Compound liposolubility prediction method based on multi-level feature fusion graph neural network

Abstract

The application discloses a compound liposolubility prediction method based on a multi-level feature fusion graph neural network, relates to the field of computational pharmacology and chemical informatics, and comprises the following steps: acquiring an atomic-level graph structure, a molecular-level global feature and a group-level substructure feature of a target molecular compound; performing feature extraction on the atomic-level graph structure to obtain an embedding representation of each atom in the target molecular compound, and performing global average pooling on the embedding representation of each atom to obtain an atomic vector; performing normalization on the molecular-level global feature to form a molecular vector; splicing the atomic vector, the molecular vector and a substructure vector to obtain a global vector; predicting the liposolubility of the target molecular compound according to the global vector; and constructing a multi-level feature system of 'atoms-molecules-substructures', and combining the features with different granularities and different properties organically to perform joint prediction, so that more comprehensive molecular representation and more accurate liposolubility prediction of the molecular compound are realized.

Description

Technical Field

[0001] This application relates to the fields of computational pharmacology and cheminformatics, and in particular to a method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network. Background Technology

[0002] In drug discovery and development, the lipophilicity of a compound (usually expressed as the octanol-water partition coefficient LogP) is a crucial physicochemical parameter. It directly affects the absorption, distribution, metabolism, and excretion (ADME) processes of the compound in vivo, and is one of the key factors determining the druggability of candidate drugs. Therefore, in the early stages of drug development, the ability to quickly and accurately predict the lipophilicity of a compound is of great significance for screening lead compounds with ideal pharmacokinetic properties and reducing development costs and failure rates.

[0003] In recent years, models such as Message Passing Neural Networks (MPNN) and AttentiveFP have demonstrated certain capabilities in the task of predicting lipid solubility. However, existing deep learning-based lipid solubility prediction methods typically only use molecular graph structure information as the data foundation for deep model networks. This prevents deep learning models from comprehensively capturing the chemical factors that affect lipid solubility, making it difficult to fully learn the rules and knowledge that have a clear impact on lipid solubility. Ultimately, this results in low accuracy of the prediction results from deep learning models. Summary of the Invention

[0004] The purpose of this application is to provide a compound lipophilicity prediction method based on a multi-level feature fusion graph neural network. This method comprehensively considers the atomic-level graph structure, molecular-level global features, and group-level substructure features of molecular compounds, constructing a multi-level feature system of "atom-molecule-substructure". Through a feature fusion module, these features of different granularities and properties are organically combined for joint prediction, thereby achieving more comprehensive molecular characterization and more accurate prediction of the lipophilicity of molecular compounds. This solves the problem of low prediction accuracy of existing deep learning-based lipophilicity prediction methods.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] In a first aspect, this application provides a method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network. The graph neural network includes a graph representation learning module, a feature fusion module, and a prediction layer. The method for predicting the lipophilicity of compounds includes:

[0007] The atomic-level diagram structure, molecular-level global features, and group-level substructure features of the target molecular compound are obtained. The nodes and edges of the atomic-level diagram structure represent the atoms and chemical bonds of the target molecular compound, respectively. The molecular-level global features include multiple molecular descriptors. The group-level substructure features include a substructure vector, where each element in the substructure vector characterizes the frequency of occurrence of its corresponding target substructure in the target molecular compound. The target substructure is a substructure that affects the lipophilicity of the compound.

[0008] The graph representation learning module extracts features from the atomic-level graph structure to obtain the embedding representation of each atom in the target molecule compound, and performs global average pooling on the embedding representation of each atom to obtain an atom vector.

[0009] The molecular-level global features are normalized to form molecular vectors;

[0010] The feature fusion module concatenates the atomic vector, the molecular vector, and the substructure vector to obtain a global vector.

[0011] The prediction layer predicts the lipid solubility of the target molecular compound based on the global vector.

[0012] In one embodiment, obtaining the atomic-level structure, molecular-level global features, and group-level substructure features of the target molecular compound specifically includes:

[0013] Obtain the SMILES string of the target molecule compound;

[0014] The atomic-level diagram structure, molecular-level global features, and group-level substructure features of the target molecular compound are determined based on the SMILES string.

