A method and apparatus for predicting molecular properties based on multi-layer graph cue learning

Abstract

This invention discloses a method and apparatus for predicting molecular properties based on multi-level graph cue learning, belonging to the field of cheminformatics. The invention constructs molecular graph data based on molecular structural data and quantizes it. Then, it constructs atomic-level cues based on node features, chemical bond cues and functional group cues based on chemical bond features and functional graphs as structure-level cues, and constructs task-level cues. These cues are integrated into node features to obtain an optimized molecular representation vector, which is then input into a neural network model for molecular property prediction. This invention better predicts molecular properties by considering molecular features at multiple levels.

Description

Technical Field

[0001] This invention belongs to the field of cheminformatics, specifically relating to a method and apparatus for predicting molecular properties based on multi-layer graph cue learning. Background Technology

[0002] Molecular property prediction plays a crucial role in AI-driven drug design and discovery. Accurate property prediction can assess the therapeutic potential of drug candidate compounds and even directly predict whether a compound will receive FDA approval, thus greatly accelerating the drug development process. However, due to the complex structure and diversity of molecules, traditional graph neural network (GNN)-based molecular property prediction methods perform poorly when faced with challenges such as insufficient labeled data and inconsistencies between pre-training and fine-tuning objectives. Specifically, in the pre-training phase, the self-supervised task focuses on distinguishing similar patterns in general patterns to reconstruct the basic molecular map. The fine-tuning process, however, differs from the pre-training objective, striving to minimize task-specific loss guided by high-quality labels. This inconsistency makes it difficult for the fine-tuning process to seamlessly transfer knowledge acquired during pre-training, potentially leading to catastrophic forgetting and thus limiting overall performance.

[0003] To address these challenges, numerous graph-based cueing learning methods have emerged in recent years, attempting to resolve the inconsistency between pre-training and fine-tuning by transforming the fine-tuning task into a pre-training format. However, existing methods neglect the unique properties of molecular graphs, such as the differences between atomic and structural levels and the importance of functional groups. This leads to a series of specific challenges when these methods are applied to molecular graphs, limiting their effectiveness and performance in predicting molecular properties. Summary of the Invention

[0004] The purpose of this invention is to propose a molecular property prediction method and apparatus based on multi-level graph cue learning, which can better predict molecular properties by considering molecular features from multiple levels.

[0005] The technical solution adopted by the present invention to achieve the above objectives is as follows: A molecular property prediction method based on multi-layer graph cue learning includes the following steps: Based on the molecular structure data, construct molecular graph data with atoms as nodes and bonds as edges; The molecular graph data is vectorized to obtain the initial molecular graph representation; Atomic-level cues are constructed based on node features in the initial molecular graph representation, and the atomic-level cues are integrated into the node features to obtain the binding vector; Based on the chemical bond features and functional diagram in the initial molecular diagram representation, chemical bond hints and functional group hints are constructed and together constitute a structure-level hint. The structure-level hint is integrated into the above binding vector to obtain a new binding vector. Task-level hints are constructed to aggregate the new binding vectors and integrate them with functional group hints to obtain the optimized molecular representation vectors; The optimized molecular representation vector is input into a trained neural network model to predict molecular properties, and the predicted molecular properties are obtained.

[0006] Furthermore, before constructing molecular graph data, the molecular structural data needs to be preprocessed to remove duplicate data and handle missing values.

[0007] Furthermore, when constructing atomic-level cues, a set of decomposed cues is built for different attributes in the atom, and specific cues are built for different types of atoms. These two types of cues are then integrated with the original node features to obtain a combined vector.

[0008] Furthermore, integrating atomic-level cues into node features yields the following formula for the combined vector: in, Represents the associative vector. Represents the original node features. Indicates the first j A breakdown hint, This indicates a set of decomposition hints. L Indicates the number of prompts for decomposition. This indicates a specific hint for an atom. Indicates a prompt For the current node i Node features Influence coefficient, express Zhongyu Different other decomposition hints r .

