A toxicity pathway prediction method based on small sample transfer reasoning

Abstract

This invention discloses a toxicity pathway prediction method based on small-sample transfer reasoning, belonging to the field of toxicity prediction technology. Addressing the problems of redundant interference from candidate pathways, lack of mechanism constraints in key pathway identification, and poor prediction stability due to sparse samples of weak toxicity categories in existing technologies, this invention first aggregates candidate pathways associated with toxicity targets, completing mechanism feature characterization and target condition encoding. Then, it calculates a comprehensive contribution score by combining pathway importance, necessity, sufficiency, and AOP mechanism priors to screen key toxicity pathways. Subsequently, it constructs a key pathway structure view and a reasoning path sequence view, and generates a unified reasoning representation through gating fusion. Finally, it decomposes the unified reasoning representation into a shared mechanism representation and a specific representation, and performs small-sample compensation through cross-toxicity transfer, outputting joint prediction results for five toxicity categories. This invention can improve the reliability of key pathway identification and the prediction performance of weak toxicity categories, and is applicable to the prediction and mechanism explanation of multiple toxicities of traditional Chinese medicine.

Description

Technical Field

[0001] This invention belongs to the field of toxicity prediction technology, specifically relating to a toxicity pathway prediction method based on small sample transfer reasoning. Background Technology

[0002] In recent years, with the rapid development of high-throughput screening (HTS) technology, toxicogenomics, and the Adverse Outcome Pathway (AOP) framework, toxicity prediction research is gradually shifting from a "single target-single pathway" guided model to a mechanism-driven, multi-pathway comprehensive assessment model. Especially in the field of traditional Chinese medicine (TCM) toxicity evaluation, the multi-component and multi-target characteristics of TCM compound formulas mean that toxic effects often involve the synergistic or antagonistic effects of multiple pathways, making accurate assessment through a single target or pathway difficult.

[0003] The Aspect-Oriented Programming (AOP) framework has become a cornerstone of 21st-century toxicology, providing a structured approach to understanding the causal relationships between tissue molecular events and adverse outcomes. In recent years, the combination of the AOP framework and deep learning has shown promising promise in toxicity prediction. For example, the ToxReason benchmark framework, based on AOP, assesses the ability to infer toxicity across multiple organs; the AOP-guided machine learning framework links molecular initiation events with neurotoxicity mechanisms. In multimodal modeling, AOP-based quantitative modeling methods (such as qAOP) have been used to improve in vitro-to-in vivo prediction translations, potentially reducing the need for animal experiments.

[0004] In terms of pathway modeling, existing methods mainly adopt the following technical approaches:

[0005] 1. Identify significantly enriched pathways based on pathway enrichment analysis, and then use these pathways as features to input into a machine learning model for toxicity prediction.

[0006] 2. Single-path independent modeling is adopted, and a prediction model is built for each path before the results are fused.

[0007] 3. Based on a multi-view learning framework, integrate multi-source information such as drug, side effects, and molecular structure to predict toxicity.

[0008] However, existing technologies still have the following problems:

[0009] 1. The candidate pathway processing is too simplistic, making it difficult to distinguish between key pathways and weakly related pathways. Existing methods often perform equal-weighted input or simple aggregation of candidate pathways, lacking a layer-by-layer screening mechanism from candidate pathways to key pathways. This can easily lead to the weakening of key pathway features and the amplification of redundant information.

[0010] 2. Lack of causal constraints in key pathway identification. Existing methods mostly rank pathways based on statistical significance or correlation, without incorporating mechanisms such as necessity analysis, sufficiency analysis, and counterfactual reasoning, resulting in insufficient stability of key pathway identification results.

[0011] 3. Insufficient multi-view reasoning mechanisms related to key pathways, resulting in limited explanatory power. Existing methods lack view designs corresponding to key pathway structural information and reasoning path sequence information, and do not fully utilize mechanism priors to guide view fusion, leading to insufficient explanatory power of the reasoning process.

[0012] 4. Insufficient predictive ability for weak toxicity categories, lack of compensation mechanisms for small samples. Existing methods have not established specific compensation strategies for weak toxicity categories such as hepatotoxicity, neurotoxicity, and reproductive toxicity, resulting in insufficient predictive stability for weak toxicity categories. Summary of the Invention

[0013] To address the problems mentioned in the background art, this invention provides a toxicity pathway prediction method based on small sample transfer reasoning, which solves the problems of existing technologies that make it difficult to distinguish between critical pathways and weakly related pathways, lack of causal mechanism constraints in critical pathway identification, insufficient multi-view reasoning mechanism related to critical pathways, and insufficient prediction ability of weak toxicity categories.

[0014] To achieve the above objectives, the present invention provides the following technical solution:

[0015] A toxic pathway prediction method based on small-sample transfer reasoning includes the following steps:

[0016] S1: For the toxicity target to be predicted, gather all candidate pathways that are associated with the toxicity target to form a candidate pathway set, and obtain the original feature vectors corresponding to each candidate pathway to construct the original feature matrix of the candidate pathway.

[0017] S2: Characterize the mechanism features of each candidate pathway in the candidate pathway set, fuse the global features corresponding to the toxicity target with the reason summary features to obtain the target context representation, encode the target condition of each candidate pathway through the target context representation, output the candidate pathway encoding matrix and input it into S3.

[0018] S3: Using the candidate pathway coding matrix as input, calculate the importance weight of each candidate pathway. At the same time, calculate the necessity score and sufficiency score of each candidate pathway through counterfactual analysis. Combine the prior score of the AOP mechanism and the general pathway correction factor to generate a comprehensive contribution score. Sort and screen the key toxic pathways according to the comprehensive contribution score and input them into S4.

