embodiment uses a multi-layer perceptron to achieve nonlinear fusion of features. The multi-layer perceptron can learn the complex relationship between features and generate high-quality preliminary fusion features through multiple hidden layers and nonlinear activation functions.

[0084] In addition, this embodiment further optimizes the fusion process by using correlation analysis between features. Correlation analysis can be performed by calculating mutual information between features.

[0085] Therefore, by analyzing the correlation between different features, highly correlated feature pairs can be identified, and appropriate feature dimensionality reduction processing can be performed to reduce redundant information and improve the quality of fused features.

[0086] Step S440: Perform feature dimensionality reduction processing based on the preliminary fusion features, and obtain the final fusion feature set by reducing the dimension of the feature set while retaining the most informative feature components.

[0087] Specifically, this step first reduces the dimension of the preliminary fused features, and then uses nonlinear dimensionality reduction methods such as t-SNE to further optimize the feature dimensionality reduction results. This method can better preserve the local and global structures of the data in the low-dimensional space. At the same time, t-SNE converts the similarity of high-dimensional data into a probability distribution in a low-dimensional space and minimizes the difference between the two, which is suitable for visualizing and processing data with complex nonlinear structures. The final fused feature set not only retains the global and local information of the data, but also significantly reduces the feature dimension, ensuring the efficiency and accuracy of the subsequent model.

[0088] Embodiment 3:

[0089] The difference between this embodiment and embodiment 1 or 2 is that step S500 includes steps S510 to S540:

[0090] Step S510, performing feature conversion processing according to the fused feature set, sorting and combining different compounds by using a multi-layer perceptron, and converting the combined features to obtain a combined feature set;

[0091] In this step, the multilayer perceptron consists of an input layer, multiple hidden layers, and an output layer. Figure 3 As shown in the figure, GSD represents the global structure descriptor, LSD represents the local structure descriptor, FTD represents the functional group descriptor, PA represents the probability that the compound has the expected biological activity, and NA represents the probability that the compound does not have the expected biological activity. The input layer receives the fused feature set, and the hidden layer performs nonlinear transformation through the activation function (such as ReLU, Sigmoid or Tanh) to extract the complex relationship between the features.

[0092] Each hidden layer consists of multiple neurons, which receive input from the previous layer in a fully connected manner and perform a weighted sum operation. During the training process, the multilayer perceptron sorts and combines different compounds. The sorting operation ranks each compound according to its characteristic value, and the combination operation crosses and recombines the features of the top-ranked compounds to generate new combined features. The above operations can simulate the interaction of compounds in different environments and reveal their potential synergistic or antagonistic effects. Subsequently, the combined features are further transformed.

[0093] The conversion operation is implemented through the hidden layer of the multilayer perceptron, which can capture the nonlinear relationship between features and generate combined features with high expressive power. For example, through multi-layer nonlinear transformation, the multilayer perceptron can identify the synergistic effect pattern between compounds, such as the ability to enhance or inhibit specific biological processes. Through this nonlinear transformation, the expression form of features can be greatly enriched, so that the combined features can better reflect the biological activity of the compounds. Finally, the output layer generates a combined feature set, which contains high-dimensional feature data after sorting, combination and transformation. It not only retains the key information in the original fusion features, but also enhances the expressive power of the features through nonlinear transformation.

[0094] Step S520, performing interaction simulation processing according to the combination feature set, and obtaining simulated interaction features by using a generative adversarial network to simulate the mutual promotion or counteraction effects of different compound combinations in a cell environment;

[0095] Specifically, the combined feature set is input into the generative adversarial network for training. The generative adversarial network consists of two main parts: the generator and the discriminator. The task of the generator is to generate simulated interactive features based on the input data, while the task of the discriminator is to distinguish between the generated data and the real data.

