System and optimization method for treating UC efficacy of fritillaria delavayi based on multi-omics joint evaluation

Abstract

The application discloses a fritillaria delavayi treatment UC efficacy system and optimization method based on multi-omics joint evaluation, belongs to the field of medicine, and comprises a sample collection module, a quality control processing module, an extraction labeling module, a graph construction module, an individual simulation module, a treatment management module, a disturbance identification module, a learning update module, a driving identification module, a therapeutic effect prediction module, an optimization simulation module, and a clinical return module; the application effectively solves the problem that the therapeutic effect of fritillaria delavayi has obvious individual differences but is difficult to explain, avoids the risk that a traditional model lacks biological rationality, improves the credibility and interpretability of a conclusion, provides clear target clues for mechanism research, pharmacological interpretation and subsequent experimental verification, reduces the application obstacles of uninterpretable model conclusions, and has high general value and platform potential.

Description

Technical Field

[0001] This invention relates to the field of medicine, and in particular to a system and optimization method for the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) based on multi-omics evaluation. Background Technology

[0002] Ulcerative colitis (UC) is an immune-mediated disease characterized by chronic inflammation and recurrent attacks of the intestinal mucosa. Its pathogenesis is highly complex, involving multiple factors such as gut microbiota imbalance, host immune abnormalities, impaired epithelial barrier function, and disruption of multi-level molecular regulatory networks. Although existing immunosuppressants, biologics, and small-molecule targeted drugs have improved the clinical outcomes of UC patients to some extent, significant differences in efficacy, high risk of adverse reactions, and difficulties in long-term maintenance remain, necessitating the exploration of safer, more effective, and individually adaptable treatment strategies. Fritillaria cirrhosa, a traditional medicinal herb long used in Tibetan and Traditional Chinese Medicine, has potential pharmacological basis in anti-inflammatory, immunomodulatory, and intestinal function improvement effects. However, its mechanism of action in UC treatment exhibits significant multi-target, multi-pathway, and systemic regulatory characteristics, making comprehensive evaluation difficult using a single molecule or indicator. Traditional efficacy evaluation methods centered on endpoint clinical scores or single biochemical indicators fail to capture the complex biological responses induced by Fritillaria cirrhosa in different patients, limiting its precise application and systematic explanation of its scientific implications.

[0003] Existing systems and optimization methods for the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) still suffer from significant individual variability in efficacy that is difficult to explain. Furthermore, they risk a lack of biological rationality, reducing the credibility and interpretability of conclusions. They fail to provide clear target clues for mechanism research, pharmacological interpretation, and subsequent experimental verification, increasing the application obstacles of unexplained model conclusions. Moreover, they lack high universal value and platform potential. Therefore, we propose a system and optimization method for the efficacy of Fritillaria cirrhosa in treating UC based on multi-omics joint evaluation. Summary of the Invention

[0004] The purpose of this invention is to address the deficiencies in the existing technology by proposing a system and optimization method for the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) based on multi-omics joint evaluation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The system for evaluating the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) based on multi-omics evaluation includes a sample collection module, a quality control processing module, an extraction and annotation module, an atlas construction module, an individual simulation module, a treatment management module, a perturbation identification module, a learning and updating module, a driver identification module, an efficacy prediction module, an optimization simulation module, and a clinical return module. The sample collection module executes patient screening, informed consent, and standardized sample collection procedures, specifies the collection time points, and standardizes the sampling type; The quality control processing module is used to perform standardized preprocessing and quality control on the raw multi-omics data and images or clinical values, while generating metadata and performing batch effect correction and missing value processing. The extraction and annotation module is used to extract biologically significant features from various omics levels and to structure clinical data and imaging features into standardized phenotypic labels. The graph construction module is used to integrate public knowledge bases and domain expert knowledge to construct a prior knowledge graph. The individual simulation module constructs an individualized virtual biological model based on the patient's baseline multi-omics input and prior knowledge graph; The treatment management module is used to collect continuous multi-omics and clinical data at designed time points under the current Fritillaria cirrhosa dosing regimen, and generate actual observation values. The perturbation identification module compares the actual observed values ​​with the predictions of the individualized virtual biology model at the corresponding time points to identify molecules or pathways that deviate from the model's prior expectations. The learning update module adaptively adjusts the network topology and dynamic parameters of the individualized virtual biology model based on each newly measured actual observation value; The driver identification module reprograms the network based on the update results of the individualized virtual biology model and the quantitative results of the perturbation, and generates a list of driver sub-lists. The efficacy prediction module defines and calculates the patient's multidimensional efficacy phenotype vector and performs cross-scale quantitative prediction based on a personalized virtual biology model. The optimization simulation module performs simulations based on a personalized virtual biology model and outputs candidate personalized adjustment schemes and their expected benefit assessments. The clinical report module integrates various information into a corresponding patient's clinical report and displays it visually, while also recording the doctor's adoption and actual therapeutic effect.

[0006] As a further aspect of the present invention, the specific steps of the extraction and annotation module in structuring clinical data and imaging features into standardized phenotypic labels are as follows: S1.1: Normalize the raw readings or peak areas of each patient sample according to the total amount in the sample, then summarize the total count of all features of each patient sample, and divide the count of each feature by the total count of the sample to obtain the relative abundance of each patient sample. Screen out patient feature data below the preset threshold, and calculate the geometric mean of the features in each patient sample and perform a centered logarithmic ratio transformation. S1.2: Calculate the diversity index within each patient sample and the distance matrix between each patient sample. Then, perform bootstrap sampling for each patient sample and estimate the confidence interval of the diversity index of each patient sample. After that, calculate the effect size and statistical test of each patient characteristic data to identify up- or down-regulated molecules, metabolites or proteins. At the same time, use the BH method to calculate the false discovery rate (FDR) and retain patient characteristic data with FDR below a preset threshold and effect size above a preset threshold as candidate biomarkers. S1.3: Map gene or functional annotations to a predefined set of pathways. For multiple genes or functional items within a pathway, calculate the activity score of the pathway in each sample by weighted summation. Then, standardize the pathway scores between patient samples and output the pathway activity matrix for each patient sample. Based on known metabolic pathways and response relationships, calculate the ratio between each metabolite in each patient sample and perform logarithmic transformation. At the same time, construct corresponding composite metabolic indices for multiple ratios. S1.4: Preprocess endoscopic images, then use semantic segmentation algorithm to identify key structures in the mucosa, ulcer and bleeding areas, and manually or automatically verify the identification results. Then, calculate the morphological indicators of each structure based on the segmentation results, and extract the texture features and color histogram statistics in the segmented area to establish the image features of the corresponding patients. Then, perform scale normalization and repeatability tests on the image features of each patient, and generate the final image quantitative feature set. S1.5: Convert clinical questionnaires, physician scores and laboratory indicators into numerical fields based on a predefined dictionary or mapping. If there are classification options, they are mapped to ordered values. Then, scale transformation is performed on indicators of different dimensions, and min-max normalization is performed on each transformed indicator. Based on the preset clinical discrimination threshold, corresponding binary labels are generated. S1.6: Summarize the acquired features into a sample-feature matrix and record the meta-information of each feature. Then, filter and reduce the dimensionality of each set of features, and establish a standardized phenotypic label vector containing continuous variables and discrete labels. Record the unit, range and missing value handling strategy of each standardized phenotypic label vector. Finally, write the multi-omics data features of each patient sample into a unified structured format.

