Molecular property prediction modeling method based on dynamic and static feature fusion pre-training architecture

Abstract

The invention belongs to the technical field of food compound analysis, drug design and synthesis and molecular property prediction, and particularly relates to a molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture. Comprising the following steps: step 1, constructing a double-flow graph representation structure of molecules; 2, designing a static feature encoder based on an isotropic graph attention module; 3, designing a dynamic information encoder based on continuous filter convolution; and 4, designing a comprehensive loss function and a hyper-parameter optimization method based on a fuzzy mechanism. The method has the beneficial effects that a static feature encoder is designed by embedding an equivariant graph attention mechanism on the basis of a Transform architecture, and a dynamic information encoder is constructed by embedding a multi-scale filtering attention mechanism on the basis of the Transform architecture, so that the intermolecular connection rule and structural constraint can be extracted, the dynamic interaction force between adjacent nodes can be mined, and the accuracy of the dynamic interaction force is improved. The comprehensive characterization and efficient extraction of molecular properties are improved.

Description

Technical Field

[0001] This invention belongs to the field of food compound analysis, drug design and synthesis, and molecular property prediction technology, specifically involving a molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture. Background Technology

[0002] In the fields of food compound analysis and drug design and synthesis, molecular properties (such as permeability, toxicity, and flavor characteristics) are key indicators determining whether candidate compounds can proceed to further development. Traditional experimental-based R&D models are typically costly and time-consuming, severely limiting the ability to rapidly screen a wide range of candidate molecules. In recent years, deep learning methods have demonstrated superior performance in molecular representation learning. By analyzing molecular structures and extracting features, they can effectively predict whether a molecule possesses a specific molecular property. Compared to traditional laboratory methods, deep learning methods can effectively reduce R&D cycles and experimental costs.

[0003] Molecular property prediction methods aim to predict the performance of molecules in terms of physicochemical properties, biological activity, or pharmacokinetics through models. Machine learning-based prediction methods primarily rely on quantitative structure-property (QSPR / QSAR) models, transforming molecular structures into numerical descriptors and then using various machine learning algorithms to build predictive models. Commonly used methods include Support Vector Machines (SVM), Random Forests (RF), and K-Nearest Neighbors (KNN). These methods are simple and direct, but struggle with large datasets and cannot support large-scale data filtering. With the development of deep learning technology, molecular property prediction is gradually shifting towards an end-to-end learning paradigm. This type of method can directly input the raw molecular representation (such as SMILES sequences) into a neural network, automatically learning task-related feature representations. Sequence-based models often employ recurrent neural networks or Transformer architectures; while molecular graph-based models widely use graph neural networks (GNNs) and their variants to naturally capture the topological relationships between atoms (nodes) and chemical bonds (edges). To enhance the characterization of complex molecular structures, recent research has further introduced three-dimensional spatial information to enhance the model's understanding of local substructures and global interactions. Current mainstream methods are mostly based on graph neural networks (GNNs) and employ a "pre-training-fine-tuning" strategy. This involves first learning a general representation on large-scale unlabeled molecular data, and then fine-tuning it for downstream tasks. This effectively reduces the dependence on labeled data and significantly improves the model's generalization ability. Despite significant progress, some limitations remain. Firstly, most work focuses on atomic-level features (nodes and edges), neglecting the chemical semantics inherent in higher-level molecular motifs (such as functional groups, ring systems, or chain motifs). These motifs often carry complex chemical functions beyond paired atoms and are crucial for property prediction. Although recent studies have attempted to incorporate substructure information into the representation framework, the high complexity of chemical space and the diversity of substructure types mean that current methods lack a unified and flexible feature extraction architecture. Secondly, current molecular characterization learning is primarily based on observations of equilibrium structures, ignoring the fact that molecules are actually dynamic entities in continuous motion in three-dimensional Euclidean space. The properties of small molecules stem from their complex interactions and chemical reactions, which can be described by their potential energy surfaces and force fields. Summary of the Invention

[0004] The purpose of this invention is to provide a molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture, which can be applied to the prediction of molecular properties in food compound processing and drug design and development. It extracts multi-scale static structural features and dynamic force field features, realizes the fusion and complementarity of static features and dynamic information, generates more comprehensive molecular characterization codes, and improves the robustness and reliability of molecular property and functional group prediction.

