Metabolic stability prediction method based on pharmacophore-oriented multi-modal representation framework and contrast hypergraph learning

Abstract

The invention provides a metabolic stability prediction method based on a pharmacophore-oriented multi-mode representation framework and contrast hypergraph learning, and belongs to the technical field of biological information, and the method comprises a pharmacophore extraction and representation module which is used for extracting hydrogen bond donor and acceptor functional groups and converting the functional groups into numerical features; the multi-mode molecular encoder module is composed of a molecular fingerprint encoder, a molecular map encoder, a pharmacophore encoder and a text encoder; the hypergraph comparative learning module is used for constructing and executing comparative learning based on the molecular graph and the pharmacophore hypergraph; and the attention fusion and prediction module is used for integrating multi-modal features through a multi-head attention mechanism and outputting a prediction result and pharmacophore importance analysis. According to the method, by integrating multi-modal molecular representation and hypergraph contrast learning, the accuracy and generalization ability of compound metabolism stability prediction are effectively improved, and specific structure guidance can be provided for compound optimization in drug research and development.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a pharmacophore-guided multimodal representation framework and a method for predicting metabolic stability based on contrastive hypergraph learning. Background Technology

[0002] Currently, methods for predicting metabolic stability mainly include those based on traditional machine learning, deep learning, and graph neural networks. Traditional machine learning methods, such as random forests, support vector machines, and Bayesian models, utilize molecular descriptors and fingerprint features to predict the metabolic stability of compounds. However, these methods may overlook the complexity of molecular structures and pharmacophore information when using hand-designed features to predict metabolic stability. On the other hand, graph neural network-based methods, such as graph convolutional networks (GCNs) and graph attention networks (GATs), have been successfully applied to extract molecular graph structure and semantic features to achieve metabolic stability prediction.

[0003] In existing research, graph contrastive learning strategies (CMMS-GCL) and gated recurrent units (GRUs) are used to encode molecular SMILES sequences and topological features, generating cross-modal molecular representations for metabolic stability prediction. To address the limitation of traditional methods lacking contrastive learning mechanisms when learning molecular features, bond graph enhancement and contrastive learning are also employed to improve metabolic stability prediction. Existing techniques also include using graph convolutional networks (GCNs) to convert molecular SMILES sequences into graph structures for metabolic stability prediction; however, these methods fail to effectively utilize the pharmacophore, a crucial functional group.

[0004] In addition, existing methods for predicting metabolic stability have some limitations, mainly in the following aspects: First, existing methods usually ignore the key functional group information of the pharmacophore, which may not be able to fully capture the key structural features that affect metabolic stability; Second, these existing methods mainly focus on the overall molecular structure information, but ignore multiple different feature sets of the pharmacophore; Third, the interpretability of existing methods is limited, and they cannot identify the pharmacophore that plays a key role in metabolic stability, thus affecting the practical value of the model in drug optimization. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the present invention aims to provide a pharmacophore-oriented multimodal representation framework and a metabolic stability prediction method based on contrastive hypergraph learning. By constructing a pharmacophore-oriented multimodal representation framework and combining hypergraph contrastive learning with multi-head attention fusion mechanism, the invention systematically integrates multimodal information such as pharmacophores, molecular graphs, fingerprints, and text. Furthermore, by using Shapley value analysis to provide pharmacophore importance assessment, the invention effectively improves the accuracy and generalization ability of compound metabolic stability prediction, enhances model interpretability, and can provide specific structural guidance for compound optimization in drug development.

[0006] To achieve the above objectives, the present invention provides the following solution: A pharmacophore-oriented multimodal representation framework includes: The pharmacophore extraction and representation module is used to extract hydrogen bond donor and acceptor functional groups and convert them into numerical features. The multimodal molecular encoder module consists of four dedicated encoders: a molecular fingerprint encoder, a molecular graph encoder, a pharmacophore encoder, and a text encoder, and is connected to the pharmacophore extraction and representation module. The hypergraph contrast learning module is used to receive the molecular graph representation and pharmacophore hypergraph representation output by the multimodal molecular encoder module, construct positive sample pairs based on the two representations of the same molecule, and use the contrast learning loss function to narrow their distance in order to learn a more discriminative molecular representation. The attention fusion and prediction module integrates multimodal features through a multi-head attention mechanism and outputs prediction results and pharmacophore importance analysis, and is connected to the hypergraph contrastive learning module.