[0015] In one embodiment, in the atomic-level graph structure, each atom is encoded as a feature vector containing the atom's type, degree, number of hydrogen atoms, aromaticity marker, formal charge, hybridization state, and mass.

[0016] In one embodiment, in the atomic-level diagram structure, each chemical bond is encoded as a feature vector containing the type of the chemical bond, an aromaticity marker, and an intra-ring marker.

[0017] In one embodiment, the plurality of descriptors include the molecular weight, number of atoms, number of chemical bonds, number of rotatable chemical bonds, number of aromatic rings, number of aliphatic rings, topological polar surface area, number of hydrogen bond donors, number of hydrogen bond acceptors, and the counting characteristics of each element in the target molecular compound.

[0018] In one embodiment, the plurality of descriptors further include the Balaban index, Bertz complexity index, molecular volume, number of heavy atoms, and number of rings of the target molecular compound.

[0019] In one embodiment, the plurality of target substructures include a benzene ring, a hydroxyl oxygen atom, a nitrogen atom, a carboxyl group, a carbonyl group, a sulfur atom, a trifluoromethyl group, a halogen atom, at least one heterocyclic structure, and unsaturated bonds and ether bonds.

[0020] In one embodiment, at least one of the heterocyclic structures includes a pyridine ring, a pyrimidine ring, a furan ring, and a thiophene ring.

[0021] In one embodiment, the graph representation learning module includes two stacked attention message passing layers, wherein the update formula for each node in the attention message passing layer is:

[0022]

[0023]

[0024]

[0025] in, This represents the embedding of the i-th node. m ij This represents the message between the i-th node and the j-th node. This represents the message passing weight between the i-th node and the j-th node. Q i This represents a query for the i-th node. K j This indicates that the j-th node corresponds to Q i The key vector, d Indicates the projection dimension. and These represent the i-th and j-th nodes, respectively. This represents the edge between the i-th and j-th nodes.

[0026] In one embodiment, the loss function of the graph neural network for:

[0027]

[0028] in, This represents the mean squared error loss of the prediction result. This represents the projection layer weights of the substructure vector.

[0029] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the compound lipophilicity prediction method based on a multi-level feature fusion graph neural network as described above.

[0030] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the compound lipophilicity prediction method based on a multi-level feature fusion graph neural network as described above.

[0031] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the compound lipophilicity prediction method based on a multi-level feature fusion graph neural network as described above.

[0032] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0033] This application provides a method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network. The atomic-level graph structure, molecular-level global features, and group-level substructure features of the target molecular compound are used as inputs to the graph neural network. The graph neural network learns the relationship between the lipophilicity of the molecular compound and the atomic-level graph structure, molecular-level global features, and group-level substructure features, and finally predicts the lipophilicity of the target molecular compound.

[0034] Among them, the atomic-level graph structure provides rich local chemical environment information for graph neural networks. The graph neural network learns the embedded representation of atomic nodes from the atomic-level graph structure to capture local structural information such as atom type and bonding environment.

[0035] Meanwhile, the molecular-level global features include multiple molecular descriptors. Molecular descriptors are descriptive languages ​​used to describe the macroscopic properties of molecular compounds, starting from the overall molecular structure. These molecular descriptors provide the graph neural network with macroscopic information about the target molecular compound, enabling the calculation of a series of physicochemical descriptors highly correlated with lipid solubility, thus supplementing the global properties of the molecular compound.

[0036] Furthermore, the group-level substructure features include a substructure vector. Each element in the substructure vector represents the frequency of occurrence of its corresponding target substructure in the target molecular compound. The target substructure is the substructure that affects the lipophilicity of the compound. The substructure of a molecular compound is its functional group or chemical pattern. Based on prior knowledge, target substructures that significantly affect lipophilicity can be identified, including both positive and negative effects. Each element in the substructure vector represents the frequency of occurrence of the corresponding target substructure in the target molecular compound. The substructure vector is provided as input features to the graph neural network, enabling the model to more accurately predict molecular lipophilicity based on the distribution characteristics of different target substructures.

