A precise comparison method for consistency of animal and non-animal toxicity evaluation results

Abstract

This invention relates to the field of computational toxicology, and in particular to a precise method for comparing the consistency of animal and non-animal toxicity evaluation results. This method acquires in vivo baseline data from animals and in vitro test data from non-animals. It employs a dynamic time warping algorithm and radial basis function kernel function to construct correction coefficients and perform equivalent mapping to generate a converted sequence. Based on a binary classification mapping of toxicity thresholds, a confusion matrix is ​​constructed to calculate sensitivity, specificity, and predicted values. Each indicator is treated as an independent source of evidence, and the Mahalanobis distance is calculated using the covariance matrix to generate a basic probability allocation function. The D-S evidence theory combination rule is applied for fusion, and a secondary factor allocation is introduced when there is conflict. Finally, the result is compared with a preset threshold to determine the feasibility of substitution. This invention adaptively eliminates the nonlinear misalignment difference between in vivo and in vitro dose responses, providing an objective quantitative judgment standard for toxicological substitution verification.

Description

Technical Field

[0001] This invention relates to the field of computational toxicology, and in particular to a precise method for comparing the consistency of animal and non-animal toxicity evaluation results. Background Technology

[0002] With the deepening implementation of the "3R" principle in toxicology, the use of non-animal in vitro cell systems to replace traditional animal experiments has become an inevitable trend in the industry. To obtain stringent regulatory approval for in vitro alternatives, it is essential to precisely quantify the consistency between in vitro dose-cell response sequences and in vivo exposure concentration-target organ toxicity sequences. This requires evaluation systems capable of deeply analyzing the complex absorption and metabolic kinetics of real organisms, thereby constructing highly statistically significant equivalent benchmarks from multidimensional, discrete biological test data.

[0003] Existing technologies mostly employ static point-to-point direct comparisons, ignoring the inherent nonlinear misalignment and hysteresis effects in the dose response patterns of in vivo and in vitro systems. They cannot adaptively eliminate natural metabolic kinetic differences, resulting in highly subjective evaluation results and a lack of rigorous mathematical comparability. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a precise comparison method for the consistency of animal and non-animal toxicity evaluation results, aiming to improve the problem of existing static comparisons ignoring in vivo and in vitro nonlinear hysteresis effects.

[0005] This invention provides the following technical solution: a precise comparison method for the consistency of animal and non-animal toxicity evaluation results, comprising: S1. Obtain the baseline data sequence in the animal body and the test data sequence in the non-animal body, both sets of sequences including dose-response discrete sampling points; S2. The dynamic time warping algorithm is used to calculate the morphological matching path of the two sets of sequences. The radial basis kernel function is used to construct the correction coefficients to perform equivalent mapping on the non-animal in vitro test data sequence, and generate the converted data sequence. S3. Based on a preset toxicity threshold, perform a binary classification mapping between the animal in vivo baseline data sequence and the converted data sequence to generate an in vivo baseline and an in vitro test tag set. Confusion matrix is ​​constructed based on the tag set, and sensitivity, specificity and predicted value are calculated. S4. Using the sensitivity, specificity, and predicted value as independent evidence sources, retrieve the covariance matrix of the preset reference distribution space, calculate the Mahalanobis distance of each independent evidence source, and generate the basic probability allocation function of the independent evidence source based on the monotonically decreasing Mahalanobis distance. S5. Apply the DS evidence theory combination rule to fuse the basic probability allocation function, and introduce a conflict allocation factor for secondary allocation when there is a conflict among the independent evidence sources, and output the comprehensive consistency confidence degree. S6. Compare the overall consistency confidence level with the preset consistency threshold. If it is not less than the preset consistency threshold, then determine that the test method corresponding to the non-animal in vitro test data has alternative feasibility.

[0006] Preferably, in step S1, the step of obtaining the in vivo baseline data sequence and the in vitro test data sequence includes: Extract raw in vivo and in vitro test data of specific compounds from a pre-set multi-source toxicology database to obtain initial data pairs covering biological response intensity and corresponding doses; The initial data pairs are denoised using a smoothing filtering algorithm to filter out random fluctuations caused by the test system. Interpolation algorithms are used to resample the denoised discrete data in dimensions, forcibly aligning the sampling coordinate sets of in vivo and in vitro data, and constructing a baseline data sequence of the animal in vivo and a test data sequence of the non-animal in vitro data with uniform resolution.

[0007] Preferably, in step S2, the step of calculating the morphological matching path of the two sets of sequences using the dynamic time warping algorithm includes: Calculate the spatial distance between the animal in vivo baseline data sequence and the non-animal in vitro test data sequence at each sampling point, and construct a two-dimensional local matching cost matrix; Based on the local matching cost matrix, dynamic programming recursive calculation is performed in combination with preset boundary constraints and monotonic step size limits to generate a cumulative cost matrix characterizing the global deformation error. A reverse search strategy is adopted to backtrack from the end point of the cumulative cost matrix to the starting point, extract the set of coordinate mappings that minimize the overall matching cost, and generate the morphological matching path that crosses the morphological differences.

