A virtual rat experimental drug hepatotoxicity evaluation method and device based on a modal perception hybrid expert and dose-time-dependent generative adversarial network

Abstract

This invention discloses a method and apparatus for assessing drug hepatotoxicity in virtual rat experiments based on a modality-aware hybrid expert and a dose-time-dependent generative adversarial network. The method includes: extracting QSAR chemical descriptors, QSIIR bioactivity features, and PSMD pathway scoring features to construct a multimodal fusion feature set; constructing a treatment condition matrix based on dose and exposure time; inputting the multimodal fusion feature set and treatment condition matrix into a virtual rat experimental model constructed based on a modality-aware hybrid expert and a dose-time-dependent generative adversarial network to generate a probability density distribution of liver function biomarkers; and achieving dose-time-effect causal inference and risk stratification assessment. This invention solves the problems of high cost, long cycle, high ethical pressure, difficulty in observing rare events in traditional animal experiments, and insufficient multimodal fusion and weak distribution generation ability of existing models, achieving efficient and interpretable virtual assessment of drug hepatotoxicity.

Description

Technical Field

[0001] This invention belongs to the field of medicine and public health technology, specifically involving the intersection of toxicology, data science and artificial intelligence, and particularly the application of generative deep learning in drug hepatotoxicity prediction and virtual animal experiment construction. Specifically, it relates to a method and device for evaluating drug hepatotoxicity in virtual rat experiments based on modality-aware hybrid experts and dose-time dependent generative adversarial networks. Background Technology

[0002] Drug-induced liver injury (DILI) is a significant cause of drug development failures and post-marketing withdrawals. The liver, as a core metabolic organ, is a primary target organ for the toxic effects of exogenous chemicals. In the early stages of drug development, in vivo toxicology studies, including but not limited to SD rats, are typically conducted to assess the risk of hepatotoxicity.

[0003] Traditional animal experiments are limited by the following bottlenecks: high cost and long cycle; increasing ethical pressure on laboratory animals; significant individual variability in biomarker distribution; difficulty in covering extreme doses or prolonged exposure scenarios; limited sample size and insufficient statistical stability. Often, oral medications that trigger hepatotoxicity warnings are only observed after clinical trials and drug launch, with an incidence of approximately 1-20 cases per 200,000 people. Meanwhile, existing in vitro models or traditional machine learning methods, including but not limited to random forests, support vector machines, and single neural networks, typically rely on limited structural features or single-modal data, exhibiting the following shortcomings: inability to handle multimodal inputs, including but not limited to chemical structures, dosage information, biochemical indicators, and time factors; insufficient ability to model complex nonlinear toxicity mechanisms; difficulty in generating continuous biomarker data that conforms to the true distribution; low probability of generating extreme toxicity events, including but not limited to severe DILI; and distribution shrinkage issues.

[0004] Generative Adversarial Networks (GANs) have advantages in data generation, but traditional GANs suffer from problems such as mode collapse, insufficient fusion of multimodal information, and unbalanced weights in different feature spaces, making them difficult to use directly for the stable generation of high-dimensional toxicology data.

[0005] Therefore, it is necessary to construct a method and system for assessing hepatotoxicity in virtual rat experiments that can: process multimodal input information, automatically perform modality weight perception, allocate specialized expert networks for different toxicity patterns, generate virtual experimental data that conforms to real biostatistical distributions, and simultaneously determine toxicity levels and risk stratification, for use in large-scale virtual rat experiments. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a virtual rat experimental drug hepatotoxicity assessment method and device based on a modality-aware hybrid expert dose-time dependent generative adversarial network. This method can process multimodal input information, automatically perform modality weight perception, allocate specialized expert networks for different toxicity patterns, generate virtual experimental data that conforms to real biostatistical distributions, and simultaneously achieve toxicity level determination and risk stratification. It is suitable for the preclinical evaluation stage in the early stages of oral drug development.

[0007] The present invention adopts the following technical solution.

[0008] The first aspect of this invention provides a method for assessing the hepatotoxicity of experimental drugs in virtual rats based on modality-aware hybrid experts and dose-time dependent generative adversarial networks, comprising:

[0009] QSAR chemical descriptors, QSIIR bioactivity features, and PSMD pathway score features were calculated using the simplified molecular linear input norms of the target oral drug, and a multimodal fusion feature set was constructed.

[0010] Obtain the user-set dose and exposure time, and construct a treatment condition matrix based on the dose and exposure time;

[0011] The multimodal fusion feature set and processing condition matrix are input into a pre-constructed virtual rat experimental model, and the output is large-scale sample data for drug hepatotoxicity experiments. The sample data includes the probability density distribution of liver function biomarkers. The virtual rat experimental model is constructed based on a modality-aware hybrid expert and a dose-time dependent generative adversarial network.

[0012] Based on the probability density distribution of liver function biomarkers and the preset rules for judging liver injury events, it is determined whether the sample data has experienced a preset liver injury event or abnormal liver function biomarkers.

[0013] Using the occurrence of a pre-specified liver injury event or abnormal liver function biomarkers as the endpoint, probability curves were plotted for different dose groups to maintain a normal liver state, and the hepatotoxicity of the drug was assessed based on the probability curves.

[0014] Optionally, the generative adversarial network includes a modality-aware hybrid expert and a dose-time dependent generator and a conditional discriminator; the multimodal fusion feature set and processing condition matrix are input as conditions and random noise into the generator to generate sample data for drug hepatotoxicity experiments; under the same conditions, the conditional discriminator evaluates the consistency of the distribution of real experimental data and generated sample data; the generator includes a gating network and a hierarchical expert network, the weights of the hierarchical expert networks are dynamically allocated through the gating network, and the outputs of each hierarchical expert network are weighted and summed based on the weights to obtain the sample data.

[0015] Optionally, the conditional discriminator includes a two-stage processing unit, namely a feature extraction unit and a true / false discrimination unit. The feature extraction unit is used to concatenate the original features of the three modalities with the exposure time and dose into a one-dimensional vector, and input it into the first three-layer fully connected network for feature extraction. The true / false discrimination unit is used to concatenate the extracted true / false features with the original biomarker measurement values ​​to be discriminated, and input them into the second three-layer fully connected network to output a scalar discrimination score.

[0016] Optionally, the method further includes:

[0017] Based on QSAR chemical descriptors, QSIIR bioactivity characteristics, and PSMD pathway scoring characteristics, the original feature matrices of QSAR, QSIIR, and PSMD modes were constructed, respectively.

[0018] The corresponding encoding networks are constructed for the original feature matrices of QSAR, QSIIR, and PSMD modes, respectively, and the corresponding mode encoding vectors are output. A multimodal fusion feature set is constructed based on the corresponding mode vectors. The mode encoding vectors have the same dimension. Each encoding network includes an input layer, a first fully connected layer, a ReLU activation function, a Dropout regularization layer, a second fully connected layer, and an output layer connected in sequence.

[0019] Optionally, constructing a treatment condition matrix based on dose and exposure time includes:

[0020] After logarithmically normalizing the exposure time, Fourier transform is performed to obtain Fourier features. The Fourier features are then mapped to the temporal embedding space through a two-layer fully connected network.

[0021] After the dose is broadcast and extended into a one-dimensional vector, it is nonlinearly transformed through a two-layer fully connected network and then mapped to a dose embedding space with the same dimension as the time embedding space.

[0022] The processing condition matrix is ​​obtained by concatenating the time embedding space and the dose embedding space.

[0023] Optionally, PSMD pathway score features can be calculated using the simplified molecular linear input specification of the obtained target oral drug, including:

[0024] Based on CTD, liver injury-related targets associated with the target oral drug were identified and the protein structures of these targets were predicted.

[0025] The molecular docking of multiple potential active cavities with the highest probability of binding to the relevant target protein structure with the target oral drug was performed to obtain the binding energy matrix, and the interaction spectrum between the drug and the target was constructed based on the binding energy.

[0026] The comprehensive score of each pathway is extracted based on the drug-target interaction spectrum and used as the PSMD pathway scoring feature.

[0027] Optionally, the comprehensive score of each pathway extracted based on the drug-target interaction spectrum includes:

[0028] Based on the drug-target interaction spectrum, pathway enrichment analysis was performed to obtain standardized enrichment scores. The significance of the observed pathway enrichment in the interaction spectrum was calculated to obtain statistical significance. ;

[0029] Based on standardized enrichment scores and statistical significance Calculate the overall score of the corresponding pathway. :

[0030] ,

[0031] in, It indicates the degree to which the pathway is activated or inhibited.

[0032] Optionally, constructing a multimodal fusion feature set includes:

[0033] Calculate the Person correlation coefficients between features in the feature candidate library and liver function biomarkers;

[0034] The probability of a current feature being correlated with liver function biomarkers is calculated based on the Person correlation coefficient, assuming the assumption that the feature is not related to liver loss. This probability serves as the significance level for the correlation test of the feature.

[0035] Features that have a significance level of less than a preset significance level threshold and a Person correlation coefficient greater than a preset correlation threshold are selected from the feature candidate library to form a feature subset;

[0036] After standardizing the feature subset, a multimodal fusion feature set is constructed.

[0037] Optionally, the method further includes:

[0038] The virtual rat experimental model outputs sample data corresponding to various white blood cell ratios; sample data whose total white blood cell ratio falls within a preset threshold range are determined to be valid sample data.

[0039] A second aspect of the present invention provides a virtual rat experimental drug hepatotoxicity assessment device based on modality-aware hybrid experts and dose-time-dependent generative adversarial networks, realizing the above-mentioned virtual rat experimental drug hepatotoxicity assessment method based on modality-aware hybrid experts and dose-time-dependent generative adversarial networks, the device comprising:

[0040] The multimodal feature extraction module is used to calculate QSAR chemical descriptors, QSIIR bioactivity features, and PSMD pathway score features using the simplified molecular linear input specification of the acquired target oral drug, and to construct a multimodal fusion feature set;

[0041] The dynamic exposure condition setting module is used to obtain the dose and exposure time set by the user, and construct a treatment condition matrix based on the dose and exposure time;

[0042] The generation module is used to input the multimodal fusion feature set and processing condition matrix into a pre-built virtual rat experimental model and output sample data for drug hepatotoxicity experiments. The sample data includes the probability density distribution of liver function biomarkers. The virtual rat experimental model is built based on a modality-aware hybrid expert and dose-time dependent generative adversarial network.

[0043] The determination module is used to determine whether a preset liver injury event or abnormal liver function biomarkers have occurred in the sample data based on the probability density distribution of liver function biomarkers and preset liver injury event determination rules.

[0044] The evaluation module is used to plot probability curves of the liver maintaining a normal state in different dosage groups, with the occurrence of a preset liver injury event or abnormal liver function biomarkers as the endpoint, and to evaluate the hepatotoxicity of the drug based on the probability curves.