[0019] S4: Based on the set of key toxic pathways, extract the structural features and reason path sequence features of the key pathways in the set of key toxic pathways respectively, construct the structural view and the sequence view, and at the same time aggregate the mechanism prior information of the key pathways to generate the mechanism bias vector. After gating fusion, the unified reasoning representation of the dual view fusion is obtained and input into S5.

[0020] S5: Decompose the unified reasoning representation into a shared mechanism representation and specific representations of various toxicities, calculate the basic discriminant values ​​of various toxicities, and then perform small-sample compensation through cross-toxicity transfer terms. Finally, output the joint prediction results of multiple toxicities including cardiotoxicity, hepatotoxicity, nephrotoxicity, neurotoxicity and reproductive toxicity.

[0021] Preferably, in S1, the candidate pathway set consists of all biological pathways associated with the current toxicity target, and the original features of each candidate pathway are arranged in rows to form an original feature matrix of dimension n×d, where n is the number of candidate pathways and d is the feature dimension of a single pathway.

[0022] Preferably, in S2, the mechanism features of a single candidate pathway are composed of basic evidence features, database source features, a priori features of AOP mechanisms of five types of toxicity, pathway general correction factors, and a priori summary features; wherein the pathway general correction factors are calculated based on the pathway support scale statistics, source breadth statistics, and general keyword suppression terms.

[0023] Preferably, the target conditional encoding of S2 specifically involves: mapping the mechanism features of candidate pathways to a unified latent space to obtain an initial pathway representation; fusing global features and reason summary features to obtain a target context representation; modulating the initial pathway representation element-wise using gating vectors and offset vectors to obtain the target conditional encoding representation of each candidate pathway, and arranging them in rows to form a candidate pathway encoding matrix; wherein the modulation calculation satisfies: ;

[0024] In the formula, For the first The target condition encoding representation of each candidate pathway. Let be the gate vector of the i-th candidate path. Let be the offset vector of the i-th candidate path. Let be the initial path representation of the i-th candidate path, and ⊙ denotes element-wise multiplication. The preferred method combines candidate pathway encoding representations with target context representations to calculate the importance score of each candidate pathway. After normalization, the importance weight of each candidate pathway is obtained. Based on the importance weight, the encoding representations of all candidate pathways are weighted and aggregated to obtain the candidate pathway aggregated representation of the toxicity target, and the initial multitoxicity prediction results are output.

[0025] Preferably, in S3, the necessity score is calculated by the difference in prediction results before and after removing a single candidate pathway, and the sufficiency score is calculated by the mean of prediction results when only a single candidate pathway is retained; the mechanism prior score is calculated by weighting the AOP prior vector of the candidate pathway with the initial multitoxicity prediction results; the general pathway correction factor is modified by combining the mechanism prior score to obtain the modified general factor.

[0026] Preferably, in S3, the necessity score, sufficiency score, mechanism prior score, and importance weight are weighted and summed according to their corresponding coefficients, and then multiplied by a modified general factor to obtain the comprehensive contribution score of each candidate pathway; the top k candidate pathways with the highest comprehensive contribution scores are selected to form a set of key toxicity pathways, where k is the preset number of key pathways to be retained; wherein the comprehensive contribution score satisfies:

[0027] ;

[0028] In the formula The overall contribution score for the i-th candidate pathway is... For necessary fractions, For sufficiency fractions, For the mechanism's prior score, As importance weight, For the corrected general factor, , , , These are the weighting coefficients for the corresponding components.

[0029] Preferably, in S4, the encoded representation of the key pathway is concatenated with the original features and mapped to obtain the structural view input, and the structural view representation is output through the structural view encoder; the reason sequence unit features are mapped to obtain the sequence view input, and the sequence view representation is output through the sequence view encoder; the AOP features of the key pathway are aggregated with the general correction factor to obtain the mechanism bias vector, and the structural view representation, the sequence view representation and the mechanism bias vector are fused through the gating mechanism to output the unified inference representation.

[0030] Preferably, in S5, the unified reasoning representation is mapped to obtain the shared mechanism representation and the toxicity-specific representation corresponding to each type of toxicity. After adding the shared mechanism representation and the specific representation of the same toxicity category, the basic discriminant value of that type of toxicity is calculated.

[0031] Preferably, in step S5, based on the migration relationship coefficients between various toxicities, the basic discriminant values ​​of other toxicity categories are migrated and superimposed onto the current toxicity category to obtain a corrected discriminant value; temperature correction is applied to the discriminant values ​​of various toxicities, and after obtaining the predicted probability through Sigmoid mapping, the final toxicity discriminant label is output through a category-specific discriminant threshold; wherein the corrected discriminant value satisfies:

[0032] ;

[0033] In the formula This is the corrected discriminant value for the toxicity target t in the class c toxicity. This serves as the basic discriminant value for Class c toxicity. denoted as the migration coefficient from class j toxicity to class c toxicity.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] 1. Candidate pathway set modeling: For each toxicity target, a candidate pathway set is constructed, and pathway evidence, database sources, prior knowledge of toxicity mechanisms and general pathway correction information are uniformly encoded to replace the traditional single-target or single-pathway direct prediction method, thereby improving the completeness and stability of toxicity mechanism characterization.

