Deep learning-based method for screening effective parts of compound mengbrand colon sustained-release tablets

Abstract

The invention relates to the technical field of screening of effective components of traditional Chinese medicine compound preparations, and discloses a deep learning-based method for screening effective parts of compound mengbrand colon sustained-release tablets, which comprises the following steps of: constructing a time sequence pharmacodynamic pathway library, and recording biological pathway activity data of each component of the compound mengbrand colon sustained-release tablets at a plurality of time points; designing a causal intervention traceability network, and analyzing and quantifying the causal relationship of each component to a path node through anti-factual intervention; developing a key node identification algorithm, and identifying key regulation and control nodes; establishing a pathway evolution tracking model, and predicting the long-term influence of the drug components on the pathway; generating a mechanism interpretability score, and evaluating the pharmacodynamic mechanism reliability of each component; effective components with a definite action mechanism are screened out; according to the method, by constructing the time sequence pharmacodynamic pathway library, the complete change process of the biological pathway at multiple time points after drug intervention can be recorded, and the limitation that a traditional static end point observation method cannot track a complete causal chain is overcome.

Description

Technical Field

[0001] This invention relates to the field of screening technology for effective components of traditional Chinese medicine compound preparations, and more specifically, to a deep learning-based method for screening effective components of compound Bupai colonic sustained-release tablets. Background Technology

[0002] Compound Bupai Colonic Sustained-Release Tablets are a traditional Chinese medicine compound preparation containing multiple active ingredients, widely used in the treatment of colon-related diseases. Due to its complex formulation, involving the synergistic effects of multiple drug components, its pharmacodynamic mechanism is characterized by multiple pathways and multiple targets, making it an important subject of current pharmacological research.

[0003] However, existing pharmacological analysis methods primarily rely on static endpoint observations, making it difficult to dynamically capture the temporal changes in biological pathways after drug intervention and to reconstruct the complete causal chain from initial action to final physiological response. This leads to difficulties in accurately identifying and locating key nodes playing crucial roles in compound drug research. Specifically, it manifests as an inability to effectively distinguish between necessary conditions and accompanying phenomena in the process of drug efficacy, resulting in ambiguity in mechanism explanation and insufficient screening criteria.

[0004] Therefore, overcoming the limitations of traditional analytical methods and achieving dynamic tracking and key node screening of the multi-component, multi-pathway mechanism of action of Compound Bupai Colonic Sustained-Release Tablets has become a pressing technical challenge in this field. This is of great significance for improving the scientific research and development level and clinical application value of compound drugs. Summary of the Invention

[0005] This invention provides a deep learning-based method for screening the effective components of compound supplemental colonic sustained-release tablets, which solves the technical problems in related technologies such as the inability to dynamically track the temporal changes of biological pathways after drug intervention, the difficulty in distinguishing causal relationships from accompanying phenomena, and the lack of objective quantitative screening standards.

[0006] This invention provides a deep learning-based method for screening the effective components of compound supplement colonic sustained-release tablets, including: Construct a time-series pharmacodynamic pathway library to record the biological pathway activity data of each component of Compound Bupai Colonic Sustained-Release Tablets at multiple time points; Based on the data output from the time-series pharmacodynamic pathway library, a causal intervention tracing network was designed, and the causal relationship between each component and pathway nodes was quantified through counterfactual intervention analysis. Using the results of causal relationship analysis, a key node identification algorithm was developed to identify key control nodes; Based on the identified key regulatory nodes, a pathway evolution tracking model is established to predict the long-term effects of drug components on pathways. Based on the prediction results of the pathway evolution tracking model, a mechanism interpretability score is generated to evaluate the reliability of the pharmacodynamic mechanism of each component. Based on the importance ranking of the causal chain, effective components with clear mechanisms of action were screened out.

[0007] Furthermore, the time-series pharmacodynamic pathway library has time-resolved structure, multi-level organization, intervention response mapping and data quality control features. By recording the activity value sequence of each pathway node at multiple time points, a hierarchical association mapping function is constructed, the intervention response function of drug components to pathway nodes is recorded, and an outlier detection function is introduced for data quality control.

[0008] Furthermore, the key node identification algorithm includes the following steps: calculating the centrality index set of each node, including betweenness centrality, proximity centrality, degree centrality, and eigenvector centrality; applying the node necessity scoring function to calculate the comprehensive score; defining the network function influence degree to quantify the impact of node removal on network function; calculating the time sensitivity to quantify the rate of change of importance of a node within a time window; and generating a key node set by combining the node necessity score, network function influence degree, and time sensitivity.

[0009] Furthermore, the pathway evolution tracking model includes a temporal coding unit, a state transition unit, a long-range dependency unit, and an association discovery unit. The temporal coding unit maps pathway state data at different time points into a hidden state sequence. The state transition unit calculates the pathway state transition probability. The long-range dependency unit captures pathway state changes over long time scales. The association discovery unit calculates the association strength between drug components and pathway nodes.

[0010] Furthermore, the interpretability score of the mechanism includes a causal chain integrity index, an evidence support index, and a temporal consistency index. The causal chain integrity index quantifies the proportion of identified pathway nodes to complete pathway nodes in the mechanism. The evidence support index quantifies the similarity between the mechanism and known literature evidence. The temporal consistency index quantifies the rationality of the order of occurrence of temporal events.