[0009] Furthermore, the chemical bond hints are integrated into the binding vector formula as follows: in, Represents the associative vector. This indicates the similarity weight between chemical bond hints and adjacent atomic nodes. This indicates that chemical bond hints are generated from atomic nodes. i neighboring nodes j Integrate into atomic nodes i The weight parameters, Represents atomic nodes i and neighboring nodes j Attention weights between them Indicates chemical bond hints, k It is an index. K Indicate the number of chemical bonds. Indicates except the first k Other chemical bond hints besides these, This indicates that chemical bond hints will be sent from nodes other than neighboring nodes. j Other i neighboring nodes r Integrate into atomic nodes i The weight parameters, Represents atomic nodes i The set of all neighboring nodes; This represents a linear mapping operation. This represents a vector aggregation operation.

[0010] Furthermore, the functional group cues are integrated into the associative vector formula as follows: in, Represents the associative vector. For the first A functional group suggests, Indicates except the first k Other functional groups besides these provide clues. For containing atomic nodes i The aggregation of functional group hints, atomic nodes i The set of associated functional groups, F Indicates the number of functional group prompts.

[0011] Furthermore, the expression for task-level hints is as follows: in, Represents the associative vector. Indicates aggregation operation, Represented as a molecular diagram, V This represents the set of nodes in the molecular graph. This represents the collection of functional groups appearing in the molecular diagram. This indicates a sensory group suggestion.

[0012] Furthermore, molecular graph data with real labels of molecular properties are used as the training set. The optimized molecular representation vectors of the training set are obtained and then input into the neural network model for training to obtain the trained neural network model.

[0013] Furthermore, the cross-entropy loss function is used to measure the difference between the predicted result and the true label when training the neural network model. The cross-entropy loss function is as follows: in, N Indicates the number of training samples. C Indicates the number of category labels. Indicates training samples i The first of the real labels i Category labels, Indicates training samples i Belongs to the j The predicted probability of each category.

[0014] A molecular property prediction device based on multi-layer graph cue learning includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0015] The beneficial effects achieved by this invention are as follows: 1. This invention employs data preprocessing steps, including data collection, cleaning, molecular graph construction, and attribute extraction, to obtain a high-quality, standardized dataset.

[0016] 2. This invention designs atomic-level and structural-level cue templates, bridging the gap between atomic and structural levels in molecular diagrams. By learning the characteristics of atoms and chemical bonds, it effectively improves the performance of the model.

[0017] 3. This invention designs a task-level prompt template, which solves the inconsistency between the pre-training and fine-tuning processes by integrating functional group information, and further improves the robustness and generalization ability of the model.

[0018] 4. This invention utilizes neural networks for molecular property prediction. Through model training and optimization, it achieves accurate prediction of the properties of new molecules. The method of this invention not only improves the accuracy and efficiency of molecular property prediction but also provides an important tool and method for the fields of molecular design and drug discovery. Attached Figure Description

[0019] Figure 1 This is a flowchart of a molecular property prediction method based on multi-layer graph cue learning according to the present invention. Detailed Implementation

[0020] To make the various technical features, advantages, or effects of the present invention more apparent and understandable, a detailed description is provided below in conjunction with the accompanying drawings.

[0021] This invention specifically discloses a molecular property prediction method based on multi-layer graph cue learning, the process of which is as follows: Figure 1 As shown, the specific steps include: Step 1: Data preprocessing, including key steps such as molecular data collection, data preprocessing, molecular map construction, attribute extraction, and data formatting. 1. Collect samples (molecular structure files) containing molecular data from publicly available large-scale molecular databases or datasets (such as PubChem, ChEMBL, etc.).

[0022] 2. Preprocess the collected data, including removing duplicate data and handling missing values, to ensure data consistency and integrity.

[0023] 3. Parse the atomic and bond information in the molecular structure file into nodes and edges to construct molecular graph data.

[0024] 4. Extract relevant chemical attribute information such as atom type, bond type, and chemical bond length from the molecular diagram data for subsequent prompt template construction.

[0025] 5. Save the constructed molecular map data in a uniform format for easy subsequent processing and analysis.

[0026] 6. Construct a training set based on the saved molecular graph data for model training. This training set includes annotations of molecular properties corresponding to molecule names, serving as true labels. In practice, the molecular graph dataset should not only be divided into a training set as needed, but also a validation set and a test set to test and evaluate the trained model.

