A molecular feature extraction method based on substructure graph convolution
By constructing molecular graphs and substructure graphs, combining hybrid multi-order graph convolution and attention models, and fusing atomic features and functional group information, the problem of inaccurate molecular feature extraction in existing technologies is solved, achieving higher accuracy and precision.
Patent Information
- Application Number
- CN202211607104.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-14
- Publication Date
- 2025-09-16
- Estimated Expiration
- 2042-12-14
AI Technical Summary
Existing graph convolution methods fail to effectively consider the atomic features and functional group information in molecules, resulting in inaccurate molecular feature extraction.
A method based on substructure graph convolution is adopted to construct molecular graphs and substructure graphs, combine hybrid multi-order graph convolution and attention model, fuse atomic features and functional group information, perform multi-order graph convolution and cross convolution processing, and extract molecular features.
It improves the accuracy and precision of molecular feature extraction, provides a reliable basis for the prediction of molecular properties, and avoids the loss of key group information.
Smart Images
Figure CN116230112B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of molecular feature extraction, and in particular to a molecular feature extraction method based on substructure graph convolution. Background Art
[0002] The emergence of computer-aided drug design (CADD) has improved the efficiency of drug development and has become an indispensable tool in the new drug design process. Computer algorithms can predict the three-dimensional structure of proteins, extract molecular pharmacophores, and predict molecular properties.
[0003] With the advent of the big data era, the number of chemical compounds is increasing. Computer-assisted methods are needed to pre-screen effective compounds from this large amount of data for biological testing. However, due to the complex structure of molecules, finding an appropriate model to describe the molecular structure is crucial for feature extraction. Molecular structures are primarily composed of connected atoms and have a certain degree of flexibility, making traditional convolutional neural networks inapplicable to such structures. With the emergence of graph convolutional neural networks, molecular structures are converted into graph structures (atoms as points, connections as edges) and convolution is performed to effectively mine the relationships between atoms and extract molecular features.
[0004] However, existing methods for extracting molecular features using graph convolutions only treat single atoms as nodes, without considering the role of functional groups in molecules. They neither address molecular substructures nor embed functional group information. Therefore, there is an urgent need to propose a molecular feature extraction method based on substructure graph convolution to achieve accurate molecular feature extraction. Summary of the Invention
[0005] In response to the problem that existing methods are difficult to accurately extract molecular features, the present invention proposes a molecular feature extraction method based on substructure graph convolution. This method fully considers the atomic features and functional group information in the molecule, and uses the cross-graph convolution method of fused substructures to fuse the molecular graph structure information of the drug molecule with the substructure graph information, effectively mining the correlation relationship between the atoms in the molecule, improving the accuracy of molecular feature extraction, and providing a basis for accurately predicting molecular properties.
[0006] The present invention adopts the following technical solutions:
[0007] A molecular feature extraction method based on substructure graph convolution specifically comprises the following steps:
[0008] Step 1: Based on the molecular structure of the drug molecule, a molecular graph structure of the drug molecule is constructed, the adjacency matrix and feature matrix of the drug molecule are obtained, and the atomic features corresponding to each atomic point in the molecular graph structure are obtained;
[0009] Step 2: According to the atomic features of each order corresponding to each atomic point in the molecular graph structure, the multi-order atomic features of each atomic point in the molecular graph structure are extracted based on the mixed multi-order graph convolution method, and the attention is calculated in combination with the attention model to determine the mixed multi-order molecular features of each atomic point in the molecular graph structure, and obtain the atomic mixed multi-order vector;
[0010] Step 3: extract the cyclic molecular structure and non-cyclic molecular structure in the drug molecule according to the molecular graph structure of the drug molecule, obtain the substructure graph of the drug molecule, extract all the bond features in the substructure graph, and determine the node features of each node in the substructure graph;
[0011] Step 4: Based on the substructure graph and molecular graph structure of the drug molecule, the substructure information is integrated into the molecular graph convolution based on the cross convolution method of fusion substructure, and the node features of each node in the substructure graph are processed by first-order convolution to construct the substructure vector Then, for each atomic point in the molecular graph structure, the atomic mixed multi-order vector and substructure vector are used to obtain the mixed multi-order molecular features and first-order node features of each layer within the atomic point. The mixed multi-order molecular features and first-order node features are convolved and spliced to obtain the molecular graph convolution value of each layer within each atomic point in the molecular graph structure.