[0007] Preferably, the pharmacophore extraction and representation module is specifically used to: identify one or more of hydrogen bond donors, hydrogen bond acceptors, hydrophobic regions, aromatic rings, cations, anions, and polymeric hydrophobic groups in the molecular structure, and construct a pharmacophore-molecule correlation matrix to convert the pharmacophore into a numerical feature representation.

[0008] Preferably, the multimodal molecular encoder module includes a molecular fingerprint encoder that encodes one or more of ECFP, MACCS, PubChem, and Pharmacophore ErG fingerprints using a multilayer perceptron to obtain a molecular fingerprint representation.

[0009] Preferably, the multimodal molecular encoder module includes a molecular graph encoder that uses a graph isomorphic network to learn the topological structure representation of the molecular graph and uses a combination strategy of global max pooling and average pooling to aggregate node embedding vectors.

[0010] Preferably, the multimodal molecular encoder module includes a text encoder that uses a pre-trained SciBERT model to process molecular description text in order to extract semantic information; the molecular description text is generated by a large language model, specifically by first converting the SMILES sequence of the molecule into IUPAC names, and then using these as input, and driving the large language model to generate text describing the structural features and functions of the molecule through prompts.

[0011] Preferably, the hypergraph contrastive learning module is specifically used to: represent pharmacophores as hyperedges connecting multiple atoms to construct a hypergraph structure; construct positive sample pairs based on the molecular graph representation and hypergraph representation of the same molecule, and construct negative sample pairs based on the representations of different molecules; and optimize model parameters using the InfoNCE contrastive loss function.

[0012] Preferably, the multi-head attention mechanism in the attention fusion and prediction module is specifically used to: concatenate multimodal feature representations into a matrix, and calculate Query, Key, and Value matrices accordingly; dynamically aggregate and interact feature information from different modalities through the attention mechanism, and assign differentiated weights to each modality; and finally connect and input the output of the multi-head attention to the multilayer perceptron layer to output metabolic stability prediction results.

[0013] This invention also provides a metabolic stability prediction method based on contrastive hypergraph learning, applying the pharmacophore-guided multimodal representation framework described above, and including the following steps: S1. Based on the pharmacophore extraction and representation module's identification of pharmacophore and molecular atom information, a hypergraph structure with atoms as nodes and pharmacophores as hyperedges is constructed. S2. The multimodal molecular encoder module encodes the numerical features, ECFP, MACCS, PubChem, Pharmacophore ErG fingerprint, molecular graph topology, and molecular description text output by the pharmacophore extraction and representation module using the pharmacophore encoder, molecular fingerprint encoder, molecular graph encoder, and text encoder respectively, to obtain four types of feature embeddings: pharmacophore, fingerprint, molecular graph, and text. S3. Through the hypergraph contrast learning module, perform hypergraph convolution operation on the hypergraph structure to generate hypergraph feature representation. Then, construct positive sample pairs based on the molecular graph representation and hypergraph feature representation of the same molecule, and construct negative sample pairs based on the representations of different molecules. Optimize the model parameters through the InfoNCE contrast loss function to obtain the optimized multimodal feature embedding. S4. Using the attention fusion and prediction module, for the optimized multimodal feature embedding, the fingerprint, molecular graph, pharmacophore and text feature embedding output by the multimodal molecular encoder module are fused and interacted through a multi-head attention mechanism to obtain a weighted multimodal fusion representation; S5. Input the hierarchically fused features into the multilayer perceptron layer, output the prediction results of molecular metabolic stability, and conduct pharmacophore importance assessment through Shapley value analysis to output the impact results of key pharmacophores.

[0014] Preferably, in S5, the pharmacophore importance assessment is achieved by quantifying the marginal contribution of each pharmacophore characteristic to the metabolic stability prediction results; the higher the score, the greater the influence of the pharmacophore on metabolic stability.

[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) This invention solves the problems of insufficient utilization of pharmacophore information and single molecular feature capture in existing technologies by constructing a pharmacophore-oriented multimodal representation framework, and realizes comprehensive and in-depth characterization of molecular features. In the framework, the pharmacophore extraction and representation module accurately identifies functional groups such as hydrogen bond donors and aromatic rings, and converts them into computable numerical features through the pharmacophore-molecule correlation matrix; the multimodal molecular encoder module captures multi-dimensional information through four dedicated encoders: the molecular fingerprint encoder processes complementary fingerprints such as ECFP and MACCS through MLP, the molecular graph encoder learns the topological structure with GIN and aggregates features through pooling, the pharmacophore encoder models multi-atom correlations with hypergraph convolution, and the text encoder extracts molecular descriptive semantics with SciBERT. Compared with existing methods that rely only on a single fingerprint or graph structure, this invention can capture more key information related to metabolic stability, laying the foundation for metabolic stability prediction.

