A time sequence characteristic analysis and prediction method for dynamic response of molecular toxicological pathways

Abstract

The application discloses a kind of molecular toxicological pathway dynamic response timing feature analysis prediction method, belong to molecular toxicology monitoring technical field.Acquire the multi-modal feature and multi-omics timing response data of to-be-measured compound;Through double-branch dynamic decoupling feature extraction network, extract chemical space feature, capture biological timing dependence, realize feature independent coding;With the aid of cross-modal cross attention fusion module, the deep interaction alignment of chemical feature and biological response is completed, and a joint representation vector is generated;Through the distribution offset correction layer built-in in time drift perception prediction module, the timing distribution difference is calculated and the model parameters are dynamically adjusted, the timing correction prediction is realized;The prediction result is mapped to AOP knowledge graph, and the toxicological pathway activation timing trajectory is output.The application realizes the deep fusion of multi-modal timing information and the self-adaptive correction of timing drift, improves the accuracy and biological interpretability of toxicological pathway prediction, and is suitable for early warning of compound toxicity and toxicological mechanism analysis.

Description

Technical Field

[0001] This invention relates to the field of molecular toxicology monitoring technology, specifically to a method for analyzing and predicting the time-series characteristics of dynamic responses in molecular toxicology pathways. Background Technology

[0002] The dynamic response of molecular toxicological pathways is a core element in the damage caused by exogenous compounds to the organism, with signal activation, transduction, and effect output all exhibiting significant time-dependent characteristics. Traditional toxicological evaluations rely heavily on animal experiments and endpoint detection, only obtaining pathway activation states at fixed time points. This makes it difficult to characterize the continuous evolution of pathways from early perturbations to late-stage damage, resulting in partial analyses of toxicity mechanisms and delayed predictions. With the development of toxicogenomics and high-throughput sequencing technologies, time-series data from transcriptomics, proteomics, and other omics are rapidly accumulating, providing a data foundation for dynamic pathway analysis. However, existing analytical methods are still limited to static differential screening and simple trend statistics, lacking a systematic exploration of time-dependent relationships, cascade regulatory patterns, and causal transmission logic. This makes it difficult to meet the practical needs for early warning of compound toxicity and precise mechanism analysis.

[0003] Existing academic research has explored time-series analysis of toxicological pathways and toxicity prediction, but it generally suffers from shortcomings such as insufficient utilization of time-series features and inadequate dynamic modeling capabilities. Zhang et al.'s review article "Artificial Intelligence-Driven Drug Toxicity Prediction: Advances, Challenges, and Future Directions" published in *Toxics* systematically summarizes the progress of AI-driven toxicity prediction, pointing out that current models mostly use molecular structure and static omics data as input, ignoring dynamic information in the time dimension and failing to reconstruct the true response process of the pathway. Huang et al.'s research on substructure-guided deep graph learning published in the *IEEE Journal of Biomedical and Health Informatics*, while improving the accuracy of molecular feature extraction, only focuses on toxic endpoint classification and does not involve time-series features and dynamic evolution modeling of pathways. In addition, many drug response prediction studies based on time-series gene expression, such as the REcursivePrediction framework, only use recursive structures to fit short-term trends without combining pathway topology and multi-omics fusion. They cannot achieve the analysis of multi-pathway time-series cascade relationships and extrapolation of long-term toxicity, and still have problems such as fragmented time-series features, ambiguous causal relationships, and poor predictive generalization.

[0004] Chinese patented technologies have developed several solutions in the field of molecular toxicity prediction, but none have achieved in-depth mining and accurate prediction of the temporal characteristics of dynamic responses in toxicological pathways. Patent CN116246718A proposes a method integrating pharmacology and toxicology profiles to predict endocrine disruptors, fusing molecular fingerprints and multi-omics data to construct a prediction model for compound toxicity risk assessment. However, this method only uses static feature fusion, failing to incorporate temporal dimension information, and cannot capture the temporal changes in pathways, completely lacking modeling of the dynamic response process. Patent CN115831260A discloses a small-sample molecular toxicity prediction method, based on deep learning to optimize feature extraction and model training, solving the prediction problem in small-sample scenarios. However, its input is only molecular structure data, without combining multi-omics temporal information, and does not involve dynamic pathway response analysis, failing to distinguish between transient perturbations and persistent damage, and the prediction results lack biological mechanism support. Publication No. CN119274688A1 describes a method and apparatus for predicting the developmental toxicity of chemicals based on multimodal fusion. This method uses multimodal fusion to predict developmental toxicity and integrates data such as molecular descriptors and biological activities. Although it expands the feature dimensions, it is still mainly based on static fusion and does not correct for temporal distribution shifts. The model is not stable enough in the generalization scenario of temporal data and it is difficult to achieve long-term accurate prediction of dynamic response of toxicological pathways.

[0005] Existing academic research and patented technologies share common problems: emphasizing static aspects while neglecting dynamic aspects, focusing on correlations while neglecting causality, and prioritizing fusion over temporal correction. This makes it difficult to achieve temporal feature quantification, multimodal deep interaction, adaptive correction of temporal drift, and interpretable prediction of the dynamic response of molecular toxicological pathways. Summary of the Invention

[0006] To address the aforementioned technical problems, this application discloses a method for analyzing and predicting the time-series characteristics of dynamic responses in molecular toxicological pathways, comprising the following steps:

[0007] S1. Acquire multimodal molecular feature data of the test compound, and multi-omics time-series response data of biological samples exposed to the test compound at multiple consecutive time points; the multimodal molecular feature data includes molecular graph structural features and molecular fingerprint sequence features, and the multi-omics time-series response data includes transcriptomics and proteomics data;

[0008] S2. Construct a dual-branch dynamic decoupled feature extraction network to independently encode the multimodal molecular feature data and multi-omics temporal response data; use a graph attention network (GAT) to extract molecular graph structure features, use a bidirectional gated recurrent unit (Bi-GRU) to extract molecular fingerprint sequence features, and use a temporal convolutional network (TCN) to capture the temporal dependencies of the multi-omics data, thereby outputting decoupled chemical spatial feature vectors and biological temporal feature vectors.

