Traditional Chinese medicine compound screening method based on machine learning

Abstract

The invention relates to the technical field of machine learning and traditional Chinese medicine research and development, and discloses a traditional Chinese medicine compound screening method based on machine learning, which comprises the following steps: carrying out structure entropy pre-detection on an unknown compound to filter a sample with structure loss, projecting the sample to a structured feature space which is defined by a plurality of prototype prescriptions and has monarch, minister, assistant and guide roles, generating a prototype space projection vector; according to the method, a feature representation normal form is constructed, the internal compatibility of the compound is intelligently translated into structured input which can be understood by a machine learning model, and therefore the reliability of the compound is improved, the reliability of the compound is improved, and the reliability of the compound is improved. The classical example-based analogy interpretability is realized, meanwhile, the judgment capability of the screening system on the self-cognitive boundary is also improved, and the prescription innovation opportunity can be actively identified and protected.

Description

TECHNICAL FIELD

[0001] The present application relates to a machine learning-based traditional Chinese medicine compound screening method, belonging to the technical field of machine learning and traditional Chinese medicine research and development. BACKGROUND

[0002] Currently, using machine learning methods to analyze and screen traditional Chinese medicine compound data has become a common technical direction, which aims to improve the efficiency of new formula discovery. In specific implementation, this direction forms two main technical paths, one is a data-driven path based on large-scale data statistical modeling, and the other is a knowledge-driven path that systematically encodes traditional Chinese medicine theory.

[0003] However, when dealing with traditional Chinese medicine compounds, both of the above technical paths have inherent technical limitations; the data-driven path, due to its main processing method of correlation analysis, fails to effectively utilize the unique compatibility structure information of traditional Chinese medicine compounds, resulting in a lack of sufficient interpretability of the generated model, making it difficult to evaluate the reliability of the screening results; the knowledge-driven path, due to the complexity and unstructured nature of the traditional Chinese medicine theory system, requires engineering investment to build a complete formalized knowledge base, which lacks universal applicability in terms of cost and cycle; the current state of the art has led to a long-standing technical problem in this field, namely, when screening traditional Chinese medicine compounds, a choice must be made between low interpretability of the model and high complexity of the implementation; any direct combination of the two, such as supplementing the data-driven model with limited expert rules or partially simplifying the knowledge system, has failed to fundamentally change this situation, because the core problem is not a lack of data volume or rule quantity, but rather the lack of a technical means to effectively convert the compatibility structure of the compound into a digital representation that can be directly processed by machine learning models.

[0004] Specifically, the existing technology mainly has the following deficiencies: 1. The data statistical-based path cannot reflect the internal compatibility logic of traditional Chinese medicine compounds in the model, resulting in insufficient interpretability; 2. The theory coding-based path has high complexity and cost in engineering implementation, making it difficult to apply on a large scale. Therefore, how to create a new digital representation method for traditional Chinese medicine compounds that can avoid complex coding of complete traditional Chinese medicine theory and instead generate an input feature with inherent logic and interpretability that is easy to implement in engineering, has become a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a machine learning-based traditional Chinese medicine compound screening method, which mainly aims to solve the problem that the existing technology lacks a digital input feature that can effectively convert the internal compatibility structure information of traditional Chinese medicine compounds into a machine learning model that can be directly processed and has both interpretability and engineering feasibility.

[0006] To achieve the above object, the application provides a Chinese medicine compound screening method based on machine learning, comprising the following steps: Step a, a prototype prescription base is constructed in advance, the prototype prescription base records a plurality of prototype prescriptions, and records a drug flavor set corresponding to each of the four roles of monarch, minister, assistant and messenger of each prototype prescription; Step b, the unknown compound is analyzed by role, the drug flavor set corresponding to the four roles of monarch, minister, assistant and messenger is obtained, and based on the distribution of the number of drug flavors of each role of the unknown compound, a structure entropy value representing the order degree of the compatibility structure is calculated; Step c, judge whether the structure entropy value is higher than the preset structure rationality entropy threshold value, if yes, terminate the subsequent processing of the unknown compound; Step d, if not, for each prototype prescription in the prototype prescription base, the similarity of the drug flavor set of the unknown compound in the role domain of monarch, minister, assistant and messenger corresponding to the role domain of the prototype prescription is calculated respectively, and all the calculated similarity values are spliced to generate a prototype space projection vector representing the unknown compound ; Step e, calculate the vector norm value of the prototype space projection vector , and judge whether it satisfies , wherein is a preset confidence threshold; Step f, if the judgment is satisfied, the unknown compound is marked as a high novelty sample; if the judgment is not satisfied, the prototype space projection vector is input into the efficacy prediction machine learning model to obtain the efficacy prediction result.

[0007] Preferably, the role analysis of the unknown compound in step b comprises: according to the core efficacy label of each drug flavor in the unknown compound in the preset efficacy classification system and the dose size of each drug flavor, a unique role in the four roles of monarch, minister, assistant and messenger is assigned to each drug flavor in the unknown compound by using a preset heuristic rule set, so as to obtain the drug flavor set corresponding to each role of the unknown compound.

[0008] Preferably, the similarity of the drug flavor set of the role domain of the unknown compound and the corresponding role domain of the prototype prescription in step d is calculated as follows: the drug property label set and the efficacy label set corresponding to all drug flavors in the two drug flavor sets are obtained, and the Jaccard similarity between the drug property label sets and the Jaccard similarity between the efficacy label sets of the two drug flavor sets are calculated respectively, and then the two Jaccard similarity values are weighted and summed according to the preset weight value to obtain the final drug flavor set similarity.