[0016] (2) This invention significantly improves the accuracy and generalization ability of metabolic stability prediction by combining the hypergraph contrast learning module and the prediction method. It effectively overcomes the limitation of existing models that simplify pharmacophores to low-order or paired features, making it difficult to capture their high-order structural dependencies. In the prediction method provided by this invention, a hypergraph structure with atoms as nodes and pharmacophores as hyperedges is constructed based on pharmacophore and atomic information, which is naturally adapted to the characteristic of pharmacophores connecting multiple atoms. First, hypergraph feature representation is generated by hypergraph convolution. Then, the graph and hypergraph representation of the same molecule are used as positive samples and the representations of different molecules are used as negative samples. The consistency of positive samples is maximized and the interference of negative samples is minimized by the InfoNCE loss function, so that the model can accurately capture the consistency between molecular structure and key pharmacophores. The optimized feature embedding has stronger discriminative power. Compared with existing methods such as CMMS-GCL, the prediction AUC, MCC and other indicators are significantly improved, and it shows excellent generalization ability in external test set and leave-one-out clustering verification.

[0017] (3) This invention uses attention fusion and interpretability analysis. In the attention fusion and prediction module, the multi-head attention mechanism dynamically allocates the modal weights by calculating the Q, K, and V matrices, and efficiently fuses multi-source information. It also uses Shapley value analysis to quantify the marginal contribution of each pharmacophore to the prediction results. The score directly reflects the degree of influence of the pharmacophore. Furthermore, this invention constructs a pharmacophore-oriented multimodal representation framework, combines hypergraph contrastive learning and multi-head attention fusion mechanism, and systematically integrates multimodal information such as pharmacophores, molecular graphs, fingerprints and text. It also uses Shapley value analysis to provide pharmacophore importance assessment, which effectively improves the accuracy and generalization ability of compound metabolic stability prediction, enhances model interpretability, and can provide specific structural guidance for compound optimization in drug development. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the structure of a pharmacophore-guided multimodal representation framework according to the present invention; Figure 2 This is a flowchart of a metabolic stability prediction method based on contrastive hypergraph learning according to the present invention; Figure 3 This is a schematic diagram comparing the performance of the model provided in the embodiments of the present invention on leave-one-out clustering verification on novel structural molecules. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example like Figure 1 As shown, this invention provides a pharmacophore-guided multimodal representation framework, wherein, Figure 1 A in the diagram illustrates the complete computational flow from input compound to final stability prediction. Figure 1 The C code in the figure shows the specific implementation of each module, which includes: The pharmacophore extraction and representation module is used to extract hydrogen bond donor and acceptor functional groups and convert them into numerical features. Specifically, it identifies one or more of hydrogen bond donors, hydrogen bond acceptors, hydrophobic regions, aromatic rings, cations, anions, and polymeric hydrophobic groups in the molecular structure, constructs a pharmacophore-molecule correlation matrix, and converts the pharmacophores into numerical feature representations.

[0023] The multimodal molecular encoder module comprises four dedicated encoders: a molecular fingerprint encoder, a molecular graph encoder, a pharmacophore encoder, and a text encoder, and is connected to the pharmacophore extraction and representation module. The molecular fingerprint encoder in the multimodal molecular encoder module encodes one or more of ECFP, MACCS, PubChem, and Pharmacophore ErG fingerprints using a multilayer perceptron to obtain the molecular fingerprint representation. The molecular graph encoder in the multimodal molecular encoder module learns the topological structure representation of the molecular graph using a graph isomorphism network and aggregates node embedding vectors using a combination of global max pooling and average pooling. The text encoder in the multimodal molecular encoder module processes the molecular description text generated by ChatGPT using a pre-trained SciBERT model, with the generation process as follows: Figure 1 As shown in B, semantic information is extracted.

[0024] The hypergraph contrastive learning module, connected to the multimodal molecular encoder module, is used to construct and execute contrastive learning based on molecular graphs and pharmacophore hypergraphs. Specifically, it is used to: represent pharmacophores as hyperedges connecting multiple atoms to construct a hypergraph structure; construct positive sample pairs based on the molecular graph representation and hypergraph representation of the same molecule, and construct negative sample pairs based on the representations of different molecules; and optimize model parameters using the InfoNCE contrastive loss function.