[0011] Furthermore, the screening of effective components with clear mechanisms of action includes the following steps: calculating the regulatory intensity of the component on the set of key nodes; evaluating the regulatory effect of the component on the necessary pathways; analyzing the synergy coefficient of the component combination; calculating and ranking the comprehensive score; and selecting the top-ranked components. The components are effective ingredients with a clear mechanism of action, among which... This represents the number of effective ingredients ultimately selected.

[0012] Furthermore, the counterfactual intervention index (CFI) in the causal intervention tracing network is determined by calculating the L2 norm of the difference between the original network and the network state after removing nodes, and the network state mapping function calculates the information transmission characteristics of each node through graph convolution operation.

[0013] Furthermore, the interpretability score is calculated by weighted combination of causal chain integrity index, evidence support index and temporal consistency index. Before calculation, each index is subjected to Min-Max normalization to ensure that different types of indexes can be reasonably weighted.

[0014] Furthermore, the comprehensive score is calculated by weighted combination of regulatory intensity, pathway regulatory effect and component synergy coefficient. Before calculation, each scoring indicator is standardized by Z-score to ensure that different scoring indicators are weighted and combined under the same dimension.

[0015] The present invention provides a computer storage medium, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the above-mentioned deep learning-based method for screening the effective components of compound supplement colonic sustained-release tablets.

[0016] The beneficial effects of this invention are as follows: by constructing a time-series pharmacodynamic pathway library, this invention can record the complete changes of biological pathways at multiple time points after drug intervention, overcoming the limitation of traditional static endpoint observation methods that cannot track complete causal chains. This dynamic monitoring mechanism provides continuous data support for understanding the entire process of drug efficacy from intervention to outcome, and improves the completeness of mechanism analysis.

[0017] The causal intervention tracing network of this invention adopts the counterfactual analysis method. By simulating the differences in network state before and after intervention at different nodes, it can effectively distinguish the necessary conditions and accompanying phenomena in the drug efficacy pathway. This deep learning-based causal inference mechanism solves the technical problem that traditional methods cannot determine causal relationships and provides a scientific basis for the accurate explanation of drug efficacy mechanisms.

[0018] The key node identification algorithm developed in this invention can accurately identify key sites affecting drug efficacy by calculating the weights of necessary nodes and branch points in the pathway network and combining network perturbation analysis. This structured node importance assessment method provides an objective and quantitative standard for screening the effective ingredients of Compound Bupai Colonic Sustained-Release Tablets, overcoming the subjectivity problem of traditional screening methods that rely on experience-based judgment.

[0019] This invention establishes a mechanism interpretability scoring system. By quantifying the causal chain integrity and reliability of the pharmacodynamic mechanism of each component, it provides an objective standard for the scientific evaluation of the mechanism of action of compound drugs. This comprehensive scoring mechanism based on evidence support, temporal consistency and pathway coverage solves the problem of the lack of quantitative evaluation standards in traditional mechanism analysis methods and improves the reliability of mechanism explanation.

[0020] Effective ingredient screening methods based on causal chain importance ranking can accurately screen drug components with clear regulatory effects at key pathway nodes under different pathological states, providing a scientific basis for personalized drug development. This precise screening method based on mechanism and key nodes overcomes the shortcomings of traditional drug screening that rely too much on endpoint effects and ignore the mechanism of action, thus improving the targeting and effectiveness of drug development. Attached Figure Description

[0021] Figure 1 This is a flowchart of the deep learning-based method for screening the effective components of compound supplement colonic sustained-release tablets in this invention; Figure 2 This is a bar chart comparing the performance of different network analysis methods in distinguishing between causal relationships and accompanying phenomena. Figure 3 This is a grouped bar chart comparing the prediction accuracy of pathway evolution tracking models; Figure 4 This is a diagram showing the network of pathways through which drug components affect the key mechanisms of IBS. Figure 5 This is a composite chart comparing the effects of screening methods on the treatment of irritable bowel syndrome. Detailed Implementation

[0022] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0023] At least one embodiment of the present invention discloses a deep learning-based method for screening the effective components of compound supplement colonic sustained-release tablets, such as... Figure 1 As shown, it includes: Step 1: Construct a time-series pharmacodynamic pathway library and record the biological pathway activity data of each component of Compound Bupai Colonic Sustained-Release Tablets at multiple time points; This study collects pharmacodynamic data on the effects of various components in Compound Bupai Colonic Extended-Release Tablets. Inputs include raw gene expression profiles, proteomics data, and metabolomics data; outputs a standardized multi-omics dataset. A multi-timepoint sampling mechanism is established, with inputs being preset timepoint parameters and outputs including sampling schemes for 0 hours, 2 hours, 6 hours, 12 hours, 24 hours, 48 ​​hours, and 72 hours. Deep learning time-series analysis is applied, with inputs being multi-timepoint biological pathway activity data and outputs a dynamic database of pathway activity changes over time. Data standardization is performed, with inputs being raw time-series data and outputs a standardized dataset after eliminating batch effects and noise interference.