[0027] Through this step, this embodiment obtained high-quality, standardized molecular map data, laying the foundation for subsequent feature extraction and property prediction.

[0028] Step 2: Molecular Graph Data Vectorization. For the obtained molecular graph data, the existing pre-trained model is used to vectorize the molecular graph data, resulting in an initial molecular graph representation. The purpose is to represent the molecular graph as a vector with a fixed length, thereby providing a unified data format for subsequent cue learning tasks.

[0029] Step 3: Design atomic-level hint templates to address atomic-level gaps in molecular diagrams.

[0030] Determine the node features in the effective representation molecular graph The method needs to consider the type of atom (such as carbon, nitrogen, oxygen, etc.), the type of bond (single bond, double bond, triple bond, etc.), and other related properties (such as atomic charge, bond length, etc.). Therefore, this embodiment introduces a cue encoding to encode the node features of atoms. Transformed into learnable hints through encoding This embodiment refers to it as an atomic-level hint. This atomic-level hint can be used indiscriminately on every node, ignoring the distinct characteristics of each atomic node in the molecular diagram. In a molecular diagram, different atoms differ significantly. Therefore, to distinguish atom types, this embodiment utilizes not only a set of decomposed hints... This addresses the differences in the properties of various atoms, and also includes targeted, specific prompts designed for different types of atoms. For example, corresponding prompts can be designed for common atom types such as carbon, nitrogen, and oxygen. To capture its characteristics and attributes.

[0031] 1. This embodiment treats node-level differences as feature perturbations and utilizes learnable atomic-level cues. Compensation is provided for each node's features. Atomic-level hints Integrating into the original node features In Chinese, the formula is expressed as follows: in, Representing the original node features and atomic-level hints The associative vector.

[0032] 2. Combine the hints from different atom types into a complete atomic-level hint template to fully reflect atomic-level differences and provide guidance for subsequent feature representation and property prediction.

[0033] For atomic node i, the hinting process can be described as follows: in, Represents the associative vector. Represents the original node features. The first atomic property j A breakdown hint, L Indicates the number of decomposition prompts. This indicates a specific clue about a certain atom. α For adjusting prompts Node features The coefficient of influence, formula (3) is the softmax function, Indicates the relationship with the current atomic node i A set of decomposition hints and hints Different other hints (the first) r (A breakdown hint).

[0034] Through this step, this embodiment successfully resolved the atomic-level gap in the molecular diagram, laying the foundation for improving the accuracy and reliability of property prediction.

[0035] Step 4: Design a structural-level hint template to resolve differences in structure at the molecular level, mainly including processing at the chemical bond level and the motif level.

[0036] 1. At the chemical bond level, learnable hints, referred to as chemical bond hints in this embodiment, are introduced and integrated into the corresponding atomic nodes through a message passing process. Each chemical bond characteristic is represented as a learnable chemical bond hint. , Then, these hints are integrated with adjacent atomic nodes through a message passing mechanism, and finally represented as the atomic node representation. This process can be described as follows: in, Representing the combined vector, formula (4) means Update process, The similarity weight between chemical bond hints and adjacent atomic nodes is used to control how chemical bond hints are integrated into the corresponding atomic nodes; This indicates that chemical bond hints are generated from atomic nodes. i neighboring nodes j Integrate into atomic nodes i The weighting parameters determine the degree of influence of chemical bond hints on neighboring nodes; Represents atomic nodes i and neighboring nodes j Attention weights between nodes are used to compute atomic nodes. i with neighboring nodes j The degree of integration of features between them k It is an index; Indicates except the first k Other chemical bonds besides those mentioned; This indicates that chemical bond hints will be sent from nodes other than neighboring nodes. j Other i neighboring nodes r Integrate into atomic nodesi Weight parameters; Represents atomic nodes i The set of all neighboring nodes; This represents a linear mapping operation, such as in a neural network. This represents vector aggregation operations, such as direct joins and calculating the mean.

[0037] The current steps enable us to better utilize the rich knowledge about chemical bonds in pre-trained models, thereby improving the accuracy and efficiency of molecular property predictions.