[0025] The attention fusion and prediction module integrates multimodal features through a multi-head attention mechanism and outputs prediction results and pharmacophore importance analysis. It is connected to the hypergraph contrastive learning module. Specifically, it is used to: concatenate the multimodal feature representations into a matrix and calculate the Query, Key, and Value matrices accordingly; dynamically aggregate and interact with feature information from different modalities through the attention mechanism and assign differentiated weights to each modality; and finally, connect the output of the multi-head attention and input it into the multilayer perceptron layer to output the metabolic stability prediction result.

[0026] Based on the pharmacophore-guided multimodal representation framework described above, this framework is applied to a metabolic stability prediction method based on contrastive hypergraph learning, such as... Figure 2 As shown, the specific steps include: S1. Based on the pharmacophore extraction and representation module's identification of pharmacophore and molecular atom information, a hypergraph structure with atoms as nodes and pharmacophores as hyperedges is constructed.

[0027] In the above steps, the pharmacophore identification tool of RDKit is used to identify the main pharmacophore types, including hydrogen bond donors, hydrogen bond acceptors, hydrophobic regions, aromatic rings, cations, anions, and polymeric hydrophobic groups. First, the molecular structure is resolved into a graphical representation of atoms and bonds. Then, pharmacophores are identified through pattern matching. Finally, a pharmacophore-molecule correlation matrix is ​​constructed to describe the distribution patterns and interactions of pharmacophores in the molecule.

[0028] S2. The multimodal molecular encoder module encodes the numerical features, ECFP, MACCS, PubChem, Pharmacophore ErG fingerprint, molecular graph topology, and molecular description text output by the pharmacophore extraction and representation module using the pharmacophore encoder, molecular fingerprint encoder, molecular graph encoder, and text encoder, respectively, to obtain four types of feature embeddings: pharmacophore, fingerprint, molecular graph, and text.

[0029] In the above steps, the molecular fingerprint encoder encodes four complementary fingerprints (ECFP, MACCS, PubChem, and Pharmacophore ErG) using a multilayer perceptron (MLP) to obtain rich fingerprint representations. The molecular graph encoder uses a graph isomorphic network (GIN) to learn the topological representation information of the molecular graph and employs a combination of global max pooling and average pooling strategies to aggregate the embedding vectors of nodes in the graph. Finally, it obtains the graph embedding representation of the molecule through an MLP layer. The pharmacophore encoder uses a hypergraph convolutional network to handle the multi-atom characteristics of the pharmacophore, enabling direct modeling of the complex relationships between the pharmacophore and multiple atoms. The text encoder uses a pre-trained SciBERT model to process the molecular descriptions generated by ChatGPT and extract semantic features.

[0030] S3. Through the hypergraph contrast learning module, perform hypergraph convolution operation on the hypergraph structure to generate hypergraph feature representation. Then, construct positive sample pairs based on the molecular graph representation and hypergraph feature representation of the same molecule, and construct negative sample pairs based on the representations of different molecules. Optimize the model parameters through the InfoNCE contrast loss function to obtain the optimized multimodal feature embedding.

[0031] In the above steps, each atom in the hypergraph structure corresponds to a node, and each pharmacophore corresponds to a hyperedge, which can simultaneously connect all atoms participating in that pharmacophore. Node embeddings are learned through hypergraph convolution operations, and then contrastive learning is applied. Positive sample pairs construct graph and hypergraph representations based on the same molecule, while negative sample pairs represent representations of different molecules. The InfoNCE contrastive loss function is used to optimize representation learning by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs.

[0032] S4. Using the attention fusion and prediction module, for the optimized multimodal feature embedding, the fingerprint, molecular graph, pharmacophore and text feature embedding output by the multimodal molecular encoder module are fused and interacted through a multi-head attention mechanism to obtain a weighted multimodal fusion representation.

[0033] In the steps described above, a multi-head attention mechanism is integrated into the attention fusion and prediction module to effectively integrate the multi-perspective features of molecules. By calculating the Query (Q), Key (K), and Value (V) matrices, the model constructed by this framework can assign weights to representations of different modalities. Finally, the outputs of the multi-head attention are concatenated and input into an MLP layer to obtain the predicted probability of metabolic stability and the pharmacophore importance score.

[0034] S5. Input the hierarchically fused features into the multilayer perceptron layer, output the prediction results of molecular metabolic stability, and conduct pharmacophore importance assessment through Shapley value analysis to output the impact results of key pharmacophores.