[0036] 2. Key pathway identification mechanism: The importance of a pathway is calculated within the candidate pathway set, and a comprehensive evaluation is conducted by combining necessity, sufficiency, mechanism prior and pathway generality correction, so as to realize the automatic identification and screening of key toxic pathways and reduce the interference of redundant and broad pathways on the prediction results.

[0037] 3. Dual-view transfer reasoning: Jointly model key pathway structure information with reasoning path sequence information, and introduce shared representation and cross-toxicity transfer mechanism to achieve joint prediction of multiple toxicities and enhance the discrimination ability of weak toxicity category and small sample category.

[0038] 4. Joint optimization and result correction: Integrating multi-toxicity prediction, key pathway identification, and dual-viewpoint optimization. Figure 1 Consistency and migration constraints are incorporated into a unified optimization framework, and combined with category-level probability correction and threshold determination, to improve the accuracy, stability and applicability of the overall prediction results. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0040] To facilitate understanding of the technical content of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0041] A method for predicting toxic pathways based on small-sample transfer reasoning, such as Figure 1 As shown, the analysis of toxic targets mainly includes the following steps:

[0042] S1: For the toxicity target to be predicted, gather all candidate pathways that are associated with the toxicity target to form a candidate pathway set, and obtain the original feature vectors corresponding to each candidate pathway to construct the original feature matrix of the candidate pathway.

[0043] S2: Characterize the mechanism features of each candidate pathway in the candidate pathway set, fuse the global features corresponding to the toxicity target with the reason summary features to obtain the target context representation, encode the target condition of each candidate pathway through the target context representation, output the candidate pathway encoding matrix and input it into S3.

[0044] S3: Using the candidate pathway coding matrix as input, calculate the importance weight of each candidate pathway. At the same time, calculate the necessity score and sufficiency score of each candidate pathway through counterfactual analysis. Combine the prior score of the AOP mechanism and the general pathway correction factor to generate a comprehensive contribution score. Sort and screen the key toxic pathways according to the comprehensive contribution score and input them into S4.

[0045] S4: Based on the set of key toxic pathways, extract the structural features and reason path sequence features of the key pathways in the set of key toxic pathways respectively, construct the structural view and the sequence view, and at the same time aggregate the mechanism prior information of the key pathways to generate the mechanism bias vector. After gating fusion, the unified reasoning representation of the dual view fusion is obtained and input into S5.

[0046] S5: Decompose the unified reasoning representation into a shared mechanism representation and specific representations of various toxicities, calculate the basic discriminant values ​​of various toxicities, and then perform small-sample compensation through cross-toxicity transfer terms. Finally, output the joint prediction results of multiple toxicities including cardiotoxicity, hepatotoxicity, nephrotoxicity, neurotoxicity and reproductive toxicity.

[0047] Specifically, the sub-steps of each step are as follows:

[0048] S1.1: Pooling candidate pathways associated with toxicity targets;

[0049] For any toxic target First, all candidate pathways related to the target are collected to form a candidate pathway set, represented as:

[0050] ;

[0051] in, Indicates a toxic target. This represents the set of candidate pathways corresponding to the toxicity target. Represents the first element in the set. Candidate pathways, Indicates the first Feature vectors of candidate pathways This indicates the number of candidate pathways corresponding to the toxicity target;

[0052] S1.2: Arranging the features of each candidate pathway by row yields the original feature matrix of the candidate pathway set:

[0053] ;

[0054] in, This represents the feature dimension of a single path.

[0055] S2.1: Characterize the mechanistic features of candidate pathways;

[0056] To simultaneously preserve the strength of evidence, source information, and mechanistic priors of candidate pathways, the first... The characteristics of each candidate pathway are represented as follows:

[0057] ;

[0058] in, Indicates the characteristics of basic evidence. Indicates database source characteristics, Prior characteristics of AOP mechanisms representing five classes of toxicity. Indicates the general modification factor for the pathway. [;] represents the prior summary features, and [;] represents vector concatenation. Among them, the basic evidence features describe the direct support relationship between the pathway and the current toxicity target, as well as the scale of upstream and downstream support; the database source features distinguish the database to which the pathway belongs; the AOP mechanism prior features describe the degree of mechanistic matching between the pathway and the five classes of toxicity; the pathway generality correction factor is used to suppress overly broad pathways; and the prior summary features preserve the comprehensive information after integrating external mechanism knowledge. Pathway generality correction factor It can be represented as:

[0059] ;

[0060] in, Indicates the first A general correction factor for candidate pathways. This represents the statistic of the support scale for this pathway. This represents the statistics on the breadth of sources of this pathway. This represents the generic keyword suppression term for this pathway. This represents the function for calculating the correction factor.

[0061] S2.2: Fuse target context information;

[0062] Since the importance of the same pathway varies depending on the toxicity of the target, target-level contextual information is further introduced to perform target conditional encoding on candidate pathways. This contextual information includes global features. and reason summary features Among them, global features are used to represent the overall statistical information of the current candidate pathway set, and reason summary features are used to represent the overall statistical information of the current candidate pathway set. Used to summarize toxicity targets Mechanistic evidence provided by multiple reasoning paths between candidate pathways. For toxicity targets. This involves extracting reason pathways that start from toxic targets and end with candidate pathways from the correlation data between targets and candidate pathways. The characteristics of a reason path are represented as follows:

[0063] ;

[0064] in, Indicates toxicity target The corresponding number The characteristics of the reason path; This represents the sequence encoding of relations in the path, formed according to their actual transmission order; The encoding of the candidate path corresponding to the path endpoint; The database source code representing the evidence for this path; This indicates the strength of the association along the path; This indicates the number of relations contained in the path; This represents the prior vector of the AOP mechanism corresponding to the path in terms of cardiotoxicity, hepatotoxicity, nephrotoxicity, neurotoxicity, and reproductive toxicity; the semicolon in square brackets indicates vector concatenation.

