A Multimodal Molecular Property Prediction Method Based on Prompt Graph Convolution

Abstract

The present invention belongs to the technical field of artificial intelligence-assisted drug design, and specifically relates to a multimodal molecular property prediction method based on prompt word graph convolution, comprising the following steps: data set selection and establishment: selecting a variety of molecular properties to construct an initial data set for the multimodal method, and dividing the initial data set into a training set, a validation set, and a test set, wherein the initial data set contains a variety of molecular properties and a plurality of molecules corresponding thereto; data preprocessing: the method constructs molecular property prediction. The present invention provides a new molecular property prediction method, which fuses multimodal molecular features through potential associations and highlights the importance of key structural parts of molecules. It captures the characteristics of molecules from multiple angles and improves the accuracy and reliability of predictions. At the same time, by combining information from multiple modalities, the method can learn richer and more comprehensive feature representations, thereby having better generalization capabilities on different data sets and tasks.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence-assisted drug design, and specifically relates to a multimodal molecular property prediction method based on cue word graph convolution. Background Art

[0002] Drug development is a lengthy, costly, and complex process. Molecular property prediction is a fundamental and crucial task in drug development. Early determination of molecular properties, such as bioactivity and toxicity, can accelerate drug development, minimize failures in later experiments due to property issues, and reduce the number and cost of experiments. Traditional methods of determining molecular properties through experiments and theoretical calculations, while accurate, are time-consuming and costly. With the advancement of artificial intelligence (AI), applying it to molecular property prediction to improve accuracy has become a key research direction, significantly reducing both the time and cost of prediction.

[0003] Currently, there are two main approaches to AI-assisted drug design: machine learning and deep learning. Both approaches face a critical challenge: how to select the most important descriptors associated with the target property from a large number of predefined and computable molecular descriptors. This is crucial not only for the accuracy of the method but also for its interpretability. Machine learning methods rely on the experience of experts and are somewhat subjective. Deep learning methods, on the other hand, delegate this task to neural networks, extracting the most valuable features from the raw input data and more accurately identifying the relationship between structure and properties. Most existing methods are based on a single molecular modality. However, information from a single modality alone cannot fully reveal the complex properties of a molecule. Furthermore, currently, no image readout method is designed specifically for molecular properties, and these methods tend to overlook some of the molecule's inherent chemical information. Therefore, it is necessary to provide a new property prediction method to address this issue. Summary of the Invention

[0004] In response to the technical problem that there is currently no graph readout method designed for molecular properties, the present invention provides a multimodal molecular property prediction method based on cue word graph convolution.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A multimodal molecular property prediction method based on cue word graph convolution includes the following steps:

[0007] S1. Dataset selection and establishment: Select multiple molecular attributes to construct an initial dataset for the multimodal method, and divide the initial dataset into a training set, a validation set, and a test set. The initial dataset contains multiple molecular attributes and their corresponding molecules.

[0008] S2. Data preprocessing: Perform graph embedding and preliminary chemical feature calculation operations on each molecule in the training set of the initial dataset to obtain features corresponding to two dimensions for each molecule; combine the molecules and their corresponding graph and sequence features to obtain the training set of the target dataset for training the multimodal method;

[0009] S3. Method construction: constructing a multimodal method, wherein the multimodal method includes a multimodal fusion module, a feature extraction module, and a molecular graph readout module connected in sequence; training the multimodal method using a training set in the target dataset, and verifying that the trained method is qualified using a validation set and a test set, thereby obtaining a trained multimodal method;

[0010] S4. Molecular property prediction: The molecules whose properties are to be determined are input into the trained multimodal method for property prediction, and the multimodal method outputs the molecular properties to be predicted.

[0011] The method for performing graph embedding and preliminary chemical feature calculation operations on each molecule in S2 is: using the RDKit extension package to calculate the chemical features of each molecule in the training set of the initial data set, and embedding them in the form of graphs and sequences.

[0012] The method for calculating the chemical characteristics of each molecule in the training set of the initial data set is:

[0013] S21. Use the RDKit extension package to calculate the chiral and aromatic chemical features of the molecule and embed them as node and bond features to obtain the graph features of the molecule.