[0045] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0046] (1) Multimodal deep fusion: For the first time, three-dimensional heterogeneous data integration of molecular descriptor (QSAR), bioactive sequence feature (QSIIR) and pathway score based on molecular docking (PSMD) is realized. The contribution of each modality is dynamically adjusted through a hierarchical expert network, providing a systematic computational framework for liver toxicity risk assessment that surpasses the traditional single chemical descriptor.

[0047] (2) Complex toxicokinetics characterization: comprehensively characterize the dose-time dependent response spectrum across exposure scenarios and reveal dose-time-effect interaction mechanisms that are difficult to identify by traditional computational models.

[0048] (3) Breakthrough in rare event identification: Large-scale Monte Carlo sampling combined with generative distribution modeling significantly improves the detection efficiency of low-incidence adverse liver outcomes. Through multi-batch generation, uncertainty quantification and confidence boundaries are provided, which effectively captures the heterogeneity and extreme values ​​among real biological individuals compared with traditional single-point prediction models.

[0049] (4) Regulatory transformation and application value: The reliable cross-dataset generalization ability has been independently verified by external parties and conforms to the 3Rs principle (replacement, reduction, optimization). It can directly reduce the use of approximately tens of thousands of rats, laying a methodological foundation for incorporating computational toxicology into the regulatory decision-making framework and reducing the dependence on animals for in vivo animal toxicity testing.

[0050] (5) Innovation of virtual experiment paradigm: Breaking through the limitations of traditional static point prediction in computational toxicology, realizing the methodological transformation from "discriminative" to "generative", replacing "point prediction" with "distribution prediction" and replacing "single structure description" with "multimodal mechanism modeling", while ensuring the generalization ability across datasets, it provides a scalable data-driven solution for early prediction of drug hepatotoxicity, risk assessment of rare events and animal experiment substitution.

[0051] (6) Visualization and interpretability: Construct a dual heatmap / curve visualization system to intuitively display the abnormal patterns of biomarkers and the inference results of clinical events. Support multi-dimensional collaborative interpretation mechanisms such as horizontal comparison, vertical comparison, diagonal pattern and blank area identification to improve the interpretability of toxicity mechanisms. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall process of the virtual rat experimental drug hepatotoxicity assessment method based on modality-aware hybrid experts and dose-time dependent generative adversarial networks as described in this invention.

[0053] Figure 2 This is a schematic diagram of the core structure of the virtual rat experimental model described in this invention;

[0054] Figure 3 This is a schematic diagram of the inclusion conditions and feature calculation process of training data in the virtual rat experiment described in this invention. It clarifies the data inclusion criteria such as test rats, control settings, biomarkers, and exposure time, as well as the calculation sources and pathway enrichment analysis results of the three types of features: QSAR, QSIIR, and PSMD.

[0055] Figure 4 This is a schematic diagram of the variable screening results of the virtual rat experiment described in this invention, showing the correlation between the Top 2 features of each modality of QSAR, QSIIR, and PSMD and liver function biomarkers, as well as the association between features, under different dose levels and exposure time combinations.

[0056] Figure 5 This is a schematic diagram of the training process indicators monitoring of the virtual rat experimental model described in this invention;

[0057] Figure 6This is a schematic diagram of the data quality control process and internal verification results after prediction by the virtual rat experimental model described in Embodiment 1 of the present invention. It shows the white blood cell distribution, the quality control results of various liver function biomarkers, as well as the cosine similarity between the generated values ​​and the true values, the RMSE distribution, and the consistency score of multiple batches of experiments.

[0058] Figure 7 This is a schematic diagram of the prediction results of the traditional control model described in Embodiment 1 of the present invention, showing the predictive effects of classic models such as LR, PLS, ARD, SVR, GPR, RF, AdaB, XGB, LGB, and MLP on various liver function biomarkers.

[0059] Figure 8 This is a schematic diagram of the external verification results of the virtual rat experimental model described in Embodiment 1 of the present invention, showing the comparison between the generated values ​​and the true values ​​of indicators such as ALT, ALP, TBiL, AST, LDH, RALB, and A / G, as well as the distribution verification results of cosine similarity and RMSE.

[0060] Figure 9 This is a schematic diagram of the results of 120,000 virtual simulation experiments of acetaminophen as described in Example 2 of the present invention, showing the abnormal trends of various liver function biomarkers of the drug, as well as the occurrence patterns of hepatocellular damage and cholestasis events.

[0061] Figure 10 This is a schematic diagram of the results of 120,000 virtual simulation experiments of abacavir sulfate as described in Example 3 of the present invention, showing the changes in the drug's biomarkers and the characteristics of various liver injury events;

[0062] Figure 11 This is a schematic diagram of the results of 120,000 virtual simulation experiments of felbamate described in Example 4 of the present invention, showing the abnormal biomarkers and the development trend of liver injury events corresponding to the full spectrum of hepatotoxicity of this drug.

[0063] Figure 12 This is a schematic diagram of the results of 120,000 virtual simulation experiments of propoxyphene hydrochloride as described in Example 5 of the present invention, showing the changes in biomarkers and characteristics of liver injury events corresponding to the contradictory toxicity signals of the drug.

[0064] Figure 13 This refers to the feature analysis and biological interpretation described in Embodiment 7 of the present invention;

[0065] Figure 14 Visualization of the gated network weights as described in Embodiment 7 of the present invention; the VREM model automatically adjusts the weights under different processing conditions, and the contribution of multimodal experts to the virtual experiment results changes dynamically;

[0066] Figure 15This is a schematic diagram of the module architecture of the virtual rat experimental drug hepatotoxicity assessment device based on modality-aware hybrid experts and dose-time dependent generative adversarial networks as described in this invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0068] While animal studies remain central to preclinical drug safety assessments, they have inherent limitations in reproducing hepatotoxicity phenotypes in humans. This is particularly true for rare but clinically serious hepatotoxic events, where sampling and characterization are often inadequate in traditional experimental designs. Furthermore, even under rigorously standardized conditions using animals of the same strain, age, and sex, significant inter-laboratory variability persists. These variability stems from subtle but systematic variations in husbandry conditions, experimental procedures, analytical platforms, and sampling strategies. This experimental heterogeneity significantly diminishes the reproducibility of toxicology studies and limits their generalizability in independent research.

[0069] Combination Figure 1 As shown, this invention provides a virtual rat experimental drug hepatotoxicity assessment method based on modality-aware hybrid experts and dose-time dependent generative adversarial networks. Figure 1 This demonstrates the complete S1-S6 steps from multimodal feature extraction and screening, dynamic exposure condition parameterization, dose-time-effect causal inference, and risk stratification:

[0070] Step S1, Multimodal Feature Extraction and Screening: Obtain the Simplified Molecular Input Line-Entry System (SMILES) of the target oral drug, and automatically calculate and screen three types of heterogeneous features:

[0071] QSAR (Quantitative Structure-Activity Relationship) chemical descriptors: Based on RDKit, Mordred, and MACCS (Molecular Access System) molecular fingerprints, they extract two-dimensional topological structures, physicochemical properties, and functional group distribution characteristics; among them, RDKit and Mordred are open-source cheminformatics libraries.

[0072] QSIIR (Quantitative Structure In vitro-In vivo Relationship) bioactivity profiles: Based on the Chemical Checker database, a 25-dimensional in vitro-in vivo correlation bioactivity profile was extracted.

[0073] PSMD (Pathway Scores from Molecular Docking) pathway perturbation scoring features: Based on the CTD (Comparative Toxicogenomics Database), liver injury-related targets associated with the drug are obtained. AlphaFold4 is used to predict the protein structures of these targets. Then, molecular docking is used to simulate the binding energy between the drug and the relevant target protein structures. Based on the binding energy, a list of affected genes is determined. This gene list is mapped to a biological pathway database, and pathway enrichment analysis is performed to obtain the pathway perturbation score. A low binding energy indicates that the drug may indeed affect the protein. Pathway perturbation scores obtained through pathway set analysis include calculating whether a specific pathway is significantly activated or inhibited, thus obtaining a perturbation score for that specific pathway.

[0074] The features are subjected to multiple standardizations to construct a multimodal fusion feature set.

[0075] Preferably, step S1 automatically calculates and filters QSAR chemical descriptors, QSIIR bioactivity characteristics, and PSMD pathway scoring characteristics, specifically including:

[0076] Step S11, Feature Variable Standardization Mapping: Obtain a list of standard variables for the built-in preset QSAR chemical descriptors, QSIIR bioactivity characteristics, and PSMD pathway score characteristics. The list of standard variables was determined in the early model building stage by screening a subset of features that are significantly associated with liver function biomarkers from the candidate feature library through Pearson correlation analysis combined with significance test (p < 0.05), where p represents the significance level of the correlation test, i.e., the probability that the observed correlation is caused by random sampling error. In subsequent model applications, the consistency between the feature variables and the list of standard variables is enforced to ensure the repeatability of cross-batch predictions and the comparability of model outputs.

[0077] Specifically, the subset of features that are significantly associated with liver function biomarkers is selected from the candidate feature library, including:

[0078] Calculate the Person correlation coefficients between features in the feature candidate library and liver function biomarkers;

[0079] The probability of a current feature being correlated with liver function biomarkers is calculated based on the Person correlation coefficient, assuming the assumption that the feature is not related to liver loss. This probability serves as the significance level for the correlation test of the feature.

[0080] Features that have a significance level of less than a preset significance level threshold and a Person correlation coefficient greater than a preset correlation threshold are selected from the feature candidate library to form a feature subset.

[0081] Optionally, the probability of correlation between the current feature and liver function biomarkers is calculated based on the Person correlation coefficient, assuming the assumption that the feature is not related to liver loss holds. The significance level for the correlation test of this feature includes:

[0082] The t-test was performed on the Person correlation coefficient to calculate the t-value.

[0083] The p-value is obtained by looking up the t-distribution table based on the t-value and used as the significance level for the correlation check of this feature.

[0084] Step S12, QSAR chemical descriptor calculation and screening: Based on the SMILES representation of the target drug, the RDKit library is used to calculate the two-dimensional topological descriptor and MACCS molecular fingerprint, and the Mordred library is used to calculate the ring system structure and macroscopic physicochemical descriptor; the calculation results are mapped to the standard variable list described in step S11, the feature values ​​of the corresponding dimensions are extracted, and after missing value imputation and outlier processing, a standardized QSAR feature vector is formed.

[0085] Step S13, Calculation and Screening of QSIIR Bioactivity Features: Based on the SMILES representation of the target drug, the Signaturizer tool of the Chemical Checker database is called to calculate a 25-dimensional bioactivity fingerprint covering molecular targets, signaling networks, cell phenotypes and clinical levels; the calculation results are mapped to the standard variable list described in step S11, and the feature values ​​of the corresponding dimensions are extracted to form a standardized QSIIR feature vector.