[0065] Different reasoning pathways provide varying degrees of support for the current toxicity target; therefore, normalized weights for each pathway are calculated:

[0066] ;

[0067] ;

[0068] in, Indicates the first Importance score of each reason path; This represents a learnable rating vector that maps path features to scalar scores; superscript Indicates vector transpose; The bias vector representing the path score; Represents the hyperbolic tangent function; Indicates the first Normalized weights for each reason path; Indicates toxicity target The corresponding number of reason paths; Indicates the reason path index during normalized summation; This represents an exponential function, where the sum of all path weights is 1. The weighted summation is performed based on the normalized weights of each reason path to obtain the reason summary characteristics:

[0069] ;

[0070] in, Indicates toxicity target The summary characteristics of the reasons; Indicates the first The characteristics of the reason path; This represents the normalized weight of the path; Indicates the reason path index; This represents the total number of reason paths. This weighted aggregation process ensures that reasoning paths with higher levels of evidence support account for a larger proportion of the target context representation.

[0071] First, the candidate pathway features are mapped to a unified latent space to obtain the initial pathway representation:

[0072] ;

[0073] in, Indicates the first Initial representation of candidate pathways, This represents the feature vector of the pathway. Represents the characteristic projection matrix, Indicates the bias term. This represents a nonlinear mapping function. Then, the global features and the reason-summated features are fused to obtain the toxicity target. Contextual representation:

[0074] ;

[0075] in, Indicates toxicity target The context indicates that Represents global features. This indicates the summarization feature of reason. and These represent the corresponding mapping matrices. Indicates the bias term. This represents a nonlinear mapping function.

[0076] After obtaining the initial path representation and context representation, target conditional modulation is performed on each candidate path to obtain the final encoding result:

[0077] ;

[0078] ;

[0079] ;

[0080] in, Indicates the first The gating vectors of the candidate paths, Indicates the first The offset vector of each candidate path, This represents the path representation after target condition encoding. , , , Indicates the mapping parameters. and Represents a nonlinear mapping function. This indicates element-wise multiplication. Finally, the encoding results of all candidate paths are arranged row by row to obtain the module output:

[0081] ;

[0082] in, Indicates toxicity target Candidate path coding matrix, Indicates the first The representation of candidate pathways after target condition encoding. Indicates the number of candidate pathways. This indicates the dimension of the encoded features. This will serve as the input for the subsequent critical path identification module.

[0083] S3.1: The same toxic target is often associated with multiple candidate pathways, but the contributions of different candidate pathways to the current toxic target are not the same. These include key pathways that are truly involved in the occurrence of toxicity, as well as weakly associated pathways caused by database reuse, enrichment bias, or overly generalized names. Therefore, this section further identifies key pathways within the candidate pathway set.

[0084] For toxic targets Candidate path encoding matrix Let the first The encoding of each candidate pathway is represented as follows First, the importance score of this pathway is calculated under the current toxicity target conditions:

[0085] ;

[0086] in, Indicates the first Importance scores of candidate pathways Indicates toxicity target The context indicates that and These represent the mapping matrices for path representation and context representation, respectively. Indicates the bias term. Represents the fractional projection vector. Indicates transpose. This represents a nonlinear mapping function.

[0087] The importance scores of each candidate pathway are normalized to obtain the candidate pathway weights:

[0088] ;

[0089] in, Indicates the first Normalized weights of candidate pathways for current toxicity targets This represents an exponential function. Based on the normalized weights, the toxicity target can be obtained. Candidate pathway aggregation representation:

[0090] ;

[0091] After obtaining the toxic target After candidate pathway aggregation representation, toxicity targets can be further obtained. Initial multitoxicity prediction results:

[0092] ;

[0093] in, Indicates toxicity target The initial multitoxicity prediction probability vector, This represents the output mapping matrix. Indicates the bias term. This represents the Sigmoid mapping function.

[0094] This allows us to obtain the relative strength of each candidate pathway in the overall prediction.

[0095] S3.2: Calculate the necessity, sufficiency, and prior score of the pathway;

[0096] To improve the stability of key pathway identification, counterfactual analysis is further adopted to characterize the importance of candidate pathways from two perspectives: "how much the prediction results change after removing a certain pathway" and "how much of the prediction results can be retained when only a certain pathway is kept".

[0097] use Indicates the removal of the first Prediction results after selecting candidate pathways This indicates that only the first [item] is retained. The prediction results for the candidate pathway are as follows: The necessity score of a candidate pathway is defined as:

[0098] ;

[0099] Where C represents the number of toxicity categories, When all candidate pathways are retained, the first... Predicted probability of toxicity Indicates the removal of the first The first candidate pathway Predicted probability of toxicity This indicates the operation of taking a non-negative value.

[0100] No. The sufficiency score of a candidate pathway is defined as follows:

[0101] ;

[0102] in, This indicates that only the first [item] is retained. The predicted probability of class c toxicity when there are 1 candidate pathway;

[0103] Building upon this, a priori toxicity mechanism is further introduced to refine the identification results of key pathways. Let the first... The AOP prior vectors of the candidate paths are Then, the mechanism prior score of this pathway under the current prediction conditions is expressed as:

[0104] ;

[0105] in, Indicates the first The prior scores of the mechanisms of the candidate pathways, This indicates that the pathway is in the first... AOP prior score on toxicity class This represents a small constant used to prevent the denominator from being zero.