[0007] As a further aspect of the present invention, the pathway set mentioned in S1.3 includes the KEGG pathway, metabolic pathways, or a custom gene set, etc.; the ratios between the metabolites include the ratio of pro-inflammatory to anti-inflammatory metabolites, the ratio of short-chain fatty acids, etc.

[0008] As a further aspect of the present invention, the specific steps for the graph construction module to construct the prior knowledge graph are as follows: S2.1: Manually set the resource types to be integrated and list accessible instances for each type of resource. Generate structured download entries for each resource by writing crawling or importing scripts, and put all entries into the database according to resource categories to form a source entry table. Perform naming standardization on each source entry, and then generate an initial search index for each record. After that, allocate metadata fields for each resource type, establish a corresponding candidate entity set for the collected source entries, and generate multiple sets of candidate standardized IDs for each candidate entity. S2.2: Perform multi-strategy matching on each candidate entity and generate a corresponding matching confidence score for each matching strategy. Then, weight and summarize the scores of each matching strategy according to resource weight and matching confidence to generate the comprehensive confidence of each standard ID. Mark the standardized IDs with a comprehensive confidence score lower than the preset threshold as pending manual verification. At the same time, record the evidence chain of each mapping. Then, extract explicit relationships directly from the structured database and extract potential relationships from the literature through the information extraction pipeline. S2.3: Collect evidence fragments from different sources and methods for each pair of entities, record the source resources, evidence type, timestamp and credibility of each evidence fragment, generate an initial edge and its evidence score for each pair of entities through weighted aggregation of evidence, if there is evidence conflict, record both positive and negative evidence and score them separately, then label the edge with the original direction or type, and construct an evidence list for each pair of entities. S2.4: Load the disease ontology, phenotype ontology, and pathway ontology into a hierarchical structure, and implement hierarchical propagation for each edge. At the same time, retain the propagation path and attenuation coefficient during the propagation process of each ontology, and dynamically adjust the confidence of the corresponding edge based on the propagation results. Construct a set of functional annotations for each entity, use set theory or embedding space metrics to calculate the semantic similarity between each entity pair, and perform cluster analysis based on the obtained semantic similarity. Mark the inconsistent edges in the analysis results for manual review. S2.5: Count the number of positive and negative data points for each entity pair from the literature and experimental annotations, and perform Bayesian smoothing on each statistical result. Calculate the directional confidence probability of each statistical result. If the positive probability is higher than 0.7, mark the edge as inclined to positive causality and assign a directional confidence score. If the difference between the positive and negative probabilities is lower than a preset threshold, keep it undirected or mark it as bidirectional / uncertain. Based on the directional confidence score and evidence score, establish a directed edge weight that includes direction and strength. S2.6: Submit the automatically extracted list of edges and nodes to the domain expert group for annotation. Each expert gives their approval, disapproval or correction opinions for each edge, and can provide additional evidence or reasons for rejection. At the same time, based on the pre-assigned reputation coefficients of each expert, the annotation results are weighted and summarized. The confidence of evidence that is consistent with the expert opinions is increased, and the confidence of rejected edges is reduced or marked as excluded, so as to generate the final edge confidence. S2.7: Sort all edges according to the final edge confidence from high to low, and delete the corresponding edges of each node whose final edge confidence is lower than the preset threshold. Then, normalize the weights of the remaining directed edges to generate a prior knowledge graph. Export the prior knowledge graph in a standardized format. Each record in the prior knowledge graph includes: source node ID, target node ID, relation type, directional probability, final confidence, evidence list, expert comments, version number and timestamp.

[0009] As a further aspect of the present invention, after the prior knowledge graph described in S2.7 is constructed, timestamp information is added to each piece of evidence in the prior knowledge graph, and the freshness of the evidence is subjected to exponential decay processing. In the evidence aggregation or final confidence calculation, a time decay factor is introduced to adjust the evidence score of each entity pair and the final edge confidence, so that the graph can automatically reflect the latest knowledge over time. Afterwards, external resources are periodically re-fetched and the update process is repeated to update the prior knowledge graph version in real time, and a change difference report is generated for each version.

[0010] As a further aspect of the present invention, the specific steps for the individual simulation module to construct an individualized virtual biological model are as follows: S3.1: Based on the preset node type catalog, map all multi-omics data features of the patient's baseline to the candidate node set, collect the original observation set corresponding to each candidate node, compress the multi-source observations corresponding to each node by PCA projection according to the linear projection from the feature matrix to the node vector to generate the initial node feature vector, then perform scale correction on the feature vectors of each node, and evaluate the measurement reliability of each node. S3.2: Match the selected candidate node set with the prior knowledge image, extract the prior subgraph containing each node in the candidate node set and its direct neighbors, and use it as the initial prior topological skeleton of the current patient. Collect the evidence strength of each extracted prior edge in the prior knowledge graph and convert it into the corresponding prior connection weight. Then, adjust the corresponding prior weight according to the reliability of each node. S3.3: Based on the feature vectors of each node, establish a sample-level node feature matrix and calculate the empirical covariance corresponding to the matrix to obtain the joint variability among nodes. Then, based on each empirical covariance, use sparsified inverse covariance estimation to establish the corresponding precision matrix, infer potential direct interaction edges, and convert the precision matrix into a symmetric data-driven edge weight matrix, and perform significance or stability screening on the edges. S3.4: For each pair of nodes in the prior subgraph, use the prior weights and the data-driven edge weight matrix to perform weighted fusion, calculate the initial adjacency weights of each pair of nodes to establish the corresponding adjacency matrix, and adaptively adjust the fusion weights according to the node reliability, the credibility of the prior source, and the stability score of the edges in the data. After the fusion is completed, perform local normalization and sparsification on the generated adjacency matrix to generate the initial network topology skeleton. S3.5: Collect the initial unsigned adjacency weights of each group, and combine the relation types in the prior knowledge graph and the directional evidence in the data to calculate the initial symbol and its directed strength for each prior subgraph. Then, based on node attributes or biological common sense, assign corresponding directional priors to edges without clear direction, map the directed coupling strength to the dynamic coupling parameter space, generate the coupling matrix, and record the initial symbol confidence of each edge. S3.6: Set an initial baseline activity value for each node to represent the state of the node when it is not affected by coupling. Based on the time-varying index or measurement uncertainty of each node in the baseline data, set the response speed of each node. Then set the self-loop coefficient for each node. Use the node baseline activity value, response speed and self-loop coefficient as its node dynamic parameters and record the parameter confidence interval. S3.7: Combine the coupling matrix with the dynamic parameters of each node into a linearized dynamic operator and evaluate its spectral radius. If the spectral radius exceeds the preset critical value, scale the directional coupling strength according to the set ratio until the spectral radius is lower than the preset critical value. Then check all parameters within a reasonable range. After the check is passed, establish the corresponding patient's individualized virtual biological model based on the initialization results.

[0011] As a further aspect of the present invention, the node type catalog described in S3.1 includes genes / transcripts, proteins, metabolites, microbial species / functions, pathways, and clinical phenotypes.