[0037] In graph neural networks, on the one hand, the graph representation learning module obtains the embedded representation of each atom based on the atomic-level graph structure, and then obtains the atomic vector through global convergence to represent local structural information. On the other hand, the molecular-level global features are normalized to form molecular vectors, which contain the macroscopic property information of the target molecular compound. Then, the feature fusion module concatenates the atomic vectors, molecular vectors, and substructure vectors to obtain the global vector, which contains the global information of the target molecular compound. Finally, the prediction layer can predict the lipid solubility of the target molecular compound relatively accurately based on the global vector.

[0038] In summary, compared to existing deep learning-based methods for predicting the lipophilicity of compounds that only use atomic-level diagram structures, the lipophilicity prediction method in this application comprehensively considers the atomic-level diagram structure, molecular-level global features, and group-level substructure features of molecular compounds, constructing a multi-level feature system of "atoms-molecules-substructures." Through a feature fusion module, these features of different granularities and properties are organically combined for joint prediction, thereby achieving more comprehensive molecular characterization and more accurate prediction of the lipophilicity of molecular compounds. This solves the problem of low prediction accuracy in existing deep learning-based lipophilicity prediction methods. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart illustrating a compound lipophilicity prediction method based on a multi-level feature fusion graph neural network, provided in an embodiment of this application;

[0041] Figure 2This is a schematic diagram illustrating the prediction of the lipophilicity of compounds using a multi-level feature fusion graph neural network, as provided in an embodiment of this application. Detailed Implementation

[0042] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0043] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] This application provides a method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network, which can be applied to various terminal devices or servers.

[0045] The terminal can be, but is not limited to, various desktop computers, laptops, smartphones, and tablets. The server can be a standalone server, a server cluster consisting of multiple servers, or a cloud server.

[0046] In one exemplary embodiment, such as Figure 1 As shown, a method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network is presented. (Refer to...) Figure 2 The graph neural network includes a graph representation learning module, a feature fusion module, and a prediction layer. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is described using a terminal device as an example, including steps 110 to 150.

[0047] Step 110: Obtain the atomic-level diagram structure (2D molecular diagram), molecular-level global features, and group-level substructure features of the target molecular compound; wherein, the nodes and edges of the atomic-level diagram structure are the atoms and chemical bonds of the target molecular compound, respectively; the molecular-level global features include multiple molecular descriptors; the group-level substructure features include substructure vectors; each element in the substructure vector is used to characterize the frequency of occurrence of its corresponding target substructure in the target molecular compound; the target substructure is the substructure that affects the lipophilicity of the compound.

[0048] Step 120: The atomic-level graph structure is feature extracted using the graph representation learning module to obtain the embedding representation of each atom in the target molecule compound. Global average pooling is then performed on the embedding representation of each atom to obtain the atom vector.

[0049] Step 130: Normalize the global features at the molecular level to form molecular vectors.

[0050] Step 140: The atomic vector, molecular vector and substructure vector are concatenated by the feature fusion module to obtain the global vector.

[0051] Step 150: Predict the lipophilicity of the target molecule compound based on the global vector using the prediction layer.

[0052] In steps 110 to 150 above, the atomic-level graph structure, molecular-level global features, and group-level substructure features of the target molecular compound are used as inputs to the graph neural network. The graph neural network learns the relationship between the lipophilicity of the molecular compound and the atomic-level graph structure, molecular-level global features, and group-level substructure features. Finally, the graph neural network predicts the lipophilicity of the target molecular compound.

[0053] Among them, the atomic-level graph structure provides rich local chemical environment information for graph neural networks. The graph neural network learns the embedded representation of atomic nodes from the atomic-level graph structure and captures local structural information such as atomic type and bonding environment.

[0054] Meanwhile, the molecular-level global features include multiple molecular descriptors. Molecular descriptors are descriptive languages ​​used to describe the macroscopic properties of molecular compounds, starting from the overall molecular structure. These molecular descriptors provide the graph neural network with macroscopic information about the target molecular compound, enabling the calculation of a series of physicochemical descriptors highly correlated with lipid solubility, thus supplementing the global properties of the molecular compound.