[0106] Because some pathways are highly generalized, indiscriminate screening can easily lead to a bias in the identification of key pathways, resulting in a narrow focus. Therefore, a generalized pathway correction factor is applied. Prior score of the mechanism The combined effect yields the corrected general factor:

[0107] ;

[0108] in, Indicates the first The general factor after modification of candidate pathways. The prior adjustment coefficient of the mechanism is indicated. This represents the lower bound of a general factor. This represents a truncation function used to restrict the result to a given interval.

[0109] S3.3: Generate a comprehensive contribution score and screen key pathways;

[0110] After obtaining the necessity score, sufficiency score, mechanism prior score, and the corrected general factor, construct the comprehensive contribution score of the i-th candidate pathway:

[0111] ;

[0112] in, Indicates the normalized weights. , , and This represents the weighting coefficient of each component.

[0113] All candidate pathways are ranked according to their overall contribution scores, and the top k candidate pathways with the highest scores are selected as the key pathway set.

[0114] ;

[0115] Where TopK() represents obtaining the top k candidate paths. This indicates the number of critical pathways that are retained.

[0116] The above method simultaneously considers the necessity of candidate pathways for the overall prediction results, the sufficiency of individual pathways supporting the prediction results, the degree of mechanism prior Pixit, and general pathway correction factors, thereby making the key pathway identification results more stable.

[0117] S4.1: Dual-view toxicity characteristic representation;

[0118] The key pathways obtained in the preceding steps mainly reflect structural-level mechanistic information, while toxicity assessment also requires combining sequence evidence in the reasoning path.

[0119] Extract key pathway structural features;

[0120] Let the critical pathway set be defined. The encoding representation of each key pathway in the middle is as follows: The corresponding pathway characteristics are First, a structural view representation is constructed based on the key pathways. For the first... The key pathway, whose structural view input is represented as follows:

[0121] ;

[0122] in, Indicates the first Structural view features of key pathways The structure-view mapping function is represented by [;], and vector concatenation is represented by [;]. The structure-view feature representations of all key pathways are fed into the structure encoder to obtain the toxicity target. Structural view representation:

[0123] ;

[0124] in, Represents the structure view encoder. This represents the aggregate marker vector of the structure view. The subscript 0 indicates that the position of the corresponding aggregate marker vector in the encoding result is taken and output.

[0125] In addition to the structural view, this invention reuses the reason path features extracted in S2.2 to form a sequence view. For toxic targets... , will the corresponding The reason paths are weighted according to normalized values. Arranged from high to low, forming a reason path sequence. The first [item] in the sequence... Each unit consists of a complete reason path, and its characteristics are... It includes relation sequence encoding, endpoint path encoding, database source encoding, association strength, path length, and AOP mechanism prior vector. The first... The features of each reason path are mapped to the sequence encoder to obtain:

[0126] ;

[0127] in, This represents the feature representation of the j-th reason sequence unit. This represents the sequence view mapping function. All sequence view feature representations are fed into the sequence encoder to obtain the toxic target. The sequence view is represented as:

[0128] ;

[0129] in, Indicates toxicity target The sequence view is represented as follows: Indicates a sequence view encoder. This represents the aggregate marker vector of the sequence view. The subscript 0 indicates that the position of the corresponding aggregate marker vector in the encoding result is taken and output.

[0130] To further preserve prior information on toxicity mechanisms in key pathways, AOP features and general correction factors in the key pathway set are aggregated to construct a mechanism bias vector. Let the... The AOP feature vectors corresponding to the key pathways are The corresponding general correction factor is The mechanism bias vector is then expressed as:

[0131] ;

[0132] ;

[0133] ;

[0134] in, Indicates toxicity target The key path AOP mean vector, This represents the mean of the general correction factor for the key pathway. Represents the mechanism bias vector. The representation mechanism bias mapping function.

[0135] S4.2: After obtaining the structural view representation, sequence view representation, and mechanism bias in the same quantity, a unified inference representation is formed using a gating fusion method:

[0136] ;

[0137] ;

[0138] in, Indicates toxicity target The gate vector, This represents the unified reasoning representation after the fusion of two views. and Indicates the gating mapping parameters, Indicates the bias weight of the mechanism. This represents the Sigmoid mapping function. [;] indicates element-wise multiplication, and [;] indicates vector concatenation.

[0139] S5.1: Construct shared representations and toxicity-specific representations;

[0140] After obtaining a unified inference representation Next, the issue of imbalanced sample sizes across different toxicity categories is considered. For toxicity categories with more samples, the model learns stable representations more easily; however, for toxicity categories with fewer samples, directly using completely independent classification methods can easily lead to insufficient representation or unstable predictions. Therefore, the unified inference representation is decomposed into shared representations and toxicity-specific representations, allowing different toxicity categories to share common mechanism information while retaining their own differences. For toxicity targets... Let the first The category index corresponding to toxicity is ,in Then, the shared representation and the toxicity-specific representation are defined as follows:

[0141] ;

[0142] in, Indicates toxicity target The sharing mechanism indicates that, Indicates toxicity target In the Specific expression of toxicity, This represents the unified reasoning representation after the fusion of two views. This represents a shared representation mapping function. Indicates the first A specific representation mapping function for toxicity. Based on this, the toxicity target is calculated. In the Baseline discriminant value for toxicity:

[0143] ;

[0144] in, and They represent the first Classification parameters corresponding to toxicity class, Indicates to Perform a transpose operation.