[0024] The time points and time windows are uniformly scaled to eliminate differences in time scales between different experiments; baseline normalization is performed on the activity sequences of each pathway node to improve cross-node comparability; categorical attributes such as component name, pathway name, node name and type are encoded as numerical features for model use; missing activity values ​​are interpolated or appropriately filled with time series data, and replacement or truncation is completed after outlier detection.

[0025] Time-series pharmacodynamic pathway libraries have the following characteristics: Time-resolved structure: for each path node Record its activity value sequence at the set of time points: ; in A set of time points, , , They represent the 1st, 2nd, and 3rd respectively. At a certain point in time, This represents the number of time points.

[0026] node At all points in time The activity value sequence is as follows: ; in For nodes At all points in time The activity value sequence, , , For nodes At the point of time , , The activity value, For path nodes, This represents the number of time points.

[0027] Multi-level organization: Based on the tissue pathway data at the molecular, cellular, tissue, and systemic levels, a hierarchical association mapping function is constructed: ; in This is a hierarchical association mapping function. Indicates the first Hierarchical data, For the first Hierarchical data. Molecular layer, cellular layer, tissue layer, and system layer represent different hierarchical structures in biology.

[0028] Intervention response mapping: Recording drug component sets: ; in It is a collection of drug components. , , They represent the 1st, 2nd, and 3rd respectively. One drug component, This refers to the quantity of components.

[0029] For path nodes Intervention response function , Let be the intervention response function, representing the components. In time For nodes The intensity of the impact; For path nodes, It is a drug ingredient. For a point in time.

[0030] The specific implementation process of the intervention response function is as follows: First, it is necessary to calculate the change in the activity value of the drug component at a specific time point on the pathway node. Then, this change is processed using a commonly used normalization method, limiting the result to the range of -1 to 1. A positive value after normalization represents that the component has an activating effect on the node, while a negative value represents an inhibitory effect; the larger the absolute value, the stronger the regulation. Finally, a time-decreasing factor is incorporated to adjust the response strength, reflecting the characteristic that the drug effect gradually weakens over time.

[0031] In terms of data quality control, an outlier detection mechanism is employed. Specifically, for each node at each time point, the deviation of its activity value from the historical average activity value of that node is calculated. If a data point deviates from the historical mean by more than three standard deviations, it is marked as an outlier and processed accordingly. This process helps to eliminate or correct outlier data, ensuring the accuracy and reliability of subsequent analyses.

[0032] Step 2: Based on the data output from the time-series pharmacodynamic pathway library, design a causal intervention tracing network, and quantify the causal relationship between each component and pathway nodes through counterfactual intervention analysis; A causal inference framework based on graph neural networks is constructed. The input is gene or protein data in biological pathways and their regulatory relationships, and the output is the causal relationship strength score between nodes.

[0033] A counterfactual intervention mechanism is introduced; the intervention effect is quantified by simulating the difference in network state before and after intervention at a specific node; and the network state vector is normalized using the L2 norm to ensure that comparisons between different states are comparable.

[0034] The system employs a temporal attention mechanism; the input is pathway activity data at multiple time points, and the output is the causal association strength between different time points; before processing the temporal data, the activity data at each time point is normalized by a time window to eliminate the influence of time scale differences.

[0035] like Figure 2 As shown in the figure, the technical advantages of the causal intervention tracing network designed in this step compared with traditional methods are quantitatively demonstrated by the bar chart, which verifies the core innovative value of the counterfactual intervention mechanism and the temporal attention mechanism in distinguishing causal relationships from accompanying phenomena.

[0036] Before performing temporal attention calculations, the features of each time slice need to be scaled uniformly to ensure comparability between features from different time slices. Then, when calculating the attention weights between time points, the sum of all weights must be 1, meaning the attention distribution across all time points must be normalized. This ensures a reasonable distribution of time points the model focuses on. Simultaneously, for categorical data features such as components, nodes, pathway types, and regulatory directions, they need to be converted into standardized numerical codes so that these features can be directly recognized and utilized by the model. For continuous features related to network state, standardization methods should be used to improve the stability and convergence speed of the model training process.

[0037] Establish a multi-level causal inference module to gradually trace the mechanism of drug action.

[0038] The causal intervention tracing network includes the following functional layers: Input layer, which receives feature vectors of biological pathway nodes: ; in The node feature matrix, For the number of nodes, For feature dimension, express OK A column of real numbers.

[0039] Counterfactual intervention layer, calculate the counterfactual intervention index: ; in For the first Nodes The counterfactual intervention index For the network The state mapping vector, For the original network, To remove the first Nodes The network after that, It is the vector norm.

[0040] network State mapping to This refers to the network topology. This is transformed into the liveness state vector of the nodes. In the specific implementation, the information transmission features of each node are first calculated using graph convolution operations. This process includes the following key steps: first, using the adjacency matrix... Sum-degree matrix Model the connections between nodes and the degree of each node, and then model the degree matrix. Perform inverse square root processing to achieve normalization, then process the node feature matrix. With trainable weight matrix Perform a linear transformation, and finally pass through the activation function. The results are then subjected to a non-linear mapping. The overall process is as follows: first, the node features are normalized; then, they are combined with network structure information; and finally, the state feature vector of each node is output through an activation function.