[0012] As a further aspect of the present invention, the specific steps of the perturbation identification module in identifying molecules or pathways that deviate from the prior expectations of the model are as follows: S4.1: Collect the actual observation values ​​of each actual sampling of the patient, as well as the predicted values ​​generated by the corresponding individualized virtual biology model at the prediction time point. Align the actual observation values ​​and predicted values ​​to a unified time grid. Estimate the approximate value of the actual observation on the model prediction time grid through spline interpolation and record the interpolation uncertainty. At the same time, establish the aligned observation matrix and prediction matrix, and label each alignment point with the original observation or interpolation estimation label. S4.2: Based on the analysis objectives and sampling frequency, select multiple time windows, summarize the residual sequences within each time window into multiple statistics, calculate the weighted mean residual of each node within each time window to obtain the corresponding window-level residual index, calculate the residual summary vector of each node in each time window and save the weight and sample number information, calculate the original p value of each node through a weighted paired t test, if the deviation of the original p value from 0 is higher than the preset deviation threshold, then it is determined that the actual observed value is significantly different from the model prediction; S4.3: Collect the original p-values ​​of all nodes and perform multiple test correction on them. At the same time, generate the original p-values ​​and the corrected q-values. Filter out nodes whose q-values ​​are less than the preset threshold and whose absolute weighted mean residuals are lower than the preset effect size lower limit. The remaining nodes are selected as candidate significant deviation nodes and a list of significant deviation nodes is generated. S4.4: Calculate the temporal difference of the weight of each edge in the patient's individualized virtual biological model between adjacent time windows, and perform statistical tests on the weight change of each edge. Edges with weight changes exceeding a preset change threshold are identified as significantly changed edges. At the same time, each significantly changed edge is mapped back to the path or pathway level, and the perturbed subnetwork is identified. S4.5: Use interrupted time series regression to fit a piecewise regression model ITS for each node, and in the ITS, construct a corresponding regression equation for each node, calculate the regression coefficients of instantaneous level change and slope change to obtain the corresponding causal effect, and estimate its significance. If there is potential confounding, adjust the corresponding covariates, and then confirm the consistency between the causal effect of each node and the correlation test results. S4.6: Map the causal effects obtained at the node level to their respective pathway sets based on pathway annotations, and then use the corresponding combination test method to integrate the node-level significance into the pathway-level significance. At the same time, calculate the summary statistic for each pathway and perform a statistical significance assessment. Based on the assessment results, report the set of pathways that are significantly perturbed. S4.7: Integrate the residual strength, significance, causal effect, edge weight change and path significance at the node level into a perturbation fingerprint report. At the same time, indicate the source and uncertainty of each piece of evidence in the report. Then, sort the nodes in the perturbation fingerprint report from high to low according to the strength of evidence.

[0013] As a further aspect of the present invention, the time window in S4.2 includes the baseline period, the early stage of treatment, the middle stage of treatment, and the late stage of treatment; the statistics include the mean within the window, the median within the window, the variance within the window, and the peak value within the window.

[0014] As a further aspect of the present invention, the specific steps for the optimization simulation module to output candidate personalized adjustment schemes and evaluate their expected returns are as follows: S5.1: Based on the observable description of each patient's individualized virtual biological model, establish the corresponding patient's state space, set the state vector dimension and standardization method, record the state naming convention and scale, and establish the corresponding action set based on the adjustable treatment parameter set. Then, use the patient's individualized virtual biological model as an environment simulator, set the current state and action as input, and the next state and immediate observation as output. Based on the efficacy prediction, side effect risk and treatment cost, design the corresponding immediate reward function. S6.2: Input each source of uncertainty into the environment simulator, and set the constraints of the environment simulator based on known biological knowledge. At the same time, run the simulation multiple times for the preset test actions, and statistically analyze the distribution of the next state to verify the behavior of the environment simulator. After the verification is passed, establish a target network based on the DQN architecture, extract multiple sets of historical patient data and input them into the environment simulator. Perform forward propagation through the target network to calculate the Q value of the corresponding action. S6.3: Based on the Q value of the target network output, calculate the loss value of the current behavior network parameters through the weighted batch loss function, and update the behavior network parameters using the gradient descent method. At the same time, periodically synchronize the behavior network parameters to the target network parameters. During the training process, put the interaction samples into the experience replay pool for the next round of training, and set the sampling priority according to the TD error of each experience from high to low. S6.4: An ε-greedy strategy is adopted as the behavior strategy. An action is randomly selected from the action space with probability ε for exploration. At the same time, the greedy action estimated by the current target network is selected with probability (1−ε). The exploration trajectory is recorded. After training, the current greedy strategy is run multiple times using an independent validation simulation set. The cumulative reward trajectory is recorded, and the expected long-term return of the strategy and its volatility are calculated. S6.5: Based on expected benefits, risk assessment and clinical constraints, output multiple sets of candidate personalized adjustment plans, and attach expected benefits, risk assessment and model confidence intervals to each candidate plan. At the same time, generate an interpretable chain of evidence for each final candidate plan, and output the candidate plan and expected benefits in the form of a structured report.

[0015] As a further aspect of the present invention, the patient-specific virtual biological model described in S6.1 can be observably described including node activity, pathway scores, and image quantification components.

[0016] An optimization method for the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) based on multi-omics evaluation is described below: I. Screen candidate patients according to the pre-established inclusion criteria, obtain written informed consent from patients, record various clinical data, and perform endoscopic examinations and collect images and endoscopic scores. II. Design and implement unified sampling, preprocess the sampled multi-omics raw data, and construct a unified feature matrix and metadata table based on the processed data of each group; III. Construct individualized virtual biological models for patients, formulate fritillaria cirrhosa dosing regimens based on clinical guidelines and individual risks, and sample actual observations after dosing in real time; IV. Compare the actual observed values ​​at the same time point with the predicted values ​​of the patient's individualized virtual biology model, and generate a list of candidate targets; V. Based on actual observations, update the topology and dynamics parameters of the patient's individualized virtual biological model and establish a mapping from the current patient simulation state to the future therapeutic phenotype; VI. Parallel simulations are performed using individualized virtual biological models of patients to output multiple sets of feasible personalized strategy recommendations; VII. Integrate the data into clinical reports for the corresponding patients, and clearly indicate uncertainties, potential side effects, and recommended monitoring indicators in the reports; VIII. Record the reasons for doctors' acceptance and rejection, and continue to sample according to the set time sequence after the adjustment is implemented to monitor the actual efficacy and side effects, and update the patient's individualized virtual biological model.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention involves manually defining the types of resources to be integrated, writing scripts to automatically crawl or import multi-source databases and literature resources, generating structured source entries and storing them in the database. The entries undergo naming standardization, index construction, and metadata annotation to form a candidate entity set. Subsequently, through multi-strategy matching and weighted confidence fusion, entity standardization is performed while preserving the evidence chain. Explicit and latent relationships are extracted from the databases and literature, and initial edges with confidence and directionality are generated by integrating multi-source evidence. Combined with ontology hierarchical propagation, semantic similarity analysis, and Bayesian directionality evaluation, a highly reliable prior knowledge graph is formed. After expert weighted review, the final version is output, mapping multi-omics features to knowledge graph nodes. Node feature vectors are generated through dimensionality reduction and scaling correction, individualized prior subgraphs are extracted, and prior weights and data-driven covariance structures are fused to construct an initial network topology, assigning attributes to nodes and edges. After verifying the system stability by defining the orientation, sign, and kinetic parameters, a personalized virtual biological model of the patient is formed. During treatment, the actual time-series observations are aligned with the model predictions, residuals are calculated, and statistical tests are used to screen for significantly deviating nodes and change edges. Combined with interrupted time series and pathway integration analysis, an interpretable perturbation fingerprint is generated. Finally, the personalized model is used as the environment for simulation training to evaluate the long-term benefits and risks of different treatment actions. Multiple sets of personalized optimization schemes with evidence chains and uncertainty assessments are output, effectively solving the problem of significant individual differences in the efficacy of Fritillaria cirrhosa that are difficult to explain. This avoids the risk of traditional models lacking biological rationality, improves the credibility and interpretability of the conclusions, provides clear target clues for mechanism research, pharmacological interpretation, and subsequent experimental verification, reduces the application barriers of uninterpretable model conclusions, and has high general value and platform potential. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0019] Figure 1 This is a system block diagram of the system for evaluating the efficacy of Fritillaria cirrhosa in treating ulcerative colitis based on multi-omics evaluation proposed in this invention; Figure 2 This is a flowchart of the method for optimizing the efficacy of Fritillaria cirrhosa in treating ulcerative colitis based on multi-omics evaluation proposed in this invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Reference Figure 1The system for evaluating the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) based on multi-omics joint evaluation includes a sample collection module, a quality control processing module, an extraction and annotation module, an atlas construction module, an individual simulation module, a treatment management module, a perturbation identification module, a learning and updating module, a driver identification module, an efficacy prediction module, an optimization simulation module, and a clinical return module.