[0009] S3. Input the chemical spatial feature vector and biological temporal feature vector obtained in S2 into the cross-modal cross-attention fusion module. By calculating the association weight between chemical substructures and biological pathway nodes, the deep interaction and alignment of molecular features and biological responses are realized, and a cross-modal joint representation vector is generated.

[0010] S4. Input the cross-modal joint characterization vector into the preset time drift sensing prediction module; the time drift sensing prediction module has a built-in time distribution offset correction layer, which is used to calculate the distribution difference between the current test compound and the historical training data in the time dimension, and dynamically adjust the parameter weights of the prediction model based on the difference, so as to output the time-corrected dynamic response prediction result of the toxicological pathway.

[0011] S5. Map the prediction results of step S4 to the pre-constructed adverse outcome pathway AOP knowledge graph. By calculating the causal consistency scores between the prediction results and the molecular initiation event (MIE) and key event (KE) in the AOP knowledge graph, the prediction results are verified by biological mechanisms and explained, and the activation timeline of the toxicological pathway is output.

[0012] Preferably, in step S2, the extraction of molecular graph structural features using a graph attention network (GAT) specifically includes:

[0013] Representing the molecular diagram structure as a graph Where V is the set of atomic nodes and E is the set of chemical bond edges; for each atomic node Its initial eigenvector It is composed of its atomic type, hybridization state, aromaticity, and formal charge;

[0014] Feature aggregation is performed through at least one graph attention layer. Layer nodes The features are updated as follows: ,in, For nodes The set of neighboring nodes, For the first Layer-learnable weight matrix, It is a non-linear activation function; attention coefficient Calculated using the following formula: ,in, For the first Layer-learnable attention vectors This represents a vector concatenation operation; the feature vectors of all nodes are then subjected to global pooling to obtain the graph structure feature vectors of the molecule. .

[0015] Preferably, step S2 utilizes a bidirectional gated recurrent unit (Bi-GRU) to extract molecular fingerprint sequence features, specifically including:

[0016] The molecular fingerprint sequence is represented as ,in Let be the d-dimensional fingerprint feature vector at time t;

[0017] The sequence is encoded using a forward GRU layer and a backward GRU layer, respectively. The hidden state of the forward GRU at time t is... The hidden state of the backward GRU at time t The calculation is as follows: , The forward and backward hidden states are concatenated in a dimension to obtain the final hidden state at time t. ;

[0018] Average pooling is performed on the hidden states at all time points in the sequence to obtain the feature vector of the molecular fingerprint sequence. .

[0019] Preferably, in step S2, the use of a Temporal Convolutional Network (TCN) to capture the temporal dependencies of multi-omics data specifically includes:

[0020] For the observation sequence of the k-th omics data at T consecutive time points ,in For time t 3D feature vector;

[0021] Processing is performed using a one-dimensional causal convolutional layer, and the output of the m-th convolutional layer... for: Where K is the kernel size and d is the dilation coefficient. and These are the weights and bias terms of the convolutional kernel in the m-th layer, respectively. For the upper layer Input at any moment;

[0022] The TCN contains at least one residual block, each containing two of the aforementioned causal convolutional layers, and incorporates weight normalization and Dropout operations. The output of the last time step is mapped through a fully connected layer to obtain the temporal feature vector of the omics data. By concatenating the time-series feature vectors of all omics data, the biological time-series feature vector is obtained. .

[0023] Preferably, the cross-modal attention fusion module in S3 specifically includes:

[0024] The chemical space feature vector output by S2 As the query vector Q, Depend on With molecular fingerprint sequence feature vector This is a concatenation of biological time-series feature vectors. As the key vector K and the value vector V;

[0025] The cross-modal attention weight matrix A is calculated using the following formula: ,in, The dimension of the key vector;

[0026] The fused feature representation is obtained by weighting the value vector V using the attention weight matrix A. : ,Will Compared with the original chemical space feature vector By performing residual connections and layer normalization, a cross-modal joint representation vector is obtained. .

[0027] Preferably, the cross-modal cross-attention fusion module employs a multi-head attention mechanism, specifically including:

[0028] The above cross-modal attention calculation process is executed in parallel for h times, and the output of the i-th attention head is: ;

[0029] The outputs of h heads are concatenated and then subjected to a learnable linear transformation. Projecting yields the final fused features: ,in, , For each head, the value vector dimension, This represents the dimension of the model's hidden layers.

[0030] Preferably, in the time drift sensing and prediction module of S4, the time distribution offset correction layer calculates the distribution difference in the following manner:

[0031] For the cross-modal joint characterization vector of the current test compound Calculate the set of cross-modal joint representation vectors of it and all samples in the historical training dataset. Maximum mean difference (MMD) between them: ,in, To map features to the reproducing kernel Hilbert space The mapping function.

[0032] Preferably, the time drift sensing and prediction module is based on the calculated MMD value. Dynamically adjust the parameter weights of the prediction model:

[0033] Let the basic parameters of the prediction model be Adjusted parameters for: ,in, For learning rate, The regularization loss term based on MMD is given by the formula: .

[0034] Preferably, the adverse outcome pathway AOP knowledge graph pre-constructed in S5 includes nodes such as molecular initiation event (MIE), key event (KE), and adverse outcome (AO), with edges representing causal relationships between events;

[0035] The causal consistency scores between the calculated prediction results and the molecular initiation events (MIE) and key events (KE) in the AOP knowledge graph specifically include:

[0036] The toxicological pathway dynamic response prediction results output by S4 are mapped to the node space of the AOP knowledge graph to obtain the set of event nodes for prediction activation. ;

[0037] for Each event node in Calculate its relationship with known causal paths in the knowledge graph. Match score : ,in, For indicator functions, For event nodes Semantic similarity with the corresponding event in path p.