[0038] 2. At the basic structural level, functional groups, such as carboxyl and amino groups, are extracted using cheminformatics tools (e.g., RDKit), and corresponding learnable cues are designed for each functional group. , , F The dimension represents the number of functional groups used for cues, which this invention refers to as functional group cues. Its dimension matches the number of molecular properties. When an atom in a molecule belongs to a certain functional group, additional functional group cues are considered to more accurately reflect the properties of that atom. This embodiment employs a cue learning method, that is, for the case of a specific atom in a specific functional group, additional functional group cue vectors are used for cues. This process can be described as follows: in, Representing the associative vector, formula (8) means Update process, For the first Functional group hints within a functional group Indicates except the first k Other functional groups besides these provide clues. For containing atomic nodes i The aggregation of functional group hints, atomic nodes i The set of associated functional groups.

[0039] This step, through hints at the chemical bond level and the basic structure level, allows for a more detailed analysis of the structural features of molecules, thereby improving the accuracy and efficiency of molecular property prediction.

[0040] Step 5: Design a Task-Level Hint Template. Besides addressing the gap between the atomic and structural levels, it's also necessary to consider the inconsistencies between pre-training and fine-tuning processes in traditional pre-training methods. The inventors believe that when functional groups are considered, the gaps in molecular prediction tasks can be naturally adjusted. Therefore, they designed a learnable hint, referred to in this embodiment as a task-level hint template, which aims to integrate functional group information and incorporate it into the molecular graph representation, transforming the fine-tuning task into a pre-training task to resolve this inconsistency.

[0041] Specifically, this embodiment first employs aggregation operations, such as average pooling, to aggregate the vectors of all nodes obtained in the above steps. Polymerization is performed to obtain a complete molecular diagram representation. Then, each functional group will be prompted. Molecular diagram representation By adding these elements, the information about the functional groups can be incorporated into the entire molecular diagram representation, resulting in an optimized molecular representation vector. The task-level hint template is as follows: in, Representing the combined vector, formula (11) represents Update process, Indicates aggregation operation, V This represents the set of nodes in the molecular graph. This is the set of functional groups that appear in the molecular diagram.

[0042] Following these steps, you can address inconsistencies in molecular prediction tasks using task-level prompt templates, thereby improving model performance and accuracy.

[0043] Step Six: Predict Molecular Properties Using a Neural Network Model. After obtaining the molecular representation vectors optimized through cue learning, a neural network model is used to predict molecular properties. Specifically, the representation vector of each molecule is input into a fully connected neural network (e.g., a deep neural network), which includes multiple hidden layers and an output layer. The softmax activation function is used in the output layer to output the probability distribution of each molecule belonging to different categories.

[0044] To train the aforementioned neural network model, the cross-entropy loss function is used to measure the difference between the predicted result and the true label. The parameters of the neural network are updated using the backpropagation algorithm to minimize the loss function. During training, optimization algorithms such as stochastic gradient descent are used to update the parameters, and the learning rate and other hyperparameters can be adjusted as needed. After training, the neural network model can be used to predict the properties of new molecules. Given a molecule with unknown properties, its representation vector is obtained through the above steps and then input into the trained neural network model to predict its molecular property category. By comparing the predicted results with the true label, the model's performance can be evaluated, and the neural network model can be further tuned and improved. The formula for the cross-entropy loss function is as follows: in, N Indicates the number of training samples. C Indicates the number of category labels. Indicates training samples i The first of the real labels i Category labels, Indicates training samples i Belongs to the j The predicted probability of each category.

[0045] Through the steps outlined above, this embodiment successfully established a correlation between the molecular representation vector and its properties using designed prompt templates at various levels, achieving accurate prediction of molecular properties. This method not only improves prediction accuracy but also provides an effective tool for the fields of molecular design and drug discovery.