[0096] Through adversarial training of the two networks, the generator gradually learns to generate realistic interaction features. During the training process, the generator receives a set of combined features and generates simulated interaction features through a series of nonlinear transformations. These transformations can capture the interactions between different combinations of compounds in the cellular environment, including mutually reinforcing effects and counteracting effects. For example, when two compounds are combined, they may jointly enhance the activation of a certain signaling pathway, thereby improving the biological activity of the cell; conversely, they may offset each other through different mechanisms, weakening the overall effect. The discriminator receives both the interaction features generated by the generator and the real interaction data, and provides feedback to the generator by evaluating their authenticity. The training goal of the discriminator is to maximize the accuracy of distinguishing generated data from real data, while the goal of the generator is to minimize the discriminator's ability to distinguish. Through this adversarial training mechanism, the generator gradually improves its generation ability, making the simulated interaction features closer and closer to the real data.

[0097] Step S530: Optimize the simulated interaction features by calculating the importance score of each feature and selecting the feature with the highest score to obtain optimized interaction features;

[0098] It should be noted that in this step, the simulated interaction features are used as input data to complete model training, preferably such as a neural network or a decision tree, and calculate the importance score of each feature. The importance score is calculated based on the sensitivity of the feature to the model prediction loss. The specific formula is as follows:

[0099] ;

[0100] in, Indicates the sample number, represents the total number of samples, represents the loss function, Indicates The characteristics of the samples, Indicates the actual value, represents the predicted value, Representation characteristics The importance score of the feature. Through the above calculation process, the most important features for model prediction can be effectively identified and retained. Features with high importance scores contribute significantly to model performance, while features with low scores may be noise or redundant information and should be removed.

[0101] Step S540: Perform model training processing according to the optimized interaction features, and obtain an activity prediction model by using a support vector regression algorithm to simulate the interaction between different compounds in the cell environment and predict the overall activity of the drug under different compound compositions.

[0102] The support vector regression algorithm maps the input features to a high-dimensional space by finding the optimal hyperplane in the high-dimensional feature space, making the nonlinear relationship between input and output linear. The activity prediction model obtained through training can be used to screen and evaluate the potential effects of new drugs, thereby improving the efficiency and success rate of drug research and development.

[0103] Embodiment 4:

[0104] The only difference between this embodiment and any one of the embodiments 1-3 is that step S600 includes steps S610 to S640.

[0105] Step S610, performing preprocessing according to the third information, obtaining a preprocessed spectrum by performing baseline correction and noise filtering on the nuclear magnetic resonance spectrum data;

[0106] Baseline drift is a common phenomenon in NMR spectroscopy data, manifested as a non-horizontal shift of the spectral baseline, which interferes with the accurate detection of characteristic peaks. The purpose of baseline correction is to adjust the baseline to a horizontal state by mathematical methods. The purpose of noise filtering is to remove random noise in NMR spectroscopy data to improve the signal-to-noise ratio.

[0107] Step S620, performing peak recognition processing on the preprocessed spectrum, identifying characteristic peaks in the spectrum by performing multi-scale analysis on the spectrum using a multi-resolution wavelet transform, and marking the chemical shift and peak intensity of each characteristic peak to obtain characteristic peak data;

[0108] Specifically, the preprocessed spectrum is firstly subjected to wavelet transform using Daubechies wavelet. Wavelet transform can decompose the spectrum into details and approximate components at different scales, facilitating multi-scale analysis. Then, the spectrum is analyzed at different scales to identify the characteristic peaks in the spectrum. The characteristic peak refers to the chemical shift signal of the compound in the nuclear magnetic resonance spectrum, and its height and position reflect the structure and environment of the compound. Then, by analyzing the detail components after wavelet transform, the position and intensity of the characteristic peak are detected.

[0109] Preferably, a peak detection algorithm (such as extreme value detection or thresholding) is used to identify significant peaks in the spectrum. The identified characteristic peaks are then marked with chemical shift (chemical shift is the horizontal coordinate of the nuclear magnetic resonance spectrum signal, expressed in ppm) and peak intensity (peak intensity reflects the relative intensity of the signal). This step ensures that the data of each characteristic peak is complete and easy to process later. Finally, all the identified and marked characteristic peak data are organized into a structured form, including the chemical shift and intensity of each characteristic peak, to generate a characteristic peak data set.