[0012] Step 5: According to the molecular graph convolution values of each layer of all atomic points in the molecular graph structure, the average value of the molecular graph convolution of each atomic point in the molecular graph structure and the maximum value of the molecular graph convolution of each atomic point are determined and spliced to obtain the molecular characteristic attributes of each atomic point in the molecular graph structure and determine the molecular characteristics of the drug molecule.
[0013] Preferably, in step 1, based on the molecular structure of the drug molecule, the atomic features of each atom in the drug molecule are obtained, and each atom in the drug molecule except hydrogen bonds is used as an atomic point and the connection relationship between the atoms of the drug molecule is used as an edge to construct the molecular graph structure of the drug molecule, as shown in formula (1):
[0014] G atom =(A atom ,X atom ) (1)
[0015] Where, is the adjacency matrix of the drug molecule, which is used to represent the connection relationship between atoms in the drug molecule. When atom i is connected to atom j in the drug molecule, A atom (i,j)=1, when atom i and atom j in the drug molecule are not connected, A atom (i,j)=0; is the characteristic matrix of the drug molecule, which is used to record the atomic characteristics of each atom in the drug molecule; M is the number of atoms in the drug molecule, and D is the dimension of the atomic characteristics in the drug molecule.
[0016] Preferably, the atomic characteristics include atom type, atomicity, free radical and aromaticity.
[0017] Preferably, the step 2 specifically includes the following steps:
[0018] Step 2.1, according to the atomic features of each order corresponding to each atomic point in the molecular graph structure, perform linear transformation on the atomic features of each order of each atomic point in the molecular graph structure, as shown in formula (2):
[0019] V atom (m)=σ(W atom X atom (m)+b atom ) (2)
[0020] Where W atom 、b atom are all learning parameters of the graph convolution model, σ is the activation function, m is the number of the atomic point in the molecular graph structure, V atom is the linear atomic feature of the atomic point;
[0021] In step 2.2, based on the multi-order graph convolution method, the graph convolution model is used to perform multi-order graph convolution calculations on each atomic point in the molecular graph structure according to the linear atomic features of each atomic point in the molecular graph structure, and the multi-order linear atomic features of all atomic points in the molecular graph structure are determined, as shown in formula (3):
[0022]
[0023] Where, is the convolution function of the p-th layer; All are training parameters; N p (m) is the set of all p-order neighbors of atomic point m; p is the order, when p = 0, N p (m) is empty, the convolution calculation only considers the atomic point itself, and does not involve the p-order neighbors of the atomic point; when p>0, all p-order neighbors of the atomic point m participate in the convolution calculation;
[0024] Then, the attention model is constructed and the attention corresponding to the multi-order linear atomic features of each atomic point is calculated using the attention model, as shown in formula (4):
[0025]
[0026] Where z p is the attention corresponding to the p-order atomic feature of the atomic point;
[0027] In step 2.3, based on the multi-order linear atomic features and attention of each atomic point, the mixed multi-order molecular features of each atomic point in the molecular graph structure are calculated and the atomic mixed multi-order vector is constructed, as shown in formula (5):
[0028]
[0029] Where, It is the mixed multi-order molecular feature of the lth layer of atomic points.
[0030] Preferably, in step 3, according to the molecular graph structure of the drug molecule, the cyclic molecular structure and the non-cyclic molecular structure in the drug molecule are extracted as substructures, the substructures are simplified into nodes to generate a substructure graph, the bond features of the drug molecule are extracted in the substructure graph, and after the atomic features of each atom in the drug molecule are fused with the bond features, the node features of each node in the substructure graph are determined, as shown in formula (6):
[0031]
[0032] Where s is the substructure, m is the node number, e is the chemical bond; [·] is the characteristic splicing function; V sub (s) is the node feature of substructure s, which is composed of the atomic features X of all atoms in the substructure atom and the original characteristics of all chemistry X bond Obtained through linear transformation.
[0033] Preferably, the substructure diagram includes not only the atomic features of each atom in the drug molecule, but also the bond features between atoms in the drug molecule, including bond type, stereo descriptors, aromatic bonds and ring states.