[0009] Preferably, the method for calculating the structural entropy value in step b is as follows: the proportions of the number of medicinal ingredients contained in the four roles of monarch, minister, assistant and guide in the unknown complex prescription to the total number of medicinal ingredients in the unknown complex prescription are used to form a probability distribution , and the structural entropy value is calculated according to the Shannon entropy formula The structural entropy value represents the role balance of the compatibility structure of the unknown complex prescription.

[0010] Preferably, the method further comprises: projecting the prototype space projection vector generated in step d to one or more risk assessment machine learning models under the condition that The risk assessment machine learning model is used to predict the risk indicators of the compatibility contraindication risk or potential toxic side effect risk associated with the unknown complex prescription; and the final screening is performed by comprehensively analyzing the performance prediction results and the risk indicators.

[0011] Preferably, for the unknown complex prescription marked as a high novelty sample, the method further comprises the following steps: determining a reference prototype formula with the highest similarity to the unknown complex prescription from the prototype formula base, the highest similarity being determined based on the prototype formula corresponding to the dimension with the largest value in the prototype space projection vector ; obtaining the standard prototype space projection vector of the reference prototype formula as an ideal projection vector; calculating the vector difference between the prototype space projection vector of the unknown complex prescription and the ideal projection vector to obtain a novelty difference vector; and analyzing the main source role domain of the novelty of the unknown complex prescription based on the component values of the dimensions corresponding to different role domains in the novelty difference vector.

[0012] Preferably, before step d, the method further comprises: receiving a domain label representing a target treatment domain; and generating the prototype space projection vector in step d, specifically comprising: calculating a domain relevance weight for each prototype formula in the prototype formula base based on the domain label; and weighting the role domain similarity values related to each prototype formula using the domain relevance weight corresponding to the prototype formula before splicing all the calculated similarity values.

[0013] Preferably, the performance prediction machine learning model and the risk assessment machine learning model are pre-trained models, and the training method is as follows: using a dataset with the prototype space projection vector of the historical Chinese medicine complex prescription as the training input, and using the known performance data or known risk data corresponding to the historical Chinese medicine complex prescription as the training label to train the dataset.

[0014] Preferably, the performance prediction machine learning model and the risk assessment machine learning model are selected from a model group comprising a gradient boosting decision tree model and a random forest model.

[0015] ​Preferably, the vector norm value in step e is L2 norm, which is calculated by taking the square root of the sum of the squares of all dimensions of the prototype space projection vector.

[0016] Compared with the prior art, the present application has the following advantages: 1. The method claimed in the present application first determines the roles of monarch, minister, assistant and envoy of each medicinal ingredient in the unknown complex prescription according to the preset heuristic rules, then calculates the similarity of the unknown complex prescription with the corresponding role domain of each prototype prescription in the pre-determined prototype prescription set in the monarch, minister, assistant and envoy role domains respectively, and finally combines the similarity values corresponding to the different role domains of the prototype prescriptions to generate a projection feature vector; this process converts an unordered medicinal ingredient list into an ordered numerical vector whose dimensions are directly related to the specific role structure of a specific prototype prescription, so that the object processed by the subsequent machine learning model is no longer the medicinal ingredient itself, but the internal correlation of the unknown complex prescription with a group of recognized examples in the compatibility structure.

[0017] 2. After generating the projection feature vector representing the unknown complex prescription, a vector norm value of the vector is further calculated, and whether the vector is input into the machine learning model for performance prediction or the unknown complex prescription is marked as a high novelty sample is determined according to whether the norm value is lower than a preset confidence threshold; this takes advantage of the additional information generated by the projection feature vector itself, i.e., the vector norm value directly represents the overall similarity of the unknown complex prescription with the prototype prescription set, thereby establishing two logically independent processing paths in the screening process, one for performance evaluation within the cognitive boundary and the other for identifying and separating samples outside the cognitive boundary, avoiding unreliable predictions by the model for structures it cannot understand.

[0018] 3. The method of the present application uses the same projection feature vector to input it into a machine learning model for predicting performance and one or more machine learning models for predicting specific risk indicators, wherein the performance prediction model and the risk assessment model are pre-trained using different labeled data sets; this way takes advantage of the fact that the projection feature vector already contains complex prescription compatibility structure information, allowing it to serve as a unified information carrier to support different dimensional evaluation tasks, thereby expanding the screening process from a single performance judgment to a comprehensive decision-making process that considers both performance and risk without increasing the complexity of feature engineering.

[0019] 4. The method of the present application receives one or more field tags representing the target treatment field before calculating the role field similarity, and calculates a field relevance weight for each prototype formula in the prototype formula set based on the field tag, and then uses the weight to weight the calculated role field similarity value; this mechanism uses external task information in the form of internal adjustment parameters to affect the generation process of the projection feature vector itself, so that the prototype formula with high relevance to the current treatment field amplifies the structural similarity information in the final vector, while the information with low relevance is suppressed, thereby dynamically adjusting the feature space to focus more on the specific screening intention. BRIEF DESCRIPTION OF DRAWINGS

[0020] Fig. 1 Flow chart of the compound screening and decision-making process of the present application based on prototype space projection; Fig. 2 Comparison chart of model key performance indicators under different feature vectors of the present application; Fig. 3 Use case diagram of the compound screening system of the present application supporting multi-dimensional analysis. DETAILED DESCRIPTION

[0021] In order to make the technical solutions and advantages of the present application clearer, the present application will be further described in detail below, but it should be understood that the specific embodiments described below are only illustrative of the present application, and are not intended to limit the protection scope of the present application.