[0022] The sample collection module executes patient screening, informed consent, and standardized sample collection procedures, specifies collection time points, and standardizes sampling types; the quality control processing module performs standardized preprocessing and quality control on raw multi-omics data and images or clinical values, generates metadata, and performs batch effect correction and missing value handling; the extraction and annotation module extracts biologically significant features from various omics levels and structures clinical data and image features into standardized phenotypic labels.

[0023] Specifically, the raw readings or peak areas of each patient sample are normalized to the total within the sample. Then, the total counts of all features in each patient sample are summed, and the count of each feature is divided by the total count of that sample to obtain the relative abundance of each patient sample. Patient feature data below a preset threshold are filtered out. The geometric mean of features within each patient sample is calculated and centered logarithmic ratio transformation is performed. The diversity index within each patient sample is calculated, and the distance matrix between patient samples is also calculated. Then, bootstrapping is performed on each patient sample, and the confidence interval of the diversity index for each patient sample is estimated. Finally, the effect of each patient feature data is calculated. Size and statistical tests were used to identify upregulated or downregulated molecules, metabolites, or proteins. The Black-Hirschman Index (BH) method was employed to calculate the false discovery rate (FDR). Patient characteristic data with FDR below a preset threshold and effect size above a preset threshold were retained as candidate biomarkers. Gene or functional annotations were mapped to a predefined pathway set. For multiple genes or functional items within a pathway, a weighted summation was performed to calculate the pathway activity score in each sample. The pathway scores were then standardized across patient samples, outputting a pathway activity matrix for each patient sample. Finally, based on known metabolic pathways and response relationships, the interaction between metabolites in each patient sample was calculated. The ratios were calculated and logarithmically transformed. A corresponding composite metabolic index was constructed for multiple ratios. Endoscopic images were preprocessed, and then a semantic segmentation algorithm was used to identify key structures in the mucosa, ulcer, and bleeding areas. The identification results were manually or automatically verified. Based on the segmentation results, morphological indicators for each structure were calculated, and texture features and color histogram statistics within the segmented areas were extracted to establish the corresponding patient's image features. Scale normalization and repeatability tests were then performed on the image features of each patient, generating the final image quantification feature set. Clinical questionnaires, physician scores, and laboratory indicators were converted into numerical values ​​based on a predefined dictionary or mapping. The data is processed by mapping the fields to ordered values ​​if classification options exist. Then, the indicators of different dimensions are scaled and the transformed indicators are normalized by min–max. Based on the preset clinical discrimination threshold, corresponding binary labels are generated. The acquired features are summarized into a sample-feature matrix, and the meta-information of each feature is recorded. The features of each group are then filtered and dimensionality reduced, and a standardized phenotypic label vector containing continuous variables and discrete labels is established. The unit, range, and missing value handling strategy of each standardized phenotypic label vector are recorded. Finally, the multi-omics data features of each patient sample are written into a unified structured format.

[0024] The graph construction module is used to integrate public knowledge bases and domain expert knowledge to build prior knowledge graphs.

[0025] Specifically, the required resource types are manually set, and accessible instances are listed for each resource type. Structured download entries for each resource are generated by writing crawling or import scripts, and all entries are stored in the database according to resource category, forming a source entry table. The naming of each source entry is standardized, and an initial search index is generated for each record. Metadata fields are then assigned to each resource type. A corresponding candidate entity set is established for the collected source entries, and multiple sets of candidate standardized IDs are generated for each candidate entity. Multi-strategy matching is performed on each candidate entity, and a corresponding matching confidence score is generated for each matching strategy. The scores of each matching strategy are then weighted and summarized according to resource weight and matching confidence to generate a comprehensive confidence score for each standardized ID. Standardized IDs with a comprehensive confidence score below a preset threshold are marked for manual verification. Simultaneously, the evidence chain for each mapping is recorded. Explicit relationships are directly extracted from the structured database, while potential relationships are extracted from literature through an information extraction pipeline. Evidence fragments from different sources and methods are collected for each pair of entities. The system records the source resources, evidence type, timestamp, and credibility of each piece of evidence. Through weighted aggregation of evidence, an initial edge and its evidence score are generated for each pair of entities. If evidence conflicts exist, both positive and negative evidence are recorded and scored separately. The edges are then labeled with their original direction or type, and an evidence list for each pair of entities is constructed. The disease ontology, phenotype ontology, and pathway ontology are loaded into a hierarchical structure, and hierarchical propagation is performed on each edge. The propagation path and attenuation coefficient of each ontology are preserved during the propagation process, and the confidence of the corresponding edge is dynamically adjusted based on the propagation results. A functional annotation set for each entity is constructed. The semantic similarity between entity pairs is calculated using set theory or embedding space metrics, and cluster analysis is performed based on the obtained semantic similarity. Inconsistent edges in the analysis results are marked for manual review. The number of positive and negative data points for each entity pair is counted from literature and experimental annotations, and Bayesian smoothing is applied to each statistical result. The directional credibility probability of each statistical result is calculated; if the positive probability is higher than 0...7. Edges are labeled as leaning towards positive causality and assigned directional confidence. If the difference between the positive and negative probabilities is below a preset threshold, the edge remains undirected or labeled as bidirectional / uncertain. Based on directional confidence and evidence scores, directed edge weights containing direction and strength are established. The automatically extracted list of edges and nodes is submitted to a domain expert panel for annotation. Each expert provides approval, disapproval, or correction opinions for each edge and can provide additional evidence or reasons for refutation. Simultaneously, based on pre-assigned expert reputation coefficients, the annotation results are weighted and summarized, and then compared with expert opinions. Consistent evidence increases confidence, while negated edges are reduced or marked for exclusion to generate final edge confidence. All edges are then sorted from highest to lowest final edge confidence, and edges for each node whose final edge confidence is below a preset threshold are deleted. The weights of the remaining directed edges are then normalized to generate a prior knowledge graph. This prior knowledge graph is exported in a standardized format, where each record contains: source node ID, target node ID, relation type, directional probability, final confidence, evidence list, expert comments, version number, and timestamp.