[0034] Preferably, in step 4, when the atomic points in the molecular graph structure correspond to multiple nodes in the substructure graph, the node features of all nodes in the substructure graph are merged based on the uniform pooling method and then convolution calculation is performed.
[0035] Preferably, in step 5, the molecular characteristic attributes of each atomic point in the molecular graph structure are as shown in formula (7):
[0036]
[0037] Where V drug is the molecular characteristic attribute of the atomic point; M is the total number of atomic points in the molecular graph structure.
[0038] The present invention has the following beneficial effects:
[0039] 1. This patented method fully considers the influence of the high-order neighbors of atoms in the molecular graph structure on their convolution calculations. Based on the hybrid multi-order graph convolution method, multi-order graph convolution is performed on each atomic point in the molecular graph structure, and different attention weights are assigned to the convolution results of different orders to obtain multi-order linear atomic features, thereby achieving effective mining of deep molecular structural information and improving the accuracy of molecular graph convolution processing.
[0040] 2. This patented method fully considers the influence of molecular substructure on molecular characteristics. It generates a substructure graph by taking the cyclic molecular structure and non-cyclic molecular structure in the molecule as substructures. The substructure information is integrated into the molecular graph convolution based on the cross convolution method of fused substructures. The molecular graph structure and the substructure graph are convolved separately, and the convolution results of the substructure graph are spliced with the convolution results of the corresponding atomic points in its molecular graph structure, thereby avoiding the loss of key group information in the molecule, improving the accuracy of molecular feature extraction, and providing a basis for accurate prediction of molecular properties. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 Schematic diagram of a molecular feature extraction method based on substructure graph convolution.
[0042] Figure 2 Schematic diagram of the molecular graph structure. In the figure, (a) is the molecular structure of the drug molecule, and (b) is the molecular graph structure of the drug molecule.
[0043] Figure 3 Schematic diagram of the hybrid multi-order graph convolution method.
[0044] Figure 4 Schematic diagram of the substructure diagram. In the figure, (a) is the molecular structure of the drug molecule, and (b) is the substructure diagram of the drug molecule. DETAILED DESCRIPTION
[0045] The specific embodiments of the present invention are further described below with reference to the accompanying drawings and a certain drug molecule as an example:
[0046] This paper proposes a molecular feature extraction method based on substructure graph convolution, such as Figure 1 As shown, the specific steps include:
[0047] Step 1: According to the molecular structure of the drug molecule, the atomic characteristics of each atom in the drug molecule are obtained, including the atomic type, atomic degree, free radical and aromaticity of each atom in the drug molecule; the atomic characteristics of the drug molecule in this embodiment are shown in Table 1.
[0048] Table 1 Atomic characteristics of drug molecules
[0049]
[0050] The molecular graph structure of the drug molecule is constructed by taking each atom in the drug molecule except hydrogen bonds as the atomic point and the connection relationship between the atoms in the drug molecule as the edge, as shown in formula (1):
[0051] G atom =(A atom ,X atom ) (1)
[0052] Where, is the adjacency matrix of the drug molecule, which is used to represent the connection relationship between atoms in the drug molecule. When atom i is connected to atom j in the drug molecule, A atom (i,j)=1, when atom i and atom j in the drug molecule are not connected, A atom (i,j)=0; is the characteristic matrix of the drug molecule, which is used to record the atomic characteristics of each atom in the drug molecule; M is the number of atoms in the drug molecule, and D is the dimension of the atomic characteristics in the drug molecule.
[0053] According to the molecular graph structure of drug molecules, such as Figure 2 As shown, the adjacency matrix and feature matrix of the drug molecule are obtained, and the atomic features corresponding to each atomic point in the molecular graph structure are obtained. In this embodiment, based on the adjacency matrix and feature matrix of the drug molecule, the atomic features in the drug molecule can be extracted and represented through spatial graph convolution or spectral convolution processing methods.