[0022] The machine learning-based traditional Chinese medicine compound screening method claimed in the present application is configured as a multi-stage serial processing data analysis system. The system receives an unknown compound in the form of a list of medicinal ingredients as initial input. First, a structure entropy pre-check step is used to filter samples that do not have basic compatibility logic. Then, the pre-checked samples are sent to a prototype space projection step, which is responsible for converting the list of medicinal ingredients into a standardized numerical feature vector that contains compatibility structure information, i.e., a prototype space projection vector. Finally, a shunt decision step determines whether to input the vector into one or more machine learning models for regular efficacy and risk assessment, or to mark it as a high novelty sample for special processing, thereby completing the comprehensive screening of unknown compounds. In the application scenario of large-scale compound screening in pharmaceutical research institutions, the existing machine learning path based on data statistics has the technical problem that the input features are usually binary vectors representing the presence or absence of medicinal ingredients. This representation method fails to utilize the unique monarch-minister-assistant-attendant compatibility structure information of traditional Chinese medicine compounds, making it difficult to interpret the model results. To solve this problem, step a of the present application is configured to construct a prototype prescription base, which is an offline preparation work. The purpose is to establish a stable and interpretable reference coordinate system. Specifically, the prototype prescription base is a structured data set that records 50 to 200 classic prescriptions recognized by the field of traditional Chinese medicine as prototype prescriptions. Each prototype prescription entity contains a unique identifier and is associated with four sets of medicinal ingredients, which record the specific names of medicinal ingredients playing the roles of monarch, minister, assistant, and attendant in the prototype prescription. For example, for the prototype prescription Mahuang Decoction, the set of medicinal ingredients corresponding to the monarch role is Mahuang, the set of medicinal ingredients corresponding to the minister role is Guizhi, the set of medicinal ingredients corresponding to the assistant role is Xingren, and the set of medicinal ingredients corresponding to the attendant role is Ganzao. The prototype prescription base constructed in this way provides a data basis for subsequent structured comparison of any unknown compound with this set of classic norms.

[0023] When processing an unknown compound formula, to enable comparison with the prototype formula base, the unknown compound formula needs to be analyzed to obtain its principal, assistant, adjuvant, and guide structure. Therefore, upon receiving the unknown compound formula, the system executes step b, which involves role-based analysis. This analysis utilizes a pre-defined set of heuristic rules. The rules are based on the core efficacy tags of each herb in a pre-defined efficacy classification system and the dosage of each herb. For example, one rule might be: in a compound formula, the one or two herbs with the highest dosage and whose core efficacy best matches the formula's therapeutic direction are identified as the principal herbs; another rule might be: after the principal herbs are determined, herbs that assist the principal herbs in enhancing the therapeutic effect are identified as assistant herbs. By applying this rule set, the unknown compound formula... Each medicinal ingredient is assigned to one of the four roles of monarch, minister, assistant, and guide, thus obtaining the set of medicinal ingredients corresponding to each role in the unknown compound. This step, with low computational cost, endows the originally disordered list of medicinal ingredients with initial structured information. When faced with a screening task containing a large number of randomly combined samples, directly performing subsequent projection calculations on all compounds that have completed role-based analysis will consume unnecessary computational resources and introduce noise. To solve this problem, the present invention sets up a pre-filtering step after role-based analysis and before projection calculation, namely, calculating a structural entropy value that characterizes the orderliness of its compatibility structure. The specific method of this calculation is to construct a probability distribution by taking the proportion of the number of medicinal ingredients contained in each of the four roles of monarch, minister, assistant, and guide in the unknown compound to the total number of medicinal ingredients in the unknown compound. And based on Shannon's entropy formula The structural entropy value was calculated. ,in, For the role The proportion of medicinal ingredients to the total number of medicinal ingredients. To calculate the structural entropy value, for example, a compound prescription containing 1 principal herb, 2 assistant herbs, 4 adjuvant herbs, and 1 guiding herb, totaling 8 herbs, has the following probability distribution: Its structural entropy value Calculated as According to step c, the system determines whether the entropy value of the structure is higher than the preset structural rationality entropy threshold. This threshold is determined by calculating the entropy values ​​of thousands of classical formulas to form a baseline distribution, and selecting the 95th percentile of this distribution as the threshold, for example, 1.90. If the calculated entropy value of 1.75 is lower than the threshold, the sample passes; if it is higher, it is judged as a sample with disordered structure and subsequent processing is terminated. This step establishes an efficient filtering mechanism for the screening system through an objectively measurable indicator to ensure the quality of data entering the core calculation stage.

[0024] For an unknown complex that has passed the structural entropy pre-detection, the system executes step d to generate a prototype space projection vector representing the unknown complex. The core of this process is to calculate the similarity of the unknown complex prescription with the corresponding role domain of the prototype prescription in the prototype prescription base for each prototype prescription in the prototype prescription base. The similarity calculation method is as follows: obtain the efficacy label set corresponding to all medicinal ingredients in the two medicinal ingredient sets to be compared, and calculate the Jaccard similarity between the two efficacy label sets. The Jaccard similarity is calculated by dividing the size of the intersection of the two sets by the size of the union. For example, if the efficacy label of the unknown complex prescription is blood-activating, stasis-removing, and pain-stopping, and the efficacy label of the prototype prescription A is stasis-removing, qi-regulating, and pain-stopping, the intersection size is 2, the union size is 4, and the Jaccard similarity is ; accordingly, the same calculation is performed on the monarch, minister, and messenger role domains, thereby obtaining a four-dimensional similarity vector of the unknown complex prescription with respect to the prototype prescription A, such as ; the unknown complex prescription is calculated with all prototype prescriptions (for example, 50) in the base, and all obtained four-dimensional vectors are sequentially spliced, thereby finally generating a dimensional prototype space projection vector Each dimension of the vector has a clear technical direction. For example, the first to fourth dimensions of the vector represent the similarity of the monarch, minister, and messenger structure of the unknown complex prescription with the first prototype prescription.