[0110] Step S630, performing chemical shift matching processing according to the characteristic peak data, and calculating the matching degree between each characteristic peak and the known compounds in the chemical shift database by using the Pearson correlation coefficient to obtain a potential compound candidate list;

[0111] The characteristic peak data includes the chemical shift and intensity information of each characteristic peak. These data are used to compare with known compounds in the chemical shift database, and the similarity between the characteristic peaks and the known compounds is quantitatively evaluated to identify possible compound components. The matching process is as follows:

[0112] Step S631: for each characteristic peak and each compound in the database , calculate the mean of its chemical shift and .

[0113] Step S632: Calculate the Pearson correlation coefficient using the following formula:

[0114] ;

[0115] in, Indicates the sequence number of the characteristic peak, Indicates the serial number of the compound, Indicates the sample number, represents the total number of samples, Characteristic peak Compounds in the database The Pearson correlation coefficient, Characteristic peak In the The chemical shift in the sample, Represents compound In the The chemical shift in the sample, Characteristic peak The average chemical shift of Represents compound The average chemical shift.

[0116] Step S633: Based on the calculated Pearson correlation coefficient , evaluate the matching degree between characteristic peaks and compounds. Correlation coefficient The value is between -1 and 1, and the closer the value is to 1, the higher the match degree is.

[0117] Step S634: According to the matching results, the compound with the highest matching degree is included in the potential compound candidate list, that is, several compounds with the highest correlation coefficient are selected as possible candidates.

[0118] Step S640: Perform structural reconstruction processing according to the potential compound candidate list and the molecular descriptors, and obtain drug component information by optimizing and confirming the structure of the candidate compounds.

[0119] Preferably, density functional theory is used to optimize the structure of the candidate compound. Density functional theory can accurately calculate molecular orbitals, energy states and other key structural parameters to optimize the three-dimensional structure of the compound. Through these calculations, the stable structure of the compound in the lowest energy conformation can be confirmed. Further, molecular dynamics simulation is used to further verify the stability of the compound structure. Molecular dynamics simulation examines the stability and changes of its structure under physiological conditions by simulating the movement of molecules under certain temperature and pressure. Through structural optimization and confirmation processing, the structure of the candidate compound is accurately reconstructed to ensure the accuracy and reliability of the drug component information.

[0120] Step S700 includes steps S710 to S740.

[0121] Step S710, performing format conversion processing according to the drug component information, and obtaining structured input data by converting the drug components and their structure information into a descriptor matrix;

[0122] Specifically, first obtain the confirmed drug component information, including chemical structure and physicochemical properties. Then, use chemical informatics tools (such as RDKit or Open Babel) to calculate molecular descriptors, including structural descriptors, topological descriptors, geometric descriptors, and physicochemical descriptors. Next, convert the descriptors into a standardized descriptor matrix format. Finally, generate a structured data file suitable for machine learning model input, providing comprehensive and detailed basic data for subsequent model training and prediction. Through this step, complex chemical structure information is effectively converted into a numerical form that can be processed by machine learning, improving the availability of data and the predictive performance of the model.

[0123] Step S720, performing feature extraction processing according to the input data, extracting molecular structure features and functional group features by using a convolutional neural network to obtain a feature vector;

[0124] Specifically, structured input data is first imported, including the descriptor matrix of drug ingredients. Then a convolutional neural network is constructed and trained. The convolutional neural network can effectively process high-dimensional data and extract important features through its convolutional layer and pooling layer. The convolutional layer performs local perception and feature extraction on the input data, and can capture local patterns and functional group information in the molecular structure; the pooling layer reduces the data dimension and retains the most significant features through downsampling operations. Through multi-layer convolution and pooling operations, the convolutional neural network can extract feature representations with high information content. Finally, after processing by the fully connected layer, feature vectors are generated, which fully describe the structural characteristics and functional group characteristics of the molecule.

[0125] Step S730, performing feature fusion processing according to the step feature vector, capturing complex feature interaction relationships by performing multi-level nonlinear transformation and fusion on the extracted features, and obtaining a fused feature vector;

[0126] Therefore, through multi-level nonlinear transformation and feature fusion, the quality and depth of feature representation are effectively improved, enabling the model to more accurately understand and utilize the complex relationships of the input data.

[0127] Step S740: Perform activity prediction processing on the fused feature vector based on the activity prediction model, and use the gradient boosting tree to train and predict the fused feature vector, calculate the activity score of each drug component, and obtain the overall activity prediction result.