[0054] Step 2: According to the atomic features of each order corresponding to each atomic point in the molecular graph structure, the multi-order atomic features of each atomic point in the molecular graph structure are extracted based on the mixed multi-order graph convolution method, and the attention is calculated in combination with the attention model to determine the mixed multi-order molecular features of each atomic point in the molecular graph structure, and obtain the atomic mixed multi-order vector, such as Figure 3 As shown, the specific steps include:
[0055] Step 2.1, according to the atomic features of each order corresponding to each atomic point in the molecular graph structure, perform linear transformation on the atomic features of each order of each atomic point in the molecular graph structure, as shown in formula (2):
[0056] V atom (m)=σ(W atom X atom (m)+b atom ) (2)
[0057] Where W atom 、b atom are all learning parameters of the graph convolution model, σ is the activation function, m is the number of the atomic point in the molecular graph structure, V atom is the linear atomic feature of the atomic point.
[0058] In step 2.2, after linearly transforming the atomic features of each order of each atomic point in the molecular graph structure, it is necessary to first consider the relationship between each atom and other atoms in the molecular graph structure before performing convolution calculations. However, graph convolution is different from conventional image convolution. The number of nodes in its receptive field is not fixed. Image convolution can specify the convolution area and set the convolution kernel to a fixed value, but similar areas cannot be found in graph convolution. In addition, not all nodes in the graph convolution receptive field play a role in molecular activity. Therefore, for graph convolution, it is also necessary to reasonably select convolution nodes. If only atoms are used as units for convolution, the influence of the molecular substructure on the properties will be destroyed. Therefore, this embodiment adopts a hybrid multi-order graph convolution method for molecular feature representation.
[0059] Since the interaction between any atomic point in the molecular graph structure and its directly connected atomic point is the most direct, followed by indirect neighbors (higher-order neighbors), the deep relationship between atomic points can be captured during graph convolutional message passing, which plays an important role in the representation of atomic point features.
[0060] According to the linear atomic features of each atomic point in the molecular graph structure, based on the multi-order graph convolution method, the graph convolution model is used to perform multi-order graph convolution calculations on each atomic point in the molecular graph structure to determine the multi-order linear atomic features of all atomic points in the molecular graph structure, as shown in formula (3):
[0061]
[0062] Where, is the convolution function of the p-th layer; All are training parameters; N p (m) is the set of all p-order neighbors of atomic point m; p is the order, when p = 0, N p (m) is empty, the convolution calculation only considers the atomic point itself and does not involve the p-order neighbors of the atomic point; when p>0, all p-order neighbors of the atomic point m participate in the convolution calculation.
[0063] However, the roles played by features of different orders are inconsistent. Therefore, this embodiment constructs an attention model and uses the attention model to calculate the attention corresponding to the multi-order linear atomic features of each atomic point, as shown in formula (4):
[0064]
[0065] Where z p is the attention corresponding to the p-order atomic feature of the atomic point.
[0066] In step 2.3, based on the multi-order linear atomic features and attention of each atomic point, the mixed multi-order molecular features of each atomic point in the molecular graph structure are calculated and the atomic mixed multi-order vector is constructed, as shown in formula (5):
[0067]
[0068] Where, It is the mixed multi-order molecular feature of the lth layer of atomic points.
[0069] Step 3. Due to the particularity of the drug molecular structure, there will be a large number of five-membered rings and six-membered rings (such as benzene rings). These group structures are extremely important for the expression of molecular chemical properties. The traditional SMILES representation and atom-based convolution method cannot consider these groups as a whole. Therefore, this patent proposes a molecular feature representation method for fusion substructures.
[0070] According to the molecular graph structure of the drug molecule, the cyclic molecular structure and the non-cyclic molecular structure in the drug molecule are extracted respectively. In this embodiment, for the non-cyclic molecular structure, the connection containing more than two atoms is extracted as the substructure, and the cyclic molecular structure and the substructure are simplified into nodes to generate a substructure graph, such as Figure 4 All bond features in the substructure graph are extracted, including bond type, stereo descriptors, aromatic bonds, and ring states. The bond features of the drug molecule in this embodiment are shown in Table 2. Node features are then determined for each node in the substructure graph. The node features include not only the atomic features of all atoms in the drug molecule but also the bond features between atoms.
[0071] Table 2 Bond characteristics of drug molecules
[0072]
[0073] After integrating the atomic features and bond features of each atom in the drug molecule, the node features of each node in the substructure graph are determined, as shown in formula (6):
[0074]
[0075] Where s is the substructure, m is the node number, e is the chemical bond; [·] is the characteristic splicing function; V sub (s) is the node feature of substructure s, which is composed of the atomic features X of all atoms in the substructure atom and the original characteristics of all chemistry X bond Obtained through linear transformation.