[0025] After obtaining the prototype space projection vector , in order to process novel complex prescriptions that are significantly different from all known prototypes, the present application introduces step e, that is, calculating the vector norm value of the prototype space projection vector , and comparing it with a preset confidence threshold . The vector norm value is preferably the L2 norm, which is calculated by taking the square root of the sum of the squares of all dimension values in the vector. For example, for an 8-dimensional vector , the L2 norm value is about , which measures the overall similarity of the unknown complex prescription with the entire prototype prescription set. The confidence threshold is determined by analyzing a validation set containing a large number of known conventional prescriptions and a small number of known novel prescriptions, observing and determining a boundary value that can effectively distinguish the norm value distribution of the two types of samples, for example, 0.5. If the judgment result of step f is If yes, the system will mark the unknown complex prescription as a high novelty sample; otherwise, if the condition is not met, the vector is input into the performance prediction machine learning model to obtain the performance prediction result. The confidence judgment and the shunt processing mechanism enable the system to distinguish whether a complex prescription is structurally inconsistent with the known paradigm or structurally similar to the known paradigm but has a lower predicted performance, thereby avoiding making unreliable performance judgments on samples outside the cognitive range.

[0026] To achieve dual evaluation of the performance and risk of a complex prescription, the method of the present application also uses a prototype space projection vector As a shared information carrier, specifically, under the condition that the vector is input into one or more risk assessment machine learning models in addition to being input into a performance prediction machine learning model. The performance prediction model and the risk assessment model are both pre-trained models. The training of the performance prediction model uses a dataset with the projection vector of a historical complex prescription as the training input and its known performance data as the training label. The training of the risk assessment model uses a dataset with the projection vector of a historical complex prescription as the training input and its risk data such as the inclusion of incompatibility or potential toxic side effects as the training label. For example, a gradient boosting decision tree model or a random forest model can be trained for performance and risk prediction, respectively. The final screening is performed by combining the performance prediction result and the risk indicator. In addition, to improve the accuracy of screening in a specific treatment field, the technical solution of the present application also includes a dynamic adjustment mechanism. Before generating the prototype space projection vector in step d, the system can first receive a field label representing the target treatment field submitted by the user, such as a relieving agent. Based on the field label, the system calculates a field relevance weight for each prototype prescription in the prototype prescription base. The calculation of the weight can be completed by comparing the semantic similarity between the user label and the pre-set field label of each prototype prescription. When generating the final Before concatenating the role domain similarity values, multiply the four-dimensional similarity vector related to each prototype prescription by its corresponding field relevance weight. Through this weighting process, the structural information of a prototype prescription with a high degree of association with the current screening task is enhanced in the final feature vector, while the information with a low degree of association is suppressed. For unknown complex prescriptions marked as high novelty samples, to analyze the source of novelty, the present application also provides a subsequent analysis method. The method first determines a reference prototype prescription with the highest similarity to the unknown complex prescription from the prototype prescription base based on the prototype space projection vector The prototype formula corresponding to the dimension with the largest median value is realized, and then the standard prototype space projection vector of the reference prototype formula itself is obtained as an ideal projection vector. It should be noted that the projection vector of a prototype formula itself is calculated by comparing its similarity with itself and all other prototype formulas. The four-dimensional component value corresponding to itself is high. Then, the vector difference between the projection vector of the unknown complex formula and the ideal projection vector is calculated to obtain a novelty difference vector. The dimensions of the difference vector quantify the degree to which the unknown complex formula deviates from its most similar paradigm in the corresponding structure. By analyzing the component values of the dimensions corresponding to different roles in the difference vector, it can be determined whether the novelty of the unknown complex formula is mainly due to the changes in the roles of monarch, minister, auxiliary or assistant drugs.

[0027] Example 1: In a drug discovery project of a pharmaceutical company, the task is to screen new formulas with potential development value from a database containing hundreds of thousands of candidate complex formulas; the company's original screening system uses the list of medicinal ingredients of each complex formula as input to train a deep learning model for efficacy prediction. There are two problems in the system: first, the prediction results of the model cannot be explained, and the researchers cannot understand the specific reasons why a candidate formula is judged to be high or low in efficiency; second, the project that tries to formalize the complete TCM theory to improve interpretability cannot meet the needs of high-throughput screening due to high engineering complexity and maintenance cost; to address this situation, the company deployed the technical solution of the present application; first, the system performed a structure entropy pre-check step. When processing a batch of 100,000 candidate complex formulas, the system calculated the structure entropy value of each complex formula according to the number distribution of the roles of monarch, minister, auxiliary and assistant drugs after role analysis ; among them, the calculated of 32,500 complex formulas is higher than the preset structure rationality entropy threshold value 1.90, and these samples are judged to be without basic compatibility logic and are directly filtered out, not entering the subsequent processing link; this step provides a preliminary purified candidate set for the subsequent projection calculation, reducing the invalid calculation load of the subsequent steps.