[0026] The specific formula for calculating the overall confidence level is as follows:

[0027] In the formula, Representing the The candidate entry is mapped to the first... The overall confidence level of each standard ID; This represents the total number of matching strategies used by the application. Representing the Each matching strategy pair Standardized score; Representing the Resources that each matching strategy depends on The resource weights are calculated using the following formula:

[0028] In the formula, Representative Resources Resource weight; Representative Resources The number of citations; Represents the resource quality coefficient; Representative Resources The time distance from the current field.

[0029] The specific calculation formula for weighted aggregation of evidence is as follows:

[0030] In the formula, Representative Entity and Aggregate evidence score between; Representative Entity and The total number of independent pieces of evidence; Representing the Evidence against the entity and The original credibility score; Representing the The weight of each piece of evidence corresponding to the resource.

[0031] The specific formula for calculating semantic similarity is as follows:

[0032] In the formula, Representative Entity and Functional semantic similarity; , Representing entities respectively and The corresponding set of functional annotations; The cardinality of the set.

[0033] The specific formula for Bayesian smoothing is as follows:

[0034] In the formula, Representative Entity point to The directional posterior probability, which ranges from (0, 1); Representative support Count of positive evidence; Representative support Or negation Evidence count; This represents the prior smoothing constant.

[0035] The specific formula for calculating the final edge confidence is as follows:

[0036] In the formula, Represents the merged entity and The confidence of the final edge connecting the edges; This represents a compression function used to map linear combinations to (0, 1); The baseline weighting coefficient representing the automatic evidence score is typically 1 and is adjustable. Represents the entity and The number of experts providing annotations for the connecting edges; Representing the The credibility rating of each expert; Representing the Experts on the entity and The annotation score of the connecting edges.

[0037] Furthermore, it should be noted that after the prior knowledge graph is constructed, timestamp information is added to each piece of evidence in the prior knowledge graph, and the freshness of the evidence is subjected to exponential decay processing. In the evidence aggregation or final confidence calculation, a time decay factor is introduced to adjust the evidence scores of each entity pair and the final edge confidence, so that the graph can automatically reflect the latest knowledge over time. Afterwards, external resources are periodically re-fetched and the update process is repeated to update the prior knowledge graph version in real time, and a change difference report is generated for each version.

[0038] The individual simulation module constructs a personalized virtual biological model based on the patient's baseline multi-omics input and prior knowledge graph.

[0039] Specifically, based on a pre-defined node type catalog, all multi-omics data features of the patient's baseline are mapped to a candidate node set. The original observation set corresponding to each candidate node is collected. The multi-source observations corresponding to each node are compressed using PCA projection, following a linear projection from the feature matrix to the node vector, to generate initial node feature vectors. Scale correction is then applied to each node feature vector, and the measurement reliability of each node is evaluated. The selected candidate node set is matched with a prior knowledge graph, and a prior subgraph containing each node and its direct neighbors is extracted as the initial prior topological skeleton for the current patient. Each extracted prior edge is collected in the prior knowledge graph. The strength of evidence is determined and converted into corresponding prior connection weights. Then, based on the reliability of each node, the prior weights are adjusted. A sample-level node feature matrix is ​​established based on the feature vectors of each node, and the empirical covariance corresponding to this matrix is ​​calculated to obtain the joint variability among nodes. Then, based on each empirical covariance, a sparsified inverse covariance estimation is used to establish the corresponding precision matrix, inferring potential direct interaction edges. The precision matrix is ​​then converted into a symmetric data-driven edge weight matrix, and the edges are screened for saliency or stability. For each pair of nodes in the prior subgraph, the prior weights are weighted and fused with the data-driven edge weight matrix to calculate the initial adjacency weights for each node pair, thus establishing corresponding adjacency... The adjacency matrix is ​​used to adaptively adjust the fusion weights based on node reliability, prior source confidence, and edge stability scores in the data. After fusion, the generated adjacency matrix is ​​locally normalized and sparsified to generate the initial network topology skeleton. Unsigned initial adjacency weights are collected for each group, and initial symbols and their directed strengths are calculated for each prior subgraph by combining relation types in the prior knowledge graph and directional evidence in the data. Then, based on node attributes or biological common sense, corresponding directional priors are assigned to edges without clear directions. The directed coupling strength is mapped to the dynamic coupling parameter space to generate a coupling matrix, and the initial symbol confidence of each edge is recorded. A corresponding initial baseline activity is set for each node. The value is used to represent the state of the node when it is not affected by coupling. Based on the time-varying index or measurement uncertainty of each node in the baseline data, the response speed of each node is set. Then, a self-loop coefficient is set for each node. The node baseline activity value, response speed and self-loop coefficient are used as its node dynamic parameters, and the confidence interval of the parameters is recorded. The coupling matrix and the dynamic parameters of each node are combined into a linearized dynamic operator, and its spectral radius is evaluated. If the spectral radius exceeds the preset critical value, the directional coupling strength is scaled by a set ratio until the spectral radius is lower than the preset critical value. Then, all parameters are checked for reasonable range. After the check is passed, the corresponding patient's individualized virtual biological model is established based on the initialization results.

[0040] The treatment management module is used to collect continuous multi-omics and clinical data at designed time points under the current Fritillaria cirrhosa dosing regimen, and generate actual observation values; the perturbation identification module compares the actual observation values ​​with the predictions of the individualized virtual biology model at the corresponding time points, and identifies molecules or pathways that deviate from the model's prior expectations.

[0041] Specifically, the actual observations from each patient sampling are collected, along with the predicted values ​​generated by the corresponding individualized virtual biology model at the predicted time points. The actual and predicted values ​​are then aligned to a unified time grid. Spline interpolation is used to estimate approximate values ​​of the actual observations on the model's predicted time grid, and the interpolation uncertainty is recorded. Simultaneously, an aligned observation matrix and prediction matrix are established, and each alignment point is labeled with either the original observation or the interpolation estimate. Based on the analysis objectives and sampling frequency, multiple time windows are selected, and the residual sequences within each time window are summarized into multiple sets of statistics. The weighted mean residuals of each node within each time window are calculated to obtain the... For window-level residual indices, the residual summary vector for each node in each time window is calculated, and the weight and sample size information are saved. The original p-value for each node is calculated using a weighted paired t-test. If the deviation of the original p-value from 0 is higher than a preset deviation threshold, the actual observed value is considered significantly different from the model prediction. The original p-values ​​of all nodes are collected and multiple test corrections are performed on them. The original p-values ​​and the corrected q-values ​​are generated simultaneously. Nodes with q-values ​​less than a preset threshold and absolute weighted mean residuals lower than a preset effect size lower limit are removed. The remaining nodes are selected as candidate significantly deviated nodes, and a list of significantly deviated nodes is generated. The individualized virtual value for each patient is calculated. In the biological model, the temporal differences in weights of each edge between adjacent time windows are analyzed, and the weight changes of each edge are statistically tested. Edges with weight changes exceeding a preset threshold are identified as significantly changed edges. These significantly changed edges are mapped back to the path or pathway level, and perturbed subnetworks are identified. Interrupted time series regression is used to fit a piecewise regression model (ITS) for each node. Within the ITS, a corresponding regression equation is constructed for each node, and the regression coefficients for instantaneous level and slope changes are calculated to obtain their corresponding causal effects and estimate their significance. If potential confounding exists, appropriate covariate adjustments are made, and each node is then re-confirmed. The consistency between the causal effect and the correlation test results of the nodes is evaluated. The causal effect obtained at the node level is mapped to the respective path set according to the path annotation. Then, the node-level significance is integrated into the path-level significance using the corresponding combination test method. At the same time, the summary statistic of each path is calculated and the statistical significance is evaluated. Based on the evaluation results, the set of significantly perturbed paths is reported. The residual strength, significance, causal effect, edge weight change and path-level significance at the node level are integrated into a perturbation fingerprint report. The source and uncertainty of each piece of evidence are noted in the report. Then, the nodes in the perturbation fingerprint table are sorted from highest to lowest according to the strength of evidence.