[0008] Preferably, in step S2, the step of constructing correction coefficients using radial basis kernel functions to perform equivalent mapping on the non-animal in vitro test data sequence includes: Extract the local deviation features of the morphological matching path, and adaptively determine the center position parameter and attenuation width parameter of the radial basis kernel function; The non-animal in vitro test data sequence is projected into a latent feature space, and the correction coefficients characterizing the dynamic differences are generated by calculating the spatial feature distance matrix; The response amplitude of the non-animal in vitro test data sequence is dynamically scaled and offset compensated using the correction coefficient, and the converted data sequence with the system distribution bias eliminated is output.

[0009] Preferably, in step S3, the step of constructing a confusion matrix based on the label set and calculating the sensitivity, specificity, and predicted value includes: Align the in vivo baseline with the in vitro test label set one by one, and statistically analyze the absolute frequencies of true positive, false positive, true negative and false negative samples in the test sequence. The absolute frequencies are used as basic matrix elements, and are mapped and filled into the internal nodes of the two-dimensional confusion matrix composed of the baseline classification dimension and the classification dimension to be tested. Extract the feature data of the main diagonal and off-diagonal lines of the confusion matrix, obtain the sensitivity and specificity by calculating the feature ratio of true positives to true negatives, and simultaneously perform posterior probability distribution calculation to obtain the predicted value.

[0010] Preferably, in step S4, the step of calculating the Mahalanobis distance of each of the independent evidence sources includes: Extract pre-stored reference standard replacement test evaluation data, construct a multidimensional reference feature distribution space, and calculate the corresponding covariance matrix and its inverse matrix; The extracted sensitivity, specificity, and predicted values ​​are aggregated into a multidimensional observation vector, and the difference vector between the multidimensional observation vector and the expected mean vector of the multidimensional reference feature distribution space is calculated. By using the difference vector and its transpose, and combining them with the inverse of the covariance matrix, a matrix dot product operation is performed to calculate the Mahalanobis distance after removing dimensionality correlation interference.

[0011] Preferably, in step S5, the step of applying the DS evidence theory combination rule to fuse the basic probability allocation function includes: Traverse each of the basic probability assignment functions, extract the subset of hypothetical propositions that are assigned non-zero initial probability mass, and define it as the focal element set for advanced reasoning computation; For the intersection propositions in the focal element set, calculate the product of the basic probability assignment values ​​of the independent evidence sources participating in the evaluation on the corresponding propositions, and perform orthogonal sum operation on the multi-source evidence space; Divide the local result of the orthogonal sum operation by a preset global normalization constant to filter out the system empty set probability caused by evidence mutual exclusion, and obtain the joint probability quality distribution after preliminary fusion.

[0012] Preferably, in step S5, the step of introducing a conflict allocation factor for secondary allocation when there is a conflict among the independent sources of evidence includes: Extract the probability mass product of each independent evidence source assigned to the mutually exclusive proposition and perform spatial accumulation to quantify and generate the overall conflict coefficient used to characterize the global conflict intensity of heterogeneous evidence; The overall conflict coefficient is compared with the set system conflict tolerance threshold. When the conflict intensity is determined to exceed the evidence fusion limit, the conflict secondary allocation process is triggered. The local confidence scores of each independent evidence source are extracted as dynamic weights to construct the conflict allocation factor. The probability quality of highly conflicting data is redirected to the system uncertainty set according to the conflict allocation factor, thus completing the robust secondary allocation.

[0013] Preferably, in step S6, the step of comparing the overall consistency trust level with a preset consistency threshold includes: Aggregate the joint probability distribution data after the secondary allocation, extract the posterior probability values ​​of the key propositions that clearly support in vivo and in vitro equivalence, and use them as the comprehensive consistency confidence level as the basis for quantifying the adjudication. The baseline acceptance threshold specified in the test specification is retrieved, and the preset consistency threshold is dynamically locked by combining the sample size and statistical confidence level parameters of the current batch of tests. The overall consistency confidence level is compared with the preset consistency threshold. After confirming that the overall consistency confidence level is not less than the preset consistency threshold, a verification result showing that the test method has alternative feasibility is output.

[0014] The present invention has the following beneficial effects: 1. In this invention, by introducing a dynamic time warping algorithm and a radial basis kernel function to construct a nonlinear correction mapping model, the limitations of static point-to-point comparison in traditional toxicological evaluation are broken. The model adaptively eliminates the natural misalignment and lag effect in absorption and metabolism kinetics and dose response of in vivo and in vitro experiments, and provides an equivalent data benchmark with mathematical comparability for consistency quantitative evaluation.

[0015] 2. In this invention, Mahalanobis distance is used to remove the correlation interference between multidimensional toxicity evaluation indicators and generate independent evidence sources. Combined with DS evidence theory with conflict allocation factors, joint probability fusion is performed, which effectively solves the problem of fusion failure caused by extreme or contradictory performance of a single evaluation indicator, and improves the objectivity of the comprehensive evaluation results of multidimensional indicators.