[0046] Experimental test: This experiment used 2 million molecules sampled from ZINC15 for pre-training. Molecular property prediction and drug selection target affinity (DTA) prediction were used as downstream tasks. The dataset used for molecular property prediction was the public dataset MoleculeNet, which contains eight widely used binary classification datasets. Details of the datasets used are shown in Table 1 below: Table 1 The results of molecular property prediction (classification) are shown in Table 2 below: Table 2 The results of the molecular property prediction (regression) are shown in Table 3 below: Table 3 Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Appropriate modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention should be covered within the protection scope of the present invention, which is defined by the claims.

Claims

1. A method for predicting molecular properties based on multi-layer graph cue learning, characterized in that, Includes the following steps: Based on the molecular structure data, construct molecular graph data with atoms as nodes and bonds as edges; The molecular graph data is vectorized to obtain the initial molecular graph representation; Atomic-level cues are constructed based on node features in the initial molecular graph representation, and the atomic-level cues are integrated into the node features to obtain the binding vector; Based on the chemical bond features and functional diagram in the initial molecular diagram representation, chemical bond hints and functional group hints are constructed and together constitute a structural-level hint. The structural-level hints are then integrated into the aforementioned binding vector to obtain a new binding vector. The formula for integrating chemical bond hints into the binding vector is as follows: in, Represents the associative vector. This indicates the similarity weight between chemical bond hints and adjacent atomic nodes. This indicates that chemical bond hints are generated from atomic nodes. i neighboring nodes j Integrate into atomic nodes i The weight parameters, Represents atomic nodes i and neighboring nodes j Attention weights between them Indicates chemical bond hints, k It is an index. K Indicate the number of chemical bonds. Indicates except the first k Other chemical bond hints besides these, This indicates that chemical bond hints will be sent from nodes other than neighboring nodes. j Other i neighboring nodes r Integrate into atomic nodes i The weight parameters, Represents atomic nodes i The set of all neighboring nodes; This represents a linear mapping operation. This represents a vector aggregation operation; The formula for integrating functional group cues into the associative vector is as follows: in, Represents the associative vector. For the first A functional group suggests, Indicates except the first k Other functional groups besides these provide clues. For containing atomic nodes i The aggregation of functional group hints, atomic nodes i The set of associated functional groups, F Indicates the number of functional group prompts; Task-level hints are constructed to aggregate the new binding vectors and integrate them with functional group hints to obtain the optimized molecular representation vectors; The optimized molecular representation vector is input into a trained neural network model to predict molecular properties, and the predicted molecular properties are obtained.

2. The method as described in claim 1, characterized in that, Before constructing molecular diagram data, the molecular structure data needs to be preprocessed to remove duplicate data and handle missing values.

3. The method as described in claim 1, characterized in that, When constructing atomic-level cues, a set of decomposed cues is built for different attributes in the atom, and specific cues are built for different types of atoms. These two types of cues are then integrated with the original node features to obtain a combined vector.

4. The method as described in claim 3, characterized in that, The formula for integrating atomic-level cues into node features to obtain the combined vector is as follows: in, Represents the associative vector. Represents the original node features. Indicates the first j A breakdown hint, This indicates a set of decomposition hints. L Indicates the number of prompts for decomposition. This indicates a specific hint for an atom. Indicates a prompt For the current node i Node features Influence coefficient, express Zhongyu Different other decomposition hints r .

5. The method as described in claim 1, characterized in that, The expression for task-level hints is as follows: in, Represents the new associative vector. Indicates aggregation operation, Represented as a molecular diagram, V This represents the set of nodes in the molecular graph. This represents the collection of functional groups appearing in the molecular diagram. This indicates a sensory group suggestion.

6. The method as described in claim 1, characterized in that, Molecular graph data with real labels of molecular properties are used as the training set. The optimized molecular representation vectors of the training set are obtained and then input into the neural network model for training to obtain the trained neural network model.

7. The method as described in claim 6, characterized in that, The cross-entropy loss function is used to measure the difference between the predicted result and the true label when training the neural network model. The cross-entropy loss function is as follows: in, N Indicates the number of training samples. C Indicates the number of category labels. Indicates training samples i The first of the real labels i Category labels, Indicates training samples i Belongs to the j The predicted probability of each category.

8. A molecular property prediction device based on multi-layer graph cue learning, characterized in that, It includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method according to any one of claims 1-7.

Images