[0055] Furthermore, the group-level substructure features include a substructure vector. Each element in the substructure vector represents the frequency of occurrence of its corresponding target substructure in the target molecular compound. The target substructure is the substructure that affects the lipophilicity of the compound. The substructure of a molecular compound is its functional group or chemical pattern. Based on prior knowledge, target substructures that significantly affect lipophilicity can be identified, including both positive and negative effects. Each element in the substructure vector corresponds to a target substructure, and the element's value represents the frequency of occurrence of the corresponding target substructure in the target molecular compound. Providing the substructure vector to the graph neural network allows the graph neural network to make more accurate predictions of lipophilicity based on the frequency of occurrence of various target substructures in the target compound.

[0056] In graph neural networks, on the one hand, the graph representation learning module obtains the embedded representation of each atom based on the atomic-level graph structure, and then performs global average pooling to obtain the atom vector, which contains the local structural information of the target molecular compound. On the other hand, the molecular-level global features are normalized to form the molecular vector, which contains the macroscopic property information of the target molecular compound. Then, the feature fusion module concatenates the atom vector, molecular vector and substructure vector to obtain the global vector, which contains the global information of the target molecular compound. Finally, the prediction layer can predict the lipophilicity of the target molecular compound relatively accurately based on the global vector.

[0057] In summary, compared to existing deep learning-based methods for predicting the lipophilicity of compounds that only use atomic-level graph structures, the lipophilicity prediction method in this embodiment comprehensively considers the atomic-level graph structure, molecular-level global features, and group-level substructure features of molecular compounds, constructing a multi-level feature system of "atoms-molecules-substructures." Through a feature fusion module, these features of different granularities and properties are organically combined for joint prediction, thereby achieving more comprehensive molecular characterization and more accurate prediction of the lipophilicity of molecular compounds. This solves the problem of low prediction accuracy in existing deep learning-based lipophilicity prediction methods.

[0058] For example, obtaining the atomic-level diagram structure, molecular-level global features, and group-level substructure features of a target molecular compound specifically includes: obtaining the SMILES (Simplified Molecular-Input Line-Entry System) string of the target molecular compound; and determining the atomic-level diagram structure, molecular-level global features, and group-level substructure features of the target molecular compound based on the SMILES string. Specifically, cheminformatics toolkits such as RDKit or OpenBabel can be used to generate the atomic-level diagram structure corresponding to the SMILES string and to calculate the molecular-level global features and group-level substructure features.

[0059] Specifically, in the atomic-level diagram structure, each atom is encoded as a feature vector containing the atom type, degree, number of hydrogen atoms, aromaticity marker, formal charge, hybridization state, and mass. For example, the type uses one-hot encoding, covering common elements such as C, N, O, S, F, Cl, Br, and I; the atom's degree, number of hydrogen atoms, and formal charge are normalized to maintain consistency in feature scale across different molecules; the hybridization state uses one-hot encoding, including four types: SP, SP2, SP3, and SP3D; and the mass uses a normalized representation.

[0060] Furthermore, in the atomic-level diagram structure, each chemical bond is encoded as a feature vector containing the bond type, aromaticity marker, and intra-ring marker. For example, the type is encoded using one-hot encoding, including single, double, and triple bonds. These bond features effectively describe the nature and strength of the interatomic connections.

[0061] Specifically, multiple descriptors include the target molecule's molecular weight, number of atoms, number of chemical bonds, number of rotatable chemical bonds, number of aromatic rings, number of aliphatic rings, topological polar surface area, number of hydrogen bond donors, number of hydrogen bond acceptors, and the counting characteristics of each element in the target molecule. Furthermore, multiple descriptors may also include the target molecule's Balaban index, Bertz complexity index, molecular volume, number of heavy atoms, and number of rings.

[0062] Among them, molecular weight reflects the overall size of the molecular compound; the number of atoms and chemical bonds describes the complexity of the molecular compound; the number of rotatable chemical bonds characterizes the flexibility of the molecular compound; the number of aromatic rings and aliphatic rings reflects the cyclic structure characteristics of the molecular compound; topological polar surface area (TPSA) measures the polarity of the molecular compound; the number of hydrogen bond donors and acceptors affects the solubility of the molecular compound; and the counting characteristics of each element (C, N, O, S, halogens) provide compositional information of the molecular compound.