[0076] Step 4: In order to better integrate the substructure features and atomic features of drug molecules, this embodiment proposes a cross-convolution model. According to the substructure graph and molecular graph structure of the drug molecules, the substructure information is integrated into the molecular graph convolution based on the cross-convolution method of fusion substructure.
[0077] Perform first-order convolution processing on the node features of each node in the substructure graph to construct the substructure vector Then, for each atomic point in the molecular graph structure, the atomic mixed multi-order vector and substructure vector are used to obtain the mixed multi-order molecular characteristics and first-order node characteristics of each layer within the atomic point. The mixed multi-order molecular characteristics and first-order node characteristics are convolved and spliced to obtain the molecular graph convolution value of each layer within each atomic point in the molecular graph structure.
[0078] Moreover, when the atomic points in the molecular graph structure correspond to multiple nodes in the substructure graph, the node features of all nodes in the substructure graph are merged using the uniform pooling method, and then spliced with the mixed multi-order molecular features of the atomic mixed multi-order vector. At the same time, a fully connected layer is added to achieve semantic consistency of the features. At this time, the output is And the convolution result after the substructure graph is fully connected It is directly used as the input of the next layer to obtain the molecular graph convolution of the atomic points.
[0079] Step 5: According to the molecular graph convolution values of each layer in all atomic points in the molecular graph structure, the average value of the molecular graph convolution of each atomic point in the molecular graph structure and the maximum value of the molecular graph convolution of each atomic point in the molecular graph structure are determined and spliced to obtain the molecular characteristic attributes of each atomic point in the molecular graph structure, as shown in formula (7):
[0080]
[0081] Where V drug is the molecular characteristic attribute of the atomic point; M is the total number of atomic points in the molecular graph structure.
[0082] The molecular characteristic attributes in this embodiment include both global attributes and attributes of key nodes. In the cross-convolution calculation process based on the fused substructure, the result of the substructure graph convolution is not used for output. Instead, the substructure graph information is fused during the convolution calculation process of each layer of each node, thereby reducing the complexity of the calculation and achieving accurate extraction of the characteristic attributes of the drug molecules.
[0083] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.
Claims
1. A molecular feature extraction method based on substructure graph convolution, characterized in that: The specific steps include: Step 1: Based on the molecular structure of the drug molecule, a molecular graph structure of the drug molecule is constructed, the adjacency matrix and feature matrix of the drug molecule are obtained, and the atomic features corresponding to each atomic point in the molecular graph structure are obtained; Step 2: According to the atomic features of each order corresponding to each atomic point in the molecular graph structure, the multi-order atomic features of each atomic point in the molecular graph structure are extracted based on the mixed multi-order graph convolution method, and the attention is calculated in combination with the attention model to determine the mixed multi-order molecular features of each atomic point in the molecular graph structure, and obtain the atomic mixed multi-order vector; Step 3: extract the cyclic molecular structure and non-cyclic molecular structure in the drug molecule according to the molecular graph structure of the drug molecule, obtain the substructure graph of the drug molecule, extract all the bond features in the substructure graph, and determine the node features of each node in the substructure graph; Step 4: Based on the substructure graph and molecular graph structure of the drug molecule, the substructure information is integrated into the molecular graph convolution based on the cross convolution method of fusion substructure, and the node features of each node in the substructure graph are processed by first-order convolution to construct the substructure vector Then, for each atomic point in the molecular graph structure, the atomic mixed multi-order vector and substructure vector are used to obtain the mixed multi-order molecular features and first-order node features of each layer within the atomic point. The mixed multi-order molecular features and first-order node features are convolved and spliced to obtain the molecular graph convolution value of each layer within each atomic point in the molecular graph structure. Step 5: According to the molecular graph convolution values of each layer of all atomic points in the molecular graph structure, the average value of the molecular graph convolution of each atomic point in the molecular graph structure and the maximum value of the molecular graph convolution of each atomic point are determined and spliced to obtain the molecular characteristic attributes of each atomic point in the molecular graph structure and determine the molecular characteristics of the drug molecule.