[0038] Preferably, the biological mechanism verification and interpretability reasoning further include:

[0039] Based on the calculated causal consistency score The prediction results are sorted, and those with scores higher than a preset threshold are selected. The event nodes constitute the core activation pathway. ;

[0040] Backtracking in the AOP knowledge graph using graph traversal algorithms The upstream MIE nodes of each event node generate a complete causal chain explanation from MIE to AO.

[0041] Compared with the prior art, the technical solution of this application has the following technical effects:

[0042] This invention, through the joint acquisition of multimodal molecular features and multi-omics time-series data and the extraction of dual-branch dynamic decoupling features, can completely extract the intrinsic structural information of compounds and the time-series response information of biological samples, forming mutually independent and comprehensive chemical spatial features and biological time-series features, providing a stable and reliable data foundation for dynamic response analysis of toxicological pathways.

[0043] This invention employs a cross-modal cross-attention fusion module to achieve deep interaction and alignment between chemical features and biological temporal features. It can automatically learn the intrinsic relationship between chemical substructures and biological pathway nodes, generate a highly compact cross-modal joint representation, improve the adaptability and synergistic expression ability between different modal information, and enable the model to more accurately capture the intrinsic relationship between molecular interactions and pathway responses.

[0044] This invention, through a time drift sensing prediction module and a time distribution offset correction layer, can adaptively identify and correct distribution differences caused by time series data, dynamically optimize model parameter weights, improve the adaptability of the prediction process to different durations and exposure conditions, ensure the stability and consistency of dynamic response prediction results of toxicological pathways, and output reliable time series correction prediction results.

[0045] This invention combines prediction results with AOP knowledge graphs, and completes mechanism verification and interpretability reasoning through causal consistency scores. It can clearly output the activation time sequence trajectory and complete causal chain of toxicological pathways, improve the biological credibility and interpretability of prediction results, and achieve accurate toxicological mechanism analysis and toxicity risk assessment.

[0046] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0047] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0049] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0050] Figure 1 A schematic diagram of the overall process steps for predicting the dynamic response of molecular toxicological pathways;

[0051] Figure 2A schematic diagram of a dual-branch dynamic decoupling feature extraction and cross-modal fusion architecture;

[0052] Figure 3 Flowchart of distribution correction and parameter adjustment for time drift sensing and prediction module;

[0053] Figure 4 A visualization comparing t-SNE dimensionality reduction clustering of different models in the toxicological feature space;

[0054] Figure 5 Cross-modal attention mechanism ablation experimental model training loss convergence comparison curves;

[0055] Figure 6 Comparison of the adaptation effects of UMAP feature distribution domain with and without time drift correction;

[0056] Figure 7 A visualization of the causal path relationships between AOP knowledge graph nodes in the embedding space;

[0057] Figure 8 Bar chart ranking the importance of key features that influence the prediction of toxicological pathways (SHAP). Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0059] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0060] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0061] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0062] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0063] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0064] Example 1

[0065] This embodiment mainly describes a method for analyzing and predicting the time-series characteristics of dynamic responses in molecular toxicological pathways, such as... Figure 1 As shown, it specifically includes:

[0066] S1. Acquire multimodal molecular feature data of the compound to be tested, and multi-omics time-series response data of biological samples exposed to the compound at multiple consecutive time points; the multimodal molecular feature data includes molecular graph structural features and molecular fingerprint sequence features, and the multi-omics time-series response data includes transcriptomics and proteomics data;

[0067] S2. Construct a dual-branch dynamic decoupled feature extraction network to independently encode the multimodal molecular feature data and multi-omics temporal response data. Specifically, a graph attention network (GAT) is used to extract molecular graph structure features, a bidirectional gated recurrent unit (Bi-GRU) is used to extract molecular fingerprint sequence features, and a temporal convolutional network (TCN) is used to capture the temporal dependencies of the multi-omics data, thereby outputting decoupled chemical spatial feature vectors and biological temporal feature vectors.

[0068] S3. Input the chemical spatial feature vector and biological temporal feature vector obtained in S2 into the cross-modal cross-attention fusion module. By calculating the association weight between chemical substructures and biological pathway nodes, the deep interaction and alignment of molecular features and biological responses are realized, and a cross-modal joint representation vector is generated.

[0069] S4. Input the cross-modal joint characterization vector into the preset time drift sensing prediction module; the time drift sensing prediction module has a built-in time distribution offset correction layer, which is used to calculate the distribution difference between the current test compound and the historical training data in the time dimension, and dynamically adjust the parameter weights of the prediction model based on the difference, so as to output the time-corrected dynamic response prediction result of the toxicological pathway.

[0070] S5. Map the prediction results of S4 to the pre-constructed adverse outcome pathway AOP knowledge graph. By calculating the causal consistency scores between the prediction results and the molecular initiation event (MIE) and key event (KE) in the AOP knowledge graph, the prediction results are verified by biological mechanisms and explained. Finally, a prediction report containing the activation timeline of toxicological pathways and the corresponding mechanism explanations is output.

[0071] Furthermore, the raw data used for subsequent model training and prediction is obtained. Specifically, multimodal molecular feature data of the compound under test is obtained, which describes the compound from two levels:

[0072] Molecular graph structural features: Representing molecules as undirected graphs , where the node set Represents atoms, edge sets Represents chemical bonds. Each atomic node An initial feature vector is assigned. This vector is constructed by mapping discrete features such as atomic number, hybridization type (e.g., sp, sp2, sp3), whether it is within an aromatic ring, formal charge, and the number of hydrogen atoms connected to it into a continuous vector through an embedding layer, and then concatenating them. Each chemical bond... Also assigned feature vectors It includes information such as bond type (single bond, double bond, triple bond, aromatic bond) and whether it is conjugated.