[0016] 3. In this invention, by combining the test sample size and the statistical confidence level to dynamically lock the consistency decision threshold, a strict decision mechanism that adaptively adjusts with the test scale is established, which effectively controls the false positive risk in the verification process of non-animal testing alternative methods, and provides a quantitative decision standard with rigorous statistical significance for promoting the regulatory recognition of toxicological alternative methods. Attached Figure Description

[0017] Figure 1This is a flowchart of a precise comparison method for the consistency of animal and non-animal toxicity evaluation results proposed in this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In embodiments of the present invention, the present invention provides a precise method for comparing the consistency of animal and non-animal toxicity evaluation results, such as... Figure 1 As shown, it includes: S1. Obtain the baseline data sequence in the animal body and the test data sequence in the non-animal body. Both sets of sequences include dose-response discrete sampling points. Further, in step S1, the step of obtaining the in vivo baseline data sequence and the non-animal in vitro test data sequence includes: extracting the original in vivo and in vitro test data of a specific compound from a preset multi-source toxicology database to obtain initial data pairs covering the intensity of biological response and corresponding doses; using a smoothing filtering algorithm to denoise the initial data pairs and filter out random fluctuation interference caused by the test system; using an interpolation algorithm to resample the denoised discrete data in dimensions, forcibly aligning the sampling coordinate sets of the in vivo and in vitro data, and constructing an in vivo baseline data sequence and a non-animal in vitro test data sequence with uniform resolution.

[0020] Specifically, in the initial stage of acquiring baseline data sequences from animals and test data sequences from non-animal in vitro systems, the system is connected to a pre-defined multi-source toxicology database. For specific compounds, raw test data are extracted from animal target organs to determine toxicity exposure concentrations and corresponding biomarker response intensities, as well as from non-animal in vitro cell culture systems to determine drug dosage and cell viability. The extracted in vivo and in vitro test data are then used to construct initial data sets covering biological response intensity and corresponding dosage.

[0021] To address random fluctuations caused by the testing system, a smoothing filter algorithm is used to denoise the initial data pairs. Specifically, a multinomial sliding window smoothing filter algorithm is employed to perform local least squares fitting on the biological response intensity data sequence. The denoising calculation formula is as follows: ; in the formula Indicates the number after noise reduction processing Correction value for biological response intensity corresponding to each dose sampling point This represents the observed biological response intensity within the sliding window in the original noisy sequence. This indicates the set width of the local smoothing data window. This represents the pre-calculated filtering weighting coefficients obtained through polynomial least squares fitting. This processing removes high-frequency random noise caused by the hardware of biochemical detection instruments while fully preserving the key morphology and peak characteristics of the dose response data.

[0022] Because the dosing gradients in in vivo animal experiments and in vitro cell experiments often have significant physical boundaries and asymmetries, interpolation algorithms must be used to resample the denoised discrete data to force alignment between the in vivo and in vitro data sampling coordinate sets. A cubic spline interpolation function is selected to construct a continuous dose-response model. For the interval between two adjacent known denoised dose sampling points, a cubic polynomial is constructed, and its interpolation calculation formula is as follows: ; in the formula Indicates the first Specific target dose points within each interpolation interval The corresponding estimated value of continuous biological reaction intensity, This indicates the x-coordinate of the starting known dose sampling point in this interval, which is the actual test dose value. as well as and besides The coefficients are undetermined polynomial coefficients. These coefficients are obtained by solving for the equality of function values ​​at each discrete dose data point and the continuous and smooth boundary conditions. The system then uses this continuous interpolation model to re-extract equidistant response values ​​in the global dose domain according to a unified standard dose scale resolution, thereby constructing an in vivo reference data sequence and a non-animal in vitro test data sequence with a consistent coordinate reference and equal length dimensions.

[0023] This step achieves fundamental standardization of heterogeneous toxicology test data in terms of sampling dimension and signal purity, effectively eliminating dimensional barriers and instrument random error interference, and providing a scale-consistent data foundation for subsequent nonlinear calculation matching.

[0024] S2. The dynamic time warping algorithm is used to calculate the morphological matching path of the two sets of sequences. The radial basis kernel function is used to construct the correction coefficients to perform equivalent mapping on the non-animal in vitro test data sequence, and generate the converted data sequence. Further, in step S2, the step of using the dynamic time warping algorithm to calculate the morphological matching path of the two sets of sequences includes: calculating the spatial distance between the animal in vivo reference data sequence and the non-animal in vitro test data sequence at each sampling point, and constructing a two-dimensional local matching cost matrix; based on the local matching cost matrix, performing dynamic programming recursive calculation in combination with preset boundary constraints and monotonic step size limits to generate a cumulative cost matrix characterizing the global deformation error; and using a reverse search strategy to backtrack from the end point of the cumulative cost matrix to the starting point, extracting the coordinate mapping set that minimizes the overall matching cost, and generating a morphological matching path that crosses morphological differences.

[0025] Further, in step S2, the step of constructing correction coefficients using the radial basis function to perform equivalent mapping on the non-animal in vitro test data sequence includes: extracting local deviation features of the morphological matching path, adaptively determining the center position parameter and attenuation width parameter of the radial basis function; projecting the non-animal in vitro test data sequence onto the latent feature space, generating correction coefficients characterizing the dynamic differences by calculating the spatial feature distance matrix; and performing dynamic scaling and offset compensation on the response amplitude of the non-animal in vitro test data sequence using the correction coefficients, outputting a reduced data sequence that eliminates the system distribution bias.