[0014] S22. Use the RDKit extension package to calculate the MACCS fingerprint, PubChem fingerprint, pharmacophore fingerprint and Morgan fingerprint of the molecule to obtain the sequence characteristics of the molecule.

[0015] The multimodal molecular property prediction method based on the cue word graph convolution in S3 is:

[0016] S31, Multimodal Fusion Module: By decoding the one-dimensional molecular fingerprint information, the substructure features of the molecule are extracted; then, these filtered substructure features are mapped onto the molecular graph to achieve the effective fusion of the one-dimensional sequence features and the two-dimensional molecular graph features;

[0017] S32, Feature Extraction Module: After fusing multi-dimensional features, a graph convolutional neural network is used to extract features. A gating mechanism and residual connections are added to control the flow of information, allowing the network to better focus on important information in the nodes.

[0018] S33. Molecular graph readout module: The present invention chooses to construct and output a complete molecular graph based on the elements contained in the molecule, and designs a convolution-based readout module; in the readout module, the molecular graph information processed by graph convolution is extracted and fused into a one-dimensional representation through a three-layer convolutional network.

[0019] The method for achieving effective fusion of one-dimensional sequence features and two-dimensional molecular graph features in S31 is:

[0020] S311. Initialize the molecular graph as an undirected graph, where atoms are represented as nodes and chemical bonds are represented as edges. Each node in the graph is represented by six atomic features, and each edge is represented by four bond features. Most of the features used to represent nodes and edges are directly calculated using RDKit. For ease of processing, these features are encoded using one-hot encoding.

[0021] S312. Using a dictionary-based Morgan fingerprint to extract substructure information from the molecule, after obtaining the dictionary containing the substructure information, a hash function in RDKit is used to map the substructure fragments back to the original molecule. To ensure the accuracy of the extracted information, only uniquely mapped substructure information is mapped back to the original molecule.

[0022] S313. Add a predefined value to the last-dimensional embedding feature of the corresponding atom and chemical bond to indicate that the atom and chemical bond are located in a substructure fragment.

[0023] The method for extracting and fusing the molecular graph information processed by graph convolution into a one-dimensional representation through a three-layer convolutional network in S33 is:

[0024] S331. Read out the molecular graph obtained by graph convolution processing according to element categories to obtain a molecular feature graph classified by elements;

[0025] S332. Use dilated convolution to perform convolution operations on each molecular feature map. The operation process is divided into three layers. The number of convolutions in each layer is inversely proportional to the number of layers. The final representation that integrates all molecular features is obtained:

[0026]

[0027] Among them, n is the number of convolutions in each layer, C i is the output of each convolutional layer, C j is the output of the jth convolution block in each layer. By using multiple identical convolutional units for feature extraction and then averaging, the model's dependence on a single convolution block is reduced, thereby increasing the model's stability. At the same time, each convolution block can extract different features from the data, making the extracted features more comprehensive and rich.

[0028] The classification method was evaluated using the area under the ROC curve (AUC), and the prediction method was evaluated using the root mean square error (RMSE).

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention provides a new method for predicting molecular properties, which integrates multimodal molecular features through potential associations and highlights the importance of key structural parts of molecules. It captures the characteristics of molecules from multiple angles and improves the accuracy and reliability of predictions. At the same time, by combining information from multiple modalities, the method can learn richer and more comprehensive feature representations, thereby having better generalization capabilities on different data sets and tasks. In addition, the present invention also designs a three-layer dilated convolution in the graph readout module based on the molecular properties. This readout method can more effectively integrate the molecular node information processed by the graph neural network. These strategies enable deep learning methods to focus on important atomic information in molecules, which not only helps to understand complex chemical structures, but also can effectively improve the accuracy of molecular property predictions, thereby promoting the drug development process. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can, without inventive effort, derive other implementation drawings based on the provided drawings.

[0032] The structures, proportions, sizes, etc. illustrated in this specification are intended solely to complement the contents disclosed herein and to facilitate understanding and reading by persons skilled in the art. They are not intended to limit the conditions under which the present invention may be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportions, or adjustments in sizes, without affecting the efficacy and objectives of the present invention, shall remain within the scope of the technical contents disclosed herein.