[0073] Molecular fingerprint sequence characteristics: Using extended connectivity fingerprinting (ECFP) or similar algorithms, the molecular structure is converted into a fixed-length binary sequence. ,in The fingerprint length is specified (e.g., 1024 or 2048 bits). This sequence can be viewed as sequence data used to capture information about substructure fragments of a molecule.

[0074] Simultaneously, biological samples were obtained after exposure to the compound. consecutive time points ( This method primarily integrates transcriptomics and proteomics data, collecting multi-omics data. For transcriptomics, the data is represented as a matrix. ,in The number of genes detected. Indicates a point in time Gene The expression level (e.g., FPKM or TPM value). For proteomics, the data are represented as a matrix. ,in To determine the amount of protein to be detected. Indicates a point in time protein The abundance of these compounds. These time-series data collectively constitute a dynamic response map of biological systems to compound perturbations;

[0075] Furthermore, such as Figure 2 As shown, this application constructs a dual-branch neural network to process chemical and biological data separately, achieving independent feature extraction and decoupling.

[0076] First branch: Chemical feature extraction

[0077] This branch contains two parallel subnetworks that process molecular graphs and molecular fingerprints, respectively.

[0078] Graph Attention Network (GAT) processes molecular graphs:

[0079] GAT aggregates neighbor node information through an attention mechanism. In its implementation, we use a multi-layered GAT stack. For the... Layers, nodes The feature update formula is: ,in, It is a node In the output features of the previous layer, It is the learnable weight matrix of this layer. It is a non-linear activation function (such as ELU), and the attention coefficient is... The calculation formula is: ;

[0080] To stabilize training and enhance the model's expressive power, multi-head attention is employed. (The following is a partial sentence fragment: "Setting...") Each attention head independently computes a set of output features. In the intermediate layer, we concatenate the outputs of these headers: ;

[0081] In the final layer, we use an averaging operation to merge the multi-head outputs to obtain the final node embedding: ;

[0082] After all GAT layers have been processed, we use global add pooling to aggregate the features of all nodes into a graph structure feature vector for the entire molecule. : ;

[0083] Bi-Gated Recurrent Unit (Bi-GRU) for processing molecular fingerprints:

[0084] Molecular fingerprint sequence As input, Bi-GRU consists of a forward GRU and a backward GRU, capable of simultaneously capturing past and future contextual information of the sequence. For time steps... (Here, (referring to the position in the fingerprint sequence), forward hidden state and backward hidden state Calculated using the following formulas respectively:

[0085]

[0086]

[0087] in, The specific calculations involve updating the gate. Reset door and candidate hidden state :

[0088]

[0089]

[0090] The forward hidden state of the last time step and the backward hidden state of the first time step are concatenated to form the feature representation of the entire fingerprint sequence. : ;

[0091] Chemical spatial eigenvectors:

[0092] The output of GAT and Bi-GRU output The data are concatenated and then subjected to dimensionality transformation and nonlinear activation through a fully connected layer to obtain the final chemical space feature vector. :

[0093]

[0094] Second branch: Extraction of biological temporal features

[0095] This branch utilizes Temporal Convolutional Networks (TCNs) to process multi-omics time-series data. TCNs capture long-term dependencies through causal convolution and dilated convolution.

[0096] Data input and preprocessing:

[0097] For the Omics data (such as transcriptomics or proteomics), in The observation sequence at each time point is ,in Before inputting into the TCN, the feature vector at each time point is standardized using Z-score.

[0098] Temporal Convolutional Network (TCN) structure:

[0099] TCN is composed of multiple residual blocks stacked together. Each residual block contains two identical convolutional layers, followed by weight normalization (WeightNorm), an activation function (such as ReLU), and Dropout. The convolutional layers employ one-dimensional causal convolution, and their output... It depends only on the input from the current and past moments. The output of a convolutional layer is calculated as follows: ,in, It is the kernel size. It is the expansion coefficient, which increases exponentially with network depth (e.g., the first...). The expansion coefficient of each residual block is This allows the network to achieve a large receptive field with fewer layers.

[0100] Residual connection:

[0101] The input for each residual block is The output is The final output is If you input Dimensions and If the dimensions are inconsistent, a 1x1 convolutional shortcut connection is used to match the dimensions.

[0102] Biological time-series feature vector:

[0103] Transcriptome data and proteomic data The inputs are fed into two independent TCN networks. The output of the last time step of each TCN network is taken and dimensionality reduced by a fully connected layer to obtain their respective temporal feature vectors. and Finally, these two vectors are concatenated to form a biological time-series feature vector. :

[0104]

[0105] Furthermore, through a cross-modal cross-attention mechanism, deep fusion of chemical and biological features is achieved, and the module employs multi-head cross-attention.

[0106] Query, key, and value generation:

[0107] Chemical space feature vector Generate the query vector through a linear transformation. Biological time-series feature vectors Generate key vectors using two different linear transformations. Sum value vector .

[0108]

[0109]

[0110]

[0111] in, It is a learnable weight matrix.

[0112] Multi-head cross-attention computation: parallel execution The first point of attention. For the first Size, its attention weight matrix and output The calculation formula is:

[0113]

[0114]

[0115] in, yes The segmentation obtained after linear projection The subvector corresponding to each head. It is the dimension of the key vector.