[0028] Then, for the 67,500 candidate complex formulas that pass the pre-check, the system performs a prototype space projection step to convert each complex formula into a 200-dimensional prototype space projection vector ; in the screening results, a candidate complex formula X obtains a high score given by the efficacy prediction machine learning model, and the researchers check its corresponding , it is found that the vector has a similarity value of 0.85 and 0.91 in the dimension component corresponding to the prototype prescription Ma Huang Tang, which is the monarch and minister role domain, respectively. At the same time, the vector has a similarity value of 0.88 and 0.82 in the dimension component corresponding to the prototype prescription Gui Zhi Tang, which is the assistant and making role domain. This information enables researchers to form a technical judgment that the compatibility structure of the candidate compound X may be based on the core structure of Ma Huang Tang and combined with the assistant and making compatibility idea of Gui Zhi Tang. This screening result is no longer a value that cannot be interpreted, but a technical hypothesis with a clear structural analogy reference that can be further verified, thereby solving the problem of having to choose between model interpretability and engineering complexity. In this screening project, another unknown compound Y generates a prototype space projection vector with a vector norm value of 0.41, which is lower than the preset confidence threshold of 0.5. Therefore, the system does not input the vector into the efficacy prediction model, but automatically marks it as a high novelty sample. Subsequently, through the novelty source analysis, it is found that the compound Y has a low structural similarity with all prototype prescriptions, but its novelty difference vector mainly reflects the difference in the monarch role domain. This process prevents the model from making a false low score rating for a prescription with a completely new use of monarch medicine outside its cognitive range, reflecting the synergistic effect of the two steps of structural entropy pre-check and confidence judgment: the former ensures that the samples entering the evaluation system have the most basic compatibility logic, and the latter identifies samples with compatibility logic but structural novelty. The combined effect of the two ensures that the entire screening process can identify both optimization opportunities similar to classical methods and opportunities different from classical methods. The application of this method changes the original screening work method, and the research and development process no longer relies on reverse engineering interpretation of the prediction model itself, but shifts to structural analysis of the prototype space projection vector. The output of the screening process changes from a ranked list to a multi-dimensional decision support report containing efficacy prediction, risk warning, and novelty assessment.

[0029] Example 2: To verify the effectiveness of using prototype space projection vectors as input features of machine learning models in the technical solutions of the present application, a comparative test was designed and performed in this example. The purpose of this test was to quantitatively compare the differences in various performance evaluation indicators between the test group using the prototype space projection vectors of the present application and the control group using a commonly used feature table ENG method in the field in training the performance prediction machine learning model. The data set used in the test was derived from the publicly available traditional Chinese medicine database TCMSP, from which a data set containing 5000 compound samples was screened and organized. Each compound had a clear list of medicinal ingredients, as well as a verified binary performance label indicating whether it was effective for a specific indication, with 1 representing effectiveness and 0 representing ineffectiveness, and the positive and negative sample ratio being 1:1. The computing environment for the test was a computer equipped with a floating point operation processor, and the software environment used Python programming language and standard data processing and modeling libraries such as scikit-learn and XGBoost.

[0030] The test was set up as follows: The data set of 5000 compound samples was randomly divided into a training set and a test set in a ratio of 80% to 20%. An experimental group and a control group were established. The experimental group used the method claimed in the present application to process the compound samples, i.e., first performing structural entropy pre-detection on all samples, then generating a 200-dimensional prototype space projection vector for the samples that passed the pre-detection as input features; the control group used a conventional TF-IDF method based on the bag-of-words model to vectorize the medicinal ingredient list of the compound, generating a TF-IDF vector with a dimension equal to the total number of non-redundant medicinal ingredients in the data set as input features; both groups used the same performance prediction machine learning model, specifically the XGBoost model, whose internal parameters were optimized through cross-validation in each group; the training of the model was performed on the training set, and the performance evaluation was performed on the test set that the model had not seen before, with evaluation indicators including accuracy, precision, recall, F1 score, and area under the ROC curve AUC; during the test, an effective compound sample in the test set was taken as an example, with its medicinal ingredients being {A, B, C, D, E, F}; in the control group, this sample was converted into a high-dimensional sparse TF-IDF vector, most of whose dimensions were 0, with only the TF-IDF values of the dimensions corresponding to medicinal ingredients A to F being non-zero, and this vector mainly reflected the frequency information of the medicinal ingredients; in the experimental group, this sample first passed the structural entropy pre-detection, and then was converted into a 200-dimensional, dense prototype space projection vector , the numerical value of each dimension of the vector respectively represents the structural similarity of the sample and 100 prototype prescriptions in the monarch, minister, assistant and messenger role domains, and the numerical value distribution reflects the correlation degree of the sample in the compatibility structure with the classical prescription group; the characteristic vectors of the two groups are respectively input into the XGBoost model trained respectively, and after prediction on the test set, the performance indicators of each group are obtained.

[0031] Table 1: Comparison of model performance indicators of test group and control group on test set.

[0032]

[0033] The test data shows that, referring to Table 1, the performance evaluation indicators of the test group using the method of the present application are all higher than those of the control group using the TF-IDF vector, and the AUC index of the test group is 0.94, which is obviously higher than that of the control group, which is 0.82; the test results show that, compared with the TF-IDF vector which only considers the frequency information of medicinal ingredients, the prototype space projection vector generated by the present application can provide higher information density and stronger performance correlation to the machine learning model by encoding the compatibility structure information of the compound into the features, so as to obtain higher prediction accuracy and model generalization ability in the performance prediction task.