[0026] Specifically, a preprocessed in vivo baseline data sequence of length N and a non-animal in vitro test data sequence of length M are obtained. The Euclidean distance between the bioresponse intensity at the i-th dose sampling point in the baseline data sequence and the bioresponse intensity at the j-th dose sampling point in the test data sequence is calculated to construct a two-dimensional local matching cost matrix. The formula for calculating the element values ​​is as follows: ; in the formula This represents the value of the element in the i-th row and j-th column of the local matching cost matrix. This represents the magnitude of the biological response intensity at the i-th target dose sampling point in the baseline data sequence within the animal body. This represents the magnitude of the biological response intensity at the j-th target dose sampling point in the non-animal in vitro test data sequence.

[0027] Based on the local matching cost matrix combined with preset boundary constraints and monotonic step size limits, dynamic programming is used to recursively calculate and generate the cumulative cost matrix representing the global deformation error. The recursive calculation formula is as follows: ; in the formula This represents the minimum cumulative deformation cost from the initial dose node to the position in the i-th row and j-th column of the cumulative cost matrix. as well as and These represent the cumulative matching costs of their adjacent preceding nodes on the dose response time axis.

[0028] A reverse search strategy is adopted to backtrack from the end point of the cumulative cost matrix to the starting point. Each time, the coordinate position with the minimum cumulative cost among the adjacent nodes is selected for path connection to extract the coordinate mapping set that minimizes the overall matching cost, thereby generating a morphological matching path that crosses the morphological differences of in vivo and in vitro dose response.

[0029] The difference in bioreactivity intensity between the in vivo baseline sequence and the in vitro test sequence at each mapping node in the morphological matching path is extracted as a local deviation feature to adaptively determine the center position parameter and attenuation width parameter of the radial basis function kernel. The non-animal in vitro test data sequence is projected onto the latent feature space, and a correction coefficient characterizing the difference in in vivo and in vitro pharmacokinetics is generated by calculating the spatial feature distance matrix. The calculation formula is as follows: ; in the formula This represents the correction coefficient for the j-th dose sampling point in a non-animal in vitro test data sequence. This represents the total number of mapped coordinate nodes in the shape matching path. The difference in biological response intensity at the k-th mapping node represents the feature deviation weight. This represents the center location parameter determined by the external response amplitude of the k-th mapping node. This represents the attenuation width parameter, which is adaptively generated based on the variance of the local response fluctuation.

[0030] Dynamic scaling and offset compensation are performed on the response amplitude of non-animal in vitro test data sequences using correction coefficients to output a reduced data sequence that eliminates system distribution bias. The mapping calculation formula is as follows: ; in the formula This represents the corrected biological response intensity at the j-th dose sampling point in the converted data sequence generated after the equivalent transformation. This sequence has been directly aligned to the animal's internal reference coordinate system in terms of morphological features.

[0031] This step eliminates the natural distribution misalignment in metabolic mechanisms and dose conversion between in vivo and in vitro experiments by nonlinear spatiotemporal warping and kernel function mapping, thus constructing an equivalent data benchmark with mathematical comparability.

[0032] S3. Based on a preset toxicity threshold, perform binary classification mapping on the animal in vivo baseline data sequence and the converted data sequence to generate in vivo baseline and in vitro test tag sets. Confusion matrix is ​​constructed based on the tag sets, and sensitivity, specificity and predicted value are calculated. Further, in step S3, the steps of constructing a confusion matrix based on the label set and calculating sensitivity, specificity, and predicted values ​​include: aligning the class states of the in vivo baseline and the in vitro test label set one by one, and statistically analyzing the absolute frequencies of true positive, false positive, true negative, and false negative samples in the test sequence; using the absolute frequencies as basic matrix elements, mapping and filling them into the internal nodes of the two-dimensional confusion matrix composed of the baseline classification dimension and the test classification dimension; extracting the main diagonal and off-diagonal feature data of the confusion matrix, obtaining sensitivity and specificity by calculating the feature ratio of true positive to true negative, and simultaneously performing posterior probability distribution calculations to obtain predicted values.

[0033] Specifically, a pre-defined toxicity threshold representing substantial toxicological damage is obtained. A binary classification mapping is performed by comparing the amplitude of the biological response at each target dose sampling point in the baseline data sequence of the animal with the corrected biological response in the corresponding converted data sequence. The mapping calculation formula is as follows: ; in the formula This represents the discrete binary label state generated after mapping the i-th dose sampling point. This represents the magnitude of the biological response intensity at the i-th point in the baseline or converted data sequence within the animal. This indicates a preset threshold for determining toxicity, representing cell viability or the concentration of a toxicity marker. A label value of 1 indicates that a positive toxicological reaction was triggered at this dose, while a label value of 0 indicates that no toxicological harm was observed at this dose.

[0034] After traversing the mapping, an in vivo baseline label set and an in vitro test label set consisting of binary numbers are generated respectively.