[0145] S5.2: Cross-toxicity migration reasoning and multi-toxicity output;

[0146] To enhance the predictive power of the weak toxicity category, a cross-toxicity transfer term is further introduced, enabling toxicity categories with fewer samples to absorb shared mechanism information from other toxicity categories. Let... Indicates the type j toxicity towards the type j The correlation coefficient for toxicity transmission is then expressed as the corrected discriminant value:

[0147] ;

[0148] in, Indicates toxicity target In the Corrected discriminant value for toxicity class. Indicates the first The basic discriminant value for toxicity-like substances, Indicates the effect of class j toxicity on class j. Migration coefficients of toxicity. Ultimately, toxicity targets. In the The predictive probability of toxicity is expressed as:

[0149] ;

[0150] in, This represents the Sigmoid mapping function.

[0151] By combining the prediction results of all toxicity categories, the toxicity target can be obtained. The final multitoxicity output vector:

[0152] ;

[0153] in, Indicates the first Predicted probability of toxicity This indicates the total number of toxicity categories.

[0154] This approach combines shared mechanism representation, toxicity-specific representation, and cross-toxicity transfer mechanism, enabling toxicity categories with fewer samples to learn without relying solely on their limited sample size. Instead, they can utilize shared mechanism information between different toxicities to achieve compensation.

[0155] After completing candidate pathway set encoding, key pathway identification, and dual-view transfer reasoning, the candidate pathway conditional encoding module, counterfactual key pathway screening module, structure-sequence dual-view mechanism reasoning module, shared-specific cross-toxicity transfer module, and multi-toxicity output module are sequentially connected to form a multi-stage, multi-branch, multi-task deep neural network customized for toxic pathway prediction tasks in this application. This network does not directly reuse external pre-trained models with independent version numbers. Instead, it uses fully connected feature mapping layers, standard Transformer encoding units, attention computation layers, gated fusion layers, and multi-task classification layers as basic computational components, and structurally reorganizes them in the order of "candidate pathway encoding—key pathway screening—dual-view mechanism fusion—cross-toxicity transfer compensation—multi-toxicity output correction." During the training phase, a unified optimization objective is used to jointly train the above modules end-to-end, enabling the candidate pathway representation, key pathway contribution, dual-view mechanism information, and cross-toxicity transfer relationship to converge synergistically, thereby improving the stability of key pathway identification and the consistency of multi-toxicity prediction.

[0156] The overall training objective is defined as the main task loss, auxiliary supervision loss, sparse constraint loss, and dual-view loss. Figure 1 A weighted combination of toxicity loss and migration constraint loss. Let the toxicity target be... The final prediction result is The initial multitoxicity prediction results are as follows: The number of critical pathways is The overall loss function is then expressed as:

[0157] ;

[0158] Where L represents the overall loss function, Indicates loss of the main task. Indicates the loss of auxiliary supervision. Represents the sparse constraint loss. Indicates binocular vision Figure 1 Sexual damage, This represents the migration constraint loss. , , , and This represents the weighting coefficient of each loss term.

[0159] The main task loss is used to constrain the consistency between the final multi-toxicity prediction results and the true labels. Let the total number of toxicity categories be C, and the true labels be... The loss of the main task can then be expressed as:

[0160] ;

[0161] in, This indicates the main task loss, and C represents the total number of toxicity categories. Indicates toxicity target In the The true label of toxicity, Indicates the final predicted probability. Indicates the first Category weights of toxicity This represents the adjustment coefficient for difficult samples.

[0162] The auxiliary supervision loss is used to constrain the initial prediction results before key pathway identification, ensuring that the candidate pathway importance calculation stage also has a clear supervision signal. It is defined as follows:

[0163] ;

[0164] in, Indicates the loss of auxiliary supervision. Indicates toxicity target In the Initial predicted probability on toxicity class.

[0165] The sparsity constraint loss is used to suppress too many candidate pathways from entering the key pathway set at the same time, thereby improving the compactness of the key pathway identification results. The sparsity constraint loss is expressed as:

[0166] ;

[0167] in, Represents the sparse constraint loss. Indicates the number of candidate pathways. Indicates the first The retention indicator variable for each candidate pathway, when retained ,otherwise , Indicates the first Importance weights of candidate pathways.

[0168] Binocular vision Figure 1 Consistency loss is used to constrain the structural view representation and the sequence view representation to maintain consistency under the same toxicity target, thereby reducing the characterization bias between the two views. It is defined as:

[0169] ;

[0170] in, Indicates binocular vision Figure 1 Sexual damage, Indicates toxicity target The structural view is shown. Indicates toxicity target The sequence view is represented as follows: This represents the L2 norm.

[0171] The migration constraint loss is used to limit the excessive perturbation of the basic discrimination result by the cross-toxicity migration term, so that the weak toxicity category can maintain the stability of the category discrimination while obtaining mechanism compensation. It is defined as:

[0172] ;

[0173] in, This represents the migration constraint loss. Indicates the first Corrected discriminant value for toxicity-like properties. Indicates the first The basic discrimination value for toxicity class, where C represents the total number of toxicity classes.