[0042] The learning update module adaptively adjusts the network topology and dynamic parameters of the individualized virtual biology model based on each newly measured actual value; the driver identification module reprograms the network based on the updated results of the individualized virtual biology model and the quantitative results of the perturbation, and generates a list of driver sub-items; the efficacy prediction module defines and calculates the multidimensional efficacy phenotype vector of patients, and performs cross-scale quantitative prediction based on the individualized virtual biology model; the optimization simulation module performs simulation based on the individualized virtual biology model, and outputs candidate personalized adjustment schemes and their expected benefit assessments.

[0043] Specifically, based on the observable descriptions of each patient's individualized virtual biological model, a state space is established for each patient, with state vector dimensions and standardization methods set. State naming conventions and scales are recorded. Simultaneously, a corresponding action set is established based on an adjustable set of treatment parameters. The patient's individualized virtual biological model is then used as an environment simulator, with the current state and action as inputs and the next state and immediate observation as outputs. Based on efficacy prediction, side effect risk, and treatment cost, a corresponding immediate reward function is designed, inputting various sources of uncertainty into the environment simulator. Constraints are set for the environment simulator based on known biological knowledge, and multiple simulations are run on preset test actions to statistically analyze the next state distribution and verify the environment simulator's behavior. After successful verification, a target network is established based on a DQN architecture. Multiple sets of historical patient data are extracted and input into the environment simulator. Forward propagation is performed through the target network to calculate the Q-value of the corresponding action. Based on the Q-value output by the target network, a weighted batch loss function is used to calculate... The loss value of the current behavior network parameters is calculated, and the behavior network parameters are updated using gradient descent. At the same time, the behavior network parameters are periodically synchronized to the target network parameters. During training, interaction samples are placed in the experience replay pool for use in the next round of training. The sampling priority is set according to the TD error of each experience from high to low. An ε-greedy strategy is adopted as the behavior strategy. Actions are randomly selected from the action space with probability ε for exploration. At the same time, the greedy action estimated by the current target network is selected with probability (1−ε). The exploration trajectory is recorded. After training, the current greedy strategy is run multiple times using an independent validation simulation set. The cumulative reward trajectory is recorded, and the expected long-term return and its volatility of the strategy are calculated. Based on the expected return, risk measurement and clinical constraints, multiple sets of candidate personalized adjustment schemes are output. Each candidate scheme is accompanied by the expected return, risk assessment and model confidence interval. An interpretable evidence chain is generated for each final candidate scheme. The candidate schemes and expected returns are output in the form of a structured report.

[0044] The clinical reports module integrates various information into a corresponding patient's clinical report and displays it visually, while also recording the doctor's adoption and actual therapeutic effect.

[0045] Reference Figure 2An optimization method for the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) based on multi-omics joint evaluation is described, and the specific steps of this optimization method are as follows: Candidate patients were screened according to pre-established inclusion criteria, and written informed consent was obtained from the patients. Various clinical data were recorded, and endoscopic examinations were performed to collect images and endoscopic scores.

[0046] Design and implement unified sampling, preprocess the sampled multi-omics raw data, and construct a unified feature matrix and metadata table based on the processed data of each group.

[0047] We constructed a patient-specific virtual biological model, formulated a fritillaria cirrhosa dosing regimen based on clinical guidelines and individual risks, and sampled actual observations after dosing in real time.

[0048] The actual observed values ​​at the same time point are compared with the predicted values ​​of the patient's individualized virtual biology model, and a list of candidate targets is generated.

[0049] Based on actual observations, the topological and dynamic parameters of the patient-specific virtual biological model are updated to establish a mapping from the current patient simulation state to the future therapeutic phenotype.

[0050] Parallel simulations are performed using patient-specific virtual biological models to output multiple sets of feasible personalized strategy recommendations.

[0051] The data were integrated into clinical reports for the corresponding patients, and the uncertainties, potential side effects, and recommended monitoring indicators were clearly marked in the reports.

[0052] Record the reasons for doctors' acceptance and rejection, and continue to sample according to the set time sequence after adjustments are made to monitor the actual efficacy and side effects, and update the patient's individualized virtual biological model.

Claims

1. A system for evaluating the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) based on multi-omics assessment, characterized in that: It includes a sample collection module, a quality control processing module, an extraction and annotation module, an atlas construction module, an individual simulation module, a treatment management module, a perturbation identification module, a learning and updating module, a driver identification module, an efficacy prediction module, an optimization simulation module, and a clinical reward module; The sample collection module executes patient screening, informed consent, and standardized sample collection procedures, specifies the collection time points, and standardizes the sampling type; The quality control processing module is used to perform standardized preprocessing and quality control on the raw multi-omics data and images or clinical values, while generating metadata and performing batch effect correction and missing value processing. The extraction and annotation module is used to extract biologically significant features from various omics levels and to structure clinical data and imaging features into standardized phenotypic labels. The graph construction module is used to integrate public knowledge bases and domain expert knowledge to construct a prior knowledge graph. The individual simulation module constructs an individualized virtual biological model based on the patient's baseline multi-omics input and prior knowledge graph; The treatment management module is used to collect continuous multi-omics and clinical data at designed time points under the current Fritillaria cirrhosa dosing regimen, and generate actual observation values. The perturbation identification module compares the actual observed values ​​with the predictions of the individualized virtual biology model at the corresponding time points to identify molecules or pathways that deviate from the model's prior expectations. The learning update module adaptively adjusts the network topology and dynamic parameters of the individualized virtual biology model based on each newly measured actual observation value; The driver identification module reprograms the network based on the update results of the individualized virtual biology model and the quantitative results of the perturbation, and generates a list of driver sub-lists. The efficacy prediction module defines and calculates the patient's multidimensional efficacy phenotype vector and performs cross-scale quantitative prediction based on a personalized virtual biology model. The optimization simulation module performs simulations based on a personalized virtual biology model and outputs candidate personalized adjustment schemes and their expected benefit assessments. The clinical report module integrates various information into a corresponding patient's clinical report and displays it visually, while also recording the doctor's adoption and actual therapeutic effect.