[0086] Step S14, Calculation and Screening of PSMD Pathway Scoring Features—A Pathway Perturbation Scoring Method Based on Molecular Docking: Specifically includes:

[0087] S141. Construction of a liver injury target library: Using "Chemical and Drug Induced Liver Injury" as the search term, a set of liver injury-related genes with direct evidence (marker / mechanism / therapeutic) was obtained from the Comparative Toxicology Genomics Database (CTD) to construct a hepatotoxicity target library containing 302 core targets.

[0088] S142. Target protein structure acquisition and preprocessing: Download the three-dimensional structure of the protein corresponding to the target library from the AlphaFold protein structure database, and select the predicted conformation with the highest pLDDT score; perform structure preprocessing using PyMOL to retain single polypeptide chains and remove solvent molecules and co-crystal ligands, and convert it into a PDBQT format receptor file after hydrogenation processing using MGLTools; that is, process the downloaded raw protein structure file into a standard input file that can be recognized.

[0089] S143. Potential binding cavity prediction and parameterization: Based on the CB-Dock algorithm, the center coordinates and spatial size parameters of the five potential active cavities with the highest prediction scores for each target protein are extracted, and docking parameter configuration files are constructed.

[0090] S144. High-throughput molecular docking and binding energy matrix construction: The three-dimensional conformation of the target drug constructed by RDKit and optimized by MMFF force field is molecularly docked with the five active cavities of the 302 target sites to obtain the binding energy matrix; the optimal binding energy in the five cavities of each target site is retained to construct the drug-target interaction spectrum.

[0091] S145. Pathway enrichment analysis and score calculation: Based on the drug-target interaction spectrum, KEGG pathway and Gene Ontology enrichment analysis is performed using the GSEApy tool to obtain standardized enrichment scores, i.e., pathway perturbation scores. The combined score of each pathway is extracted as a quantitative characterization of pathway activation intensity. The calculation results are mapped to the standard variable list described in step S11 to form a standardized PSMD pathway score feature vector.

[0092] The comprehensive score for each pathway is calculated as follows:

[0093]

[0094] in, This indicates the overall score. Represents the standardized enrichment score. This indicates the degree to which the pathway is activated or inhibited. When the pathway is activated, A value of 1 indicates that the pathway is inhibited. -1 is used in pathway enrichment analysis. The positive and negative directions of the enrichment result allow the comprehensive score to reflect both the statistical significance of the enrichment result and the direction of the biological effect of the pathway being activated or inhibited. This indicates the statistical significance of the enrichment analysis.

[0095] Step S15, Feature-level Standardization—Eliminating Intramodal Dimensional Differences: Perform Min-Max Scaling on QSAR chemical descriptors, QSIIR bioactivity features, and PSMD pathway score features respectively, linearly mapping each modal feature value to the [0,1] interval:

[0096]

[0097] in, These are the original eigenvalues. and These are the minimum and maximum values ​​of the feature in the training set, respectively.

[0098] Step S16, Target Variable Transformation – Mitigating Biomarker Distribution Skewness: Perform log-space transformation on the maximum-minimum standardization of the true measurements of the nine liver function biomarkers to compress the dynamic range of high-value regions and mitigate the interference of right-skewed distribution on model training.

[0099]

[0100]

[0101] in, This represents the logarithmic spatial transformation value of the original biomarker measurement. The original biomarker measurements are used, and the constant 1 is introduced to ensure non-negativity constraints; the model is trained using the transformed values. As a monitoring signal; This represents the minimum value of the logarithmic space transformation. This represents the maximum value of the logarithmic space transformation.

[0102] Step S17, Inverse Transformation and Truncation—Ensuring the Physiological Rationality of the Predicted Output: The Model Generates Predicted Values Then, an inverse exponential transform is performed and a non-negativity truncation is applied to ensure that the final output conforms to the physiological constraints of the biomarker:

[0103]

[0104] in, This represents the final predicted value after inverse transformation and truncation. This represents the predicted drug hepatotoxicity value generated by the model.

[0105] Step S18: The model input includes three types of heterogeneous compound feature matrices: quantitative structure-activity relationship (QSAR), quantitative structure-in vitro-in vivo association (QSIIR), and pathway perturbation score (PSMD). The original dimensions and distribution characteristics of different modalities are significantly different. Direct splicing may result in scale inconsistency and feature space misalignment. To solve this problem, this invention has specially designed encoding networks for each modality and mapped them to the same d-dimensional feature space. In this invention, d=256 is chosen.

[0106] Each encoding network consists of two fully connected layers connected by a ReLU activation function and a Dropout regularization layer (with a dropout rate of 0.25). The structure of the encoding network includes, in sequence, an input layer, a first fully connected layer, a ReLU activation function, a Dropout regularization layer, a second fully connected layer, and an output layer.

[0107] Taking QSAR mode as an example, the encoding process is expressed as follows:

[0108]

[0109] in, Represents the QSAR mode coding vector. This represents the weight matrix of the second fully connected layer. Represents the ReLU activation function. This represents the weight matrix of the first fully connected layer. This represents the original feature matrix of the QSAR modes. This represents the bias vector of the first fully connected layer. This represents the bias vector of the second fully connected layer. This represents the original feature matrix of the QSIIR modes. This represents the QSIIR mode coding network. This represents the QSIIR mode coding vector. This represents the original feature matrix of the PSMD mode. This represents the PSMD modal coding network. This represents the PSMD modality encoding vector. The three encoding networks have the same structure, but their parameters are independent of each other. This preserves the information specific to each modality and achieves alignment of the feature space.

[0110] As one of the most prominent and substantial features of this invention, the PSMD pathway scoring feature is an original toxicity mechanism characterization method. It provides a novel molecular interaction feature description for existing computational toxicology modeling. Unlike traditional pathway enrichment methods based on gene expression or literature mining, this invention does not rely on microarray sequencing, thus reducing experimental costs.

[0111] Step S2, Dynamic Exposure Condition Parameterization Setting: Set the drug treatment condition matrix: dose level (low, medium, high) × exposure time (4 days, 8 days, 15 days, 29 days), forming 12 standardized treatment combinations, and map the dose scalar and time scalar to a high-dimensional embedding space.

[0112] Preferably, the parameterization setting of the dynamic exposure conditions in step S2 specifically includes:

[0113] Step S21, Construction of the treatment condition matrix: Based on the toxicological repeated-dose experimental specifications, the treatment condition matrix is ​​set as shown in Table 1:

[0114] Table 1

[0115] Twelve standardized treatment combinations were developed, each corresponding to an independent experimental group in real animal experiments.

[0116] Step S22, Temporal Embedding – Log-Compressed Fourier Feature Encoding: To address the problem of insufficient short-term time point discrimination caused by the large time span (4-29 days) in toxicology experiments, a logarithmic transformation is used to compress the dynamic range of the long-tailed distribution, combined with Fourier basis functions to achieve periodic feature encoding. This allows the model to distinguish subtle changes in the early stages and grasp the long-term trends in the later stages, solving the problems of linear bias and difficulty in learning the intrinsic semantics of time points caused by insufficient short-term time point discrimination. The logarithmic transformation relatively amplifies the time distance in the logarithmic space, making it easier for the model to capture rapid changes in the early stages: For acute liver injury signals appearing 4-7 days after drug administration, the model can effectively capture and amplify them, without ignoring subtle changes in the early stages due to the 29-day time span. For chronic fibrosis or steatosis signals at 21-29 days, the model can characterize this stable, long-term change through low-frequency components.

[0117] S221. Logarithmic normalization: Normalizes the original time... Mapping to normalized intervals:

[0118]

[0119] in, For normalized time, The maximum exposure time is used; logarithmic transformation ensures that the model has higher resolution during shorter exposure periods (4-8 days) while avoiding severe oscillations of high-frequency components caused by long-term input.

[0120] S222. Fourier Feature Calculation: Based on a predefined frequency basis ωk = 2^k·π (k = 0, 1, …, K-1, K = 16), where k represents the total number of predefined frequency bases, i.e., there are 16 different frequencies, calculate the sine-cosine Fourier features. :

[0121]

[0122] In this way, Fourier encoding transforms scalars into high-dimensional vectors, making time points that are difficult to distinguish in low-dimensional space mapped to completely different positions in high-dimensional vectors, thus making them easier for the model to distinguish.

[0123] S223, Nonlinear Projection: Fourier features are mapped to... through a two-layer fully connected network. 3D time-embedded space:

[0124]

[0125] in, The temporal embedding vector represents a d-dimensional continuous vector that encodes temporal information and is used to inject time step information into the model. The temporal multilayer perceptron is a nonlinear transformation network consisting of two fully connected layers (containing SiLU activation functions), which is responsible for further mapping Fourier features to the target embedding space. represents the Fourier feature, which is a high-dimensional feature vector after sinusoidal / cosine positional encoding of the original time scalar; t represents the time step, which represents the time variable; d represents the embedding dimension, which is the vector dimension shared by the time embedding and the dose embedding.

[0126] Step S23, Dose Embedding – Broadcast Extended Nonlinear Transformation: Transforming the dose scalar (Corresponding to low, medium, and high levels respectively) After performing broadcast extension, a nonlinear transformation is performed through two fully connected layers to map to the dose embedding space of the same dimension as the time embedding:

[0127]

[0128] in, The dose embedding vector represents a d-dimensional continuous vector encoding the dose level, consistent with the time embedding dimension for subsequent fusion. () represents a dose multilayer sensing machine, a nonlinear transformation network consisting of two fully connected layers (containing SiLU activation functions), responsible for mapping dose features to... Same-dimensional embedding space, d represents the dose tensor after broadcast expansion; d represents the embedding dimension, which is the vector dimension shared by the temporal embedding and the dose embedding.

[0129] Step S3: Large-scale virtual experiment based on modality-aware hybrid expert generative adversarial network: Input the multimodal fusion feature set described in Step S1 and the dynamic exposure conditions described in Step S2 into the pre-trained Virtual Rat Experiment Model (VREM), the model comprising:

[0130] Modality-aware hybrid expert generator: The activation weights of 7 hierarchical experts are dynamically allocated through a gating network, multi-source information is adaptively fused, and random noise is combined to generate a 9-dimensional liver function biomarker vector; the 7 hierarchical experts include 3 single-modal experts, 3 bimodal interactive experts, and 1 trimodal joint expert.

[0131] Conditional discriminator: evaluates the consistency between the distribution of generated data and real experimental data;

[0132] For each treatment combination, 10,000 Monte Carlo samplings were performed, resulting in a total of 120,000 iterative predictions, generating the complete probability density distribution of liver function biomarkers.