[0035] The absolute frequencies of true positive, false positive, true negative, and false negative samples in the test sequences were statistically analyzed by aligning the in vivo baseline label set and the in vitro test label set at the same dosage node. Specifically, the total number of sample points with both the in vivo baseline label and the in vitro test label set set being 1 was counted as the true positive frequency; the total number of sample points with both the in vivo baseline label and the in vitro test label set being 1 was counted as the false positive frequency; the total number of sample points with both the in vivo baseline label and the in vitro test label set being 0 was counted as the true negative frequency; and the total number of sample points with both the in vivo baseline label and the in vitro test label set being 1 was counted as the false negative frequency. These four absolute frequencies were then used as basic matrix elements and mapped and filled into the internal nodes of a two-dimensional confusion matrix composed of the animal in vivo baseline true classification dimension and the non-animal in vitro predicted classification dimension.

[0036] The feature data of the main diagonal and off-diagonal lines of the two-dimensional confusion matrix are extracted. The sensitivity and specificity of the in vitro alternative testing method are obtained by calculating the feature ratio of true positives to true negatives. Simultaneously, posterior probability distribution calculations are performed to obtain positive and negative predictive values. The performance calculation formula is as follows: ; ; ; in the formula This indicates the sensitivity of non-animal in vitro test data in accurately identifying positive in vivo toxicity samples. This indicates the specificity for accurately identifying negative samples that are actually non-toxic in vivo. This represents the positive predictive value, or conditional posterior probability, of a sample that is judged to be toxic in vitro but actually exhibits a true toxic reaction. This represents the true positive frequency extracted from the confusion matrix. Indicates the frequency of true negatives. Indicates the frequency of false positives. This indicates the frequency of false negatives.

[0037] This step transforms continuous dose-response feature vectors into discrete clinical toxicology criteria, establishing a quantitative efficacy evaluation basis for non-animal testing alternatives.

[0038] S4. Using sensitivity, specificity, and predicted values ​​as independent evidence sources, retrieve the covariance matrix of the preset reference distribution space, calculate the Mahalanobis distance of each independent evidence source, and generate the basic probability allocation function of the independent evidence source based on the monotonically decreasing Mahalanobis distance. Further, in step S4, the step of calculating the Mahalanobis distance of each independent evidence source includes: extracting pre-stored reference standard substitute test evaluation data, constructing a multidimensional reference feature distribution space and calculating the corresponding covariance matrix and its inverse matrix; aggregating the currently extracted sensitivity, specificity and predicted values ​​into a multidimensional observation vector, calculating the difference vector between the multidimensional observation vector and the expected mean vector of the multidimensional reference feature distribution space; and using the difference vector and its transpose vector, combined with the inverse matrix of the covariance matrix, performing matrix multiplication to calculate the Mahalanobis distance after removing dimensional correlation interference.

[0039] Specifically, reference standard alternative test evaluation data are extracted from a pre-stored database. This reference data covers historical indicators of the efficacy evaluation of standardized non-animal toxicology alternative test methods that have been internationally certified. A multidimensional reference feature distribution space is constructed based on the sensitivity and specificity of the historical extractions and the predicted values. Statistical analysis is performed on all historical sample indicators within this multidimensional reference feature distribution space to calculate the expected mean vector of the distribution space, and the covariance matrix of the interrelationships between the indicators of each dimension is calculated. The inverse matrix of this covariance matrix is ​​then derived by differentiation.

[0040] The sensitivity, specificity, and predicted value obtained from the current non-animal in vitro test data are aggregated and concatenated into a multidimensional observation vector in a fixed order. The difference vector between this multidimensional observation vector and the expected mean vector of the multidimensional reference feature distribution space is calculated. Using this difference vector and its transpose, along with the inverse of the covariance matrix, a matrix multiplication operation is performed to calculate the Mahalanobis distance after removing dimensionality-correlation interference. The distance calculation formula is as follows: ; in the formula The calculated Mahalanobis distance is used to objectively quantify the statistical distance between the current combination of in vitro toxicity evaluation indicators and the distribution center of historical acceptable standards. This represents a multidimensional observation vector consisting of the sensitivity, specificity, and predicted values ​​of the current in vitro testing method to be validated. This represents the expected mean vector of the multidimensional reference feature distribution space. The matrix representing the inverse of the covariance matrix of the multidimensional reference feature distribution space. It represents the transpose of the difference vector between the multidimensional observation vector and the expected mean vector.

[0041] The current testing system with extracted features is set as an independent source of evidence, and a basic probability assignment function for the in vivo-in vitro toxicity equivalence proposition is generated based on the monotonically decreasing Mahalanobis distance. The probability mapping formula is as follows: ; in the formula This indicates that the basic probability allocation function for assigning the independent source of evidence to the core proposition of the feasibility of non-animal testing as an alternative falls within a continuous interval of zero to one. This indicates that the preset distance attenuation adjustment coefficient is used to control the evaluation system's tolerance sensitivity to deviations in test performance. This represents the Mahalanobis distance obtained after performing matrix dot product and square root operations in the previous step.

[0042] This step effectively eliminates the evaluation interference caused by the inconsistency of dimensions among multidimensional toxicological evaluation indicators and the intrinsic correlation between indicators, and transforms the multiple mathematical deviations of the test results into standardized probability quality parameters.