[0063] Specifically, all the target substructures constitute a SMARTS dictionary containing multiple chemical patterns. These chemical patterns are carefully selected based on expert knowledge in the field of medicinal chemistry, and are common functional groups and structural units that affect lipophilicity, including benzene rings (c1ccccc1), heterocyclic structures (pyridine ring c1ccncc1, pyrimidine ring c1cccnc1, furan ring c1ccoc1, thiophene ring c1ccsc1), polar functional groups (carboxyl groups (C(=O)O), carbonyl groups (C(=O))) and halogen-containing substituents (Cl, Br, F, I), etc.

[0064] The above provides a complete description of the atomic-level graph structure, molecular-level global features, and group-level substructure features of the target molecule compound. Below, we provide an illustrative example of the specific structure of a graph neural network.

[0065] Specifically, the graph representation learning module consists of two stacked attention message passing layers. In the attention message passing layer, the update formula for each node is:

[0066]

[0067]

[0068]

[0069] in, This represents the embedding of the i-th node. m ij This represents the message between the i-th node and the j-th node. This represents the message passing weight between the i-th node and the j-th node. Q i This represents a query for the i-th node. K j This indicates that the j-th node corresponds to Q i The key vector, d Indicates the projection dimension. and These represent the i-th and j-th nodes, respectively. This represents the edge between the i-th and j-th nodes.

[0070] The graph representation learning module consists of two stacked attention message-passing layers for learning high-level structural representations of atomic-level graph structures. The first layer captures initial local interactions; the second layer further refines the representation, capturing more complex long-range dependencies (such as indirect influences through multi-hop neighbors). This enables the graph neural network to learn structural cues reflecting hydrogen bond donor-acceptor environments and aromaticity, thereby indirectly capturing the effects of these chemical interactions. Molecules are represented as graph structures, and the attention mechanism dynamically captures chemical interactions between atoms (such as the influence of neighboring atoms and bond properties) by calculating attention weights between nodes (atoms). Specifically, each atomic node generates query, bond, and value vectors through linear projection, and edge features are used as conditional quantities in the message function through projection, thus completing joint message passing based on nodes and edges. The attention mechanism allows the graph representation learning module to adaptively assign influence weights between different atoms according to the chemical environment.

[0071] exist In the process, the source node (the j-th node), the target node (the i-th node), and the edge features are combined. After splicing, the data is processed by a multilayer perceptron (MLP) and multiplied by attention weights. The message is then aggregated to the target node, and the node embedding is updated via GRU cells and LayerNorm.

[0072]

[0073] The feature fusion module aims to fuse features from different levels of abstraction.

[0074] For example, firstly, global average pooling is performed on the embedding representations of all atoms to obtain a 256-dimensional vector representing the entire atomic-level graph structure. V graph (Atomic vectors) are vectors formed after normalizing global features at the molecular level. V global (Molecular vector), the vector representation of group-level substructure features is as follows: V sub (Substructure vectors) concatenate the vectors from the three levels to form the final hybrid molecular representation:

[0075]

[0076] The prediction layer inputs the mixed molecular representation into a 3-layer multilayer perceptron (using the ReLU activation function in the middle layers and the linear activation function in the output layer), and regresses to predict the final lipophilicity (which can be represented by the octanol-water partition coefficient LogP):

[0077]

[0078]

[0079]

[0080] in, Indicates batch normalization, Represents the ReLU activation function. Represents a linear activation function. , and All represent weights. , and Both represent bias. This represents the predicted value.

[0081] To enable a graph neural network to make predictions, it needs to be trained beforehand. During training, the loss function of the graph neural network... for:

[0082]

[0083] in, This represents the mean squared error loss of the prediction result. The projection layer weights represent the substructure vectors. Represents the coefficient. , Indicates a label, Indicates the predicted value. B This represents the total number of samples.