[0034] Embodiment 3: This embodiment combines Figs. 1 to 3 to illustrate a machine learning-based traditional Chinese medicine compound screening method, as Fig. 1 shown, the process starts with receiving an unknown compound input in the form of an unordered medicinal ingredient list. First, it enters the structural entropy pre-check step, which performs role-based analysis on the compound and filters out samples with structural anomalies. Then, it determines whether the structural entropy is higher than the rationality threshold. If the answer is yes, the process is terminated, and the combination is determined as invalid or unordered. If the answer is no, the sample enters the prototype space projection step, which translates the internal compatibility wisdom of the compound into a structured feature vector based on a prototype prescription base containing classical prescription structure information, generating a prototype space projection vector . Subsequently, the process enters the confidence judgment link, which determines whether the norm value of the vector is lower than a confidence threshold. If the answer is yes, it means that the sample has high novelty, and it is marked as a high novelty sample to identify and protect innovation opportunities. If the answer is no, it means that the sample is regular, and it is input into the performance prediction machine learning model and the risk assessment machine learning model to obtain the performance prediction result and the risk indicator respectively. Finally, a multi-dimensional decision support report is output by integrating these information.

[0035] As Fig. 2As shown, the horizontal coordinate of the figure is the performance index, specifically including accuracy, precision, recall, F1 score and AUC, and the vertical coordinate is the performance index value. It is clearly shown in the figure that for all evaluation indexes, the striped column representing the test group is significantly higher than the diagonal column representing the control group, which confirms the superiority of the feature representation paradigm proposed in the present application compared with the traditional method.

[0036] As shown in the figure, as a participant, the drug researcher can interact with the system to perform the operation of configuring the screening task parameters, such as specifying a target treatment field, and the core interaction is to perform compound screening, which triggers two core function branches of the system, namely analyzing the general compound efficacy and risk and identifying high novelty samples. In addition, the drug researcher can further perform the operation of analyzing the source of novelty for the identified high novelty samples, thereby comprehensively supporting the research and development decision. Fig. 3

[0037] Embodiment 4: In a specific system deployment scenario, in order to apply the screening method of the present application to a new treatment field, the system performs a set of offline calibration and initialization procedures before formal deployment; the procedure first prepares a data set dedicated to calibration, which contains three types of compounds: the first type is 500 classical prescriptions related to the target treatment field obtained from public literature; the second type is 5000 noise prescriptions generated by randomly combining the commonly used medicinal ingredients in this field; the third type is a validation set containing 200 new prescriptions verified by experiments to be effective but different in structure from classical prescriptions; the first step of the procedure is to determine the structure rationality entropy threshold; the system performs the role analysis and structure entropy value calculation in the specific embodiment for each compound in the first set of classical prescriptions, thereby obtaining a distribution containing 500 entropy values; statistical analysis shows that the mean of the distribution is 1.65 and the standard deviation is 0.12; the system sets the 95th percentile of the distribution, i.e. 1.85, as the structure rationality entropy threshold of the treatment field according to the preset rule.

[0038] The second step of the procedure is to determine the confidence threshold ; the system first generates the prototype space projection vector of all three types of compounds in the calibration data set using the 100 prototype prescriptions that have been determined, and calculates the L2 norm value of each vector; the analysis result shows that the norm values of the first type of classical prescriptions are distributed between 1.2 and 3.5, the norm values of the second type of noise prescriptions are concentrated in the interval of 0.1 to 0.4, and the norm values of the third type of new prescriptions are mostly distributed between 0.4 and 0.8; the system determines the lower quartile of the norm value distribution of the third type of new prescriptions, i.e. 0.5, as the confidence threshold of this deployment ​​Meanwhile, in the procedure, the heuristic rule set relied on for role analysis of unknown complex prescriptions is embodied as a set of ordered logical judgment steps: first, identify the drug with the largest dosage in the complex prescription as the candidate monarch drug; second, query the core efficacy label of the candidate monarch drug, if its semantic similarity with the target therapeutic domain label of the current screening task is greater than 0.7, it is determined as the monarch drug, otherwise the drug with the second largest dosage is taken as the new candidate monarch drug and the step is repeated; third, after the monarch drug is determined, the core efficacy labels of the remaining drugs are compared with that of the monarch drug, if the similarity is greater than 0.6, it is determined as the minister drug; fourth, among the remaining drugs, if the efficacy of a drug is opposite to that of the monarch drug or can treat the combined symptoms, it is determined as the auxiliary drug, and the rest are determined as the minister drug; for the efficacy prediction machine learning model, the procedure defines its life cycle management mechanism; the initialization training of the model uses the total of 700 prescriptions of the first and third classes in the calibration data set as positive samples, and a small amount of noise prescriptions remaining after structure entropy pre-examination as negative samples for training; after the model is deployed, a performance monitoring mechanism is established, the system regularly inputs the prescriptions and their real efficacy data verified by subsequent experiments as new blind test samples into the model for prediction, and calculates the prediction accuracy of the model on these new samples; once the accuracy of the model on the new samples is lower than the preset 80% baseline in the rolling period of three consecutive months, the system triggers the retraining process, and the new data verified is included in the training set to update the model, so as to ensure the timeliness of the prediction ability of the model.