[0133] Preferably, the MA-MoE Generator (Modality-Aware Mixture-of-Experts Generator) in step S3 specifically includes:

[0134] Step S31, Multimodal Feature Encoding and Alignment: Receive the three-modal feature representations encoded in step S18. And the time embedding output of steps S22 and S23 Dosage embedding ;in, ;

[0135] Step S32, Gating Network – Dynamic Expert Weight Calculation: Construct a gating network to achieve sparse activation for expert selection. The gating network takes the concatenated vector of modal features and conditional embeddings as input, and deliberately excludes noise vectors to ensure the stability and interpretability of weight allocation.

[0136]

[0137] in, The gating input vector is a high-dimensional vector composed of multimodal features and conditional embeddings, serving as the input representation for the gating network. This represents the QSAR modal feature vector, a modality-specific characterization extracted from the drug molecule structure-activity relationship branch. This represents the QSIIR modal feature vector, a modality-specific characterization extracted from the molecular structure-immune response relationship branch. This represents the PSMD modal feature vector, a characterization extracted from the third modal branch, such as protein structure or molecular dynamics.

[0138] The N=7 dimensional weight vector is calculated using a two-layer fully connected network, and the expert activation probability distribution is obtained by Softmax normalization.

[0139]

[0140] in, Represents the gating weight vector. Indicates a gating network. T The sum of all components must be 1 and non-negative; the gate weights Characterizing the effect of the current drug-condition combination on the first The degree of dependence on expert networks provides a quantitative basis for subsequent interpretability analysis.

[0141] In this two-layer fully connected network, the hidden layer output of the first fully connected layer has a dimension of 256, and the GELU activation function is used between the two fully connected layers.

[0142] Step S33, Hierarchical Expert Network – Three-Level Granularity Information Processing: Construct N=7 expert networks with identical structures but independent parameters, divided into three levels according to the granularity of information fusion:

[0143] S331, Single-modal experts (E1-E3) – Specific information preservation: Three single-modal experts each use a single modality feature as input, aiming to preserve and enhance the independent predictive ability of each modality, as shown in Table 2:

[0144] Table 2

[0145]

[0146] When the information quality of a certain modality is significantly better than that of other modalities, the gating network automatically increases the weight of the corresponding single-modality expert to avoid noise interference introduced by low-quality modalities. Each single-modality expert concatenates the modal features with time embedding, dose embedding, and noise vector, and then maps them to the measurement data space through an independent two-layer fully connected network (the hidden layer dimension is 256, and the GELU activation function is used between the two fully connected layers).

[0147]

[0148] in, Indicates the expert and their number. Indicates random noise;

[0149] S332, Bimodal Interaction Experts (E4-E6) – Cross-modal Collaborative Capture: Three bimodal experts, each taking pairwise modal combinations as input, model interaction effects that cannot be expressed by a single modality, as shown in Table 3:

[0150] Table 3

[0151]

[0152] The network structure of each bimodal expert is consistent with that of the unimodal expert, but the input dimension is expanded to 2d + 2d + dz; d represents the expanded dimension of the Fourier feature of the dose or time univariate, and dz represents the dimension of the random noise. The outputs of the three bimodal experts are as follows:

[0153]

[0154] S333, Trimodal Joint Expert (E7) – High-Order Global Integration: Using feature concatenation from all three modalities as input, it captures global interaction information that cannot be fully expressed by either unimodal or bimodal levels.

[0155]

[0156] The input dimension of the three-modal expert is 3d+2d+dz, which is the largest among the seven experts. The output layer of all expert networks adopts the Tanh activation function to constrain the output value range within the interval [−1, 1] to match the range of standardized clinical pathological measurements.

[0157] Step S34, Adaptive Feature Fusion: Based on the weight coefficients calculated by the gating network, the outputs of each expert are weighted and summed to obtain the final generated prediction.

[0158]

[0159] Step S35, Sparse Activation and Computational Efficiency Optimization: The softmax output of the gated network naturally induces sparsity: the weight distribution of different samples is usually concentrated in a few experts, and the model can selectively focus on activating the most relevant expert combination, effectively guiding the gradient flow to the most relevant experts and improving parameter utilization efficiency; when the data quality of a certain modality is poor, the gated network can automatically reduce the weights of the corresponding single-modality and related bimodal experts, and instead rely on other modal compensation, so that the model can still maintain stable performance under data missing or noisy conditions.

[0160] The gradient refers to the gradient of the loss function of the generator in an adversarial network, which guides the gradient flow towards the most relevant experts, including:

[0161] Calculate the partial derivatives of the generator's overall loss with respect to all trainable parameters of the generator to obtain the global gradient matrix;

[0162] The global gradient matrix is ​​split according to the parameter dimensions of the seven expert networks. The parameter gradient of each expert network is multiplied by the corresponding gating weight to obtain the gradient specific to each expert system.

[0163] Each expert network updates its own network parameters through gradient descent, while the parameters of the gated network are updated directly by the overall loss, ensuring the accuracy of weight allocation.

[0164] Step S36, Random Noise Injection and Distribution Modeling: During the forward inference process, the model samples noise vectors from a standard normal distribution:

[0165]

[0166] in, Represent a random vectors, express A 3D identity matrix, where each dimension follows a standard normal distribution.

[0167] The noise vector, along with the time embedding and dose embedding, is injected into each expert network as a source of randomness in the generation process. Noise injection enables the model to capture the inherent variability of experimental data—the clinical pathological indicators of the same compound often exhibit biological fluctuations in different experimental batches under the same dosing conditions. Through multiple sampling, the uncertainty range of the prediction can also be estimated, providing more comprehensive risk information for toxicological assessment.

[0168] Preferably, the conditional discriminator in step S3 specifically includes:

[0169] Step S37, Conditional Discriminator Architecture: Construct a conditional discriminator to evaluate the authenticity of the generated data, including two stages: feature extraction and authenticity discrimination.

[0170] S371. In the feature extraction stage, the original feature matrices (flattened) of the three modalities are combined with the drug administration time t and the scalar dose. The vectors are concatenated into a one-dimensional vector and then input into a three-layer fully connected network for nonlinear transformation.

[0171]

[0172] in, The conditional information compression encoding vector represents an intermediate hidden layer representation that maps the high-dimensional original features of the three drug modalities and the scalar of the dosing regimen to a space of the same dimension as the biomarkers. () represents a nonlinear transformation function consisting of three fully connected layers. () represents the flattened original feature matrix. This represents the original feature matrix of the QSAR modes. This represents the original feature matrix of the QSIIR modes. The original feature matrix of the PSMD modality is represented by the following: the dimensions of each layer of the network are 512, 256 and M (M=9, corresponding to 9 liver function biomarkers), and the LeakyReLU activation function is used with a negative slope α=0.2.

[0173] S372. In the true / false discrimination stage, the extracted features are concatenated with the measurement data y to be judged, and then input into a three-layer fully connected network to output a scalar discrimination score.

[0174]

[0175] in, Indicates a fully connected network. This represents the original biomarker measurement value. Represents a scalar discrimination fraction.

[0176] The two-stage design allows the discriminator to first compress and encode the high-dimensional conditional information into an intermediate representation of the same dimension as the measurement data, and then perform joint discrimination, thus avoiding the information asymmetry problem caused by directly splicing high-dimensional conditional features with low-dimensional measurement data.

[0177] Step S38, Adversarial Training Strategy: An alternating optimization strategy is adopted, and the discriminator and generator parameters are updated sequentially in each training iteration:

[0178] S381. Discriminator Optimization: The discriminator aims to maximize its ability to distinguish between real and generated data, and its loss function is:

[0179]

[0180] in, This represents the loss function of the discriminator. This indicates that the discriminator judges given condition c( The truth value of sample x under (t) The generator is derived from random noise z and conditional... Generate fake samples, This indicates the exposure time, which is also the time of drug administration. Indicates the dosage. This represents the original multimodal features.

[0181] S382. Generator Optimization: The generator aims to maximize the false positive rate of the discriminator, and its loss function is:

[0182]

[0183] Where c represents the conditional input (covering compound multimodal characteristics, dosing time, and dosing dose). For the true data distribution, The noise prior distribution is given.

[0184] Step S4, Data Quality Control: Simultaneously predict the composition ratio of 5 white blood cell subtypes and construct a "white blood cell composition checker" as an internal quality control indicator: samples with the total white blood cell ratio within the tolerance range of 95%-105% are judged as valid predictions, and abnormal generated samples that exceed this range are removed to ensure the biological rationality of the virtual experimental data.

[0185] like Figure 6 As shown, Figure 6 In the diagram, 'a' represents quality control of the generated data based on leukocyte composition. The virtual rat experiment, while predicting liver function biomarkers, also generates the proportions of five leukocyte subtypes. If the sum of the five leukocyte proportions is within the range of 95%-105%, it is considered within the allowable error range; samples exceeding this range are deemed invalid predictions. 'b' represents the distribution of leukocyte composition in the validation set before and after quality control. Box plots show the distribution of the original leukocyte measurements, generated leukocyte proportion data, and retained samples after quality control in the validation set; samples exceeding the allowable error range are removed. The boxes represent the first and third quartiles, and must be 1.5 times the interquartile range (IQR). 'c' represents the comparison between the virtual rat experiment's generated results and the background values. Box plots show the distribution of cosine similarity and root mean square error (RMSE) between the generated data and the real experimental data in the validation set under various treatment conditions. In the validation set (n = 1,596; 321 treatment combinations), a two-tailed Wilcoxon rank-sum test is used to assess statistical differences. The background value is the cosine similarity or RMSE between any two sample biomarker vectors (n = 1,272,810, calculated by 1,596 × 1,595 / 2).

[0186] A preferred embodiment of the data quality control mechanism for the "White Cell Quality Checker" in step S4 includes:

[0187] Step S41, Simultaneous Generation of Leukocyte Subtypes: While generating 9 core liver function biomarkers (ALT, AST, ALP, TBiL, DBiL, GTP, LDH, RALB, A / G), the MA-MoE GAN model simultaneously generates the proportions of 5 leukocyte components as internal quality control indicators: Neutrophil percentage (Neu); Eosinophil percentage (Eos); Basophil percentage (Bas); Mono; Lymocyte percentage.

[0188] Step S42, Verification of the total white blood cell composition: Calculate the sum of the proportions of the five white blood cell components generated in each virtual experiment. :

[0189]

[0190] Step S43, Allowable error range determination: Samples with the total white blood cell ratio within the range of 95% - 105% are determined as valid predictions, and samples outside this range are regarded as invalid and excluded.