[0005] The technical solution of this invention is as follows: A molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture, comprising the following steps:

[0006] Step 1: Based on the fuzzy partitioning method, each molecule is decomposed into the smallest atomic and bond links and the smallest ring structure, and a two-flow graph characterization structure of the molecule is constructed.

[0007] Step 2: Based on the Transformer architecture, embed the equivariant graph attention mechanism to design a static feature encoder to extract the connection rules and structural constraints between molecules;

[0008] Step 3: Based on the Transformer architecture, embed a multi-scale filtering attention mechanism to construct a dynamic information encoder, and learn molecular representation by capturing the changes in orientation vectors and scalar features between adjacent nodes induced by the rotation of the node center;

[0009] Step 4: Based on the fuzzy mechanism, design a method that integrates the compliance and loss functions and hyperparameter optimization, and construct a molecular property prediction modeling method based on the dynamic and static feature fusion pre-training architecture to predict molecular properties and functional groups;

[0010] Step 5: The pre-training dataset is from the ZINC public dataset. The pre-training dataset is screened and deduplicated to retain non-repeating small molecules. The downstream task data comes from BBBP, ToxCast and Sider. An additional flavor molecule dataset is constructed as an extension task. A joint pre-training strategy of mask reconstruction-denoising is designed to complete the model pre-training process and obtain the molecular property prediction model. Fine-tuning is then performed on the downstream task.

[0011] Step 1 includes constructing a motif dictionary and introducing a fuzzy segmentation method. The motif dictionary construction process is as follows: First, all subgraphs are extracted based on interatomic chemical bonds and ring structures. The subgraphs are divided into three categories: ring structures, bonds that are not directly connected to ring structures, and bonds that are directly connected to ring structures. The motifs are sorted using a term frequency-inverse document frequency algorithm to generate a dictionary index.

[0012] For a molecule composed of n atoms, the atomic hierarchy diagram is defined as G. a =(X a E a ,R a ), where the set A set of atoms, where each atom is represented by its one-hot encoding of its atomic number. Chemical bonds within a molecule, including single bonds, double bonds, triple bonds, and aromatic bonds, are collectively referred to as bonds. Represents three-dimensional atomic coordinates;

[0013] For a molecule composed of k motifs, its motif hierarchy diagram Gm =(X m E m ,R m ), element set Represents all radixes, where each node is one-hot encoded using a radix dictionary, and the set... An edge is defined between two motifs when they share at least one atom within the molecule. (Set) The coordinates of the motif nodes are represented by the average of the coordinates of all atoms within the pattern. Ultimately, the molecule is described as a bigraph structure G. Dual =(G a G m ).

[0014] Step 2, in the molecular representation task, involves representing the molecular structure as a geometric graph, with atoms as nodes and chemical bonds as edges. The process for generating a static molecular representation is as follows: The static feature encoder first generates node embeddings by aggregating node content and geometric information derived from the relative positions of connected nodes through an embedding layer, where the node content x... i Obtained through linear layer mapping Geometric information of node i The relative position vectors between adjacent nodes j connected by edges within their neighborhood. The content mapping and geometric information of the nodes are added together to obtain the node embedding information of the static feature encoder. The node embedding information is fed into the isomorphic graph attention mechanism, where it is first fused with the embeddings of neighboring nodes to generate a hybrid feature f. ij Based on the Wigner-Eckart theorem, the transmitted message is decomposed into invariant scalars. With Euclidean vectors Invariant scalars are used to calculate attention weights, while Euclidean vectors are used to process the geometric information being propagated. Attention is updated through multi-layer isomorphic graphs, and static structural representations are obtained at the atomic, primitive, and molecular levels through multi-level aggregation methods.

[0015] Step 3 involves introducing continuous filtering convolution, dynamically creating filters through a generator function, directly processing data at arbitrary positions, and finally obtaining updated node features through concatenation and linear transformation.