[0039] Example 5: This example describes a set of pre-constructing and verifying procedures to ensure the robustness and consistency of the core knowledge base in the technical solution before the system is formally deployed in a multi-center, multi-task production environment; before the system is deployed, in order to deal with the possible differences in the records of the same prescription in different historical classics and the heterogeneity of the efficacy description of the same drug, the system performs a set of prototype prescription base construction procedures; first, all the recorded prescriptions from five authoritative Chinese medicine classics recognized by the field are preliminarily extracted as a candidate set; then, the system applies a set of filtering rules to each prescription in the candidate set, which requires that only the prescriptions recorded in at least three independent classics, whose core indications are undisputed, and whose monarch-minister-auxiliary structure is identifiable, are included in the final prototype prescription base; through this multi-source cross verification and filtering step, a base consisting of 150 prototype prescriptions is finally generated.

[0040] To ensure the consistency of the drug efficacy label system relied on by the similarity calculation, the system also performs a set of label semantic relationship quantification procedures based on computational linguistics models; the procedure traverses all drug herbs in the prototype prescription base and the extended drug herb library, extracts all efficacy descriptive words associated with each drug herb from the above-mentioned multiple authoritative classics, and forms a word library containing thousands of efficacy labels; then, the system uses a pre-trained word vector model to vectorize all efficacy labels in the word library, and pre-calculates the cosine similarity between any two efficacy labels in the word library; all calculated similarity values are stored in a two-dimensional similarity matrix as the system's internal knowledge base for direct calling by subsequent role analysis or field relevance weight calculation steps.

[0041] Embodiment 6: This embodiment describes a set of pre-constructing and verifying procedures performed to ensure the robustness and consistency of the core knowledge base of the technical solution before it is formally deployed in a multi-center, multi-task production environment; before a specific system deployment, to deal with the possible differences in the records of the same prescription in different historical classics, and the heterogeneity of the same drug efficacy description, the system performs a set of standardized prototype prescription base construction procedures; the procedure first extracts all recorded prescriptions from five authoritative Chinese medicine classics recognized by the field as a candidate set; then, the system applies a set of filtering rules to each prescription in the candidate set, which requires that only the prescriptions recorded in at least three independent classics, whose core indications are undisputed, and whose monarch, minister, assistant and messenger structures are identifiable, are included in the final prototype prescription base; through this multi-source cross-validation and filtering step, a base consisting of 150 prototype prescriptions is finally generated.

[0042] To ensure the consistency of the drug efficacy label system relied on by the similarity calculation, the system also performs a set of label semantic relationship quantification procedures based on computational linguistics models; the procedure traverses all drug herbs in the prototype prescription base and the extended drug herb library, extracts all efficacy descriptive words associated with each drug herb from the above-mentioned multiple authoritative classics, and forms a word library containing thousands of efficacy labels; then, the system uses a pre-trained word vector model to vectorize all efficacy labels in the word library, and pre-calculates the cosine similarity between any two efficacy labels in the word library; all calculated similarity values are stored in a two-dimensional similarity matrix as the system's internal knowledge base for direct calling by subsequent role analysis or field relevance weight calculation steps.

[0043] To further verify the technical means of converting the complex compound structure into a digital representation that can be directly processed by the machine learning model in the method claimed in the present application, the advantages over the prior art, especially in the complex scenario of the same medicine but different compatibility structure and treatment direction, are particularly supplemented as follows.

[0044] The purpose of this example is to verify whether the conventional technical path in the art can be used to effectively distinguish and identify a group of traditional Chinese medicine compounds with the same composition but different dosage, leading to different compatibility structure and core efficacy.

[0045] The test conditions, calculation environment and machine learning model (XGBoost model) used in this example as the performance prediction target are strictly consistent with those described in Example 2. Instead of using the feature vector generation method based on prototype space projection claimed in the present application, this example uses a conventional TF-IDF method based on the bag-of-words model mentioned in Example 2 to vectorize the medicine list of the compound. The test process is as follows: prepare a special test set containing 1000 compound samples. The construction of this test set aims to simulate a common and key challenge in traditional Chinese medicine research and development: deriving several modified prescriptions from a master prescription. The test set contains 100 pairs of core compounds, a total of 200 core compounds, each pair of compounds contains exactly the same 5 medicines, but one of the core medicines has a significant difference in dosage between the two compounds, resulting in a shift in the monarch medicine role, thus changing the treatment efficacy of the two compounds from relieving the exterior and dispelling cold to warming the lungs and resolving fluid. The remaining 800 compounds in the test set are other irrelevant interference prescriptions. The test task is to train a classification model to accurately distinguish between these two types of samples with different treatment efficacy.

[0046] Using the test group of this example, all compound samples in the test set are converted into a high-dimensional sparse TF-IDF vector. Since the generation of TF-IDF vector is mainly based on the word frequency of the medicine, it is not sensitive to the absolute dosage of each medicine in the compound. Therefore, for the two compounds in the above core test pair, the feature vectors generated by this method are almost identical, and theoretically have no distinguishing ability. The TF-IDF feature vectors generated are input into the pre-configured XGBoost model for training and testing. The classification performance on the special test set and the engineering problems exposed during processing are recorded in Table 2 below.

[0047] Table 2: Performance index comparison table of Example 1 and the method of the present application in the same medicine but different compatibility structure and treatment direction identification task.

[0048]

[0049] The test results show that the machine learning model using the conventional TF-IDF vector as the feature has no significant difference in classification accuracy (53.5%) from random guessing (50%) when facing the compound prescriptions with the same flavor composition but different compatibility structure and efficacy due to the change of dosage, which objectively confirms that the feature representation method reflecting only the frequency information of the medicinal ingredients and ignoring the internal compatibility structure information of the compound prescriptions indeed cannot handle the screening task which is highly sensitive to the change of compatibility structure in the technical principle.