[0041] Temporal aggregation layer, based on attention weights At different time points and Establish temporal dependencies and define temporal causality strength: ; in For time points and Temporal causal strength between them For time points and Attention weights between them For nodes The counterfactual intervention index For summation notation. Attention weights. The implementation process is as follows: First, calculate the similarity between any two time points, using methods such as cosine similarity. Then, combine this similarity with a decay factor reflecting the influence of temporal distance; this factor gradually decreases as the interval between the two time points increases. Finally, normalize all calculated weights so that their sum equals 1; a common method is softmax normalization. The larger the normalized weight value, the stronger the temporal dependency between the two time points.

[0042] Output layer: Generates node importance ranking and causal path graph.

[0043] Step 3: Utilize the causal relationship analysis results to develop a key node identification algorithm to identify key control nodes; The algorithm employs a network centrality calculation method, taking network structure data as input and node centrality indices as output. It also applies a node necessity scoring function, taking multiple centrality indices as input and a comprehensive node importance score as output. Furthermore, it performs network perturbation analysis, taking network structure as input and a quantitative indicator of network function changes after node removal as output. Finally, it integrates time-series change data to identify key control nodes at different time stages.

[0044] The key node identification algorithm includes the following specific steps: calculating each node The set of centrality indices, including betweenness centrality. Proximity centrality Degree centrality and eigenvector centrality ; The application node necessity scoring function calculates the comprehensive score: ; in For nodes Necessity rating , , , These represent the weight coefficients for betweenness centrality, proximity centrality, degree centrality, and eigenvector centrality, respectively. The optimal values ​​are determined through cross-validation. For nodes betweenness centrality, For nodes Proximity centrality For nodes Degree centrality, For nodes eigenvector centrality.

[0045] Before applying this function, each centrality metric needs to be Min-Max normalized. Specifically, for each centrality metric, first find its minimum and maximum values ​​across all nodes. Then, subtract the minimum value from the original centrality value of each node, and divide by the difference between the maximum and minimum values. This scales the range of all centrality metric values ​​to between 0 and 1. This process ensures comparability between different centrality metrics, which is beneficial for subsequent comprehensive score calculations.

[0046] Define the network function impact: ;in For nodes The impact of network functionality To remove a node Post-network functional evaluation value, For vector norm, This is a network function evaluation function, calculated by the network. The key topological feature combination is used to quantify the functional integrity of the network, defined as: ; in The network connectivity coefficient. The average shortest path length, For network information entropy, , , These represent the weighting coefficients for connectivity, average shortest path length, and information entropy, respectively.

[0047] In calculating network function evaluation values Previously, it was necessary to first determine the network connectivity coefficients. Average shortest path length and network information entropy Standardization is performed separately for each indicator. Specifically, for each indicator, the mean and standard deviation of that indicator across all nodes are first calculated. Then, the original indicator value for each node is subtracted from the mean and divided by the standard deviation, thus transforming it into a standardized value. This eliminates differences in dimensions and scales between different indicators, making them comparable in subsequent comprehensive calculations.

[0048] Before integrating centrality and network function metrics, data from different sources are scaled and missing values ​​are filled appropriately. Before calculating time sensitivity, the time window length is scaled to avoid bias caused by time scale differences. Normalization strategies are used for the hyperparameters used for weighting to avoid bias caused by inconsistent dimensions.

[0049] Network connectivity coefficient Specifically, this is achieved by calculating the ratio of the actual number of edges in the network to the theoretically maximum number of edges, reflecting the network's connection density and overall connectivity level; the average shortest path length... Specifically, it is implemented by calculating the average of the shortest path lengths between any two nodes in the network, reflecting the efficiency of information transmission in the network; network information entropy The specific implementation involves calculating Shannon entropy based on node degree distribution to quantify the complexity and uncertainty of the network structure; and calculating time sensitivity. ; in For nodes In the time window Rate of change of importance within For nodes The change in necessity ratings over the time window. The time window length, This is for absolute value operations.

[0050] comprehensive , and Generate a set of key nodes: ; in For the set of key nodes, For nodes Node necessity score, For nodes The impact of network functionality For nodes The time sensitivity, For the preset threshold, This is a set filtering condition.

[0051] Key Node Set Each node All meet To enhance robustness across datasets, a threshold is set. It can be adaptively set according to the data distribution (e.g., based on quantiles or validation set performance) to avoid the inapplicability of fixed thresholds on different networks.

[0052] Step 4: Based on the identified key regulatory nodes, establish a pathway evolution tracking model to predict the long-term effects of drug components on the pathway. Applying recurrent neural networks for time series prediction, with the input being... arrive Pathway activity data for different time periods, output as Predicted pathway activity at specific time points; integrate drug component feature vectors with pathway state vectors; predict long-term intervention effects; construct a pathway state transition matrix, with drug components and time points as inputs. The path status is output as a time point. The pathway state transition probability is calculated; an attention mechanism is applied, with drug component features and pathway node features as inputs and the correlation strength between them as outputs.