[0016] Step 4 includes, for a given molecular graph G∈{G a G m A graph with n nodes and corresponding coordinates R is used in a self-supervised learning task. The node content and coordinates in the bi-graph are randomly masked at a fixed ratio to generate masked data, which is then input into a static structure encoder. The indexes of the masked nodes are denoted as U. content with U coordThe content of the mask node is replaced with a special mask marker, and its coordinates are replaced with random values ​​in the range of [-1,1]. The node content reconstruction uses cross-entropy loss for atomic type prediction classification task, while the coordinate reconstruction uses smooth L1 loss for coordinate regression task.

[0017] The beneficial effects of this invention are as follows: First, this invention designs a static feature encoder based on the Transformer architecture with an embedded isomorphic graph attention mechanism, and constructs a dynamic information encoder based on the Transformer architecture with an embedded multi-scale filtering attention mechanism. This can extract the connection rules and structural constraints between molecules, and also mine the dynamic interaction forces between adjacent nodes, thereby improving the comprehensive characterization and efficient extraction of molecular properties.

[0018] II. This invention proposes a molecular property prediction modeling method, MolSGPE, based on a pre-trained architecture that integrates static and dynamic features. By fusing static structural features and dynamic force field information, it significantly improves the generalization ability of molecular property characterization. Faced with the complex changes in molecular activity, it can more deeply mine fine-grained information about molecules, thereby improving prediction accuracy.

[0019] The overall framework diagram of the MolSGPE molecular property prediction modeling method of this invention is as follows: Figure 1 As shown, it has higher detection accuracy and stronger robustness in actual molecular property prediction and functional group analysis tasks, and has excellent performance in predicting multiple types of properties. Attached Figure Description

[0020] Figure 1 This is the overall model framework diagram of MolSGPE;

[0021] Figure 2 This is a structural diagram of the attention module in the isomorphic graph;

[0022] Figure 3 This is a structural diagram of a continuous filtering convolution module;

[0023] Figure 4 This is a diagram of the residual update layer network structure;

[0024] Figure 5 This is a flowchart of the backpropagation process for the loss function;

[0025] Figure 6 It is a pie chart of ROC-AUC scores;

[0026] Figure 7 It is a visualization of the classification performance;

[0027] Figure 8 It is a visualization of the differences in molecular attention weights;

[0028] Figure 9It is an interpretability analysis of the structure-function relationship based on attention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] This invention provides a molecular property prediction modeling method based on a dynamic-static feature fusion pre-training architecture. Through a flexible motif partitioning strategy, a dual-flow graph representation structure encompassing atomic and molecular levels is constructed. A static feature encoder and a dynamic information encoder are designed to mine molecular dynamic interaction information and integrate it into the molecular static structure, thereby generating a more comprehensive and high-quality molecular representation. Based on a fuzzy mechanism, a novel integrated compliance and hyperparameter optimization method is designed to construct a molecular property prediction modeling method based on a dynamic-static feature fusion pre-training architecture, achieving high prediction performance and generalization ability for different molecular properties and interacting functional groups.

[0031] The features of this invention are as follows:

[0032] I. This invention designs a static feature encoder based on the Transformer architecture with an embedded isomorphic graph attention mechanism, and constructs a dynamic information encoder based on the Transformer architecture with an embedded multi-scale filtering attention mechanism. This can extract the connection rules and structural constraints between molecules, and also mine the dynamic interaction forces between adjacent nodes, thereby improving the comprehensive characterization and efficient extraction of molecular properties.

[0033] II. This invention proposes a molecular property prediction modeling method, MolSGPE, based on a pre-trained architecture that integrates static and dynamic features. By fusing static structural features and dynamic force field information, it significantly improves the generalization ability of molecular property characterization. Faced with the complex changes in molecular activity, it can more deeply mine fine-grained information about molecules, thereby improving prediction accuracy.

[0034] The overall framework diagram of the MolSGPE molecular property prediction modeling method of this invention is as follows: Figure 1 As shown, it has higher detection accuracy and stronger robustness in actual molecular property prediction and functional group analysis tasks, and has excellent performance in predicting multiple types of properties.