[0033] Figure 1 Schematic diagram of the prediction method flow in an embodiment of the present invention;

[0034] Figure 2 Schematic diagram of the prediction method structure of an embodiment of the present invention.

[0035] Figure 3 This is a schematic diagram of the fusion module structure of an embodiment of the present invention.

[0036] Figure 4 Schematic diagram of the feature extraction module structure of an embodiment of the present invention.

[0037] Figure 5 Schematic diagram of the attention mechanism module structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of this application, not all the embodiments. These descriptions are only to further illustrate the features and advantages of the present invention, rather than to limit the claims of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0039] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following embodiments are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0040] The purpose of this embodiment is to propose a new multimodal molecular property prediction method based on prompt word graph convolution, which effectively solves the problem that existing methods do not fully consider the potential correlation between molecular features, the characteristics of molecular graphs, and other factors, and realize the prediction or classification of molecular properties. Figure 1-5 As shown, the program includes:

[0041] S1. Dataset selection and establishment: Select multiple molecular attributes to construct an initial dataset for the multimodal method, and divide the initial dataset into a training set, a validation set, and a test set. The initial dataset contains multiple molecular attributes and their corresponding molecules.

[0042] S2, data preprocessing: perform graph embedding and preliminary chemical feature calculation operations on each molecule in the training set of the initial dataset to obtain features corresponding to two dimensions for each molecule; combine the molecules and their corresponding graph and sequence features to obtain the training set of the target dataset for training the multimodal method.

[0043] In a specific embodiment, for ease of description, feature calculation and multimodal embedding are performed for any molecule in the training set of the initial dataset. On this basis, when using the RDKit extension package to calculate the chemical features of each molecule in the training set of the initial dataset and embedding them in the form of graphs and sequences, the S21 may include the following steps S211 and S212:

[0044] S211 uses the RDKit extension package to calculate chemical features such as chirality and aromaticity of the molecule, as shown in Table 1. Atoms are represented as nodes, and chemical bonds are represented as edges connecting the nodes, thus embedding the molecule in the form of a graph. The initial embedding features of atoms are 28-dimensional, and the initial embedding features of chemical bonds are 7-dimensional, which are the graph features of the molecule.

[0045] S212, using the RDKit extension package to calculate the MACCS fingerprint, PubChem fingerprint, pharmacophore fingerprint (Pharmacophore ErG fingerprint) and Morgan fingerprint of the molecule, which are sequence characteristics of the molecule.

[0046] S3, Method Construction: Construct a multimodal method, comprising a multimodal fusion module, a feature extraction module, and a molecular graph readout module, connected in sequence. The multimodal method is trained using a training set from the target dataset. The trained method is then validated using a validation set and a test set, resulting in a trained multimodal method.

[0047] Specifically, when the multimodal method is trained using the training set in the target dataset, the multimodal method is trained using a plurality of molecules in the training set of the target dataset and their corresponding graph and sequence features.

[0048] Furthermore, when S3 uses the training set in the target dataset to train the multimodal method, it may include the following steps S31 to S33:

[0049] S31, Multimodal Fusion Module: This embodiment designs an algorithm to fuse the fingerprint and graph features of each molecule. By decoding the one-dimensional molecular fingerprint information, the substructure features of the molecule can be extracted. These filtered substructure features are then mapped onto the molecular graph, achieving an effective fusion of the one-dimensional sequence features and the two-dimensional molecular graph features.

[0050] Furthermore, when S31 is specifically implemented, it includes the following steps S311 to S313:

[0051] In step S311, the molecular graph is initialized as an undirected graph, with atoms represented as nodes and chemical bonds as edges. As shown in Table 1, each node in the graph is represented by six atomic features, and each edge is represented by four bond features. Most of the features used to represent nodes and edges are directly calculated using RDKit. For ease of processing, these features are encoded using one-hot encoding.

[0052] S312 uses a dictionary-based Morgan fingerprint to extract substructure information from the molecule. After obtaining the dictionary containing the substructure information, the hash function in RDKit is used to map the substructure fragments back to the original molecule. To ensure the accuracy of the extracted information, only uniquely mapped substructure information is mapped back to the original molecule.