[0043] S5. Apply the DS evidence theory combination rule to fuse the basic probability allocation function. When there is a conflict between independent evidence sources, introduce a conflict allocation factor for secondary allocation and output the comprehensive consistency confidence level. Further, in step S5, the step of applying the DS evidence theory combination rule to fuse the basic probability assignment functions includes: traversing each basic probability assignment function, extracting a subset of hypothetical propositions assigned non-zero initial probability mass, and defining it as the focal element set for advanced reasoning calculations; for the intersection propositions in the focal element set, calculating the product of the basic probability assignment values ​​of the independent evidence sources participating in the evaluation on the corresponding propositions, and performing an orthogonal sum operation on the multi-source evidence space; dividing the local result of the orthogonal sum operation by a preset global normalization constant, filtering out the system empty set probability caused by evidence mutual exclusion, and obtaining the joint probability mass distribution after preliminary fusion.

[0044] Furthermore, in step S5, the step of introducing a conflict allocation factor for secondary allocation when there is conflict among independent evidence sources includes: extracting the probability quality product of each independent evidence source to mutually exclusive propositions and performing spatial accumulation to quantify and generate an overall conflict coefficient to characterize the global conflict intensity of heterogeneous evidence; comparing the overall conflict coefficient with the set system conflict tolerance threshold, and triggering the conflict secondary allocation process when the conflict intensity is determined to exceed the evidence fusion limit; extracting the local confidence of each independent evidence source as a dynamic weight to construct a conflict allocation factor, and redirecting the probability quality of highly conflicting evidence to the system uncertainty set according to the conflict allocation factor to complete the robust secondary allocation.

[0045] Specifically, the basic probability allocation function for independent evidence sources is obtained, derived from sensitivity, specificity, and predicted values. In the consistency evaluation system for toxicological alternative methods, a system identification framework is defined, comprising three basic propositions: feasible alternatives, infeasible alternatives, and undetermined alternative feasibility. A subset of hypothetical propositions with non-zero initial probability mass is extracted by traversing each basic probability allocation function and defined as the focal element set for advanced inference calculations.

[0046] For the intersection propositions in the focal element set, the product of the basic probability assignments of the independent evidence sources participating in the evaluation on the corresponding propositions is calculated, and an orthogonal sum operation of the multi-source evidence space is performed. The local result of the orthogonal sum operation is divided by a preset global normalization constant to filter out the system empty set probability caused by evidence mutual exclusion, obtaining the preliminary fused joint probability quality distribution. The calculation formula is as follows: ; in the formula This refers to the toxicity equivalence proposition generated by fusing two independent sources of evidence. The new fundamental probability assignment function takes values ​​of . This indicates the first independent source of evidence for the proposition. The basic probability allocation value, This indicates that the second independent source of evidence is relevant to the proposition. The basic probability allocation value, This means that traversing all intersections equals the target proposition. The combination of focal elements, The overall conflict coefficient, which represents the global conflict intensity of heterogeneous evidence, is used as a core parameter in the calculation of the global normalization constant.

[0047] Extract the probability-mass product of each independent source of evidence assigned to mutually exclusive propositions and perform spatial summation quantization to generate the overall conflict coefficient. The quantization calculation formula is as follows: ; in the formula This represents the combination of propositions that are completely mutually exclusive and have an empty intersection in the set of focal elements. The result of this multiplication and summation intuitively reflects the intensity of the opposite orientation of different evaluation indicators such as sensitivity and specificity towards the same toxicological proposition.

[0048] The overall conflict coefficient generated by quantification The system's conflict coefficient is numerically compared with the set system conflict tolerance threshold. When the overall conflict coefficient is not greater than the system conflict tolerance threshold, the joint probability quality distribution is directly output. When the overall conflict coefficient exceeds the system conflict tolerance threshold (i.e., the conflict intensity exceeds the evidence fusion limit), a secondary conflict allocation process is triggered. The local confidence scores of each independent evidence source are extracted as dynamic weights to construct a conflict allocation factor. This redirects highly conflicting probability qualities to the system uncertainty set, completing a robust secondary allocation. The redistribution calculation formula is as follows: ; in the formula This represents the overall consistency confidence level of the final output after conflict reassignment. This represents a conflict allocation factor constructed based on the reliability of local evidence sources. This factor forces the overall conflict coefficient to be adjusted. The contradictory probabilistic qualities they represent are stripped away and transferred to the set of propositions with uncertain alternative feasibility in the system identification framework to avoid the system making erroneous toxicological decisions with high confidence.

[0049] This step addresses the issue of fusion failure caused by extreme performance of a single evaluation indicator by introducing a quantitative conflict coefficient and a probability redirection mechanism, thereby improving the robustness of the joint evaluation results of multiple indicators.

[0050] S6. Compare the overall consistency confidence level with the preset consistency threshold. If it is not less than the preset consistency threshold, it is determined that the test method corresponding to the non-animal in vitro test data has alternative feasibility.

[0051] Furthermore, in step S6, the step of comparing the overall consistency confidence level with the preset consistency threshold includes: aggregating the joint probability distribution data after secondary allocation, extracting the posterior probability values ​​of the focus propositions that clearly support in vivo and in vitro equivalence, and using them as the basis for quantifying the overall consistency confidence level; retrieving the benchmark acceptance threshold specified in the test specification, and dynamically locking the preset consistency threshold by combining the sample size and statistical confidence level parameters of the current batch of tests; performing a numerical comparison between the overall consistency confidence level and the preset consistency threshold, and outputting a verification result showing that the test method has alternative feasibility after confirming that the overall consistency confidence level is not less than the preset consistency threshold.