[0035] A molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture includes the following steps:

[0036] Step 1: Constructing a two-flow graph to characterize the molecular structure

[0037] This invention constructs a motif dictionary and introduces a fuzzy segmentation method. The motif dictionary construction process is as follows: First, all subgraphs are extracted based on interatomic chemical bonds and ring structures. The subgraphs are divided into three categories: ring structures, bonds that do not directly connect ring structures, and bonds that directly connect ring structures. The motifs are sorted using a term frequency-inverse document frequency algorithm, which is based on the frequency of occurrence of the motif in a single molecule and its prevalence in molecules throughout the entire dataset, thereby generating a dictionary index.

[0038] For a molecule composed of n atoms, the atomic hierarchy diagram is defined as G. a =(X a E a ,R a ). Among them, the set This represents a set of atoms, where each atom is represented by its one-hot encoding based on its atomic number. (Set) Chemical bonds within a molecule, including single, double, triple, and aromatic bonds. (Set) Represents three-dimensional atomic coordinates.

[0039] For a molecule composed of k motifs, its motif hierarchy diagram G m =(X m E m ,R m ). Element set Represents all radixes, where each node is one-hot encoded using a radix dictionary. Set An edge is defined between two motifs when they share at least one atom within the molecule. (Set) The coordinates of the motif nodes are represented by the average of the coordinates of all atoms within the pattern. Ultimately, the molecule is described as a bigraph structure G. Dual =(G a G m ).

[0040] Step 2: Design a static feature encoder based on the isomorphic graph attention module;

[0041] In molecular representation tasks, molecular structures are represented as geometric graphs (atoms as nodes and chemical bonds as edges). The traditional Transformer architecture struggles to handle geometric structures existing in a continuous three-dimensional space. Therefore, an equivariant graph attention mechanism is introduced into the Transformer to satisfy physical constraints and ensure that the model output has corresponding equivariance under transformations such as rotation and translation in three-dimensional space. This module can effectively capture the spatial geometric relationships between atoms.

[0042] The process for generating static molecular representations is as follows: The static feature encoder first generates node embeddings by aggregating node content and geometric information derived from the relative positions of connected nodes through an embedding layer. This embedding generation method directly integrates local structural information into the node embeddings, enhancing local structure perception. Wherein, the node content x... i Obtained through linear layer mapping Geometric information of node i The relative position vectors between adjacent nodes j connected by edges within their neighborhood. The content mapping and geometric information of the nodes are added together to obtain the node embedding information of the static feature encoder. The specific formula is as follows:

[0043]

[0044] The node embedding information is fed into the isomorphic graph attention mechanism, the structure of which is as follows: Figure 2 As shown, the feature is first fused with the embeddings of neighboring nodes to generate a hybrid feature. Based on the Wigner-Eckart theorem, the transmitted message is decomposed into invariant scalars. With Euclidean vectors Invariant scalars are used to calculate attention weights, while Euclidean vectors are processed as propagated geometric information. This split feature processing approach ensures that scalar and vector features do not mutually destroy each other, guaranteeing that features have corresponding covariance under transformations such as rotation and translation, while improving processing efficiency. Finally, attention is updated through multi-layer isovariant graphs, and a static structural representation is obtained through a multi-level aggregation method.

[0045] Finally, attention updates are performed using multi-layer isomorphic graphs, and static structural characterizations are obtained at the atomic, elementary, and molecular levels through a multi-level aggregation method. The specific formulas are as follows:

[0046]

[0047] in This indicates that atomic-level diagrams and motif-level diagrams have the same aggregation method. and The atomic-level diagram and the order-level diagram are represented respectively, H static This represents the final output of the static encoder.

[0048] Step 3: Design a dynamic information encoder based on continuous filter convolution;

[0049] Traditional Transformers typically rely on discrete tokens or grid inputs, while atomic positions are continuous coordinates in three-dimensional space. Discretizing these coordinates results in the loss of precise distance and angle information between atoms, disrupting the continuity of the molecular structure. Therefore, continuous filtering convolution is introduced, which dynamically creates filters through a generator function to directly process data at arbitrary positions, avoiding the information loss caused by gridding. Its network structure is as follows: Figure 3 As shown.