2. The molecular feature extraction method based on substructure graph convolution according to claim 1, characterized in that: In step 1, the atomic features of each atom in the drug molecule are obtained according to the molecular structure of the drug molecule, and each atom in the drug molecule except hydrogen bonds is used as an atomic point and the connection relationship between the atoms of the drug molecule is used as an edge to construct the molecular graph structure of the drug molecule, as shown in formula (1): G atom =(A atom ,X atom ) (1) Where, is the adjacency matrix of the drug molecule, which is used to represent the connection relationship between atoms in the drug molecule. When atom i is connected to atom j in the drug molecule, A atom (i,j)=1, when atom i and atom j in the drug molecule are not connected, A atom (i,j)=0; is the characteristic matrix of the drug molecule, which is used to record the atomic characteristics of each atom in the drug molecule; M is the number of atoms in the drug molecule, and D is the dimension of the atomic characteristics in the drug molecule.
3. The molecular feature extraction method based on substructure graph convolution according to claim 2, characterized in that: The atomic characteristics include atom type, atomicity, radicality and aromaticity.
4. The molecular feature extraction method based on substructure graph convolution according to claim 2, characterized in that: The step 2 specifically includes the following steps: Step 2.1, according to the atomic features of each order corresponding to each atomic point in the molecular graph structure, perform linear transformation on the atomic features of each order of each atomic point in the molecular graph structure, as shown in formula (2): V atom (m)=σ(W atom X atom (m)+b atom ) (2) Where W atom 、b atom are all learning parameters of the graph convolution model, σ is the activation function, m is the number of the atomic point in the molecular graph structure, V atom is the linear atomic feature of the atomic point; In step 2.2, based on the multi-order graph convolution method, the graph convolution model is used to perform multi-order graph convolution calculations on each atomic point in the molecular graph structure according to the linear atomic features of each atomic point in the molecular graph structure, and the multi-order linear atomic features of all atomic points in the molecular graph structure are determined, as shown in formula (3): Where, is the convolution function of the p-th layer; All are training parameters; N p (m) is the set of all p-order neighbors of atomic point m; p is the order, when p = 0, N p (m) is empty, the convolution calculation only considers the atomic point itself, and does not involve the p-order neighbors of the atomic point; when p>0, all p-order neighbors of the atomic point m participate in the convolution calculation; Then, the attention model is constructed and the attention corresponding to the multi-order linear atomic features of each atomic point is calculated using the attention model, as shown in formula (4): Where z p is the attention corresponding to the p-order atomic feature of the atomic point; In step 2.3, based on the multi-order linear atomic features and attention of each atomic point, the mixed multi-order molecular features of each atomic point in the molecular graph structure are calculated and the atomic mixed multi-order vector is constructed, as shown in formula (5): Where, It is the mixed multi-order molecular feature of the lth layer of atomic points.
5. The molecular feature extraction method based on substructure graph convolution according to claim 4, characterized in that: According to the molecular graph structure of the drug molecule, the cyclic molecular structure and the non-cyclic molecular structure in the drug molecule are extracted as substructures, and the substructures are simplified into nodes to generate a substructure graph. The bond features of the drug molecule are extracted from the substructure graph. After the atomic features of each atom in the drug molecule are integrated with the bond features, the node features of each node in the substructure graph are determined, as shown in formula (6): Where s is the substructure, m is the node number, e is the chemical bond; [·] is the characteristic splicing function; V sub (s) is the node feature of substructure s, which is composed of the atomic features X of all atoms in the substructure atom and the original characteristics of all chemistry X bond Obtained through linear transformation.
6. The molecular feature extraction method based on substructure graph convolution according to claim 5, characterized in that: In addition to the atomic features of each atom in the drug molecule, the substructure diagram also includes the bond features between atoms in the drug molecule, including bond type, stereo descriptors, aromatic bonds and ring states.
7. The molecular feature extraction method based on substructure graph convolution according to claim 1, characterized in that: In step 4, when the atomic points in the molecular graph structure correspond to multiple nodes in the substructure graph, the node features of all nodes in the substructure graph are merged based on the uniform pooling method and then convolution calculation is performed.
8. The molecular feature extraction method based on substructure graph convolution according to claim 5, characterized in that: In step 5, the molecular characteristic attributes of each atomic point in the molecular graph structure are as shown in formula (7): Where V drug is the molecular characteristic attribute of the atomic point; M is the total number of atomic points in the molecular graph structure.