[0174] A joint training approach is used to optimize the overall model. The training sample set is D, where each training sample includes a toxicity target. The set of candidate pathways corresponding to this toxicity target Candidate pathway features, target-level contextual features, reason pathway sequence features, and true multitoxicity markers Let the set of all trainable parameters of the model be:

[0175] ;

[0176] in, This represents the complete set of trainable parameters for the model. Indicates the parameters of the candidate path coding module. This indicates the parameters of the critical path identification module. Indicates the parameters of the dual-view fusion module. Indicates parameters for cross-toxicity migration modules. This represents the parameters of the multi-toxicity output module. The model parameters are initialized before training begins. For each mapping matrix or weight matrix W, a random initialization method based on the input and output dimensions is used:

[0177] ;

[0178] in, This represents the initialized weight matrix. Indicates uniform distribution. Indicates the input dimension. This represents the output dimension. The bias parameter b is initialized to 0. Features derived from external knowledge or statistical rules, such as prior features of the AOP mechanism and general path correction factors, are not considered. It participates in the calculation as a fixed input or constraint information during the training process.

[0179] During training, candidate pathway encoding, key pathway contribution calculation, key pathway screening, dual-view feature fusion, cross-toxicity transfer inference, and multi-toxicity prediction are sequentially performed on each batch of training samples to obtain the final prediction results. Subsequently, based on the main task loss, auxiliary supervision loss, sparse constraint loss, and dual-view loss... Figure 1 Consistency loss and transfer constraint loss are used to calculate the overall loss function L. For a training batch B, its batch loss is expressed as:

[0180] ;

[0181] in, Let B represent the average loss of the current training batch, and let B represent the sample set of the current training batch. Indicates the number of samples in the current batch. Indicates toxicity target In the current parameters The corresponding overall loss.

[0182] During the parameter update phase, the backpropagation algorithm is used to calculate the gradient of the batch loss with respect to the model parameters, and the model parameters are updated based on the gradient results.

[0183] ;

[0184] in, Indicates the first Model parameters during round iterations, This represents the updated model parameters. Indicates the learning rate. This represents the gradient of the current batch loss with respect to the set of model parameters.

[0185] Repeat the above training process until the overall loss function converges or the preset number of training rounds is reached, and the trained toxic pathway prediction model is obtained.

[0186] After training is completed, the information of the toxic targets to be predicted and their candidate pathways is input during the inference phase, and the key pathway set, the prediction probabilities of five types of toxicity and the corresponding key pathway interpretation results are obtained in sequence.

[0187] During the inference phase, considering the differences in sample size and discrimination difficulty among the five toxicity categories, the output probability scale and discrimination threshold for each toxicity category are adjusted to improve the predictive applicability of weak toxicity categories such as neurotoxicity and reproductive toxicity, as well as small sample categories, as detailed below:

[0188] S5.3: Perform small-sample toxicity category transfer compensation;

[0189] Even after joint training, the output probabilities of different toxicity categories may still exhibit scale differences. For toxicity categories with a large number of samples, the model output is usually more stable; for toxicity categories with fewer samples or more complex boundaries, the model output probability may be too high or too low. Therefore, category-level probability correction and category-level threshold determination are further introduced during the inference phase. Let the... The modified discriminant value for toxicity is The discriminant value after temperature correction is expressed as:

[0190] ;

[0191] in, Temperature-corrected discriminant value of the toxicity target t for Class c toxicity. Indicates the first Corrected discriminant value for toxicity-like properties. Indicates the first Temperature parameters related to toxicity. Accordingly, the first... The corrected predictive probability of toxicity is expressed as:

[0192] ;

[0193] in, Indicates the first Corrected predicted probability of toxicity This represents the discriminant value after temperature correction. This represents the Sigmoid mapping function. After obtaining the corrected predicted probabilities, independent discrimination thresholds are set for different toxicity categories. And obtain the final judgment result:

[0194] ;

[0195] in, Indicates toxicity target In the The final label for class toxicity, Indicates the first Corrected predicted probability of toxicity Indicates the first Threshold for toxicity discrimination.

[0196] By applying differential correction to the output results of different toxicity categories, the same model can maintain relatively stable output performance when facing toxicity categories with different sample sizes and different discrimination boundaries.

[0197] This invention constructs a candidate pathway set for toxicity targets and identifies key toxicity pathways within the set, thereby effectively reducing interference from redundant, generalized, and weakly correlated pathways in the toxicity prediction process and improving the reliability and interpretability of toxicity target-pathway prediction results.

[0198] By jointly modeling key pathway structural information with inference path sequence information and introducing a cross-toxicity migration compensation mechanism, this invention can improve the prediction ability of weak toxicity categories such as hepatotoxicity, neurotoxicity, and reproductive toxicity, as well as small sample categories, while ensuring overall prediction performance.

[0199] Furthermore, in addition to providing multi-toxicity prediction results, this invention can simultaneously provide key pathway identification results, offering a more compact, clear, and biologically significant basis for subsequent toxicity mechanism analysis, candidate pathway verification, and traditional Chinese medicine safety evaluation.

[0200] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting toxic pathways based on small-sample transfer reasoning, characterized in that, Includes the following steps: S1: For the toxicity target to be predicted, gather all candidate pathways that are associated with the toxicity target to form a candidate pathway set, and obtain the original feature vectors corresponding to each candidate pathway to construct the original feature matrix of the candidate pathway. S2: Characterize the mechanism features of each candidate pathway in the candidate pathway set, fuse the global features corresponding to the toxicity target with the reason summary features to obtain the target context representation, encode the target condition of each candidate pathway through the target context representation, output the candidate pathway encoding matrix and input it into S3. S3: Using the candidate pathway coding matrix as input, calculate the importance weight of each candidate pathway. At the same time, calculate the necessity score and sufficiency score of each candidate pathway through counterfactual analysis. Combine the prior score of the AOP mechanism and the general pathway correction factor to generate a comprehensive contribution score. Sort and screen the key toxic pathways according to the comprehensive contribution score and input them into S4. S4: Based on the set of key toxic pathways, extract the structural features and reason path sequence features of the key pathways in the set of key toxic pathways respectively, construct the structural view and the sequence view, and at the same time aggregate the mechanism prior information of the key pathways to generate the mechanism bias vector. After gating fusion, the unified reasoning representation of the dual view fusion is obtained and input into S5. S5: Decompose the unified reasoning representation into a shared mechanism representation and specific representations of various toxicities, calculate the basic discriminant values ​​of various toxicities, and then perform small-sample compensation through cross-toxicity transfer terms. Finally, output the joint prediction results of multiple toxicities including cardiotoxicity, hepatotoxicity, nephrotoxicity, neurotoxicity and reproductive toxicity.