[0050] The embedding layer of a dynamic feature encoder is similar to that of a static structure encoder, but instead of relying on chemical bonds when considering neighboring nodes, it considers all nodes in the neighborhood. Finally, the updated node features are obtained through concatenation and linear transformation.

[0051]

[0052] in This represents the characteristics of nodes within a node's neighborhood.

[0053] Perform linear transformations on the features of node i and node j respectively to obtain the matrix. and Filter D K The aim is to calculate the attention weight α using Q and K in a continuous energy field. ij Using the cosine truncation function Ensure that interactions between nodes beyond a distance d are ignored.

[0054]

[0055]

[0056] in This represents the relative distance between node i and node j after radial basis function mapping.

[0057] Filter D V The value vector V is transformed into a continuous energy field. Subsequently, the representation of V is divided into three equal parts. and vector Through attention weight α ij The aggregation involves interactions between nodes, and the aggregation of information from all heads generates an aggregated vector. in and Reserved for updating the direction vector

[0058]

[0059] Where L represents the number of parallel attention heads. The aggregate vector generated by the attention module. It is split into three feature vectors: and and Used to calculate scalar feature update amount Calculate the update amount of vector features.

[0060] Direction vector And the transformation is performed through linear layers U1, U2, and U3, where and Performing a dot product operation converts the equidistant vector features into invariant scalars. The dot product result is used to update the scalar features. The residual update layer network structure is used to update vector features, as follows: Figure 4 As shown.

[0061]

[0062] Node features With direction vector Updates are performed through an iterative loop, which contains updates. and The process involves operations and attention mechanisms to generate node representations. After multiple levels of aggregation, the final output of the dynamic encoder is generated.

[0063]

[0064]

[0065] Where M represents each node in the graph, The L2 norm represents the features of a vector.

[0066] Step 4: Design a comprehensive loss function and hyperparameter optimization method based on fuzzy mechanism;

[0067] For a given molecular graph G∈{G a G m This framework uses a bi-graph representation with n nodes and corresponding coordinates R to apply to a self-supervised learning task. To learn the static structure of molecules, the node content (atom type / motif index) and coordinates (3D position) in the bi-graph are randomly masked at a fixed ratio (15%), generating masked data which is then input into the static structure encoder. The masked node indices are denoted as U... content with U coordThe content of the masked node is replaced with a special mask marker (set to 0), and its coordinates are replaced with random values ​​in the range [-1, 1]. Node content reconstruction uses cross-entropy loss for atomic type prediction in a classification task, while coordinate reconstruction uses smoothed L1 loss for coordinate regression. The specific calculations are as follows:

[0068]

[0069] in This represents the predicted node content. This represents the predicted node coordinates. x i Represents the actual node content, This represents the probability predicted by the model.

[0070] To learn the force field characteristics of the molecular system, noise is introduced to perturb the original molecular structure. This is achieved by adding Gaussian noise ε ~ N(0, σ) to the coordinates of each node. 2 I3) is used to obtain the perturbed structure R'. The dynamic force encoder, after training, can predict and eliminate introduced noise. From a statistical physics perspective, the molecular structure R can be considered as a random quantity drawn from the Boltzmann distribution. Since low-energy structures have a high probability of occurrence, the molecular equilibrium state can be approximated by minimizing the potential energy function E(R), where... This refers to the nodal force field. A denoising loss function is defined based on the scoring matching principle. as follows:

[0071]

[0072] in, It represents the expectation under a given distribution.

[0073] The loss function for the pre-training phase is defined as follows:

[0074]

[0075] Among them, γ1, γ2 and γ3 are weights that are dynamically adjusted using the Bayesian optimization method.