[0052] Specifically, the joint probability distribution data of the secondary allocation after conflict allocation factor processing is aggregated. Within the constructed toxicology system identification framework, the posterior probability values ​​of key propositions that explicitly support in vivo-in vitro equivalence between non-animal tests and animal benchmark tests are extracted. The posterior probability value of this key proposition is directly used as the quantitative basis for the comprehensive consistency confidence degree, and its numerical extraction formula is as follows: ; in the formula This represents the overall consistency confidence level of the system's final output used to evaluate the reliability of in vitro alternative methods. This represents the set of core propositions within the system identification framework that demonstrate the feasibility and equivalence of alternative testing methods. This represents all proper subsets of propositions belonging to the core focus set. This indicates that the proposition falls within the corresponding subset after a secondary distribution. The final probability mass assignment value.

[0053] The baseline acceptance threshold specified in the toxicology testing guidelines is retrieved, and the sample size and statistical confidence level parameters of the current batch of tests are combined to dynamically lock the preset consistency threshold. The dynamic calculation formula is as follows: ; in the formula This represents the preset consistency threshold that is finally locked after correction for sample size and confidence level. This refers to the static baseline acceptance threshold pre-set by international toxicology verification centers or relevant regulations and standards. This indicates the total effective sample size covered by the current batch of non-animal in vitro tests, i.e., the total number of dose response discrete sampling points. This represents the standard normal distribution quantile parameter corresponding to the target statistical confidence level. This formula allows the system to dynamically raise the decision threshold when the validation sample size is small, thereby strictly controlling the risk of false positives in surrogate validation.

[0054] The system performs a numerical comparison between the overall consistency confidence level and a preset consistency threshold. When the overall consistency confidence level is confirmed to be greater than or equal to the preset consistency threshold, the system determines that the test data is sufficient and the evaluation indicators meet the regulatory equivalence requirements, triggering a verification pass command and outputting a formal verification result indicating that the test method corresponding to the non-animal in vitro test data has alternative feasibility. If the overall consistency confidence level is less than the preset consistency threshold, the system directly outputs a conclusion that there is no alternative feasibility or suggests that the test sample size needs to be expanded.

[0055] This step completes the transformation from multidimensional probabilistic reasoning to explicit verification conclusions, establishes a dynamic adjudication mechanism that adapts to the size of the test sample, and improves the statistical rigor of toxicity alternative methods evaluation.

[0056] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A precise comparative method for the consistency of animal and non-animal toxicity evaluation results, characterized in that, include: S1. Obtain the baseline data sequence in the animal body and the test data sequence in the non-animal body, both sets of sequences including dose-response discrete sampling points; S2. The dynamic time warping algorithm is used to calculate the morphological matching path of the two sets of sequences. The radial basis kernel function is used to construct the correction coefficients to perform equivalent mapping on the non-animal in vitro test data sequence, and generate the converted data sequence. S3. Based on a preset toxicity threshold, perform a binary classification mapping between the animal in vivo baseline data sequence and the converted data sequence to generate an in vivo baseline and an in vitro test tag set. Confusion matrix is ​​constructed based on the tag set, and sensitivity, specificity and predicted value are calculated. S4. Using the sensitivity, specificity, and predicted value as independent evidence sources, retrieve the covariance matrix of the preset reference distribution space, calculate the Mahalanobis distance of each independent evidence source, and generate the basic probability allocation function of the independent evidence source based on the monotonically decreasing Mahalanobis distance. S5. Apply the DS evidence theory combination rule to fuse the basic probability allocation function, and introduce a conflict allocation factor for secondary allocation when there is a conflict among the independent evidence sources, and output the comprehensive consistency confidence degree. S6. Compare the overall consistency confidence level with the preset consistency threshold. If it is not less than the preset consistency threshold, then determine that the test method corresponding to the non-animal in vitro test data has alternative feasibility.

2. The precise comparison method for the consistency of animal and non-animal toxicity evaluation results according to claim 1, characterized in that, Step S1, the step of acquiring the in vivo baseline data sequence and the in vitro test data sequence of the animal, includes: Extract raw in vivo and in vitro test data of specific compounds from a pre-set multi-source toxicology database to obtain initial data pairs covering biological response intensity and corresponding doses; The initial data pairs are denoised using a smoothing filtering algorithm to filter out random fluctuations caused by the test system. Interpolation algorithms are used to resample the denoised discrete data in dimensions, forcibly aligning the sampling coordinate sets of in vivo and in vitro data, and constructing a baseline data sequence of the animal in vivo and a test data sequence of the non-animal in vitro data with uniform resolution.