2. The system for treating ulcerative colitis with Fritillaria cirrhosa based on multi-omics evaluation according to claim 1, characterized in that, The specific steps of the extraction and annotation module in structuring clinical data and image features into standardized phenotypic labels are as follows: S1.1: Normalize the raw readings or peak areas of each patient sample according to the total amount in the sample, then summarize the total count of all features of each patient sample, and divide the count of each feature by the total count of the sample to obtain the relative abundance of each patient sample. Screen out patient feature data below the preset threshold, and calculate the geometric mean of the features in each patient sample and perform a centered logarithmic ratio transformation. S1.2: Calculate the diversity index within each patient sample and the distance matrix between each patient sample. Then, perform bootstrap sampling for each patient sample and estimate the confidence interval of the diversity index of each patient sample. After that, calculate the effect size and statistical test of each patient characteristic data to identify molecules, metabolites or proteins that are upregulated or downregulated. At the same time, use the BH method to calculate the false discovery rate (FDR) and retain patient characteristic data with FDR below a preset threshold and effect size above a preset threshold as candidate biomarkers. S1.3: Map gene or functional annotations to a predefined set of pathways, weight and summarize multiple genes or functional items within a pathway, calculate the pathway activity score of the pathway in each sample, standardize the pathway activity score between samples, output the pathway activity matrix for each patient sample, and then calculate the ratio between each metabolite based on known metabolic pathways and response relationships and perform logarithmic transformation. Weight and integrate multiple ratios to construct the corresponding composite metabolic index. S1.4: Preprocess endoscopic images, then use semantic segmentation algorithm to identify key structures in the mucosa, ulcer and bleeding areas, and manually or automatically verify the identification results. Then, calculate the morphological indicators of each structure based on the segmentation results, and extract the texture features and color histogram statistics in the segmented area to establish the image features of the corresponding patients. Then, perform scale normalization and repeatability tests on the image features of each patient, and generate the final image quantitative feature set. S1.5: Convert clinical questionnaires, physician scores and laboratory indicators into numerical fields based on a predefined dictionary or mapping. If there are classification options, they are mapped to ordered values. Then, scale transformation is performed on indicators of different dimensions, and min-max normalization is performed on each transformed indicator. Based on the preset clinical discrimination threshold, corresponding binary labels are generated. S1.6: Summarize the acquired features into a sample-feature matrix and record the meta-information of each feature. Then, filter and reduce the dimensionality of each set of features, and establish a standardized phenotypic label vector containing continuous variables and discrete labels. Record the unit, range and missing value handling strategy of each standardized phenotypic label vector. Finally, write the multi-omics data features of each patient sample into a unified structured format.

3. The system for treating ulcerative colitis with Fritillaria cirrhosa based on multi-omics evaluation according to claim 2, characterized in that, The specific steps for the graph construction module to construct the prior knowledge graph are as follows: S2.1: Manually set the resource types to be integrated and list accessible instances for each type of resource. Generate structured download entries for each resource by writing crawling or importing scripts, and put all entries into the database according to resource categories to form a source entry table. Perform naming standardization on each source entry, and then generate an initial search index for each record. After that, allocate metadata fields for each resource type, establish a corresponding candidate entity set for the collected source entries, and generate multiple sets of candidate standardized IDs for each candidate entity. S2.2: Perform multi-strategy matching on each candidate entity, including string matching, semantic matching, and ontology matching. Calculate the corresponding matching confidence score based on the matching rules, matching coverage, and consistency of each matching strategy. Then, weight and summarize the scores of each matching strategy according to resource weight and matching confidence to generate the comprehensive confidence score of each standard ID. Mark standardized IDs with a comprehensive confidence score lower than a preset threshold as requiring manual verification. Simultaneously, record the evidence chain for each mapping. Then, directly extract explicit relationships from the structured database and extract potential relationships from the literature through the information extraction pipeline. S2.3: Collect evidence fragments from different sources and methods for each pair of entities, record the source resources, evidence type, timestamp and credibility of each evidence fragment, generate an initial edge and its evidence score for each pair of entities through weighted aggregation of evidence, if there is evidence conflict, record both positive and negative evidence and score them separately, then label the edge with the original direction or type, and construct an evidence list for each pair of entities. S2.4: Load the disease ontology, phenotype ontology, and pathway ontology into a hierarchical structure, and implement hierarchical propagation for each edge. At the same time, retain the propagation path and attenuation coefficient during the propagation process of each ontology, and dynamically adjust the confidence of the corresponding edge based on the propagation results. Construct a set of functional annotations for each entity, use set theory or embedding space metrics to calculate the semantic similarity between each entity pair, and perform cluster analysis based on the obtained semantic similarity. Mark the inconsistent edges in the analysis results for manual review. S2.5: Count the number of positive and negative data points for each entity pair from the literature and experimental annotations, and perform Bayesian smoothing on each statistical result. Calculate the directional confidence probability of each statistical result. If the positive probability is higher than 0.7, mark the edge as inclined to positive causality and assign a directional confidence score. If the difference between the positive and negative probabilities is lower than a preset threshold, keep it undirected or mark it as bidirectional / uncertain. Based on the directional confidence score and evidence score, establish a directed edge weight that includes direction and strength. S2.6: Submit the automatically extracted list of edges and nodes to the domain expert group for annotation. Each expert gives their approval, disapproval or correction opinions for each edge, and can provide additional evidence or reasons for rejection. At the same time, based on the pre-assigned reputation coefficients of each expert, the annotation results are weighted and summarized. The confidence of evidence that is consistent with the expert opinions is increased, and the confidence of rejected edges is reduced or marked as excluded, so as to generate the final edge confidence. S2.7: Sort all edges according to the final edge confidence from high to low, and delete the corresponding edges of each node whose final edge confidence is lower than the preset threshold. Then, normalize the weights of the remaining directed edges to generate a prior knowledge graph. Export the prior knowledge graph in a standardized format. Each record in the prior knowledge graph includes: source node ID, target node ID, relation type, directional probability, final confidence, evidence list, expert comments, version number and timestamp.

4. The system for treating ulcerative colitis with Fritillaria cirrhosa based on multi-omics evaluation according to claim 3, characterized in that, The specific steps for constructing a personalized virtual biological model using the individual simulation module are as follows: S3.1: Based on the preset node type catalog, map all multi-omics data features of the patient's baseline to the candidate node set, collect the original observation set corresponding to each candidate node, compress the multi-source observations corresponding to each node by PCA projection according to the linear projection from the feature matrix to the node vector to generate the initial node feature vector, then perform scale correction on the feature vectors of each node, and evaluate the measurement reliability of each node. S3.2: Match the selected candidate node set with the prior knowledge image, extract the prior subgraph containing each node in the candidate node set and its direct neighbors, and use it as the initial prior topological skeleton of the current patient. Collect the evidence strength of each extracted prior edge in the prior knowledge graph and convert it into the corresponding prior connection weight. Then, adjust the corresponding prior weight according to the reliability of each node. S3.3: Based on the feature vectors of each node, establish a sample-level node feature matrix and calculate the empirical covariance corresponding to the matrix to obtain the joint variability among nodes. Then, based on each empirical covariance, use sparsified inverse covariance estimation to establish the corresponding precision matrix, infer potential direct interaction edges, and convert the precision matrix into a symmetric data-driven edge weight matrix, and perform significance or stability screening on the edges. S3.4: For each pair of nodes in the prior subgraph, use the prior weights and the data-driven edge weight matrix to perform weighted fusion, calculate the initial adjacency weights of each pair of nodes to establish the corresponding adjacency matrix, and adaptively adjust the fusion weights according to the node reliability, the credibility of the prior source, and the stability score of the edges in the data. After the fusion is completed, perform local normalization and sparsification on the generated adjacency matrix to generate the initial network topology skeleton. S3.5: Collect the initial unsigned adjacency weights of each group, and combine the relation types in the prior knowledge graph and the directional evidence in the data to calculate the initial symbol and its directed strength for each prior subgraph. Then, based on node attributes or biological common sense, assign corresponding directional priors to edges without clear direction, map the directed coupling strength to the dynamic coupling parameter space, generate the coupling matrix, and record the initial symbol confidence of each edge. S3.6: Set an initial baseline activity value for each node to represent the state of the node when it is not affected by coupling. Based on the time-varying index or measurement uncertainty of each node in the baseline data, set the response speed of each node. Then set the self-loop coefficient for each node. Use the node baseline activity value, response speed and self-loop coefficient as its node dynamic parameters and record the parameter confidence interval. S3.7: Combine the coupling matrix with the dynamic parameters of each node into a linearized dynamic operator and evaluate its spectral radius. If the spectral radius exceeds the preset critical value, scale the directional coupling strength according to the set ratio until the spectral radius is lower than the preset critical value. Then check all parameters within a reasonable range. After the check is passed, establish the corresponding patient's individualized virtual biological model based on the initialization results.