[0076] Once the fine-tuning phase begins, Based on the addition of property prediction loss Losses during the fine-tuning phase:

[0077]

[0078] Where S represents the total number of molecules and C represents the number of tags. i Indicates the true label, This represents the probability that molecule i, predicted by the model, belongs to the class c. The specific loss function and backpropagation process are as follows: Figure 5 As shown.

[0079] Bayesian optimization is employed to optimize the model's hyperparameters. This method builds the model based on historical data, evaluates the performance of the hyperparameters, and then selects new hyperparameters for testing. This process is repeated continuously, dynamically adjusting the search direction through a probabilistic model. This allows for a more efficient approximation of the global optimum within a finite number of iterations, reducing the time cost and computational complexity of hyperparameter tuning.

[0080] Example:

[0081] To evaluate the effectiveness of the MolSGPE proposed in this invention, three molecular property prediction datasets (BBBP, ToxCast, and Sider) and a flavor molecule dataset were used as downstream tasks. Each dataset contains one or more properties. If a molecule possesses a certain property, it is labeled as 1; otherwise, it is labeled as 0.

[0082] This invention preprocesses the molecular graph by first traversing all molecules to generate a motif dictionary, and then further generating a motif-level graph based on the motif dictionary. Masked data and noisy data are generated according to a pre-designed training strategy, and the data are divided into training, validation, and test sets in an 8:1:1 ratio.

[0083] This experiment used the ZINC compound database for pre-training and fine-tuning on the BACE, BBBP, and ClinTox datasets. The hardware environment was as follows: Processor: Intel(R) Core i7-14700K@5.60GHz; Fan: Deepcool Captain 240EX; Motherboard: ASUS X99-A II; Memory: G.SKILL Trident Z Royal, 32GB DDR5 6000; SSD: Samsung 990PRO PCIe 4.0 NVMe. TM M.2 SSD, Graphics Card: RTX 4090*2, Case: Obsidian 750D, Power Supply: Seasonic 1200W Gold VERTEX; Software Environment: GPU computing architecture: CUDA 12.1, cuDNN 8.5.4, Deep learning framework: PyTorch 2.3.0, Python 3.9.

[0084] To demonstrate the detection performance of MolSGPE, it was compared with several models, including the non-pretrained models D-MPNN and AttentiveFP, the 3D-based models GEM and Uni-Mol, and the multi-view pretraining methods GraphMVP, MoleculeSDE, and MolGT. The evaluation metrics used in the experiments included accuracy, recall, F1 score, and ROC-AUC. Accuracy, representing the proportion of correct predictions made by the model, is expressed as:

[0085]

[0086] Recall is the percentage of correctly predicted true positive samples out of all true positive samples. It reflects the proportion of positive samples correctly predicted by the model out of the total number of positive samples. The expression is:

[0087]

[0088] The F1 score considers both the accuracy of positive sample predictions and the model's coverage capability; its expression is:

[0089]

[0090] Where TP is the number of true positives, FP is the number of false positives, and FN is the number of false negatives.

[0091] The ROC-AUC response model comprehensively distinguishes between positive and negative samples at different classification thresholds, is relatively insensitive to class imbalance, and provides a stable evaluation standard.

[0092]

[0093] Where N ps N represents the number of positive samples. ns This represents the number of negative samples. I(·) is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. P positive P represents the probability of correctly identifying a positive sample. negative This indicates the probability of misclassifying a negative sample as positive.

[0094] Regarding the experimental parameter settings, the batch size was set to 16, and the training epochs were set to 300. The experimental results are shown in Table 1.

[0095] Table 1. Comparison of experimental results for molecular property prediction datasets

[0096]

[0097] As shown in Table 1, MolSGPE achieved the best performance on all three datasets compared to several prediction methods. Additionally, as... Figure 6 As shown, a pie chart was plotted for ROC-AUC to more intuitively demonstrate the model's performance. It can be seen that the MolSGPE method also outperforms other models in this metric, proving the superior performance of MolSGPE in molecular property prediction tasks.