[0053] S313, adding a predefined value to the last-dimensional embedding feature of the corresponding atom and chemical bond to indicate that the atom and chemical bond are located in a substructure fragment.

[0054] Table 1 Description of atomic and chemical bond embedding characteristics

[0055]

[0056]

[0057] S32, Feature Extraction Module: After fusing the multimodal features, this embodiment uses a graph convolutional neural network to extract features. To avoid the problems of gradient vanishing and gradient exploding during network training, a residual connection module is added to each layer of the graph convolutional neural network. This allows the node representation learned in the previous layer to be fused with the new representation of the current layer to retain the important features of the previous layer. At the same time, this method also uses a gating mechanism to control the flow of information, allowing the network to better focus on the important information in the node.

[0058] S33, Molecular Graph Readout Module: This embodiment constructs and outputs a complete molecular graph based on the elements contained in the molecule, and designs a convolution-based readout module. In the readout module, the molecular graph information processed by graph convolution is extracted and fused into a one-dimensional representation through a three-layer convolutional network.

[0059] Furthermore, when the multimodal method is trained using the training set in the target data set, S33 includes the following steps S331 and S332:

[0060] S331, reading out the molecular graph obtained by the graph convolution process according to element categories, thereby obtaining a molecular feature graph classified by elements;

[0061] S332, use dilated convolution to perform convolution operation on each molecular feature map. The operation process is divided into three layers. The number of convolutions in each layer is inversely proportional to the number of layers. The final representation that integrates all molecular features is obtained. The formula table can be used as follows:

[0062]

[0063] Among them, n is the number of convolutions in each layer, C i is the output of each convolutional layer, Cj is the output of the jth convolution block in each layer. By using multiple identical convolutional units for feature extraction and then averaging the results, the model's reliance on a single convolutional block can be reduced, thereby increasing model stability. At the same time, each convolutional block can extract different features from the data, making the extracted features more comprehensive and richer.

[0064] S4, molecular property prediction: The molecule whose properties are to be determined is input into the trained multimodal method for property prediction, and the multimodal method outputs the molecular properties to be predicted.

[0065] During the training phase, the method uses extensive data to learn the relationships between molecular features and properties. During validation or practical application, the addition of prompts, consisting of fingerprints, can guide the method to more accurately understand the requirements of specific tasks and generate outputs that better meet expectations. These prompts are composed of structural MACCS fingerprints, pharmacophore fingerprints, and PubChem fingerprints.

[0066] Furthermore, the classification method of this embodiment is evaluated using the area under the ROC curve AUC, and the higher the index, the better the effect; the prediction method is evaluated using the root mean square error RMSE, and the lower the index, the better the effect. The results are shown in Table 2.

[0067] Table 2 Results of this embodiment on the data set

[0068]

[0069] GCN stands for graph convolution, GAT for graph attention convolution, and MPNN for information transfer. All three are deep learning methods, and GMP is the abbreviation for the method of this embodiment. It can be seen that compared with existing methods, the property prediction and classification method proposed in this embodiment has the best accuracy in most cases, demonstrating that the molecular property prediction method proposed in this embodiment has relatively good classification and prediction capabilities.

[0070] The above only describes in detail the preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the purpose of the present invention, and various changes should be included in the scope of protection of the present invention.