3. The precise comparison method for the consistency of animal and non-animal toxicity evaluation results according to claim 1, characterized in that, In step S2, the step of calculating the morphological matching path of the two sets of sequences using the dynamic time warping algorithm includes: Calculate the spatial distance between the animal in vivo baseline data sequence and the non-animal in vitro test data sequence at each sampling point, and construct a two-dimensional local matching cost matrix; Based on the local matching cost matrix, dynamic programming recursive calculation is performed in combination with preset boundary constraints and monotonic step size limits to generate a cumulative cost matrix characterizing the global deformation error. A reverse search strategy is adopted to backtrack from the end point of the cumulative cost matrix to the starting point, extract the set of coordinate mappings that minimize the overall matching cost, and generate the morphological matching path that crosses the morphological differences.

4. The precise comparison method for the consistency of animal and non-animal toxicity evaluation results according to claim 1, characterized in that, In step S2, the step of constructing correction coefficients using radial basis kernel functions to perform equivalent mapping on the non-animal in vitro test data sequence includes: Extract the local deviation features of the morphological matching path, and adaptively determine the center position parameter and attenuation width parameter of the radial basis kernel function; The non-animal in vitro test data sequence is projected into a latent feature space, and the correction coefficients characterizing the dynamic differences are generated by calculating the spatial feature distance matrix; The response amplitude of the non-animal in vitro test data sequence is dynamically scaled and offset compensated using the correction coefficient, and the converted data sequence with the system distribution bias eliminated is output.

5. The precise comparison method for the consistency of animal and non-animal toxicity evaluation results according to claim 1, characterized in that, Step S3, the step of constructing a confusion matrix based on the label set and calculating sensitivity, specificity, and predicted values, includes: Align the in vivo baseline with the in vitro test label set one by one, and statistically analyze the absolute frequencies of true positive, false positive, true negative and false negative samples in the test sequence. The absolute frequencies are used as basic matrix elements, and are mapped and filled into the internal nodes of the two-dimensional confusion matrix composed of the baseline classification dimension and the classification dimension to be tested. Extract the feature data of the main diagonal and off-diagonal lines of the confusion matrix, obtain the sensitivity and specificity by calculating the feature ratio of true positives to true negatives, and simultaneously perform posterior probability distribution calculation to obtain the predicted value.

6. The precise comparison method for the consistency of animal and non-animal toxicity evaluation results according to claim 1, characterized in that, Step S4, the step of calculating the Mahalanobis distance of each of the independent evidence sources, includes: Extract pre-stored reference standard replacement test evaluation data, construct a multidimensional reference feature distribution space, and calculate the corresponding covariance matrix and its inverse matrix; The extracted sensitivity, specificity, and predicted values ​​are aggregated into a multidimensional observation vector, and the difference vector between the multidimensional observation vector and the expected mean vector of the multidimensional reference feature distribution space is calculated. By using the difference vector and its transpose, and combining them with the inverse of the covariance matrix, a matrix dot product operation is performed to calculate the Mahalanobis distance after removing dimensionality correlation interference.

7. The precise comparison method for the consistency of animal and non-animal toxicity evaluation results according to claim 1, characterized in that, Step S5, the step of applying the DS evidence theory combination rule to fuse the basic probability assignment function, includes: Traverse each of the basic probability assignment functions, extract the subset of hypothetical propositions that are assigned non-zero initial probability mass, and define it as the focal element set for advanced reasoning computation; For the intersection propositions in the focal element set, calculate the product of the basic probability assignment values ​​of the independent evidence sources participating in the evaluation on the corresponding propositions, and perform orthogonal sum operation on the multi-source evidence space; Divide the local result of the orthogonal sum operation by a preset global normalization constant to filter out the system empty set probability caused by evidence mutual exclusion, and obtain the joint probability quality distribution after preliminary fusion.

8. The precise comparison method for the consistency of animal and non-animal toxicity evaluation results according to claim 1, characterized in that, Step S5, the step of introducing a conflict allocation factor for secondary allocation when there is a conflict among the independent sources of evidence, includes: Extract the probability mass product of each independent evidence source assigned to the mutually exclusive proposition and perform spatial accumulation to quantify and generate the overall conflict coefficient used to characterize the global conflict intensity of heterogeneous evidence; The overall conflict coefficient is compared with the set system conflict tolerance threshold. When the conflict intensity is determined to exceed the evidence fusion limit, the conflict secondary allocation process is triggered. The local confidence scores of each independent evidence source are extracted as dynamic weights to construct the conflict allocation factor. The probability quality of highly conflicting data is redirected to the system uncertainty set according to the conflict allocation factor, thus completing the robust secondary allocation.

9. The precise comparison method for the consistency of animal and non-animal toxicity evaluation results according to claim 1, characterized in that, Step S6, the step of comparing the overall consistency trust level with the preset consistency threshold, includes: Aggregate the joint probability distribution data after the secondary allocation, extract the posterior probability values ​​of the key propositions that clearly support in vivo and in vitro equivalence, and use them as the comprehensive consistency confidence level as the basis for quantifying the adjudication. The baseline acceptance threshold specified in the test specification is retrieved, and the preset consistency threshold is dynamically locked by combining the sample size and statistical confidence level parameters of the current batch of tests. The overall consistency confidence level is compared with the preset consistency threshold. After confirming that the overall consistency confidence level is not less than the preset consistency threshold, a verification result showing that the test method has alternative feasibility is output.

Images