[0098] To further demonstrate the effectiveness of this method for molecular representation learning, the molecular representations learned from three benchmark datasets are projected onto a two-dimensional space, such as... Figure 7 As shown, blue dots represent molecules possessing this property, while orange dots represent molecules lacking this property. This figure clearly demonstrates the efficient ability to distinguish between the two molecular types within the MolSGPE framework, proving that this method can effectively learn molecular features to generate high-quality molecular representations.

[0099] Multiple ablation experiments were conducted to evaluate the performance of MolSGPE, including removing the static feature encoder (w / o static), removing the dynamic feature encoder (w / o dynamic), and the complete model (MolSGPE). The specific experimental results are shown in Table 2. After removing the static or dynamic feature encoder, the model performance decreased significantly, proving that both modules have a positive effect on prediction accuracy.

[0100] Table 2 Ablation Experiment Results

[0101]

[0102]

[0103] To further explore the differences between the MolSGPE static feature encoder and the dynamic feature encoder in terms of molecular focus and feature extraction, this invention visualizes the attention weights of the final layer of the model using different datasets (e.g., Figure 8 (As shown). Taking the molecular correlation analysis of benzene and pyrazine as an example, the results show that the two encoders have significantly different focuses on the information related to the same molecule. Specifically, Figure 8 (a) and Figure 7 (b) illustrates the difference in attention between molecules in the ToxCast and Tox21 datasets: the attention weights of the static feature encoder are more significant along the diagonal and in its vicinity, indicating that it primarily focuses on the atoms themselves and their direct connections. In contrast, the attention weights of the dynamic feature encoder suggest that this module plays an active role in mining key regions around molecular atoms, indicating that the dynamic feature encoder focuses more on capturing interactions between neighboring atoms.

[0104] Table 3. Experimental results of the flavor dataset

[0105]

[0106] An additional flavor dataset was used for validation, and Table 3 shows the experimental results for the flavor dataset. The MolSGPE model demonstrated good and balanced performance across all flavor recognition tasks. Specifically, the model performed best in recognizing sweetness, achieving a high ROC-AUC score of 89.64%; however, it performed relatively weaker in recognizing nutty flavors, with a ROC-AUC score of only 77.81%.

[0107] To delve deeper into the contribution of important motifs of flavor molecules to flavor characteristics, attention scores are used to differentiate the degree of attention given to different motifs, such as... Figure 9 Attention scores for motifs in four flavor molecules are presented. All four examples are from datasets. The first example is spicy, where the furan ring and thioether group show high attention values, followed by the ester group. Previous research has shown that capsaicin analogs typically have a vanillin group (head), an amide or ester group (neck), and a fatty acid chain (tail). In the second example, thiols and carbonyl groups show higher attention. These two types of functional groups work together to create a fruity flavor. In the third example, carboxyl, ester, and peptide bonds show higher attention. The carbonyl oxygen atom (C=O) acts as a hydrogen bond acceptor, matching the hydrogen bond network of the sweet taste acceptor T1R3, which is crucial for sweet taste recognition. In the fourth example, the carbonyl, ether, and hydroxyl groups of the molecule are more noteworthy. This is because the synergistic effect between the hydroxyl and carbonyl groups activates bitterness perception.

[0108] Results Analysis: The experimental results from the implementation cases demonstrate that the MolSGPE molecular property prediction method with multiple dual-flow feature fusion proposed in this invention can improve the accuracy of property prediction and uncover important motifs with greater contributions within molecules, making it suitable for the preliminary screening of candidate compounds in actual drug development. The implementation cases also clearly demonstrate the innovation, correctness, and indispensability of this invention. Therefore, the molecular property prediction method for drug development proposed in this invention has certain research significance and engineering application value.