Claims

1. A multimodal molecular property prediction method based on prompt word graph convolution, characterized in that: The following steps are involved: S1. Dataset selection and establishment: Select multiple molecular attributes to construct an initial dataset for the multimodal method, and divide the initial dataset into a training set, a validation set, and a test set. The initial dataset contains multiple molecular attributes and their corresponding molecules. S2. Data preprocessing: Perform graph embedding and preliminary chemical feature calculation operations on each molecule in the training set of the initial dataset to obtain features corresponding to two dimensions for each molecule; Combining molecules and their corresponding graph and sequence features to obtain a training set of target datasets for training multimodal methods; S3. Method construction: constructing a multimodal method, wherein the multimodal method includes a multimodal fusion module, a feature extraction module, and a molecular graph readout module connected in sequence; training the multimodal method using a training set in the target dataset, and verifying that the trained method is qualified using a validation set and a test set, thereby obtaining a trained multimodal method; S31, Multimodal Fusion Module: By decoding the one-dimensional molecular fingerprint information, the substructure features of the molecule are extracted; then, these filtered substructure features are mapped onto the molecular graph to achieve the effective fusion of the one-dimensional sequence features and the two-dimensional molecular graph features; S32, Feature Extraction Module: After fusing multi-dimensional features, a graph convolutional neural network is used to extract features. A gating mechanism and residual connections are added to control the flow of information, allowing the network to better focus on important information in the nodes. S33, Molecular Graph Readout Module: We construct and output a complete molecular graph based on the elements contained in the molecule, and design a convolution-based readout module. In the readout module, the molecular graph information processed by graph convolution is extracted and fused into a one-dimensional representation through a three-layer convolutional network. S4. Molecular property prediction: The molecule whose properties are to be determined is input into the trained multimodal method for property prediction, and the multimodal method outputs the molecular properties to be predicted; During the training phase, the multimodal method learns the relationship between molecular features and properties through large amounts of data. During verification or practical application, prompt words consisting of fingerprint patterns are added to guide the method to more accurately understand the requirements of specific tasks and thus generate outputs that are more in line with expectations; the prompt words are composed of structural MACCS fingerprints, pharmacophore fingerprints and PubChem fingerprints.

2. A multimodal molecular property prediction method based on prompt word graph convolution according to claim 1, characterized in that: The method for performing graph embedding and preliminary chemical feature calculation operations on each molecule in S2 is: using the RDKit extension package to calculate the chemical features of each molecule in the training set of the initial data set, and embedding them in the form of graphs and sequences.

3. A multimodal molecular property prediction method based on prompt word graph convolution according to claim 2, characterized in that: The method for calculating the chemical characteristics of each molecule in the training set of the initial data set is: S21. Use the RDKit extension package to calculate the chiral and aromatic chemical features of the molecule and embed them as node and bond features to obtain the graph features of the molecule. S22. Use the RDKit extension package to calculate the MACCS fingerprint, PubChem fingerprint, pharmacophore fingerprint and Morgan fingerprint of the molecule to obtain the sequence characteristics of the molecule.

4. The multimodal molecular property prediction method based on prompt word graph convolution according to claim 1, characterized in that: The method for achieving effective fusion of one-dimensional sequence features and two-dimensional molecular graph features in S31 is: S311. Initialize the molecular graph as an undirected graph, in which atoms are represented as nodes and chemical bonds are represented as edges; in the graph, each node is represented by six atomic features and each edge is represented by four bond features; Most of the features used to represent nodes and edges are calculated directly using RDKit. For ease of processing, these features are encoded using one-hot encoding. S312. Using a dictionary-based Morgan fingerprint to extract substructure information from the molecule, after obtaining the dictionary containing the substructure information, a hash function in RDKit is used to map the substructure fragments back to the original molecule. To ensure the accuracy of the extracted information, only uniquely mapped substructure information is mapped back to the original molecule. S313. Add a predefined value to the last-dimensional embedding feature of the corresponding atom and chemical bond to indicate that the atom and chemical bond are located in a substructure fragment.

5. The multimodal molecular property prediction method based on prompt word graph convolution according to claim 1, characterized in that: The method for extracting and fusing the molecular graph information processed by graph convolution into a one-dimensional representation through a three-layer convolutional network in S33 is: S331. Read out the molecular graph obtained by graph convolution processing according to element categories to obtain a molecular feature graph classified by elements; S332. Use dilated convolution to perform convolution operations on each molecular feature map. The operation process is divided into three layers. The number of convolutions in each layer is inversely proportional to the number of layers. The final representation that integrates all molecular features is obtained: Among them, n is the number of convolutions in each layer, C i is the output of each convolutional layer, C j is the output of the jth convolution block in each layer. By using multiple identical convolutional units for feature extraction and then averaging, the model's dependence on a single convolution block is reduced, thereby increasing the model's stability. At the same time, each convolution block can extract different features from the data, making the extracted features more comprehensive and rich.

6. The multimodal molecular property prediction method based on prompt word graph convolution according to claim 5, characterized in that: The classification method was evaluated using the area under the ROC curve (AUC), and the prediction method was evaluated using the root mean square error (RMSE).

Images