[0053] like Figure 3 As shown, a grouped bar chart is used to comprehensively evaluate the performance of the pathway evolution tracking model established in this step. By standardizing the indicators of different magnitudes, the indicators of different magnitudes are converted into dimensionless relative numbers, which intuitively demonstrates the predictive accuracy of the model on multiple pathway types and time scales, and verifies the technical advantages of recurrent neural networks and attention mechanisms in predicting long-term intervention effects.

[0054] The pathway state and component features are scaled to a consistent range; masking and interpolation strategies are used for missing time points to maintain sequence integrity; the input sequence is standardized to improve prediction stability; and the classification attributes of components, nodes, and pathways are encoded with the same encoding vectors as those in the previous steps.

[0055] The pathway evolution tracing model comprises the following components: Timing coding unit: This unit encodes different time points. Path status data Mapped to a sequence of hidden states ;in , They represent the 1st and the 2nd respectively. At a certain point in time, For the first The path status at each point in time. , They represent the 1st and the 2nd respectively. Hidden states at each point in time For the feature dimension of the pathway state, Represents the set of real numbers. This represents the number of time points.

[0056] State transition unit: calculates the transition matrix ,in This indicates the transition relationship of the pathway state between adjacent time points. Feature dimensions representing pathway status Represents the set of real numbers; through the drug component feature matrix Perform conditional modulation, where A feature matrix representing all drug components, For the number of components, As a component feature dimension, It represents the set of real numbers.

[0057] Long-range dependency units: By introducing forget and remember gate mechanisms, changes in pathway states over long time scales are captured. The forget gate determines which information from the previous hidden state needs to be retained or discarded, while the remember gate decides which new information from the current pathway state needs to be written into the memory unit. The parameters of these gating mechanisms (including the weight matrix and bias terms) can be automatically learned through model training, and the activation function is typically the sigmoid function. In this way, the model can effectively integrate historical information and current input to model long-term dependencies in pathway states.

[0058] Association discovery unit: Calculates drug components With path nodes Association strength: ; in For ingredients With nodes The strength of the association, For the first The feature vector of each component For the first Feature vectors of each path node For a trainable correlation weight matrix, For normalization function, This represents the transpose symbol.

[0059] Step 5: Combine the prediction results of the pathway evolution tracking model to generate a mechanism interpretability score and evaluate the reliability of the pharmacodynamic mechanism of each component. The system calculates the causal chain integrity index, with the input being the predicted pathway activity change path and the output being the quantified value of pathway coverage from the initial intervention to the endpoint effect; it calculates the evidence support index, with the input being the predicted mechanism and known pharmacological research literature data, and the output being the mechanism reliability score; it calculates the temporal consistency index, with the input being the predicted action time sequence and biological law data, and the output being the temporal logical rationality score; and it comprehensively calculates the interpretability score to quantify the reliability of the mechanism of action of each component.

[0060] The mechanism interpretability score includes a causal chain integrity index, an evidence support index, and a temporal consistency index, where the causal chain integrity index is expressed as: ; in For mechanism Causal chain integrity index For mechanism The set of path nodes already identified in the data. A complete set of path nodes; This is for absolute value operations.

[0061] Path node set function The specific implementation is as follows: through the analysis mechanism The biological pathways involved are identified, and all gene, protein, or metabolite nodes involved in the mechanism are extracted to form a pathway node set for assessing the completeness of the biological processes covered by the mechanism. Indicators of Evidence Support: ; in For mechanism Evidence support index It is a set of known documentary evidence. For a single piece of evidence in the evidence set, As evidence The weighting coefficients, For mechanism With evidence similarity, For absolute value operations, This is the summation symbol.

[0062] Similarity function The specific implementation is as follows: First, the mechanism and evidence The features are converted into feature vectors, including the set of pathway nodes, the direction of regulation, and the time series pattern. Then, the cosine similarity between the two feature vectors is calculated. Finally, the similarity value is mapped to the interval between 0 and 1 using the sigmoid function. The higher the value, the stronger the consistency between the mechanism and the evidence. Time series consistency metrics: ; in For mechanism The time-series consistency index and They represent the first The and the first The time when the event occurred For indicator functions, From The number of combinations of any two events selected from a given set of events. This is the summation symbol.

[0063] Indicator Function The specific implementation is as follows: when the first The time of occurrence of each event Later than the event Time of occurrence The function returns 1 if the condition is met, otherwise it returns 0. It is used to detect whether there is a violation of biological logic in the temporal order of events, such as a subsequent event occurring before a preceding event. Overall interpretability score: ; in For mechanism The overall interpretability score, , , These represent the weighting coefficients for the three indicators: causal chain integrity, evidentiary support, and temporal consistency. For mechanism Causal chain integrity index For mechanism Evidence support index For mechanism The time-series consistency index.

[0064] In calculation Before, to , and Perform Min-Max normalization to ensure that different types of indicators can be reasonably weighted.

[0065] Before weighting and combining the various indicators, the weight coefficients are normalized or calibrated through cross-validation to avoid a single indicator dominating the overall score; the classification attributes such as the type of evidence source and the level of evidence are coded to reasonably set or learn the evidence weights.