[0191] Step S5, Multidimensional characterization of the liver burden spectrum: Based on the probability density distributions of the 9 liver function biomarkers (ALT, AST, ALP, TBiL, DBiL, GTP, LDH, RALB, A / G) generated in Step S3, combined with the preset determination rules for 9 types of liver injury events (hepatocellular injury, cholestatic injury, mixed injury, hepatocellular jaundice, cholestatic jaundice, mixed jaundice, mitochondrial dysfunction, elevation of ubiquitous enzymes, liver synthetic dysfunction), the heterogeneous spectrum of drug-induced liver burden is characterized through combinatorial logical reasoning.

[0192] Preferably, the multidimensional characterization of the liver burden spectrum and clinical event reasoning in Step S5 specifically includes:

[0193] Step S51, Establishment of physiological reference intervals: Based on the 95th / 5th percentiles of the biomarker distributions in the control group, establish the physiological reference intervals for each indicator:

[0194] (1) Upper limit of normal (ULN): 95th percentile of the biomarker distribution in the control group

[0195] (2) Lower limit of normal (LLN): 5th percentile of the biomarker distribution in the control group

[0196] The percentile method avoids the bias in outlier determination by the mean ± standard deviation method under the assumption of normal distribution, ensuring the comparability of the virtual experiment results with the clinical interpretation criteria of real biomarkers at the level of toxicity severity stratification.

[0197] Step S52, Abnormality determination of biomarkers and inference rules for 6 types of liver injury events: For the valid prediction data after quality control in Step S4, determine whether each biomarker exceeds the physiological reference interval; based on the preset combinatorial logical determination rules, infer clinical liver injury events from the abnormal patterns of biomarkers, as shown in Table 4:

[0198] Table 4

[0199]

[0200] Where "↑" indicates elevation (>ULN), "↓" indicates decrease (<LLN), "-" indicates normal, and "fold" indicates the fold change relative to the control group.

[0201] Step S53, Liver Burden Map Construction: Based on the calculation results of step S52, a liver burden map visualization system consisting of two heatmaps and curves is constructed.

[0202] S531. Biomarker Anomaly Heatmap: The biomarker anomaly heatmap consists of 9 sub-heatmaps / curves, corresponding to ALT, ALP, TBiL, DBiL, AST, GTP, LDH, RALB, and A / G, respectively. Each sub-heatmap has 12 cells, and the content of each cell displays the ratio of the number of anomalies of the biomarker under the corresponding treatment combination to the total number of experiments, in the format of "number of anomalies / total number of experiments". Color gradient is used for visual encoding, where light color represents low anomaly rate, dark color represents high anomaly rate, and white represents no anomaly.

[0203] S532. Clinical Event Reasoning Heatmap: The clinical event reasoning heatmap consists of 6 sub-heatmaps / curves, corresponding to hepatocellular injury, cholestatic injury, mixed injury, hepatocellular jaundice, cholestatic jaundice, and mixed jaundice, respectively. Each sub-heatmap has 12 cells, and the content of each cell shows the ratio of the occurrence of the event under the treatment combination to the total number of valid experiments, in the format of "number of events / total number of experiments". The event occurrence rate is represented by a color gradient.

[0204] S553, Dual Heatmap Synergistic Interpretation Mechanism: Horizontal Comparison: Observe time-dependent evolution within the same row to identify toxicity accumulation or recovery trends; Vertical Comparison: Observe the synergy of multiple biomarkers / events within the same column to determine the consistency of damage types; Diagonal Pattern: Abnormal clustering of high-dose-long-duration combinations (lower right region) suggests dose-time synergistic effects; Blank Area Identification: Specific events that do not occur in all treatment combinations (all white lines) suggest that the drug does not have this type of damage mechanism;

[0205] Step S54: Digital Output of Liver Burden Spectrum: The dual-heatmap data matrix is ​​stored in a structured format, supporting: inter-drug comparison: comparing the differences in spectrum characteristics of drugs with different DILI risk levels; dose optimization decision-making: identifying the maximum tolerated dose without significant abnormalities; monitoring strategy formulation: determining key monitoring time points based on time evolution patterns. For example, time-series observations show that the high-dose group exhibits significant curve separation from the low- and medium-dose groups on day 4 after administration, suggesting that long-term exposure to high doses may induce irreversible liver damage. Therefore, day 4 is identified as a key monitoring point, and it is recommended to discontinue medication and begin recovery observation at this time.

[0206] Step S6, Dose-Time-Effect Causal Inference and Risk Stratification: Using abnormal biomarkers or the occurrence of liver injury events as endpoints, plot dynamic probability curves of maintaining normal liver function in different dose groups; combine curve separation time, separation amplitude, and event accumulation patterns to infer the causal relationship of safety hazards in drug dosage use, and achieve dynamic stratified assessment of the risk level of drug-induced liver injury.

[0207] Assessing drug hepatotoxicity based on probability curves includes classifying drugs into corresponding risk levels based on the overlap of probability curves, the presence of dose threshold effects, the presence of major targets, and time dynamics.

[0208] Preferably, step S6, dose-time-effect causal inference and risk stratification, specifically includes:

[0209] Step S61, Curve Analysis: Using the occurrence of 6 types of liver injury events obtained from the abnormal determination of 9 biomarkers in step S52 as endpoint events, the dynamic evolution of liver burden and toxicity events caused by drugs at different doses over time is clearly shown; the curve separation time, separation amplitude and event accumulation pattern are observed to analyze the characteristics of drug action; the differences in curves between different dose groups are observed to quantitatively evaluate the dose-response relationship.

[0210] This invention also provides a virtual rat experimental drug hepatotoxicity assessment device based on modality-aware hybrid experts and dose-time-dependent generative adversarial networks, realizing the above-mentioned virtual rat experimental drug hepatotoxicity assessment device based on modality-aware hybrid experts and dose-time-dependent generative adversarial networks, the device comprising:

[0211] The multimodal feature extraction module is used to calculate QSAR chemical descriptors, QSIIR bioactivity features, and PSMD pathway score features using the simplified molecular linear input specification of the acquired target oral drug, and to construct a multimodal fusion feature set;

[0212] The dynamic exposure condition setting module is used to obtain the dose and exposure time set by the user, and construct a treatment condition matrix based on the dose and exposure time;

[0213] The generation module is used to input the multimodal fusion feature set and processing condition matrix into a pre-built virtual rat experimental model and output sample data for drug hepatotoxicity experiments. The sample data includes the probability density distribution of liver function biomarkers. The virtual rat experimental model is built based on a modality-aware hybrid expert and dose-time dependent generative adversarial network.

[0214] The determination module is used to determine whether a preset liver injury event or abnormal liver function biomarkers have occurred in the sample data based on the probability density distribution of liver function biomarkers and preset liver injury event determination rules.

[0215] The evaluation module is used to plot probability curves of the liver maintaining a normal state in different dosage groups, with the occurrence of a preset liver injury event or abnormal liver function biomarkers as the endpoint, and to evaluate the hepatotoxicity of the drug based on the probability curves.

[0216] Example 1: Construction and Validation of a Virtual Rat Experimental Model

[0217] 1.1 Data Inclusion Criteria

[0218] The following guidelines were strictly followed: Ⅰ The test rats must be male Sprague-Dawley rats (5 weeks old); Ⅱ A matched vector control group must be included; Ⅲ The nine liver function biomarkers (ALT, ALP, TBiL, DBiL, AST, GTP, LDH, RALB, A / G) must not be completely absent; Ⅳ The administration time was strictly limited to 4 days, 8 days, 15 days, and 29 days; After screening, a total of 10,205 records were included in the analysis.

[0219] like Figure 3 As shown, Figure 3 The 'a' in the diagram represents the training data acquisition and inclusion process. To ensure data quality, this study established strict inclusion criteria. The flowchart illustrates the standard rat experimental procedure: 5-week-old rats were selected, and after a 7-day isolation adaptation period, they were given a single oral administration. They were then sacrificed 24 hours after the last administration, including the four time points mentioned above.

[0220] 1.2 Multimodal Feature Extraction and Calculation

[0221] (1) QSAR chemical descriptor calculation: Based on the SMILES representation of the target drug, the RDKit library was used to calculate the two-dimensional topological descriptor and MACCS molecular fingerprint, and the Mordred library was used to calculate the ring structure and macroscopic physicochemical descriptor. Specifically, it includes: RDKit 2D descriptors: 208 (topological structure, functional group distribution, etc.), Mordred descriptors: 1,613 (ring structure, physicochemical properties, etc.), and MACCS fingerprints: 167 (binary structural features, etc.), for a total of 1,988 QSAR features.

[0222] (2) QSIIR bioactivity fingerprint calculation: Based on the Chemical Checker database, the Signaturizer tool calculates a 25-dimensional bioactivity fingerprint covering molecular targets, signaling networks, cell phenotypes and clinical levels.

[0223] like Figure 3 As shown, Figure 3In the figure, b represents multimodal feature calculation. In the quantitative structure-activity relationship (QSAR) modality, the RDKit 2D, Mordred, and molecular access system (MACCS) descriptors were calculated using Python's RDKit and Mordred libraries, resulting in a total of 1,988 features (208+1,613+167). In the quantitative structure-in vivo activity relationship (QSIIR) modality, 3,200 global features were calculated using the Chemical Checker tool. In the pathway scoring molecular docking (PSMD) modality, 688 pathway scoring features were finally selected.

[0224] (3) PSMD pathway scoring feature calculation: This feature is a unique toxicity mechanism characterization method of this invention. For the specific process, please refer to [link to relevant documentation]. Figure 3 The steps are as follows:

[0225] S141. Construction of a liver injury target library: Using "Chemical and Drug Induced Liver Injury" as the search term, a set of liver injury-related genes with direct evidence (marker / mechanism / therapeutic) was obtained from the CTD database to construct a hepatotoxicity target library containing 302 core targets.

[0226] S142. Obtaining target protein structure: Download the corresponding protein three-dimensional structure from the AlphaFold protein structure database and select the predicted conformation with the highest pLDDT (predicted local distance difference test) score.

[0227] S143. Prediction of active cavities: Based on the CB-Dock algorithm, the top 5 potential active cavities with the highest prediction scores for each target protein are identified; the center coordinates and spatial size parameters of each cavity are extracted to construct a docking parameter configuration file.

[0228] S144, High-throughput molecular docking: such as Figure 3 As shown in d, Figure 3 In the diagram, 'd' represents the construction of a protein-protein interaction network. A PPI network was constructed by screening target genes from the CTD (Content-to-Depth) model and visualized using Cytoscape. The target drug was docked to five active cavities at 302 target sites, retaining the optimal binding energy to construct the drug-target interaction spectrum. The docking parameters were uniformly set as follows: energy range 4 kcal / mol, number of modes 10, exhaustiveness 8, and random seed 10¹².