[0128] Specifically, first, the fused feature vector is trained using a gradient boosting tree. The gradient boosting tree is an ensemble learning method that gradually improves the predictive ability of the model by combining multiple weak learners (usually decision trees). During the training process, the gradient boosting tree gradually adds new trees, and each tree learns and adjusts by minimizing the residual of the previous tree. During the training process, the gradient boosting tree algorithm evaluates the importance of each feature and adjusts it according to its contribution to the prediction accuracy of the model, which helps to improve the accuracy and robustness of the model. After the training is completed, the fused feature vector of the new drug component is input into the trained gradient boosting tree model for prediction. The model predicts the activity score of each drug component by calculating the contribution of each feature to the overall activity score. Finally, all the prediction results are integrated, and the final activity score of each drug component is calculated to obtain the overall activity prediction result. This result can reflect the biological activity of the drug component in the target environment and provide an important basis for subsequent drug screening and optimization.

[0129] Embodiment 5:

[0130] like Figure 4 As shown, this embodiment provides a drug activity screening device, the device comprising:

[0131] Acquisition module 1, used to acquire first information, second information and third information, the first information includes drug-related compounds and their biological activity data, the second information includes the action mechanism data of the compounds at the cellular level, and the third information includes nuclear magnetic resonance spectrum data of the drug to be screened;

[0132] Extraction module 2, used for performing feature extraction processing according to the first information, obtaining a molecular descriptor by extracting global and local structural features of the compound and identifying functional groups related to biological activity;

[0133] Identification module 3, used to perform feature extraction processing according to the second information, and obtain cell mechanism feature data by identifying cell morphological changes, protein distribution and signal pathway activation status under the action of the compound;

[0134] Fusion module 4, used for fusing molecular descriptors and cell mechanism feature data, and obtaining a fusion feature set by combining feature selection and weighted average strategy;

[0135] Building module 5, for modeling according to the fusion feature set, and obtaining an activity prediction model by simulating the interactions of different compounds and compound combinations in the cell environment;

[0136] An analysis module 6 is used to perform spectrum analysis and component identification processing according to the molecular descriptor and the third information to obtain drug component information, where the drug component information includes drug components and their structural information;

[0137] The prediction module 7 is used to input the drug component information into the activity prediction model to perform activity screening processing to obtain the overall activity prediction result.

[0138] It should be noted that, regarding the device in the above embodiment, the specific manner in which each module performs the operation has been described in detail in the embodiment of the method, and will not be elaborated here.

[0139] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for screening drug activity, characterized in that: include: Acquiring first information, second information and third information, wherein the first information includes drug-related compounds and their biological activity data, the second information includes data on the mechanism of action of the compounds at the cellular level, and the third information includes nuclear magnetic resonance spectroscopy data of the drug to be screened; Performing feature extraction processing according to the first information to obtain a molecular descriptor by extracting global and local structural features of the compound and identifying functional groups associated with biological activity; Performing feature extraction processing according to the second information, obtaining cell mechanism feature data by identifying changes in cell morphology, protein distribution and signal pathway activation status under the action of the compound; Performing fusion processing according to the molecular descriptor and the cell mechanism feature data, and constructing a fusion feature set by combining feature selection and weighted average strategy; Performing modeling processing according to the fusion feature set, and obtaining an activity prediction model by simulating the interactions of different compounds and compound combinations in a cell environment; Performing spectrum analysis and component identification processing according to the molecular descriptor and the third information to obtain drug component information, wherein the drug component information includes drug components and their structural information; Inputting the drug component information into the activity prediction model to perform activity screening processing to obtain an overall activity prediction result; The feature extraction process is performed according to the first information to obtain a molecular descriptor by extracting the global and local structural features of the compound and identifying the functional groups related to the biological activity, including: Performing global structural feature extraction processing according to the first information, by converting the molecular structure of the compound into a graph, and using a molecular graph convolutional network to perform a convolution operation on the graph, extracting local information of each node and aggregating it in the entire graph, thereby obtaining a global structural feature representation of the compound; Performing local structural feature extraction processing according to the global structural features, by identifying the interactions between groups and calculating the mutual dependencies between substructures using a self-attention mechanism, thereby obtaining local structural features; Performing functional group identification processing according to the local structural features, identifying specific functional groups related to biological activity by analyzing and matching the local structural features, and comparing the local structural features with the patterns of known active groups to confirm the existence of specific functional groups in the molecule and their relative positions, thereby obtaining functional group features; Feature fusion processing is performed according to the global structural feature representation, the local structural feature and the functional group feature, and the contribution ratio of each feature is adjusted through a weighted average strategy to construct a molecular descriptor set.