[0066] Step 6: Based on the importance ranking of the causal chain, screen out the effective components with clear mechanisms of action; Calculate the regulatory intensity of each component on key nodes; the input is component characteristics and key node characteristics, and the output is the regulatory intensity score; evaluate the activation or inhibition effect of components on essential pathways; the input is component intervention data and pathway activity change data, and the output is the pathway regulation effect score; analyze the synergistic and antagonistic effects of component combinations; the input is multi-component combination intervention data, and the output is the combination effect score; based on the comprehensive score ranking, screen effective components with clear mechanisms of action and key node regulation capabilities.

[0067] The method for screening effective ingredients based on causal chain importance ranking includes the following calculation steps: Calculated components For the set of key nodes Regulation intensity: ; in For ingredients For the set of key nodes The intensity of regulation, For the set of key nodes, For set A single node in For ingredients With nodes The strength of the association, For nodes Necessity rating For absolute value operations, This is the summation symbol.

[0068] Correlation strength function The specific implementation is as follows: by calculating drug components For path nodes The magnitude of the regulatory effect was determined by analyzing the influence of this component on the node at different time points using time-series data. Finally, a correlation strength value between 0 and 1 was obtained through normalization, with higher values ​​indicating a stronger regulatory effect of the component on the node. The component was evaluated. For essential routes The regulatory effect: ; in For ingredients For essential routes The regulatory effect, A set of necessary pathways, For set A single pathway in To regulate direction ( Indicates activation. Indicates inhibition. (This indicates no significant regulation) For pathway The change in activity This is the summation symbol.

[0069] Regulation direction function The specific implementation is as follows: by analyzing drug components For pathways The function tracks the trend of activity value changes. When the activity value increases over time, it returns 1 to indicate activation; when the activity value decreases over time, it returns -1 to indicate inhibition; and when the activity value does not change significantly, it returns 0 to indicate no significant regulatory effect. The effectiveness of the regulatory direction is determined through statistical significance testing. Pathway activity change function The specific implementation involves calculating pathways before and after drug intervention. The difference in activity values ​​was analyzed by comparing the activity levels of the intervention time series and the control time series, ultimately yielding the magnitude of activity changes in this pathway under drug action. Positive values ​​indicated enhanced activity, while negative values ​​indicated inhibitory activity. The composition of the components was also analyzed. Coefficient of cooperation: ; in For ingredients and The coefficient of synergy, , The components are respectively , The individual effect, For ingredients and The combined effect and They represent the first The and the first One component.

[0070] Individual effect function The specific implementation is as follows: by calculating drug components The sum of the regulatory intensity of all key pathway nodes, combined with the mechanism interpretability score of the component, is finally weighted and averaged to obtain the individual pharmacodynamic score of the component, with a value ranging from 0 to 1, reflecting the independent contribution of the component in the compound preparation. Combined effect function The specific implementation is as follows: the efficacy of two components acting simultaneously is predicted by experimental measurement or model, the synergistic enhancement or antagonistic inhibition between components is considered, and finally the combined efficacy score is obtained to evaluate the rationality of the component combination. Calculate the overall score: ; in For ingredients Overall score , , These represent the weighting coefficients of the regulation intensity score, pathway regulation effect score, and synergy coefficient score, respectively. For ingredients Other than Components The coefficient of synergy, For the first Components Importance weights This is the summation symbol.

[0071] In calculation Before, to , and conduct Score standardization ensures that different scoring indicators are weighted and combined under the same unit of measurement. Before calculating the composite score, the coefficients are normalized or tuned based on the validation set to maintain a balance in the contributions of different scoring items; a unified coding method is used for the classification attributes of components to ensure alignment with the features of preceding modules; when estimating the combined effect, a unified scaling correction is performed on data from different experimental batches and measurement sources to improve cross-experiment comparability. Sort all ingredients in descending order and select the top-ranked ones. The components are effective ingredients with a clear mechanism of action.

[0072] like Figure 4 As shown, a relational graph is used to visualize the results of the screening of effective ingredients based on the importance ranking of causal chains in this step. The network topology structure intuitively presents the regulatory relationship between drug components and key pathway nodes, verifying the application value of the causal intervention traceability network in accurately identifying the mechanism of action of effective ingredients.

[0073] like Figure 5 As shown, the clinical application effect of the screening method in this step is quantitatively verified by combining the graphs. The differences between the traditional screening method and the patented method in multiple treatment indicators are compared, which proves the technical advantages and practical value of the effective component screening method based on the importance ranking of causal chains in personalized precision treatment.

[0074] A computer storage medium includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the above-described deep learning-based method for screening the effective components of compound supplement colonic sustained-release tablets.

[0075] Here, the present invention provides an implementation example: A pharmaceutical research institution needed to screen the most effective active ingredient in Compound Bupai Colonic Extended-Release Tablets for the treatment of irritable bowel syndrome (IBS). This compound formulation contains eight main components (curcumin, atractylodes polysaccharide, paeoniflorin, quercetin, naringenin, gypenosides, emodin, and aloe-emodin), but the contribution of each component to the improvement of IBS symptoms and their mechanisms of action remained unclear. The research team applied the method described in this patented paper, using deep learning technology to construct a causal intervention traceability network, and systematically analyzed the regulatory effects of each component on gut-related pathways, providing a scientific basis for the precise screening of effective ingredients.