[0050] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for screening traditional Chinese medicine compound based on machine learning, characterized in that, The method comprises the following steps: Step a, pre-constructing a prototype formula base, the prototype formula base recording a plurality of prototype formulas, and recording a respective corresponding medicinal ingredient set of each of the four roles of monarch, minister, assistant and envoy of each prototype formula; Step b, role analysis of the unknown complex formula to obtain the medicinal ingredient set corresponding to the four roles of monarch, minister, assistant and envoy, and calculating a structure entropy value representing the order degree of the compatibility structure of the unknown complex formula based on the distribution of the number of medicinal ingredients of each role of the unknown complex formula; Step c, judging whether the structure entropy value is higher than a preset structure rationality entropy threshold value, if yes, terminating the subsequent processing of the unknown complex formula; Step d, if no, then for each prototype formula in the prototype formula base, the similarity of the unknown complex prescription in the domain of monarch, minister, assistant and guide with the corresponding domain of the prototype formula is calculated respectively, and all the calculated similarity values are spliced to generate a prototype space projection vector representing the unknown complex prescription ; Step e, calculating the vector norm value of the prototype space projection vector and determining whether wherein is a preset confidence threshold; Step f, if the judgment is satisfied, marking the unknown complex formula as a high novelty sample; if the judgment is not satisfied, inputting the prototype space projection vector into the efficacy prediction machine learning model to obtain the efficacy prediction result.

2. The machine learning-based traditional Chinese medicine compound screening method according to claim 1, characterized in that, In step b, the role analysis of the unknown complex formula specifically comprises: according to the core efficacy label of each medicinal ingredient in the unknown complex formula in the preset efficacy classification system and the dosage size of each medicinal ingredient, assigning a unique role in the four roles of monarch, minister, assistant and envoy to each medicinal ingredient in the unknown complex formula through a preset heuristic rule set to obtain the medicinal ingredient set corresponding to each role of the unknown complex formula.

3. The machine learning-based traditional Chinese medicine compound screening method according to claim 1, characterized in that, In step d, the medicinal ingredient set similarity between the role domain of the unknown complex formula and the corresponding role domain of the prototype formula is calculated, specifically: obtaining the respective corresponding medicinal property label set and efficacy label set of all medicinal ingredients in the two medicinal ingredient sets, and respectively calculating the Jaccard similarity between the medicinal property label sets and the Jaccard similarity between the efficacy label sets of the two medicinal ingredient sets, and then weighting and summing the two calculated Jaccard similarity values according to a preset weight value to obtain the final medicinal ingredient set similarity.

4. The machine learning-based traditional Chinese medicine compound screening method according to claim 1, characterized in that, The calculation method of the structure entropy value in step b is as follows: the proportion of the number of medicinal ingredients contained in the four roles of monarch, minister, assistant and guide in the unknown complex prescription to the total number of medicinal ingredients in the unknown complex prescription forms a probability distribution According to the Shannon entropy formula The structure entropy value is calculated The structure entropy value represents the role balance of the compatibility structure of the unknown complex prescription.

5. The machine learning-based traditional Chinese medicine compound screening method according to claim 1, characterized in that, The method further comprises projecting the prototype space vector generated in step d , under the condition that the compatibility contraindication risk or the potential toxic side effect risk associated with the unknown complex prescription does not meet , into one or more risk assessment machine learning models for predicting a risk indicator of the risk, and the final screening is performed by combining the efficacy prediction result and the risk indicator.

6. The machine learning-based traditional Chinese medicine compound screening method according to claim 1, characterized in that, For the unknown complex prescription marked as high novelty sample, the method further comprises the following steps: determining a reference prototype formula with the highest similarity with the unknown complex prescription from the prototype formula base, the highest similarity being based on the prototype space projection vector of the reference prototype formula with the largest numerical value; obtaining the standard prototype space projection vector of the reference prototype formula as an ideal projection vector; calculating the vector difference between the prototype space projection vector of the unknown complex prescription and the ideal projection vector to obtain a novelty difference vector; and analyzing the main source role domain of the novelty of the unknown complex prescription based on the component values of the dimensions corresponding to different role domains in the novelty difference vector.

7. The machine learning-based traditional Chinese medicine compound screening method according to claim 1, characterized in that, Before step d, the method further comprises: receiving a domain label representing a target treatment domain; and generating the prototype space projection vector in step d specifically comprises: calculating a domain relevance weight for each prototype formula in the prototype formula base based on the domain label; and before weighting all the calculated similarity values, weighting the role domain similarity values related to each prototype formula using the domain relevance weight corresponding to the prototype formula.

8. The machine learning-based traditional Chinese medicine compound screening method according to claim 1, characterized in that, The efficacy prediction machine learning model and the risk assessment machine learning model are pre-trained models, and the training method is: using a data set with the prototype space projection vector of the historical traditional Chinese medicine complex formula as the training input, and using the known efficacy data or known risk data corresponding to the historical traditional Chinese medicine complex formula as the training label to train.

9. The machine learning-based traditional Chinese medicine compound screening method according to claim 8, characterized in that, The efficacy prediction machine learning model and the risk assessment machine learning model are selected from a model group comprising a gradient boosting decision tree model and a random forest model.

10. The machine learning-based traditional Chinese medicine compound screening method according to claim 1, characterized in that, In step e, the vector norm value is L2 norm, and the calculation method is: square root operation on the sum of squares of all dimensions in the prototype space projection vector.

Images