[0116] Feature fusion and residual connectivity:

[0117] Will The outputs of each head are concatenated and then subjected to a linear transformation. Integrate to obtain the fused features :

[0118]

[0119] Integrating fusion features with original chemical features Residual connections are performed, followed by layer normalization, to obtain the final cross-modal joint representation vector. :

[0120]

[0121] Furthermore, such as Figure 3 As shown, time drift correction is applied to the model to address the inconsistency between the distribution of new compounds and historical data.

[0122] Time distribution shift calculation: The maximum mean difference (MMD) is used to measure the joint characterization of the current analyte. Compared with historical training datasets The distribution differences between them. The formula for calculating the square of MMD is: The formula can be obtained through kernel functions. For efficient computation, its unbiased estimate is:

[0123] Among them, the Gaussian kernel function is used. .

[0124] Dynamic parameter adjustment: Calculated MMD value It is used as a regularization term to adjust the parameters of the prediction model. Let the basic parameters of the prediction model (a multilayer perceptron) be... Its original loss function is (e.g., cross-entropy loss). During the training phase, we minimize a total loss function that includes MMD regularization. :

[0125]

[0126] in, It is a hyperparameter used to control the strength of regularization. During the inference phase, for a new compound, we calculate its... The weights of the last layer of the model are then fine-tuned based on this value. The adjusted parameters... The calculation is as follows:

[0127]

[0128] in, It uses a relatively small learning rate. This process can be viewed as updating the model with a small gradient step size in the parameter space, making it more suitable for the current data distribution of the compound. Finally, the adjusted parameters are used. right Predictions were made to obtain time-corrected dynamic response predictions for toxicological pathways. .

[0129] Furthermore, prior knowledge is used to verify and interpret the biological rationality of the prediction results;

[0130] AOP knowledge graph construction:

[0131] A pre-built AOP knowledge graph is a directed graph. Node set It includes three types of entities: Molecular Initiation Events (MIEs), Key Events (KEs), and Adverse Outcomes (AOs). Edge Set It indicates a causal relationship between events, such as MIE1 causing KE2.

[0132] Causality consistency score calculation: The prediction results (A vector representing the activation probability of each pathway) is mapped onto nodes in the AOP knowledge graph, resulting in a set of event nodes for predicted activation. .for Each event node in We calculate its relationship with all known causal paths in the knowledge graph. Match score :

[0133]

[0134] in, It is an indicator function, when an event occurs. Exists in the path The value is 1 if it is in the middle, and 0 otherwise. It is an event With path The semantic similarity is calculated using a pre-trained biomedical text embedding model, such as BioBERT. Specifically, we will... Name and description and path The names and descriptions of all events are input into the model to obtain their text embedding vectors. and Then calculate their cosine similarity:

[0135]

[0136] Biological Mechanism Verification and Interpretable Reasoning: Based on Computational Causal Consistency Scores We selected those with scores higher than a preset threshold. The event nodes constitute the core activation pathway. Then, using graph traversal algorithms (such as depth-first search), the AOP knowledge graph is accessed from... Starting from each event node in the algorithm, the algorithm traces back to all its upstream nodes until all MIE nodes are found. This process generates one or more complete causal chains from MIE to AO, such as MIE_A->KE_B->KE_C->AO_D. This causal chain is the biological mechanism explanation for the predicted results. It clearly shows what molecular events might initiate a compound, triggering a series of key events that ultimately lead to an adverse outcome. The prediction report will include the predicted pathway activation probability, a list of core activation events, and the derived complete causal chain explanation.

[0137] This implementation achieves accurate prediction of dynamic responses of toxicological pathways by integrating multimodal molecular features and multi-omics time-series data and utilizing a dual-branch network and cross-modal attention mechanism. The built-in time drift sensing module effectively overcomes the data distribution offset problem, significantly improving the model's generalization ability and robustness. Combined with the causal consistency verification of the AOP knowledge graph, it not only ensures the biological rationality of the prediction results, but also generates a complete mechanism explanation.

[0138] Based on Example 1, this example details the fundamental validation of the multimodal temporal feature decoupling and cross-modal fusion model. The Tox21 challenge dataset was selected as the core validation object. This dataset contains in vitro bioactivity data for 8,012 compounds, covering nuclear receptors and stress response pathways. The experiments first standardized the compounds by removing salts and adding hydrogen atoms using the RDKit toolkit, subsequently generating molecular graph structure data and ECFP4 molecular fingerprint sequences. For the bioresponse data, transcriptome sequencing data (RNA-seq) and quantitative proteomics data from the human hepatocyte cell line (HepG2) at four consecutive time points (0h, 6h, 12h, and 24h) after exposure to different concentrations of compounds were selected. During data preprocessing, the gene expression matrix was TPM normalized and log2 transformed, and the protein abundance data was Z-score normalized to eliminate batch effects.

[0139] Regarding model construction, the specific parameter settings for the dual-branch network are as follows: The molecular graph branch uses a 3-layer graph attention network (GAT), with each layer containing 8 attention heads, a hidden layer dimension of 64, an ELU activation function, and a dropout rate of 0.5 to prevent overfitting. The molecular fingerprint branch uses a bidirectional gated recurrent unit (Bi-GRU), with 128 hidden layer units and 2 layers. The biological temporal branch uses a temporal convolutional network (TCN), with a kernel size of 3, 4 layers, and an inflation factor of [missing value]. Incrementing the receptive field. The query, key, and value matrices for the cross-modal attention module are all set to 128 dimensions. Model training uses the Adam optimizer with an initial learning rate of 0.001 and a weight decay coefficient of [missing value]. The batch size is set to 32, the number of training epochs is set to 100, and an early stopping mechanism (Patience=10) is used.