Claims

1. A molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture, characterized in that, Includes the following steps: Step 1: Based on the fuzzy partitioning method, each molecule is decomposed into the smallest atomic and bond links and the smallest ring structure, and a two-flow graph characterization structure of the molecule is constructed. Step 2: Based on the Transformer architecture, embed the equivariant graph attention mechanism to design a static feature encoder to extract the connection rules and structural constraints between molecules; Step 3: Based on the Transformer architecture, embed a multi-scale filtering attention mechanism to construct a dynamic information encoder, and learn molecular representation by capturing the changes in orientation vectors and scalar features between adjacent nodes induced by the rotation of the node center; Step 4: Based on the fuzzy mechanism, design a method that integrates the compliance and loss functions and hyperparameter optimization, and construct a molecular property prediction modeling method based on the dynamic and static feature fusion pre-training architecture to predict molecular properties and functional groups; Step 5: The pre-training dataset is from the ZINC public dataset. The pre-training dataset is screened and deduplicated to retain non-repeating small molecules. The downstream task data comes from BBBP, ToxCast and Sider. An additional flavor molecule dataset is constructed as an extension task. A joint pre-training strategy of mask reconstruction-denoising is designed to complete the model pre-training process and obtain the molecular property prediction model. Fine-tuning is then performed on the downstream task.

2. The molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture as described in claim 1, characterized in that: Step 1 includes constructing a motif dictionary and introducing a fuzzy segmentation method. The motif dictionary construction process is as follows: First, all subgraphs are extracted based on interatomic chemical bonds and ring structures. The subgraphs are divided into three categories: ring structures, bonds that are not directly connected to ring structures, and bonds that are directly connected to ring structures. The motifs are sorted using a term frequency-inverse document frequency algorithm to generate a dictionary index. For a molecule composed of n atoms, the atomic hierarchy diagram is defined as G. a =(X a E a ,R a ), where the set A set of atoms, where each atom is represented by its one-hot encoding of its atomic number. Chemical bonds within a molecule, including single bonds, double bonds, triple bonds, and aromatic bonds, are collectively referred to as bonds. Represents three-dimensional atomic coordinates; For a molecule composed of k motifs, its motif hierarchy diagram G m =(X m E m ,R m ), element set Represents all radixes, where each node is one-hot encoded using a radix dictionary, and the set... An edge is defined between two motifs when they share at least one atom within the molecule. (Set) The coordinates of the motif nodes are represented by the average of the coordinates of all atoms within the pattern. Ultimately, the molecule is described as a bigraph structure G. Dual =(G a G m ).

3. The molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture as described in claim 1, characterized in that: Step 2, in the molecular representation task, involves representing the molecular structure as a geometric graph, with atoms as nodes and chemical bonds as edges. The process for generating a static molecular representation is as follows: The static feature encoder first generates node embeddings by aggregating node content and geometric information derived from the relative positions of connected nodes through an embedding layer, where the node content x... i Obtained through linear layer mapping Geometric information of node i The relative position vectors between adjacent nodes j connected by edges within their neighborhood. The content mapping and geometric information of the nodes are added together to obtain the node embedding information of the static feature encoder. The node embedding information is fed into the isomorphic graph attention mechanism, where it is first fused with the embeddings of neighboring nodes to generate a hybrid feature f. ij Based on the Wigner-Eckart theorem, the transmitted message is decomposed into invariant scalars. With Euclidean vectors Invariant scalars are used to calculate attention weights, while Euclidean vectors are used to process the geometric information being propagated. Attention is updated through multi-layer isomorphic graphs, and static structural representations are obtained at the atomic, primitive, and molecular levels through multi-level aggregation methods.

4. The molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture as described in claim 1, characterized in that: Step 3 involves introducing continuous filtering convolution, dynamically creating filters through a generator function, directly processing data at arbitrary positions, and finally obtaining updated node features through concatenation and linear transformation.

5. The molecular property prediction modeling method based on a dynamic and static feature fusion pre-training architecture as described in claim 1, characterized in that: Step 4 includes, for a given molecular graph G∈{G a G m A graph with n nodes and corresponding coordinates R is used in a self-supervised learning task. The node content and coordinates in the bi-graph are randomly masked at a fixed ratio to generate masked data, which is then input into a static structure encoder. The indexes of the masked nodes are denoted as U. content with U coord The content of the mask node is replaced with a special mask marker, and its coordinates are replaced with random values ​​in the range of [-1,1]. The node content reconstruction uses cross-entropy loss for atomic type prediction classification task, while the coordinate reconstruction uses smooth L1 loss for coordinate regression task.

Images