5. The system for treating ulcerative colitis with Fritillaria cirrhosa based on multi-omics evaluation according to claim 4, characterized in that, The specific steps of the perturbation identification module in identifying molecules or pathways that deviate from the prior predictions of the model are as follows: S4.1: Collect the actual observation values ​​of each actual sampling of the patient, as well as the predicted values ​​generated by the corresponding individualized virtual biology model at the prediction time point. Align the actual observation values ​​and predicted values ​​to a unified time grid. Estimate the approximate value of the actual observation on the model prediction time grid through spline interpolation and record the interpolation uncertainty. At the same time, establish the aligned observation matrix and prediction matrix, and label each alignment point with the original observation or interpolation estimation label. S4.2: Based on the analysis objectives and sampling frequency, select multiple time windows, summarize the residual sequences within each time window into multiple statistics, calculate the weighted mean residual of each node within each time window to obtain the corresponding window-level residual index, calculate the residual summary vector of each node in each time window and save the weight and sample number information, calculate the original p value of each node through a weighted paired t test, if the deviation of the original p value from 0 is higher than the preset deviation threshold, then it is determined that the actual observed value is significantly different from the model prediction; S4.3: Collect the original p-values ​​of all nodes and perform multiple test correction on them. At the same time, generate the original p-values ​​and the corrected q-values. Filter out nodes whose q-values ​​are less than the preset threshold and whose absolute weighted mean residuals are lower than the preset effect size lower limit. The remaining nodes are selected as candidate significant deviation nodes and a list of significant deviation nodes is generated. S4.4: Calculate the temporal difference of the weight of each edge in the patient's individualized virtual biological model between adjacent time windows, and perform statistical tests on the weight change of each edge. Edges with weight changes exceeding a preset change threshold are identified as significantly changed edges. At the same time, each significantly changed edge is mapped back to the path or pathway level, and the perturbed subnetwork is identified. S4.5: Use interrupted time series regression to fit a piecewise regression model ITS for each node, and in the ITS, construct a corresponding regression equation for each node, calculate the regression coefficients of instantaneous level change and slope change to obtain the corresponding causal effect, and estimate its significance. If there is potential confounding, adjust the corresponding covariates, and then confirm the consistency between the causal effect of each node and the correlation test results. S4.6: Map the causal effects obtained at the node level to their respective pathway sets based on pathway annotations, and then use the corresponding combination test method to integrate the node-level significance into the pathway-level significance. At the same time, calculate the summary statistic for each pathway and perform a statistical significance assessment. Based on the assessment results, report the set of pathways that are significantly perturbed. S4.7: Integrate the residual strength, significance, causal effect, edge weight change and path significance at the node level into a perturbation fingerprint report. At the same time, indicate the source and uncertainty of each piece of evidence in the report. Then, sort the nodes in the perturbation fingerprint report from high to low according to the strength of evidence.

6. The system for treating ulcerative colitis with Fritillaria cirrhosa based on multi-omics evaluation according to claim 5, characterized in that, The specific steps for evaluating the candidate personalized adjustment schemes and their expected returns output by the optimization simulation module are as follows: S5.1: Based on the observable description of each patient's individualized virtual biological model, establish the corresponding patient's state space, set the state vector dimension and standardization method, record the state naming convention and scale, and establish the corresponding action set based on the adjustable treatment parameter set. Then, use the patient's individualized virtual biological model as an environment simulator, set the current state and action as input, and the next state and immediate observation as output. Based on the efficacy prediction, side effect risk and treatment cost, design the corresponding immediate reward function. S6.2: Input each source of uncertainty into the environment simulator, and set the constraints of the environment simulator based on known biological knowledge. At the same time, run the simulation multiple times for the preset test actions, and statistically analyze the distribution of the next state to verify the behavior of the environment simulator. After the verification is passed, establish a target network based on the DQN architecture, extract multiple sets of historical patient data and input them into the environment simulator. Perform forward propagation through the target network to calculate the Q value of the corresponding action. S6.3: Based on the Q value of the target network output, calculate the loss value of the current behavior network parameters through the weighted batch loss function, and update the behavior network parameters using the gradient descent method. At the same time, periodically synchronize the behavior network parameters to the target network parameters. During the training process, put the interaction samples into the experience replay pool for the next round of training, and set the sampling priority according to the TD error of each experience from high to low. S6.4: An ε-greedy strategy is adopted as the behavior strategy. An action is randomly selected from the action space with probability ε for exploration. At the same time, the greedy action estimated by the current target network is selected with probability (1-ε). The exploration trajectory is recorded. After training, the current greedy strategy is run multiple times using an independent validation simulation set. The cumulative reward trajectory is recorded, and the expected long-term return of the strategy and its volatility are calculated. S6.5: Based on expected benefits, risk assessment and clinical constraints, output multiple sets of candidate personalized adjustment plans, and attach expected benefits, risk assessment and model confidence intervals to each candidate plan. At the same time, generate an interpretable chain of evidence for each final candidate plan, and output the candidate plan and expected benefits in the form of a structured report.

7. A method for optimizing the efficacy of Fritillaria cirrhosa in treating ulcerative colitis (UC) based on multi-omics evaluation, used to achieve the function of the Fritillaria cirrhosa efficacy system for treating UC based on multi-omics evaluation as described in any one of claims 1-6, characterized in that, The specific steps of this optimization method are as follows: I. Screen candidate patients according to the pre-established inclusion criteria, obtain written informed consent from patients, record various clinical data, and perform endoscopic examinations and collect images and endoscopic scores. II. Design and implement unified sampling, preprocess the sampled multi-omics raw data, and construct a unified feature matrix and metadata table based on the processed data of each group; III. Construct individualized virtual biological models for patients, formulate fritillaria cirrhosa dosing regimens based on clinical guidelines and individual risks, and sample actual observations after dosing in real time; IV. Compare the actual observed values ​​at the same time point with the predicted values ​​of the patient's individualized virtual biology model, and generate a list of candidate targets; V. Based on actual observations, update the topology and dynamics parameters of the patient's individualized virtual biological model and establish a mapping from the current patient simulation state to the future therapeutic phenotype; VI. Parallel simulations are performed using individualized virtual biological models of patients to output multiple sets of feasible personalized strategy recommendations; VII. Integrate the data into clinical reports for the corresponding patients, and clearly indicate uncertainties, potential side effects, and recommended monitoring indicators in the reports; VIII. Record the reasons for doctors' acceptance and rejection, and continue to sample according to the set time sequence after the adjustment is implemented to monitor the actual efficacy and side effects, and update the patient's individualized virtual biological model.

Images