[0229] Figure 3The "c" in the diagram represents a high-throughput molecular docking process. Using "Chemical and Drug Induced Liver Injury" as the keyword, target genes were screened from the CTD (Chemical and Drug Induced Liver Injury) database, retaining only those with direct evidence (marker, mechanism, or therapeutic). The 3D structures of target proteins were obtained from EMBL-EBI (European Bioinformatics Institute), and the structure with the highest pLDDT score predicted by AlphaFold 4 was selected as the receptor. Protein preprocessing was performed using PyMOL and MGLTools. Potential binding cavities were predicted using DrugRep, and the top five cavities with the highest scores and their spatial parameters were selected. Molecular docking parameters were uniformly set as follows: energy range 4 kcal / mol, number of modes 10, exhaustiveness 8, and random seed 10¹². Docking was performed using AutoDock Vina.

[0230] S145. Pathway Enrichment Analysis: KEGG and Gene Ontology enrichment analyses were performed using the GSEApy tool. The comprehensive scores of each pathway were extracted as a quantitative characterization of pathway activation intensity. Finally, 688 PSMD pathway score features were selected. The KEGG and GO enrichment analysis results for target genes are shown below. Figure 3 As shown in e and f, e represents KEGG pathway enrichment analysis; KEGG pathway enrichment analysis of target genes was performed using clusterProfiler and related packages in R language to reveal a large number of pathways related to drug metabolism and related biological processes; f represents Gene Ontology (GO) enrichment analysis; GO enrichment analysis of target genes mainly involves pathways related to the metabolism of exogenous chemicals and enzyme activity and function.

[0231] 1.3 Feature Filtering and Standardization

[0232] Pearson correlation analysis combined with significance test (p<0.05) was used to screen a subset of features significantly associated with liver function biomarkers from the candidate feature library. A total of 34,933 biomarker-feature statistically significant associations were identified. The top 2 association feature networks for different modalities are shown in [link to network]. Figure 4 .

[0233] The multi-standardization process performs the following steps:

[0234] Step S15, Scale Normalization: Perform Min-Max normalization on the three types of features, QSAR, QSIIR, and PSMD, respectively, and map them to the [0,1] interval;

[0235] Step S16, Target variable transformation: Perform logarithmic space transformation on the true measured values ​​of the 9 liver function biomarkers to alleviate the right skewed distribution, and further perform Min-Max standardization to map to the [0,1] interval;

[0236] Step S17, Inverse Transformation and Truncation: After the model generates the predicted values, an inverse exponential transformation is performed and a non-negativity truncation is applied to ensure the physiological rationality of the predicted output.

[0237] 1.4 Construction of Modality-Aware Hybrid Expert Generative Adversarial Network

[0238] (1) Overall Network Architecture: The virtual rat experimental model was constructed based on the developed DT-cGAN (Dose and Time-depended conditional GAN) framework, and its overall structure is as follows: Figure 2 As shown, the architecture includes a modality-aware hybrid expert generator and a discriminator, as well as the input-output relationships of single-modality, dual-modality, and trimodality experts, and the logical model comparing the real and generated results. DT-cGAN consists of two core components: a generator G and a discriminator D.

[0239] Generator G: During model training, a fixed combination of dose-time dosing conditions is embedded to provide digital virtual processing for the virtual rat model. It generates simulated clinical pathological measurement data by taking conditional input c (compound multimodal features, dosing time, and dosing dose) and random noise Z as input.

[0240] Discriminator D: Under the same condition c, it performs binary classification to distinguish between real experimental data and generated data.

[0241] (2) Modality-Aware Hybrid Expert Generator (MA-MoE Generator): The core structure of the generator is shown in [link to generator]. Figure 2 It comprises five key stages: multimodal feature encoding, modality-aware hybrid expert fusion, drug administration condition embedding, noise injection, and measurement data generation. The specific steps are as follows:

[0242] Multimodal feature encoding (step S31): Receive the feature matrices of three heterogeneous compounds (QSAR, QSIIR, PSMD), and map them to a unified d=256-dimensional feature space through independent encoding networks. Each encoding network consists of two fully connected layers (hidden layer dimension 256, ReLU activation, Dropout rate 0.25).

[0243] Dynamic weight calculation of gating network (step S32): The three modal feature representations are concatenated with temporal embedding and dose embedding, and an N=7-dimensional weight vector is calculated through a two-layer fully connected network (hidden layer dimension 256, GELU activation), and then normalized by Softmax;

[0244] Hierarchical Expert Network Design (Step S33): Construct seven expert networks with identical structures but independent parameters, divided into three levels according to the granularity of information fusion: single-modal experts (E1-E3), bimodal interactive experts (E4-E6), and trimodal joint experts (E7). The inputs and functions of each expert network are detailed below. Figure 2 ;

[0245] Adaptive feature fusion (step S34): The outputs of each expert are weighted and summed based on the gate weights to obtain the final generated features;

[0246] Dosing condition embedding (steps S22-S23): Temporal embedding uses log-compressed Fourier feature encoding, and dose embedding expands the dose scalar broadcast and maps it to 256 dimensions through a two-layer fully connected network (SiLU activation).

[0247] Noise injection and output generation (step S36): Sample noise vectors from the standard normal distribution and inject them into the expert network along with each conditional embedding to finally generate a 9-dimensional liver function biomarker vector.

[0248] (3) Conditional discriminator (step S37): A two-stage conditional discriminator architecture is adopted (Figure 2). In the feature extraction stage, the original features of the three modalities (flattened) are concatenated with the dosing time and dose scalar, and input into a three-layer fully connected network (dimension 512→256→9, LeakyReLU activation, α=0.2). α represents a hyperparameter that controls the slope of the negative half axis and is used to adjust information leakage and gradient flow. In the true / false discrimination stage, the extracted features are concatenated with the measurement data y to be discriminated, and input into a three-layer fully connected network, and the scalar discrimination score is output.

[0249] (4) Adversarial Training Strategy (Step S38): The generator and discriminator parameters are updated using an alternating optimization strategy. Key performance indicators during the training process are monitored in [link to training process]. Figure 5 This includes the changing trends of metrics such as generator total loss, L1 loss, adversarial loss, discriminator loss, cosine similarity, and gradient norm under Epoch and Batch conditions, as well as the distribution and rate of change of L1 loss. For example... Figure 5As shown, 'a' represents the generator's total loss. The training set (blue) drops rapidly initially and then gradually stabilizes, while the validation set (red) shows a similar trend but converges slightly slower. 'b' represents the generator's L1 loss. The L1 loss is calculated from the pointwise error between the generator's output and the true label, measuring the absolute difference between the generated sample and the true sample, and constraining the numerical accuracy of the generated results. As shown in 'b', both the training and validation sets show a decreasing trend, with the validation set decreasing more slowly. 'c' represents the generator's adversarial loss. The adversarial loss comes from the adversarial process between the generator and the discriminator, aligning the generated distribution with the real data distribution, reflecting the generator's ability to deceive the discriminator. The training set rises rapidly and then remains at a high level, while the validation set fluctuates significantly and is lower than the training set. 'd' represents the discriminator's... Discriminator loss is used to optimize its ability to distinguish between real and generated samples. As shown by d, the training set loss is extremely low and stable. e represents cosine similarity, used to evaluate the vector similarity between generated and real samples, measuring the semantic consistency between generated and real samples in the feature embedding space, and used to evaluate the alignment of potential representations. f represents gradient norm, calculated with respect to the gradient of the loss function of the generator and discriminator parameters, used to monitor the stability and convergence of the training process. g represents generator batch loss. h represents discriminator batch loss, and i represents the rate of change of generator loss. Fluctuations around 0 indicate that the training has entered a stable stage without significant turbulence. j represents L1 loss distribution. The histogram shows the frequency distribution of L1 loss. The mean μ = 0.0561 and the standard deviation σ = 0.0209 are right-skewed, with most sample losses concentrated in the low value range.

[0250] 1.5 Model Performance Evaluation

[0251] (1) Internal validation results: The data generated by the optimized virtual rat model has a consistent distribution range with the real experimental data. Adjusting R... 2 Values ​​of multiple indicators reached 0.9 or above, R 2 The coefficient of determination is represented. The model's average cosine similarity reaches 0.9986, significantly outperforming traditional machine learning models. The quality control results of the white blood cell composition checker are as follows: Figure 6 As shown in Figure b, samples with a total white blood cell percentage between 95% and 105% are considered valid predictions, while abnormal generated samples outside this range are removed. The cosine similarity and RMSE (Root Mean Square Error) distributions between the generated data and the real data are shown below. Figure 6 As shown in Figure c, the cosine similarity between the generated data and the real data is significantly higher than that of the background distribution (p=1.28×10). -106 The RMSE was significantly lower than the background distribution (p=9.13×10). -111 ); the consistency scores of 10 independent batches of experiments are as follows Figure 6As shown in Figure d, where d represents the consistency of statistical test results among different virtual experimental batches, the scores of each batch on each biomarker are highly overlapping, indicating that the generative model has excellent batch reproducibility. Ten independent batch predictions were performed on 321 treatment combinations in the validation set, and 50 sets of predicted values ​​were generated for each batch under each condition. The bee colony diagram shows the distribution of ECS values ​​of each liver function biomarker, which is used to evaluate the stability of the results.

[0252] Figure 6 In the middle, e–m represents the comparison between the validation data and the liver function biomarkers generated from the virtual rat experiment; each point represents a treatment combination, and the corresponding value is the median under that condition; the vertical line segment represents the interquartile range of the generated data, and the horizontal line segment represents the interquartile range of the actual experimental data; except for direct bilirubin (DBiL), the vertical line segments of the other biomarkers are relatively short, indicating that the generated results have high stability; points located near the diagonal line indicate that the predicted results are more consistent with the actual experimental values.

[0253] (2) Comparison with traditional models: Ten classic regression models (LR, SVR, RF, XGB, LGB, AdaB, MLP, PLS, ARD, GPR) were compared, and the results are shown in [the table below]. Figure 7 , Figure 7 In this diagram, 'a' represents Linear Regression (LR); 'b' represents Partial Least Squares (PLS); 'c' represents Automatic Relevance Determination (ARD); 'd' represents Support Vector Regression (SVR); 'e' represents Gaussian Process Regression (GPR); 'f' represents Random Forest (RF); 'g' represents Adaptive Boosting Algorithm (AdaBoost, AdaB); 'h' represents Extreme Gradient Boosting (XGBoost, XGB); 'i' represents Lightweight Gradient Boosting Machine (LightGBM, LGB); and 'j' represents Multilayer Perceptron (MLP). Each point represents a treatment condition, with its value being the median of the corresponding dimension. Since traditional models generate only one prediction per condition, there are no vertical line segments; horizontal line segments represent the interquartile range of the actual experimental results; points on the diagonal represent perfect predictions. Predictions appearing outside the first quadrant indicate a severe model failure.