[0140] To quantitatively evaluate model performance, a 5-fold cross-validation method was employed, and the proposed method (named MDTL-Net) was compared with six existing mainstream deep learning algorithms. These six algorithms were: Graph Convolutional Networks (GCN), Deep Neural Networks (DNN), Long Short-Term Memory Networks (LSTM), Deep Belief Networks (DBN), Convolutional Neural Networks (CNN), and Graph Isomorphic Networks (GIN). All comparison models underwent hyperparameter grid search optimization to achieve their optimal performance. Experimental results are shown in Table 1, demonstrating that MDTL-Net exhibits significant advantages across all metrics.

[0141] Table 1: Performance comparison of different algorithms on the Tox21 dataset

[0142] Model AUC-ROC AUC-PR Accuracy F1-Score MCC MDTL-Net 0.894±0.012 0.765±0.021 0.852±0.015 0.813±0.018 0.724±0.025 GCN 0.841±0.015 0.692±0.024 0.795±0.018 0.751±0.021 0.642±0.028 DNN 0.812±0.018 0.654±0.028 0.763±0.022 0.718±0.025 0.591±0.031 LSTM 0.825±0.016 0.671±0.026 0.778±0.019 0.732±0.023 0.615±0.029 DBN 0.798±0.021 0.632±0.031 0.745±0.025 0.695±0.029 0.563±0.035 CNN 0.833±0.017 0.680±0.025 0.786±0.020 0.740±0.024 0.628±0.030 GIN 0.855±0.014 0.710±0.023 0.810±0.017 0.772±0.020 0.675±0.027

[0143] Table 1 clearly demonstrates the superiority of the proposed method (MDTL-Net) over existing mainstream deep learning techniques. MDTL-Net achieves an AUC-ROC of 0.894, which is approximately 3.9 percentage points higher than the second-best performing Graph Isomorphic Network (GIN) and approximately 5.3 percentage points higher than the traditional Graph Convolutional Network (GCN). This indicates that by combining a dual-branch architecture with a cross-modal attention mechanism, the model can more accurately capture the complex nonlinear relationship between molecular structure and biological response. Furthermore, MDTL-Net achieves a Matthews correlation coefficient (MCC) of 0.724, significantly higher than other comparative models, proving the robustness of this method in handling imbalanced data. In contrast, simple DNNs or DBNs, lacking effective modeling of molecular topology and temporal dynamic features, exhibit relatively weaker performance.

[0144] To further visualize the distribution of the feature space, the output features of the last fully connected layer of the model are dimensionality-reduced to generate a high-dimensional feature distribution visualization, such as... Figure 4 As shown, Figure 4 The t-SNE algorithm is used to map high-dimensional features to a two-dimensional plane, where points of different colors represent compounds with different toxicological activities. Figure 4As can be seen, the features extracted by MDTL-Net form obvious clusters in the vector space, with similar samples tightly clustered and dissimilar samples clearly demarcated. In contrast, the feature distribution of comparative algorithms (such as GCN and DNN) appears more chaotic, with more overlapping regions. This phenomenon intuitively verifies the effectiveness of the proposed method in feature decoupling and representation.

[0145] In addition, to explore the contribution of cross-modal attention mechanisms to model performance, ablation experiments were conducted in this embodiment. Three variants were set up: variant A retained only the molecular graph branch, variant B retained only the molecular fingerprint and biological temporal branches, and variant C removed the cross-modal attention module (replacing it with simple feature splicing). The experimental results are shown in Table 2.

[0146] Table 2: Analysis of Ablation Experiment Results of Model Architecture

[0147] Model variants describe AUC-ROC AUC-PR Loss value convergence speed (Epochs) Variant A Molecular diagram branching only (GAT) 0.845±0.014 0.698±0.023 65 Variant B Fingerprint + Timing Branch Only (Bi-GRU + TCN) 0.832±0.016 0.685±0.025 72 Variant C A complete architecture but lacking an attention mechanism (concatenation). 0.868±0.013 0.731±0.022 55 Complete model MDTL-Net (including attention mechanism) 0.894±0.012 0.765±0.021 42

[0148] Table 2 reveals the synergistic effect of each module. The results of variants A and B show that neither molecular structure information nor biological time-series information alone can achieve optimal prediction results, indicating that the complementarity of multimodal data is crucial for toxicology prediction. Although variant C integrates multimodal data, its performance is lower than the full model due to the lack of an effective feature interaction mechanism. The full model, by introducing a cross-modal attention mechanism, achieves convergence of the loss value in 42 epochs, far faster than variant C's 55 epochs, and improves the AUC-ROC by approximately 2.6%. This demonstrates that the attention mechanism not only improves prediction accuracy but also significantly accelerates the model's training and convergence process. Figure 5 The chart shows the loss curve during training. The curve for the complete model shows the steepest and most stable decline, further supporting the above conclusion.

[0149] Example 3, building upon Example 2, further introduces a time-drift perception module and an adverse outcome pathway (AOP) knowledge graph causal consistency check to verify the model's generalization ability and biological interpretability in complex dynamic environments. The experimental dataset was expanded to the ToxCast dynamic dataset, which includes time-series drift, and a liver toxicity AOP sub-graph was constructed containing 1,250 nodes (including MIE, KE, and AO) and 3,400 edges.

[0150] In terms of model architecture, a time drift sensing module was added based on Example 2. This module adopts a variational autoencoder (VAE) structure, with the latent variable dimension set to 32 and KL divergence weight coefficients. Set to 0.1. Simultaneously, a knowledge graph embedding module is introduced, using the TransE algorithm to map the AOP graph into vectors, with the embedding dimension set to 128. The marginal value of the marginal loss function is... Set to 1.0. The model training strategy employs a two-stage training method: the first stage freezes the knowledge graph module to train the feature extractor, and the second stage jointly fine-tunes all modules. The optimizer is still Adam, but a learning rate decay strategy is introduced with a decay coefficient of 0.95.