[0035] In the above steps, the importance assessment of pharmacophores is achieved by quantifying the marginal contribution of each pharmacophore feature to the prediction results of metabolic stability. The higher the score, the greater the impact of the pharmacophore on metabolic stability.

[0036] The feasibility of the above is verified through a specific experiment. This experiment used the HLM (Human Liver Microsomes) and RLM (Rat Liver Microsomes) datasets, which are public databases containing a large number of compounds with metabolic stability tags. The HLM dataset includes 5878 compounds, of which 3784 are stable and 2094 are unstable. The RLM dataset contains 3108 compounds, of which 1542 are stable and 1566 are unstable. This experiment employed a 10-fold cross-validation strategy to screen for valid compound structures from the database. The experiment also included SMILES data for each compound to capture structural information. Finally, the main experiments were conducted on the HLM dataset, and the model's generalization ability was verified using independent external test sets. Comprehensive experiments were performed on these datasets and data from different metabolic stability categories.

[0037] In this embodiment, a pharmacophore-guided multimodal model was constructed based on the framework provided above. This model was compared with nine other metabolic stability prediction models, including state-of-the-art methods such as FP-GBDT, XGBoost, D-MPNN, GAT, PredMS, MGCN, AttentiveFP, CMMS-GCL, and MS-BACL. Experimental results show that the model provided in this embodiment has the best performance. The specific experimental results are shown in Table 1.

[0038] Table 1. Performance comparison results of different models on the HLM dataset

[0039] Furthermore, compared to other models, this embodiment achieved the best performance in AUC, accuracy, F1 score, and MCC. On the HLM dataset, it achieved an AUC of 87.6% and an MCC of 62.6%, particularly improving the MCC metric by 6 percentage points compared to CMMS-GCL and 2.5 percentage points compared to MS-BACL. On the external test set, it achieved an AUC of 88.3%, demonstrating good generalization ability.

[0040] To verify the predictive ability of the model constructed in this embodiment for novel structural molecules, a leave-one-out clustering cross-validation experiment was also conducted to evaluate the ability of this embodiment to accurately predict the metabolic stability of each novel structural cluster. Experiments were also performed according to standard methods, clustering the dataset into five structurally similar groups using Tanimoto similarity calculation, followed by leave-one-out clustering validation. Figure 3 As shown, this embodiment significantly outperforms existing methods in all five clusters, achieving an F1 score of 92.55%, a 12.55% improvement over CMMS-GCL. Furthermore, this embodiment validates the model's interpretability through Shapley value analysis. The superior performance of this embodiment surpasses all baseline methods, and the identified key pharmacophores are highly consistent with known metabolic sites, demonstrating the biological rationale and practical value of the model's predictions.

[0041] In addition, this embodiment and other models underwent ablation experiments to compare the contributions of different technical modules to the prediction performance. The results are shown in Table 2.

[0042] Table 2 Ablation Experiment Results

[0043] Table 2 shows that this embodiment significantly outperforms the G-Model, GH-Model, and models that remove the text encoder or attention fusion module in terms of AUC, accuracy, F1 score, and MCC. This fully verifies the effective improvement in the prediction performance of molecular metabolic stability based on the synergistic effect of pharmacophore-oriented multimodal representation, hypergraph contrastive learning, and attention fusion mechanism.

[0044] Therefore, the above-mentioned pharmacophore-oriented multimodal representation framework and contrastive hypergraph learning-based metabolic stability prediction method is adopted. By constructing a pharmacophore-oriented multimodal representation framework and combining hypergraph contrastive learning and multi-head attention fusion mechanism, multimodal information such as pharmacophore, molecular graph, fingerprint and text is systematically integrated. Furthermore, by using Shapley value analysis to provide pharmacophore importance assessment, the accuracy and generalization ability of compound metabolic stability prediction are effectively improved, the interpretability of the model is enhanced, and specific structural guidance can be provided for compound optimization in drug development.

[0045] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A pharmacophore-oriented multimodal representation framework, characterized in that, include: The pharmacophore extraction and representation module is used to extract hydrogen bond donor and acceptor functional groups and convert them into numerical features. The multimodal molecular encoder module consists of four dedicated encoders: a molecular fingerprint encoder, a molecular graph encoder, a pharmacophore encoder, and a text encoder, and is connected to the pharmacophore extraction and representation module. The hypergraph contrast learning module is used to receive the molecular graph representation and pharmacophore hypergraph representation output by the multimodal molecular encoder module, construct positive sample pairs based on the two representations of the same molecule, and use the contrast learning loss function to narrow their distance in order to learn a more discriminative molecular representation. The attention fusion and prediction module integrates multimodal features through a multi-head attention mechanism and outputs prediction results and pharmacophore importance analysis, and is connected to the hypergraph contrastive learning module.