[0254] Traditional models generally exhibit distribution compression in their predictions, making it difficult to capture the distribution broadening caused by dose-time changes in real data. The model of this invention is significantly superior to traditional methods in terms of prediction accuracy, distribution fidelity, and biological consistency.

[0255] (3) Independent external validation: Independent external validation was performed based on the DrugMatrix database (n=1,453,464 processing combinations). The results are shown in […]. Figure 8 , Figure 8 In the text, 'ag' represents the comparison of liver function biomarkers between DrugMatrix experimental data and virtual rat experimental results, and 'hi' represents the comparison of RMSE and cosine similarity of virtual rat experimental results with DrugMatrix background values. In the test set (n = 1,453; 464 treatment combinations), the two-tailed Wilcoxon rank-sum test was used to assess statistical differences. The background value is the cosine similarity or RMSE between any two sample biomarker vectors (n = 1,054,878, calculated by 1,453 × 1,452 / 2).

[0256] Virtual experiments in ALT (adjusting R) 2 =0.9329), ALP (adjusted R) 2 =0.9433), RALB (adjusted R) 2 The prediction accuracy was highest on indicators such as (p=0.9955), and the median cosine similarity was ≈0.95, significantly better than background variation (p=3.36×10). -31 This confirms that the model has reliable cross-dataset prediction capabilities.

[0257] Example 2: Acetaminophen - a classic dose-threshold type of hepatotoxicity model

[0258] 2.1 Drug Background DILI Risk Level: Acetaminophen is one of the most commonly used antipyretic analgesics, with a clear dose-threshold type hepatotoxicity characteristic. It is classified as "vLess-DILI-Concern" in the DILIrank 2.0 database, and is an ideal model for verifying the identification of dose-response relationships in virtual experimental systems.

[0259] 2.2 Large-scale virtual experiment setup: Based on the virtual rat model constructed in Example 1, a virtual toxicity experiment was conducted on acetaminophen with 12 treatment combinations (4 exposure cycles × 3 dose levels). Each treatment combination underwent 10,000 independent predictions, for a total of 120,000 virtual experiments. After quality control using a leukocyte composition checker (95%-105% tolerance), the effective experimental data was 95,135 (effectiveness rate 79.3%). The quality control results are shown in […]. Figure 9 As shown in Figure a, a represents a biomarker and b represents a rare liver injury event.

[0260] 2.3 Dynamic changes in biomarkers: Results of the virtual experiment are shown in […]. Figure 9 Acetaminophen exhibited a dose-dependent hepatotoxicity pattern highly consistent with classical toxicology. ALT showed a strict dose-threshold dependence. AST was an early and sensitive indicator of mitochondrial damage. GTP was occasionally elevated, but ALP remained normal. TBiL was slightly elevated without accompanying DBiL synchronous abnormalities. RALB, A / G, and LDH remained stable, confirming that the damage did not affect liver synthetic function or cause widespread disruption of cell membrane integrity.

[0261] 2.4 Liver Injury Event Inference and Probability Curve Analysis: Based on the pre-defined rules for judging six types of liver injury events, a multidimensional characterization of the liver burden spectrum of acetaminophen was performed. Hepatocellular injury occurred in a concentrated manner in the high-dose group and was reversible. Mitochondrial dysfunction showed clear dose stratification characteristics. Cholestatic injury, mixed injury, and various jaundice-related events did not occur under all dose and time conditions.

[0262] 2.5 Summary of toxicity characteristics: The virtual experiment of acetaminophen successfully reproduced the nonlinear characteristics of "low-dose steady-state masking - high-dose metabolic saturation". The core toxicity characteristic is a clear dose threshold effect. Exceeding the threshold rapidly induces an acute toxic reaction mainly characterized by hepatocellular damage, and this damage has a significant time-dependent recovery trend.

[0263] Example 3: Characterization of hepatotoxicity of abacavir sulfate-vNo-DILI-Concern

[0264] 3.1 Drug Background and DILI Risk Level: Abacavir sulfate is a nucleoside reverse transcriptase inhibitor, classified as "vNo-DILI-Concern" in the DILIrank 2.0 database. It has good clinical liver safety and is an ideal negative control for verifying the ability of virtual experimental systems to identify safe drugs.

[0265] 3.2 Results of the large-scale virtual experiment: After quality control using a leukocyte composition analyzer, the effective experimental data was 119,091 times (effective rate 99.2%). The quality control results are shown below. Figure 10 As shown in Figure a, the high efficacy rate indicates that the drug did not trigger the removal mechanism for a large number of abnormally generated samples.

[0266] 3.3 Biomarker and Event Analysis: Results of the virtual experiment are shown in [link to results]. Figure 10 , Figure 10In the figure, 'a' represents biomarkers and 'b' represents rare liver injury events. The curves of core liver enzyme indicators ALT and ALP, as well as liver synthetic function indicators (RALB, A / G), basically overlapped across all dose groups, and no dose-related effects were observed. Some indicators (TBiL, DBiL, AST, etc.) showed slight time-dependent fluctuations under long-term exposure to high doses, but did not form a consistent abnormal pattern across indicators. Various types of liver injury and jaundice events did not occur under all dose and time conditions, and the curves almost completely overlapped.

[0267] 3.4 Safety Conclusion: Abacavir sulfate did not form the event combination that meets the criteria for hepatocellular injury or cholestasis. The results of the virtual experiment were highly consistent with clinical evidence of liver safety, verifying the accurate identification capability of this invention for drugs without DILI risk.

[0268] Example 4: Comprehensive toxicity characterization of felbamate-vMost-DILI-Concern

[0269] 4.1 Drug Background and DILI Risk Level: Felbamate is an antiepileptic drug. Due to the risk of severe liver failure, it is included in the "black box warning" category and is classified as "vMost-DILI-Concern" in the DILIrank 2.0 database. It is an ideal positive control for verifying the ability of the virtual experimental system to identify high-risk drugs.

[0270] 4.2 Results of the large-scale virtual experiment: After quality control, the effective experimental data was 118,768 times (effectiveness rate 99.0%). The quality control results are shown in [link to quality control results]. Figure 11 As shown in Figure a.

[0271] 4.3 Full spectrum toxicity characteristics: Virtual experiment results are shown in […]. Figure 11 'a' represents biomarkers and 'b' represents rare liver injury events. Felbamate exhibits a full spectrum of toxicity covering hepatocellular injury, cholestasis, mitochondrial dysfunction, and synthetic failure. The three core indicators, ALT, AST, and TBiL, decreased synchronously from day 15 in the high-dose group and also showed a progressive decrease in the medium- and low-dose groups, suggesting the lack of a safe dose threshold. Significant decreases in RALB and A / G indicate depletion of liver synthetic reserve, and widespread increases in LDH suggest general disruption of cell membrane integrity. The high-dose group also showed concurrent hepatocellular injury, mixed injury, cholestatic injury, and hepatocellular jaundice, exhibiting a "comprehensive and progressive" toxicity profile.

[0272] 4.4 High-Risk Drug Early Warning Value: The virtual toxicity profile of fentanyl hydrochloride clarified the complete connotation of vMost-DILI-Concern, providing a key computational biology marker for the early identification of drugs with black box warnings. Example 5: Quantification of uncertainty of propoxyphene hydrochloride-vAmbiguous-DILI-Concern

[0273] 5.1 Drug Background and DILI Risk Level: Propoxyphene hydrochloride is an opioid analgesic with controversial liver safety assessment. It is classified as "vAmbiguous-DILI-Concern" in the DILIrank 2.0 database, making it an ideal case for verifying the ability of virtual experimental systems to quantify uncertainty.

[0274] 5.2 Results of the large-scale virtual experiment: After quality control, the effective experimental data was 118,212 times (effectiveness rate 98.5%). The quality control results are shown in [link to quality control results]. Figure 12 As shown in Figure a.

[0275] 5.3 Paradoxical toxic signals: Results of the virtual experiment are shown in [link to experiment]. Figure 12 'a' represents biomarkers and 'b' represents rare liver injury events. Propoxyphene hydrochloride presents contradictory, inconsistent, and difficult-to-classify toxic signals: DBiL abnormality rate was higher in the early stage in the low-dose group than in the high-dose group, while TBiL remained stable throughout; GTP was sporadically abnormal but ALP was completely normal, and cholestasis signals were fragmented; hepatocellular injury only accumulated slightly in the low- and medium-dose groups and occurred zero in the high-dose group, showing a dose inversion; some index curves showed intergroup separation, but the separation pattern did not conform to the typical dose-time-response relationship.

[0276] 5.4 Quantification of Uncertainty and Regulatory Implications: This invention quantifies the "ambiguous" characteristics of propoxyphene hydrochloride, namely ambiguous liver damage, through probability density distribution and curve visualization in virtual experiments. This provides clear "uncertainty boundary" information for regulatory decisions, suggesting the need for further mechanistic studies or clinical monitoring.

[0277] Example 6: Cross-sectional comparison and risk stratification of four drugs

[0278] Based on the virtual experiment results of Examples 2-5 ( Figures 9-12 The core toxicity characteristics of the four drugs are compared in Table 5:

[0279] Table 5

[0280]

[0281] 6.2 Risk Stratification: Using abnormal biomarker activity or the occurrence of liver injury events as endpoints, dynamic probability curves of maintaining normal liver function were plotted for different dosage groups. Figures 9-12 The drugs are classified into four risk levels:

[0282] Low-risk drug (abacavir sulfate): The curves for each dose-time combination basically overlapped, and there was no significant dose-response relationship;

[0283] Medium-risk drugs (acetaminophen): a clear dose threshold effect, with significant early separation of the curves in the high-dose group, but the damage is reversible;

[0284] High-risk drug (fenamic acid): The probability curves of the full-dose group progressively separate, with no safety threshold, and multiple target damage events occur in parallel;

[0285] Drugs with uncertain risks (propoxyphene hydrochloride): The curves show atypical separation patterns and lack a consistent dose-time-effect relationship.

[0286] 6.3 Verification of the methodological advantages of the virtual experimental system: Through systematic evaluation of four representative drugs, the virtual rat experimental system of the present invention demonstrates the core advantages of dynamic risk characterization, rare event identification, and uncertainty quantification, and conforms to the 3Rs principle. A total of 120,000 virtual experiments for each of the four drugs can reduce the use of approximately 40,000 rats, while improving the resolution of risk assessment.