2. The method for screening drug activity according to claim 1, characterized in that , performing feature extraction processing according to the second information, obtaining cell mechanism feature data by identifying cell morphological changes, protein distribution and signal pathway activation status under the action of the compound, including: Performing cell morphology change analysis and processing according to the second information, identifying and extracting cell morphology change data to obtain cell morphology characteristics by processing cell image data under the action of the compound; Performing protein distribution feature extraction processing according to the cell morphology feature, identifying and extracting the distribution features of the protein in the cell, including protein concentration, distribution pattern and localization features, by processing the protein staining image data, to obtain the protein distribution features; Performing signal pathway activation feature extraction processing according to the protein distribution feature, analyzing phosphorylated protein data and using a pathway analysis algorithm to identify and extract the activation state of the cell signal pathway under the action of the compound, thereby obtaining the signal pathway activation feature, wherein the signal pathway activation feature includes the activation level of key node proteins and the activation order of the signal pathway; Principal component analysis is performed based on the cell morphology characteristics, the protein distribution characteristics and the pathway activation characteristics to construct cell mechanism characteristic data.

3. The method for screening drug activity according to claim 1, characterized in that , according to the molecular descriptor and the cell mechanism feature data, fusion processing is performed, and a fusion feature set is obtained by combining feature selection and weighted average strategy, including: Performing feature selection processing according to the molecular descriptors and the cell mechanism feature data, recursively training the model by using a recursive feature elimination algorithm and removing the features that contribute the least to the prediction, thereby obtaining an optimized feature set; Performing evaluation processing according to the optimized feature set, and obtaining a weighted feature set by calculating the importance weight of each feature to the model prediction result; Performing feature fusion processing according to the weighted feature set, by jointly training the molecular descriptor and the cell mechanism feature, and combining the correlation and complementarity between the features, to obtain a preliminary fusion feature; A feature dimension reduction process is performed based on the preliminary fusion features, and a final fusion feature set is obtained by reducing the dimension of the feature set while retaining the feature components with the most information.

4. The method for screening drug activity according to claim 1, characterized in that , modeling is performed according to the fusion feature set, and an activity prediction model is obtained by simulating the interactions of different compounds and compound combinations in the cell environment, including: Performing feature conversion processing according to the fused feature set, sorting and combining different compounds by using a multi-layer perceptron, and converting the combined features to obtain a combined feature set; Performing interaction simulation processing according to the combination feature set, and obtaining simulated interaction features by using a generative adversarial network to simulate the mutual promotion or counteraction effects of different compound combinations in a cell environment; Performing optimization processing according to the simulated interaction features, obtaining optimized interaction features by calculating the importance score of each feature and selecting the feature with the highest score; Model training is performed according to the optimized interaction features, and an activity prediction model is obtained by simulating the interaction of different compounds in a cell environment and predicting the overall activity of the drug under different compound compositions using a support vector regression algorithm.

5. The method for screening drug activity according to claim 4, characterized in that ,According to the simulated interaction features, optimization processing is performed, by calculating the importance score of each feature and selecting the feature with the highest score, to obtain the optimized interaction features, including: The calculation formula of the importance score is: Among them, n represents the sample number, N represents the total number of samples, L represents the loss function, and x i represents the feature of the nth sample, y represents the actual value, represents the predicted value, I i Represents feature x i Importance score.