[0084] Specifically, mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for supervised training on a dataset of molecular compounds with known lipophilicity. An L1 regularization term is added to the weights of the substructure projection layers to prevent overfitting. This constraint causes the model to tend to retain only substructures that significantly affect lipophilicity prediction, thereby improving prediction stability and interpretability. For example, the optimization process employs the following strategy: the optimizer uses the Adam optimizer with a learning rate of... Weight decay rate Learning rate scheduling: A cosine annealing warm-up restart strategy is used, with a restart period of T0 = 300 epochs. Gradient clipping: Before each parameter update, the norm of the gradient vector is clipped, with a threshold of 1.0. Early stopping: The validation set RMSE is monitored; if there is no improvement for 50 consecutive epochs, training is terminated.

[0085] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0086] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0087] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0088] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0089] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network, wherein the graph neural network includes a graph representation learning module, a feature fusion module, and a prediction layer, characterized in that, The method for predicting the lipid solubility of the compound includes: The atomic-level diagram structure, molecular-level global features, and group-level substructure features of the target molecular compound are obtained. The nodes and edges of the atomic-level diagram structure represent the atoms and chemical bonds of the target molecular compound, respectively. The molecular-level global features include multiple molecular descriptors. The group-level substructure features include a substructure vector, where each element in the substructure vector characterizes the frequency of occurrence of its corresponding target substructure in the target molecular compound. The target substructure is a substructure that affects the lipophilicity of the compound. The graph representation learning module extracts features from the atomic-level graph structure to obtain the embedding representation of each atom in the target molecule compound, and performs global average pooling on the embedding representation of each atom to obtain an atom vector. The molecular-level global features are normalized to form molecular vectors; The feature fusion module concatenates the atomic vector, the molecular vector, and the substructure vector to obtain a global vector. The prediction layer predicts the lipid solubility of the target molecule compound based on the global vector. The graph representation learning module includes two stacked attention message passing layers. In the attention message passing layer, the update formula for each node is: in, This represents the embedding of the i-th node. m ij This represents the message between the i-th node and the j-th node. This represents the message passing weight between the i-th node and the j-th node. Q i This represents a query for the i-th node. K j This indicates that the j-th node corresponds to Q i The key vector, d Indicates the projection dimension. and These represent the i-th and j-th nodes, respectively. This represents the edge between the i-th and j-th nodes; The loss function of the graph neural network for: in, This represents the mean squared error loss of the prediction result. This represents the projection layer weights of the substructure vector.

2. The method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network according to claim 1, characterized in that, Obtain the atomic-level structure, molecular-level global features, and group-level substructure features of the target molecule compound, specifically including: Obtain the SMILES string of the target molecule compound; The atomic-level diagram structure, molecular-level global features, and group-level substructure features of the target molecular compound are determined based on the SMILES string.

3. The method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network according to claim 1, characterized in that, In the atomic-level diagram structure, each atom is encoded as a feature vector containing the atom's type, degree, number of hydrogen atoms, aromaticity marker, formal charge, hybridization state, and mass.

4. The method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network according to claim 1, characterized in that, In the atomic-level diagram structure, each chemical bond is encoded as a feature vector containing the type of the chemical bond, an aromaticity marker, and an intra-ring marker.

5. The method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network according to claim 1, characterized in that, The descriptors include the molecular weight, number of atoms, number of chemical bonds, number of rotatable chemical bonds, number of aromatic rings, number of aliphatic rings, topological polar surface area, number of hydrogen bond donors, number of hydrogen bond acceptors, and the counting characteristics of each element in the target molecule.

6. The method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network according to claim 5, characterized in that, The descriptors also include the Balaban index, Bertz complexity index, molecular volume, number of heavy atoms, and number of rings of the target molecule compound.

7. The method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network according to claim 1, characterized in that, The target substructures include benzene rings, polar functional groups, heterocyclic structures, and halogen-containing substituents.

8. The method for predicting the lipophilicity of compounds based on a multi-level feature fusion graph neural network according to claim 7, characterized in that, The heterocyclic structure includes a pyridine ring, a pyrimidine ring, a furan ring, and a thiophene ring, and the polar functional groups include a carboxyl group and a carbonyl group.

Images