[0287] Example 7: Model Interpretation and Dynamically Gated Network Weights

[0288] Although VREM is a black-box model, the features themselves can provide interpretation, further enhancing the model's interpretability. Taking felbamate as an example, we will analyze the model interpretation and characterize the weights of the dynamically gated network. Figure 13 As shown, Figure 13 Figures a and b show the distribution of QSAR and QSIIR features, figure c shows the target sites screened during molecular docking, and figures d and e further show the distribution of pathway perturbations. Figure 14 Visualization of gated network weights: The VREM model automatically adjusts weights under different processing conditions, and the contributions of multimodal experts to the virtual experiment results change dynamically.

[0289] Example 8: Specific Implementation of the Computer Device

[0290] This embodiment provides a virtual rat experimental drug hepatotoxicity assessment device based on modality-aware hybrid experts and dose-time dependent generative adversarial networks. The device architecture is described in [link to device architecture]. Figure 15 It includes a processor, a memory, and a display. The processor reads a computer program from the memory and displays the result on the display for performing the following operations:

[0291] Module M1, Multimodal Feature Extraction Module: Receives SMILES input of the target drug, calls RDKit and Mordred to calculate QSAR chemical descriptors, calls Chemical Checker to calculate QSIIR bioactivity features, calls AutoDockVina and GSEApy to calculate PSMD pathway score features, and outputs a standardized multimodal fusion feature set.

[0292] Module M2, Dynamic Exposure Condition Setting Module: Receives the dose level and exposure time input by the user, performs log-compressed Fourier feature encoding to generate a time embedding vector, performs broadcast extended nonlinear transformation to generate a dose embedding vector, and outputs the condition embedding vector to the generation module;

[0293] Module M3, MA-MoE GAN generation module: Load a pre-trained modality-aware hybrid expert generator (7 hierarchical experts and gating networks), perform 10,000 Monte Carlo samplings, generate the probability density distribution of 9-dimensional liver function biomarkers, and simultaneously generate 5 white blood cell composition ratios;

[0294] Module M4, Quality Control Module: Calculates the total proportion of white blood cell components, performs 95%-105% tolerance error verification, removes invalid prediction samples, and outputs the quality-controlled valid prediction dataset;

[0295] Module M5, Liver Burden Spectrum Construction Module: Based on the 95th / 5th percentile of the control group distribution, establishes a physiological reference interval, executes 6 types of liver injury event judgment rules, generates a biomarker abnormal heatmap (9×12 grid) and a clinical event reasoning heatmap (6×12 grid), and outputs the dual heatmap visualization results to the display.

[0296] Module M6, Risk Stratification Module: Using the occurrence of abnormal biomarkers or liver injury events as the endpoint, the risk level of DILI (No / Less / Ambiguous / Most-Concern) is determined based on curve separation time, separation amplitude, and event accumulation pattern.

[0297] Module M7, Result Display Module: Outputs the dual heatmap visualization results on the monitor (e.g., Figures 9-12 (In the format shown), display the probability curve and generate a structured risk assessment report;

[0298] Module M8, Data Storage Module: Stores raw input data, intermediate calculation results and final evaluation reports, supports cross-drug comparison queries and historical data tracing and model iteration optimization.

[0299] Regarding the system in the above embodiments, the specific manner in which each unit performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0300] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0301] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for assessing the hepatotoxicity of drugs in virtual rat experiments based on modality-aware hybrid experts and dose-time dependent generative adversarial networks, characterized in that, include: QSAR chemical descriptors, QSIIR bioactivity features, and PSMD pathway score features were calculated using the simplified molecular linear input norms of the target oral drug, and a multimodal fusion feature set was constructed. Obtain the user-set dose and exposure time, and construct a treatment condition matrix based on the dose and exposure time; The multimodal fusion feature set and processing condition matrix are input into a pre-constructed virtual rat experimental model, and the output is large-scale sample data for drug hepatotoxicity experiments. The sample data includes the probability density distribution of liver function biomarkers. The virtual rat experimental model is constructed based on a modality-aware hybrid expert and a dose-time dependent generative adversarial network. Based on the probability density distribution of liver function biomarkers and the preset rules for judging liver injury events, it is determined whether the sample data has experienced a preset liver injury event or abnormal liver function biomarkers. Using the occurrence of a pre-specified liver injury event or abnormal liver function biomarkers as the endpoint, probability curves were plotted for different dose groups to maintain a normal liver state, and the hepatotoxicity of the drug was assessed based on the probability curves.

2. The virtual rat experimental drug hepatotoxicity evaluation method based on the modal perception hybrid expert and dose-time-dependent generative adversarial network of claim 1, wherein, The generative adversarial network (GAN) comprises a modality-aware hybrid expert and a dose-time dependent generator and a conditional discriminator. A multimodal fusion feature set and a processing condition matrix are input as conditions and random noise to the generator to generate sample data for drug hepatotoxicity experiments. Under the same conditions, the conditional discriminator evaluates the consistency of the distributions of real experimental data and generated sample data. The generator includes a gating network and a hierarchical expert network. The gating network dynamically allocates weights to the hierarchical expert networks, and the outputs of each hierarchical expert network are weighted and summed based on these weights to obtain the sample data.

3. The virtual rat experimental drug hepatotoxicity evaluation method based on the modal perception mixed expert and dose-time-dependent generative adversarial network of claim 2, wherein, The conditional discriminator includes a two-stage processing unit: a feature extraction unit and a true / false discrimination unit. The feature extraction unit concatenates the original features of the three modalities with the exposure time and dose into a one-dimensional vector, which is then input into the first three-layer fully connected network for feature extraction. The true / false discrimination unit concatenates the extracted features with the original biomarker measurement values ​​to be discriminated and then inputs the result into the second three-layer fully connected network to output a scalar discrimination score.

4. The virtual rat experimental drug hepatotoxicity evaluation method based on the modal perception mixed expert and dose-time-dependent generative adversarial network of claim 1, wherein, The method further includes: Based on QSAR chemical descriptors, QSIIR bioactivity characteristics, and PSMD pathway scoring characteristics, the original feature matrices of QSAR, QSIIR, and PSMD modes were constructed, respectively. The corresponding encoding networks are constructed for the original feature matrices of QSAR, QSIIR, and PSMD modes, respectively, and the corresponding mode encoding vectors are output. A multimodal fusion feature set is constructed based on the corresponding mode vectors. The mode encoding vectors have the same dimension. Each encoding network includes an input layer, a first fully connected layer, a ReLU activation function, a Dropout regularization layer, a second fully connected layer, and an output layer connected in sequence.

5. The method for evaluating the hepatotoxicity of drugs in virtual rat experiments based on modality-aware hybrid experts and dose-time dependent generative adversarial networks according to claim 1, characterized in that, The treatment condition matrix was constructed based on dose and exposure time, including: After logarithmically normalizing the exposure time, Fourier transform is performed to obtain Fourier features. The Fourier features are then mapped to the temporal embedding space through a two-layer fully connected network. After the dose is broadcast and extended into a one-dimensional vector, it is nonlinearly transformed through a two-layer fully connected network and then mapped to a dose embedding space with the same dimension as the time embedding space. The processing condition matrix is ​​obtained by concatenating the time embedding space and the dose embedding space.

6. The method for evaluating the hepatotoxicity of drugs in virtual rat experiments based on modality-aware hybrid experts and dose-time dependent generative adversarial networks according to claim 1, characterized in that, PSMD pathway score features were calculated using the simplified molecular linear input specification of the acquired target oral drug, including: Based on CTD, liver injury-related targets associated with the target oral drug were identified and the protein structures of these targets were predicted. The molecular docking of multiple potential active cavities with the highest probability of binding to the relevant target protein structure with the target oral drug was performed to obtain the binding energy matrix, and the interaction spectrum between the drug and the target was constructed based on the binding energy. The comprehensive score of each pathway is extracted based on the drug-target interaction spectrum and used as the PSMD pathway scoring feature.

7. The virtual rat experimental drug hepatotoxicity evaluation method based on the modal perception hybrid expert and dose-time-dependent generative adversarial network of claim 6, wherein, The comprehensive score for each pathway extracted based on the drug-target interaction spectrum includes: Based on the interaction profile of drugs and targets, pathway enrichment analysis is performed to obtain the standardized enrichment score and calculate whether the pathway enrichment observed in the interaction profile is significant, obtaining statistical significance ; Based on the standardized enrichment scores and statistical significance Computing the integrated score for the corresponding pathway : ， in, It indicates the degree to which the pathway is activated or inhibited.

8. The virtual rat experimental drug hepatotoxicity evaluation method based on the modal perception mixed expert and dose-time-dependent generative adversarial network of claim 1, wherein, Constructing a multimodal fusion feature set includes: Calculate the Person correlation coefficients between features in the feature candidate library and liver function biomarkers; The probability of a current feature being correlated with liver function biomarkers is calculated based on the Person correlation coefficient, assuming the assumption that the feature is not related to liver loss. This probability serves as the significance level for the correlation test of the feature. Features that have a significance level of less than a preset significance level threshold and a Person correlation coefficient greater than a preset correlation threshold are selected from the feature candidate library to form a feature subset; After standardizing the feature subset, a multimodal fusion feature set is constructed.

9. The virtual rat experimental drug hepatotoxicity evaluation method based on the modal perception hybrid expert and dose-time-dependent generative adversarial network of claim 1, wherein, The method further includes: The virtual rat experimental model outputs sample data corresponding to various white blood cell ratios; sample data whose total white blood cell ratio falls within a preset threshold range are determined to be valid sample data.

10. A virtual rat experimental drug hepatotoxicity assessment device based on modality-aware hybrid experts and dose-time-dependent generative adversarial networks, implementing the virtual rat experimental drug hepatotoxicity assessment method based on modality-aware hybrid experts and dose-time-dependent generative adversarial networks as described in any one of claims 1 to 9, characterized in that, The device includes: The multimodal feature extraction module is used to calculate QSAR chemical descriptors, QSIIR bioactivity features, and PSMD pathway score features using the simplified molecular linear input specification of the acquired target oral drug, and to construct a multimodal fusion feature set; The dynamic exposure condition setting module is used to obtain the dose and exposure time set by the user, and construct a treatment condition matrix based on the dose and exposure time; The generation module is used to input the multimodal fusion feature set and processing condition matrix into a pre-built virtual rat experimental model and output sample data for drug hepatotoxicity experiments. The sample data includes the probability density distribution of liver function biomarkers. The virtual rat experimental model is built based on a modality-aware hybrid expert and dose-time dependent generative adversarial network. The determination module is used to determine whether a preset liver injury event or abnormal liver function biomarkers have occurred in the sample data based on the probability density distribution of liver function biomarkers and preset liver injury event determination rules. The evaluation module is used to plot probability curves of the liver maintaining a normal state in different dosage groups, with the occurrence of a preset liver injury event or abnormal liver function biomarkers as the endpoint, and to evaluate the hepatotoxicity of the drug based on the probability curves.

Images