[0151] To verify the effectiveness of the time-drift-aware module, the temporal distribution difference between the training and test sets was simulated (i.e., training data mainly comes from 0-12h, and test data mainly comes from 12-24h and beyond). The experiment compared the baseline model without time-drift correction (i.e., MDTL-Net in Example 2) with the improved model (MDTL-Net-TD) that incorporated the time-drift-aware module. The results are shown in Table 3.

[0152] Table 3: Verification of the impact of the time drift sensing module on the model's generalization ability

[0153] Experimental group Model Configuration Source domain test set AUC-ROC Target domain (drift) test set AUC-ROC Performance degradation (Δ) Maximum Mean Difference (MMD) Loss control group MDTL-Net (without TD module) 0.891 0.742 16.7% 0.452 experimental group MDTL-Net-TD (including TD module) 0.885 0.836 5.5% 0.128

[0154] Table 3 shows that when faced with the challenge of temporal drift, the performance of the uncorrected control group model significantly declined, with AUC-ROC dropping from 0.891 to 0.742, a decrease of 16.7%. Furthermore, the maximum mean difference (MMD) loss between the source and target domain feature distributions reached 0.452, indicating a severe shift in feature distribution. Conversely, the experimental group model, which incorporated the temporal drift-aware module, maintained a high AUC-ROC of 0.836 on the target domain test set, with the performance decline controlled within 5.5%, and the MMD loss significantly reduced to 0.128. This fully demonstrates that the temporal drift-aware module can effectively align feature distributions at different time points, greatly improving the model's robustness and predictive stability in dynamically changing environments. Figure 6 The image shows a comparison of the feature distributions of the source and target domains visualized using the UMAP algorithm. The overlap between the source and target domain samples in the experimental group is significantly higher than that in the control group, which intuitively reflects the effect of domain adaptation.

[0155] To verify the improvement in the biological rationality of model predictions by AOP knowledge graph constraints, this embodiment calculates the causal consistency score (CCS) between the model's predictions and known AOP pathways. The CCS score is defined as the degree of matching between the predicted key event (KE) activation order and the order defined in the AOP graph, with a value ranging from [0, 1]. Simultaneously, the performance of models with and without knowledge graph constraints (MDTL-Net-KG) was compared in a sparse data scenario (using only 10% of the training data). The results are shown in Table 4.

[0156] Table 4: Impact of AOP knowledge graph constraints on model interpretability and small sample performance

[0157] Experimental setup Model Configuration Full data AUC-ROC 10% sparse data AUC-ROC Causality Consistency Score (CCS) False positive rate (FPR) Setting 1 MDTL-Net (without KG constraints) 0.894 0.654 0.612 0.145 Setting 2 MDTL-Net-KG (including KG constraints) 0.908 0.782 0.895 0.082

[0158] Table 4 shows that introducing AOP knowledge graph constraints not only significantly improved the causal consistency score (CCS) from 0.612 to 0.895, indicating that the model's predicted toxicological mechanisms are highly consistent with biological facts, but also improved the AUC-ROC by 12.8% compared to the unconstrained model in sparse data scenarios (from 0.654 to 0.782). This demonstrates that the prior causal relationships provided by the knowledge graph effectively regularize the model, enabling it to maintain high prediction accuracy even with insufficient data, and significantly reducing the false positive rate (FPR from 0.145 to 0.082). Figure 7 A visualization of the knowledge graph embedding space is presented. Figure 7 The semantically related toxicological pathway nodes (such as mitochondrial damage and oxidative stress) are closer in the vector space, proving that TransE embedding effectively captures toxicological semantic information.

[0159] To delve deeper into the model's decision-making process, this embodiment utilizes the SHAP (SHapley Additive ex Planations) value analysis method to perform attribution analysis on the prediction results, generating a feature importance ranking chart, such as... Figure 8 As shown. Figure 8 The study showcases the top 20 key features influencing hepatotoxicity prediction, including specific substructures in the molecular fingerprint (such as the nitrobenzene ring) and changes in gene expression at specific time points (such as the expression level of CYP3A4 at 12h). The figure shows that cross-modal attention mechanisms assign higher weight to biological temporal features, particularly in the early stages of toxicity development (6h-12h), consistent with the theory of "early biomarkers" in toxicology. Through a single-sample prediction path interpretation diagram, key paths in the AOP (Adverse Event Path) map are highlighted, demonstrating the logical progression from compound structural features to molecular initiation events (MIEs) and then to adverse outcomes (AOs).

[0160] This embodiment, by introducing time drift awareness and AOP knowledge graph constraints, not only maintains excellent predictive performance under extreme conditions of dynamic data distribution and sparse samples, but also endows the model with profound biological interpretability, achieving a leap from black-box prediction to white-box mechanism analysis.

[0161] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for analyzing and predicting the time-series characteristics of dynamic responses in molecular toxicological pathways, characterized in that, Includes the following steps: S1. Acquire multimodal molecular feature data of the test compound, and multi-omics time-series response data of biological samples exposed to the test compound at multiple consecutive time points; the multimodal molecular feature data includes molecular graph structural features and molecular fingerprint sequence features, and the multi-omics time-series response data includes transcriptomics and proteomics data; S2. Construct a dual-branch dynamic decoupled feature extraction network to independently encode the multimodal molecular feature data and multi-omics temporal response data; use a graph attention network (GAT) to extract molecular graph structure features, use a bidirectional gated recurrent unit (Bi-GRU) to extract molecular fingerprint sequence features, and use a temporal convolutional network (TCN) to capture the temporal dependencies of the multi-omics data, thereby outputting decoupled chemical spatial feature vectors and biological temporal feature vectors. S3. Input the chemical spatial feature vector and biological temporal feature vector obtained in S2 into the cross-modal cross-attention fusion module. By calculating the association weight between chemical substructures and biological pathway nodes, the deep interaction and alignment of molecular features and biological responses are realized, and a cross-modal joint representation vector is generated. S4. Input the cross-modal joint characterization vector into the preset time drift sensing prediction module; the time drift sensing prediction module has a built-in time distribution offset correction layer, which is used to calculate the distribution difference between the current test compound and the historical training data in the time dimension, and dynamically adjust the parameter weights of the prediction model based on the difference, so as to output the time-corrected dynamic response prediction result of the toxicological pathway. S5. Map the prediction results of step S4 to the pre-constructed adverse outcome pathway AOP knowledge graph. By calculating the causal consistency scores between the prediction results and the molecular initiation event (MIE) and key event (KE) in the AOP knowledge graph, the prediction results are verified by biological mechanisms and explained, and the activation timeline of the toxicological pathway is output.

2. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 1, characterized in that, The S2 step utilizes a graph attention network (GAT) to extract molecular graph structural features, specifically including: Representing the molecular diagram structure as a graph Where V is the set of atomic nodes and E is the set of chemical bond edges; for each atomic node Its initial eigenvector It is composed of its atomic type, hybridization state, aromaticity, and formal charge; Feature aggregation is performed through at least one graph attention layer. Layer nodes The features are updated as follows: ,in, For nodes The set of neighboring nodes, For the first Layer-learnable weight matrix, It is a non-linear activation function; attention coefficient Calculated using the following formula: ,in, For the first Layer-learnable attention vectors This represents a vector concatenation operation; the feature vectors of all nodes are then subjected to global pooling to obtain the graph structure feature vectors of the molecule. .

3. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 2, characterized in that, The extraction of molecular fingerprint sequence features using a bidirectional gated recurrent unit (Bi-GRU) in S2 specifically includes: The molecular fingerprint sequence is represented as ,in Let be the d-dimensional fingerprint feature vector at time t; The sequence is encoded using a forward GRU layer and a backward GRU layer, respectively. The hidden state of the forward GRU at time t is... The hidden state of the backward GRU at time t The calculation is as follows: , The forward and backward hidden states are concatenated in a dimension to obtain the final hidden state at time t. ; Average pooling is performed on the hidden states at all time points in the sequence to obtain the feature vector of the molecular fingerprint sequence. .

4. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 3, characterized in that, The S2 method utilizes a Temporal Convolutional Network (TCN) to capture the temporal dependencies of multi-omics data, specifically including: For the observation sequence of the k-th omics data at T consecutive time points ,in For time t 3D feature vector; Processing is performed using a one-dimensional causal convolutional layer, and the output of the m-th convolutional layer... for: Where K is the kernel size and d is the dilation coefficient. and These are the weights and bias terms of the convolutional kernel in the m-th layer, respectively. For the upper layer Input at any moment; The TCN contains at least one residual block, each containing two of the aforementioned causal convolutional layers, and incorporates weight normalization and Dropout operations. The output of the last time step is mapped through a fully connected layer to obtain the temporal feature vector of the omics data. By concatenating the time-series feature vectors of all omics data, the biological time-series feature vector is obtained. .

5. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 4, characterized in that, The cross-modal attention fusion module in S3 specifically includes: The chemical space feature vector output by S2 As the query vector Q, Depend on With molecular fingerprint sequence feature vector This is a concatenation of biological time-series feature vectors. As the key vector K and the value vector V; The cross-modal attention weight matrix A is calculated using the following formula: ,in, The dimension of the key vector; The fused feature representation is obtained by weighting the value vector V using the attention weight matrix A. : ,Will Compared with the original chemical space feature vector By performing residual connections and layer normalization, a cross-modal joint representation vector is obtained. .

6. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 5, characterized in that, The cross-modal attention fusion module employs a multi-head attention mechanism, specifically including: The above cross-modal attention calculation process is executed in parallel for h times, and the output of the i-th attention head is: ; The outputs of h heads are concatenated and then subjected to a learnable linear transformation. Projecting yields the final fused features: ,in, , For each head, the value vector dimension, This represents the dimension of the model's hidden layers.

7. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 1, characterized in that, The time drift sensing and prediction module in S4 calculates the distribution difference in its time distribution offset correction layer in the following manner: For the cross-modal joint characterization vector of the current test compound Calculate the set of cross-modal joint representation vectors of it and all samples in the historical training dataset. Maximum mean difference (MMD) between them: ,in, To map features to the reproducing kernel Hilbert space The mapping function.

8. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 7, characterized in that, The time drift sensing and prediction module is based on the calculated MMD value. Dynamically adjust the parameter weights of the prediction model: Let the basic parameters of the prediction model be Adjusted parameters for: ,in, For learning rate, The regularization loss term based on MMD is given by the formula: .

9. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 1, characterized in that, The pre-constructed adverse outcome pathway AOP knowledge graph in S5 includes nodes such as molecular initiation event (MIE), key event (KE), and adverse outcome (AO), and edges represent causal relationships between events. The causal consistency scores between the calculated prediction results and the molecular initiation events (MIE) and key events (KE) in the AOP knowledge graph specifically include: The toxicological pathway dynamic response prediction results output by S4 are mapped to the node space of the AOP knowledge graph to obtain the set of event nodes for prediction activation. ; for Each event node in Calculate its relationship with known causal paths in the knowledge graph. Match score : ,in, For indicator functions, For event nodes Semantic similarity with the corresponding event in path p.

10. The method for analyzing and predicting the time-series characteristics of dynamic responses of molecular toxicological pathways according to claim 9, characterized in that, The biological mechanism verification and interpretability reasoning also include: Based on the calculated causal consistency score The prediction results are sorted, and those with scores higher than a preset threshold are selected. The event nodes constitute the core activation pathway. ; Backtracking in the AOP knowledge graph using graph traversal algorithms The upstream MIE nodes of each event node generate a complete causal chain explanation from MIE to AO.

Images