2. The pharmacophore-guided multimodal representation framework according to claim 1, characterized in that, The pharmacophore extraction and representation module is specifically used to: identify one or more of hydrogen bond donors, hydrogen bond acceptors, hydrophobic regions, aromatic rings, cations, anions, and polymeric hydrophobic groups in the molecular structure, construct a pharmacophore-molecule correlation matrix, and convert the pharmacophore into a numerical feature representation.

3. The pharmacophore-guided multimodal representation framework according to claim 1, characterized in that, The multimodal molecular encoder module includes a molecular fingerprint encoder that uses a multilayer perceptron to encode one or more of ECFP, MACCS, PubChem, and Pharmacophore ErG fingerprints to obtain a molecular fingerprint representation.

4. The pharmacophore-guided multimodal representation framework according to claim 1, characterized in that, The multimodal molecular encoder module includes a molecular graph encoder that uses a graph isomorphic network to learn the topological structure representation of the molecular graph and employs a combination strategy of global max pooling and average pooling to aggregate node embedding vectors.

5. The pharmacophore-guided multimodal representation framework according to claim 1, characterized in that, The multimodal molecular encoder module includes a text encoder that uses a pre-trained SciBERT model to process molecular description text in order to extract semantic information. The molecular description text is generated by a large language model. Specifically, the SMILES sequence of the molecule is first converted into IUPAC name, and then used as input. The large language model is driven by prompts to generate text describing the structural features and functions of the molecule.

6. The pharmacophore-guided multimodal representation framework according to claim 1, characterized in that, The hypergraph contrastive learning module is specifically used to: represent pharmacophores as hyperedges connecting multiple atoms to construct a hypergraph structure; construct positive sample pairs based on the molecular graph representation and hypergraph representation of the same molecule, and construct negative sample pairs based on the representations of different molecules; and optimize model parameters using the InfoNCE contrastive loss function.

7. The pharmacophore-guided multimodal representation framework according to claim 1, characterized in that, The multi-head attention mechanism in the attention fusion and prediction module is specifically used to: concatenate multimodal feature representations into a matrix, and calculate Query, Key, and Value matrices accordingly; dynamically aggregate and interact feature information from different modalities through the attention mechanism, and assign differentiated weights to each modality; finally, connect the output of the multi-head attention and input it into the multilayer perceptron layer to output the metabolic stability prediction result.

8. A method for predicting metabolic stability based on contrastive hypergraph learning, characterized in that, The application of the pharmacophore-guided multimodal representation framework as described in any one of claims 1 to 7 includes the following steps: S1. Based on the pharmacophore extraction and representation module's identification of pharmacophore and molecular atom information, a hypergraph structure with atoms as nodes and pharmacophores as hyperedges is constructed. S2. The multimodal molecular encoder module encodes the numerical features, ECFP, MACCS, PubChem, Pharmacophore ErG fingerprint, molecular graph topology, and molecular description text output by the pharmacophore extraction and representation module using the pharmacophore encoder, molecular fingerprint encoder, molecular graph encoder, and text encoder respectively, to obtain four types of feature embeddings: pharmacophore, fingerprint, molecular graph, and text. S3. Through the hypergraph contrast learning module, perform hypergraph convolution operation on the hypergraph structure to generate hypergraph feature representation. Then, construct positive sample pairs based on the molecular graph representation and hypergraph feature representation of the same molecule, and construct negative sample pairs based on the representations of different molecules. Optimize the model parameters through the InfoNCE contrast loss function to obtain the optimized multimodal feature embedding. S4. Using the attention fusion and prediction module, for the optimized multimodal feature embedding, the fingerprint, molecular graph, pharmacophore and text feature embedding output by the multimodal molecular encoder module are fused and interacted through a multi-head attention mechanism to obtain a weighted multimodal fusion representation; S5. Input the hierarchically fused features into the multilayer perceptron layer, output the prediction results of molecular metabolic stability, and conduct pharmacophore importance assessment through Shapley value analysis to output the impact results of key pharmacophores.

9. The metabolic stability prediction method based on contrastive hypergraph learning according to claim 8, characterized in that, In S5, the pharmacophore importance assessment is achieved by quantifying the marginal contribution of each pharmacophore feature to the metabolic stability prediction results. The higher the score, the greater the impact of the pharmacophore on metabolic stability.

Images