2. The toxic pathway prediction method based on small sample transfer reasoning according to claim 1, characterized in that, In S1, the candidate pathway set consists of all biological pathways associated with the current toxicity target. The original features of each candidate pathway are arranged in rows to form an original feature matrix of dimension n×d, where n is the number of candidate pathways and d is the feature dimension of a single pathway.

3. The toxic pathway prediction method based on small sample transfer reasoning according to claim 1, characterized in that, In S2, the mechanistic features of a single candidate pathway are composed of basic evidence features, database source features, a priori features of AOP mechanisms of five types of toxicity, general pathway correction factors, and a priori summary features. The general path correction factor is calculated based on the path support size statistic, source breadth statistic, and general keyword suppression term.

4. The toxic pathway prediction method based on small sample transfer reasoning according to claim 3, characterized in that, The target conditional encoding of S2 is as follows: the mechanism features of the candidate pathways are mapped to a unified latent space to obtain the initial pathway representation; the global features and the reason summary features are mapped and fused to obtain the target context representation; the initial pathway representation is modulated element-wise by the gating vector and the offset vector to obtain the target conditional encoding representation of each candidate pathway, and the representations are arranged in rows to form the candidate pathway encoding matrix. The modulation calculation satisfies: ； In the formula, For the first The target condition encoding representation of each candidate pathway. Let be the gate vector for the i-th candidate path. Let be the offset vector of the i-th candidate path. Let be the initial path representation of the i-th candidate path, and ⊙ denotes element-wise multiplication.

5. The toxic pathway prediction method based on small sample transfer reasoning according to claim 1, characterized in that, In S3, the importance score of each candidate pathway is calculated by combining the candidate pathway encoding representation and the target context representation. After normalization, the importance weight of each candidate pathway is obtained. The encoding representations of all candidate pathways are weighted and aggregated according to the importance weight to obtain the candidate pathway aggregated representation of the toxicity target, and the initial multitoxicity prediction results are output.

6. The toxic pathway prediction method based on small sample transfer reasoning according to claim 5, characterized in that, In S3, the necessity score is calculated by the difference in prediction results before and after removing a single candidate pathway, and the sufficiency score is calculated by the mean of prediction results when only a single candidate pathway is retained; the mechanism prior score is calculated by weighting the AOP prior vector of the candidate pathway with the initial multitoxicity prediction results; the general pathway correction factor is modified by combining the mechanism prior score to obtain the modified general factor.

7. The toxic pathway prediction method based on small sample transfer reasoning according to claim 6, characterized in that, In S3, the necessity score, sufficiency score, mechanism prior score and importance weight are weighted and summed according to the corresponding coefficients, and then multiplied by the modified general factor to obtain the comprehensive contribution score of each candidate pathway. The top k candidate pathways in terms of comprehensive contribution score are selected to form a set of key toxicity pathways, where k is the preset number of key pathways to be retained. The overall contribution score satisfies: ； In the formula The overall contribution score for the i-th candidate pathway is... For necessary fractions, For sufficiency fractions, For the mechanism's prior score, As importance weight, This is the corrected general factor. , , , These are the weighting coefficients for the corresponding components.

8. The toxic pathway prediction method based on small sample transfer reasoning according to claim 1, characterized in that, In S4, the encoded representation of the key pathway is concatenated with the original features and mapped to obtain the structural view input, which is then output as a structural view representation by the structural view encoder. The reason sequence unit features are mapped to obtain the sequence view input, which is then output as a sequence view representation by the sequence view encoder. The AOP features of the key pathway are aggregated with the general correction factor to obtain the mechanism bias vector. The structural view representation, the sequence view representation, and the mechanism bias vector are fused through a gating mechanism to output a unified inference representation.

9. The toxic pathway prediction method based on small sample transfer reasoning according to claim 1, characterized in that, In S5, the unified reasoning representation is mapped to obtain the shared mechanism representation and the toxicity-specific representation corresponding to each type of toxicity. After adding the shared mechanism representation and the specific representation of the same toxicity category, the basic discriminant value of that type of toxicity is calculated.

10. The toxic pathway prediction method based on small-sample transfer reasoning according to claim 9, characterized in that, In S5, based on the migration relationship coefficients between various toxicities, the basic discriminant values ​​of other toxicity categories are migrated and superimposed onto the current toxicity category to obtain the corrected discriminant value; Temperature correction is applied to the discrimination values ​​for various toxicities. After obtaining the predicted probabilities through Sigmoid mapping, the final toxicity discrimination label is output through a category-specific discrimination threshold. The corrected discrimination values ​​satisfy the following: ； In the formula This is the corrected discriminant value for the toxicity target t in the class c toxicity. This serves as the basic discriminant value for Class c toxicity. denoted as the migration coefficient from class j toxicity to class c toxicity.

Images