6. The method for screening drug activity according to claim 1, characterized in that , performing spectrum analysis and component identification processing according to the molecular descriptor and the third information to obtain drug component information, including: Performing preprocessing according to the third information, obtaining a preprocessed spectrum by performing baseline correction and noise filtering on the nuclear magnetic resonance spectrum data; Perform peak recognition processing on the preprocessed spectrum, identify characteristic peaks in the spectrum by performing multi-scale analysis on the spectrum using multi-resolution wavelet transform, and mark the chemical shift and peak intensity of each characteristic peak to obtain characteristic peak data; Perform chemical shift matching processing according to the characteristic peak data, and calculate the matching degree of each characteristic peak with known compounds in the chemical shift database by using the Pearson correlation coefficient to obtain a potential compound candidate list; Structural reconstruction is performed according to the potential compound candidate list and the molecular descriptors, and drug component information is obtained by optimizing and confirming the structures of the candidate compounds.

7. The method for screening drug activity according to claim 6, characterized in that , the matching degree between each characteristic peak and the known compounds in the chemical shift database is calculated according to the following Pearson correlation coefficient calculation formula: Where i represents the number of the characteristic peak, j represents the number of the compound, k represents the number of the sample, N represents the total number of samples, and ρ i,j represents the Pearson correlation coefficient between characteristic peak i and compound j in the database, x i,k represents the chemical shift of characteristic peak i in the kth sample, y j,k represents the chemical shift of compound j in the kth sample, represents the average chemical shift of characteristic peak i, represents the average chemical shift of compound j.

8. The method for screening drug activity according to claim 1, characterized in that , inputting the drug component information into the activity prediction model for activity screening to obtain an overall activity prediction result, including: Performing format conversion processing according to the drug component information, and obtaining structured input data by converting the drug components and their structure information into a descriptor matrix; Performing feature extraction processing according to the input data, and extracting molecular structure features and functional group features by using a convolutional neural network to obtain a feature vector; Perform feature fusion processing according to the feature vector in step 1, and capture the complex feature interaction relationship by performing multi-level nonlinear transformation and fusion on the extracted features to obtain a fused feature vector; The fused feature vector is processed for activity prediction based on the activity prediction model, and the fused feature vector is trained and predicted using a gradient boosting tree to calculate the activity score of each drug component to obtain an overall activity prediction result.

9. A drug activity screening device, characterized in that: include: an acquisition module, configured to acquire first information, second information and third information, wherein the first information includes data on drug-related compounds and their biological activities, the second information includes data on the action mechanism of the compounds at the cellular level, and the third information includes nuclear magnetic resonance spectroscopy data of the drug to be screened; An extraction module, used for performing feature extraction processing according to the first information, obtaining a molecular descriptor by extracting global and local structural features of the compound and identifying functional groups related to biological activity; an identification module, configured to perform feature extraction processing according to the second information, and obtain cell mechanism feature data by identifying changes in cell morphology, protein distribution, and activation status of signal pathways under the action of the compound; A fusion module, used for performing fusion processing according to the molecular descriptor and the cell mechanism feature data, and obtaining a fusion feature set by combining feature selection and weighted average strategy; A construction module is used to perform modeling processing according to the fusion feature set, and obtain an activity prediction model by simulating the interactions of different compounds and compound combinations in a cell environment; An analysis module, used for performing spectrum analysis and component identification processing according to the molecular descriptor and the third information to obtain drug component information, wherein the drug component information includes drug components and their structural information; A prediction module, used for inputting the drug component information into the activity prediction model to perform activity screening processing to obtain an overall activity prediction result; Wherein, the extraction module comprises: A first extraction unit is used to perform global structural feature extraction processing according to the first information, by converting the molecular structure of the compound into a graph, and using a molecular graph convolutional network to perform a convolution operation on the graph, extracting local information of each node and aggregating it in the entire graph, so as to obtain a global structural feature representation of the compound; A second extraction unit is used to perform local structural feature extraction processing based on the global structural feature, and obtain local structural features by calculating the mutual dependence between substructures using a self-attention mechanism by identifying the interactions between groups; The first recognition unit is used to perform functional group recognition processing according to the local structural features, identify specific functional groups related to biological activity by analyzing and matching the local structural features, and compare the local structural features with the patterns of known active groups to confirm the existence of specific functional groups in the molecule and their relative positions, thereby obtaining functional group features; The first fusion unit is used to perform feature fusion processing according to the global structural feature representation, the local structural feature and the functional group feature, and to construct a molecular descriptor set by adjusting the contribution ratio of each feature through a weighted average strategy.

Images