[0076] The research team collected multi-omics data on the effects of various components of Compound Bupai Colonic Sustained-Release Tablets, and established a sampling scheme including seven time points: 0 hours, 2 hours, 6 hours, 12 hours, 24 hours, 48 ​​hours, and 72 hours. They then applied deep learning time-series analysis methods to construct a dynamic database of pathway activity changes over time. The time-series pharmacodynamic pathway library possesses four main characteristics: time-resolved structure, multi-level organization, intervention-response mapping, and data quality control. Data quality control was achieved through a Z-score standardization function, successfully identifying and handling outlier data points exceeding three standard deviations. Examples of some basic data from the time-series pharmacodynamic pathway library are shown in Table 1. Table 1: Example of basic data for time-series pharmacodynamic pathway libraries (partial) The research team applied counterfactual intervention analysis, using graph neural networks to model the impact of different intervention points on the activity of biological pathways, and utilized temporal attention mechanisms to identify causal relationships between drug components and pathway nodes. The implementation comprises four functional layers: an input layer receives feature vectors of biological pathway nodes; a counterfactual intervention layer calculates the degree of intervention impact at each node, extracting node information transmission features through graph convolution operations; a temporal aggregation layer constructs temporal dependencies between nodes based on attention mechanisms, measuring the causal strength between different time points; and an output layer generates a ranking of node importance and a causal path graph. The research team compared the differences between traditional static network analysis and causal intervention tracing networks in distinguishing causal relationships from accompanying phenomena. Table 2 shows a comparison of different network analysis methods in distinguishing causal relationships from accompanying phenomena. Table 2: Comparison of different network analysis methods in distinguishing causal relationships from concomitant phenomena By calculating various centrality indices (such as betweenness centrality, proximity centrality, degree centrality, and eigenvector centrality) for each pathway node, the research team comprehensively scored each node. Simultaneously, combining indicators such as network functional influence and temporal sensitivity, they ultimately selected a set of key nodes based on these three indices, successfully identifying 15 key regulatory nodes in irritable bowel syndrome (IBS)-related pathways. The identification results of the first five key nodes in IBS-related pathways are shown in Table 3. Table 3: Identification results of key nodes in irritable bowel syndrome-related pathways (top-5) The research team constructed a pathway evolution tracking model using recurrent neural networks. This model comprises four core units: a temporal encoding unit that maps pathway state data at different time points into a sequence of hidden states; a state transition unit that conditionally modulates pathway states based on drug component characteristics to achieve dynamic state transitions; a long-range dependency unit that captures long-term changes in pathway states by introducing forget and remember gates; and an association discovery unit that calculates the association strength between drug components and pathway nodes. Based on the pathway evolution tracking model, the research team was able to successfully predict activity values ​​at subsequent time points using pathway activity data from the previous few time periods, achieving an average prediction accuracy of 91.7%. The prediction accuracy evaluation results of the pathway evolution tracking model are shown in Table 4. Table 4: Prediction Accuracy Evaluation of Pathway Evolution Tracking Model The research team calculated three core indicators to assess the scientific reliability of the drug component's mechanism of action: causal chain integrity, used to assess the coverage of identified pathway nodes in the mechanism; evidence support, used to assess the consistency between the mechanism and existing literature evidence; and temporal consistency, used to assess the rationality of the order of events in the mechanism. After Min-Max normalization, the three indicators were weighted to obtain a comprehensive interpretability score, providing an objective basis for subsequent component screening. The prediction accuracy assessment of the pathway evolution tracking model is shown in Table 5. Table 5: Prediction accuracy evaluation of the pathway evolution tracking model The research team calculated the regulatory intensity of each component on key nodes, assessed the activation or inhibition effects of components on essential pathways, and analyzed the synergistic and antagonistic effects of component combinations. Finally, based on a comprehensive score ranking, the most effective drug component combinations were selected. The regulatory intensity and mechanism scores of drug components on key pathway nodes are shown in Table 6. Table 6: Regulatory strength and mechanism score of drug components on key pathway nodes The synergistic effect scores for different component combinations are shown in Table 7: Table 7: Synergistic effect scores of different component combinations This application example focuses on verifying the following two core technical effects: by comparing the accuracy of traditional correlation analysis, static network analysis and the causal intervention tracing network of this patent in analyzing the mechanism of action of Compound Bupai Colonic Sustained-Release Tablets, it verifies the advantages of this method in distinguishing the necessary conditions and accompanying phenomena in the pharmacodynamic pathway.

[0077] The synergistic effect scores for different component combinations are shown in Table 8: Table 8: Comparison of three methods for analyzing the mechanism of action of curcumin This method, based on the importance ranking of causal chains, can accurately screen drug component combinations with clear regulatory effects targeting key pathway nodes in different pathological states. The study compared the screening results based on traditional pharmacodynamic evaluation with those based on this patented method.

[0078] Table 9 shows a comparison of the effectiveness of different screening methods in personalized treatment of irritable bowel syndrome. Table 9: Comparison of the effectiveness of different screening methods for personalized treatment of irritable bowel syndrome The combination of curcumin, paeoniflorin, and quercetin, screened using this patented method, demonstrated superior clinical efficacy compared to traditional screening methods in the treatment of irritable bowel syndrome (IBS), validating the technical advantages of this method in personalized and precise screening. This combination achieved precise intervention targeting key nodes such as the IL-6 signaling pathway, TNF-α receptor, and COX-2 expression regulation, effectively improving patient symptoms and reducing the incidence of adverse reactions.

[0079] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A deep learning-based method for screening the effective components of compound supplemental colonic sustained-release tablets, characterized in that, include: Construct a time-series pharmacodynamic pathway library to record the biological pathway activity data of each component of Compound Bupai Colonic Sustained-Release Tablets at multiple time points; Based on the data output from the time-series pharmacodynamic pathway library, a causal intervention tracing network was designed, and the causal relationship between each component and pathway nodes was quantified through counterfactual intervention analysis. Using the results of causal relationship analysis, a key node identification algorithm was developed to identify key control nodes; Based on the identified key regulatory nodes, a pathway evolution tracking model is established to predict the long-term effects of drug components on pathways. Based on the prediction results of the pathway evolution tracking model, a mechanism interpretability score is generated to evaluate the reliability of the pharmacodynamic mechanism of each component. Based on the importance ranking of the causal chain, effective components with clear mechanisms of action were screened out.

2. The method for screening the effective components of compound supplementary colonic sustained-release tablets based on deep learning according to claim 1, characterized in that, The time-series pharmacodynamic pathway library features a time-resolved structure, multi-level organization, intervention-response mapping, and data quality control. It constructs a hierarchical association mapping function by recording the activity value sequence of each pathway node at multiple time points, records the intervention response function of drug components to pathway nodes, and introduces an outlier detection function for data quality control.

3. The method for screening the effective components of compound supplementary colonic sustained-release tablets based on deep learning according to claim 1, characterized in that, The key node identification algorithm includes the following steps: calculating the centrality index set of each node, including betweenness centrality, proximity centrality, degree centrality, and eigenvector centrality; applying the node necessity scoring function to calculate the comprehensive score; defining the network function influence degree to quantify the impact of node removal on network function; calculating the time sensitivity to quantify the rate of change of importance of a node within a time window; and generating a key node set by combining the node necessity score, network function influence degree, and time sensitivity.

4. The method for screening the effective components of compound supplementary colonic sustained-release tablets based on deep learning according to claim 1, characterized in that, The pathway evolution tracking model includes a temporal coding unit, a state transition unit, a long-range dependency unit, and an association discovery unit. The temporal coding unit maps pathway state data at different time points into a hidden state sequence. The state transition unit calculates the pathway state transition probability. The long-range dependency unit captures pathway state changes over long time scales. The association discovery unit calculates the association strength between drug components and pathway nodes.

5. The method for screening the effective components of compound supplementary colonic sustained-release tablets based on deep learning according to claim 1, characterized in that, The interpretability score of the mechanism includes a causal chain integrity index, an evidence support index, and a temporal consistency index. The causal chain integrity index quantifies the proportion of identified pathway nodes to complete pathway nodes in the mechanism. The evidence support index quantifies the similarity between the mechanism and known literature evidence. The temporal consistency index quantifies the rationality of the order of occurrence of temporal events.

6. The method for screening the effective components of compound supplementary colonic sustained-release tablets based on deep learning according to claim 1, characterized in that, The process of screening effective components with clear mechanisms of action includes the following steps: calculating the regulatory intensity of the component on the set of key nodes; evaluating the regulatory effect of the component on the essential pathway; analyzing the synergistic coefficient of the component combination; calculating and ranking the comprehensive score; and selecting the top-ranked components. The components are effective ingredients with a clear mechanism of action, among which... This represents the number of effective ingredients ultimately selected.

7. The method for screening the effective components of compound supplementary colonic sustained-release tablets based on deep learning according to claim 1, characterized in that, The counterfactual intervention index (CFI) in the causal intervention tracing network is determined by calculating the L2 norm of the difference between the original network and the network state after removing nodes. The network state mapping function calculates the information transmission characteristics of each node through graph convolution operation.

8. The method for screening the effective components of compound supplementary colonic sustained-release tablets based on deep learning according to claim 5, characterized in that, The interpretability score is calculated by weighted combination of causal chain integrity index, evidence support index and temporal consistency index. Before calculation, each index is normalized by Min-Max to ensure that different types of indexes can be reasonably weighted.

9. The method for screening the effective components of compound supplementary colonic sustained-release tablets based on deep learning according to claim 6, characterized in that, The comprehensive score is calculated by weighted combination of regulatory intensity, pathway regulation effect and component synergy coefficient. Before calculation, each scoring indicator is standardized by Z-score to ensure that different scoring indicators are weighted under the same dimension.

10. A computer storage medium, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the deep learning-based method for screening effective components of compound supplement colonic sustained-release tablets as described in any one of claims 1-9.

Images