Hierarchical geometric graph neural network construction method and device for molecular representation learning

Abstract

The application provides a hierarchical geometric graph neural network construction method and device for molecular representation learning, relates to the technical field of graph neural networks, and is characterized in that a hierarchical geometric graph neural network obtained by training through a deep learning method is used to convert the spatial coordinate information of each atom in a to-be-processed molecule into relative geometric information, the to-be-processed molecule is decomposed according to different numbers of atoms to obtain substructures at different levels, the initialization messages of each substructure are obtained, and transverse message passing and longitudinal message passing of the molecular geometric messages are performed between each level to obtain a more complete target molecular representation structure, the target molecular representation structure is used to help the hierarchical geometric graph neural network understand the molecular geometric features, and thus downstream tasks such as prediction of molecular properties and dynamics are completed. The application realizes effective and complete representation of the molecular geometric structure, is beneficial to hierarchical geometric graph neural network modeling and learning of the molecular geometric structure, and is convenient for completing various downstream tasks of computational chemistry.

Description

Technical Field

[0001] This application relates to the field of graph neural network technology, and in particular to a method and apparatus for constructing hierarchical geometric graph neural networks for molecular representation learning. Background Technology

[0002] Effective modeling and learning of molecular geometry is a crucial step in predicting molecular properties and dynamics. Molecular geometry (i.e., the spatial distribution of atoms within a molecule) is an essential characteristic of molecules, reflecting not only quantum chemical properties but also determining molecular dynamics—specifically, the forces, motions, and energies of atoms within the molecule. However, 2D molecular graph neural networks and 3D molecular graph-based invariant neural network models have limited ability to model geometry and properties, making it difficult to learn an effective and complete representation of molecular geometry. Summary of the Invention

[0003] In view of this, embodiments of this application provide a method and apparatus for constructing a hierarchical geometric graph neural network for molecular representation learning, so as to overcome the above problems or at least partially solve the above problems.

[0004] The first aspect of this application provides a method for constructing a hierarchical geometric graph neural network for molecular representation learning, the method comprising:

[0005] The spatial coordinate information of each atom contained in the multibody molecule to be processed is input into a pre-trained hierarchical geometric graph neural network. The hierarchical geometric graph neural network converts the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information.

[0006] The hierarchical geometric graph neural network decomposes the molecule to be processed according to different numbers of atoms based on the relative geometric information, resulting in multiple substructures belonging to different levels. Each level includes multiple substructures, and the number of atoms contained in the substructures of each level is the same.

[0007] The hierarchical geometric graph neural network is used to calculate the interaction messages between substructures in each level and the internal messages of the substructures to obtain the initialization messages of each substructure.

[0008] The hierarchical geometric graph neural network is used to transmit the initialization messages of each level to obtain the horizontal messages of each substructure.

[0009] Through the hierarchical geometric graph neural network, the horizontally aggregated message obtained by aggregating multiple horizontal messages is passed to the various substructures contained in other levels in a way from high level to low level, so that the substructures of each level can obtain vertical messages.

[0010] By using the hierarchical geometric graph neural network, the initialization messages of each substructure are updated and read out according to the vertical messages and the horizontal messages, so as to obtain the target molecular representation structure of the molecule to be processed.

[0011] The hierarchical geometric graph neural network is used to predict the molecular properties and dynamics of the target molecule representation structure, thereby obtaining the state of the molecule to be processed and the force situation of each atom contained in the molecule to be processed.

[0012] Optionally, the hierarchy includes a first level, a second level, a third level, and a fourth level, and the substructure includes one-body, two-body, three-body, and four-body structures. The process involves using the hierarchical geometric graph neural network to decompose the molecule to be processed according to the selected number of atoms based on the relative geometric information, resulting in multiple substructures at different levels, including:

[0013] The hierarchical geometric graph neural network decomposes the molecule to be processed into a first level based on the relative geometric information, according to the number of atoms. The first level includes multiple units, the internal message of each unit is the atomic number of the unit, and the interaction message between each unit is the distance between atoms.

[0014] The hierarchical geometric graph neural network decomposes the molecule to be processed into two atoms according to the relative geometric information to obtain a second level. The second level includes multiple binary bodies. The internal message of each binary body is the bond length between the two atoms contained in the binary body. The interaction message between each binary body is the bond angle between the bond length of the binary body.

[0015] The hierarchical geometric graph neural network decomposes the molecule to be processed into three atoms according to the relative geometric information to obtain a third level. The third level includes multiple three-body structures. The internal message of each three-body structure is the area formed between the three atoms contained in the three-body structure. The interaction message between each three-body structure is the torsion angle of the area between the three-body structures.

[0016] The hierarchical geometric graph neural network decomposes the molecule to be processed into four atoms according to the relative geometric information to obtain a fourth level. The fourth level includes multiple quadrilaterals. The internal message of each quadrilateral is the volume formed between the four atoms contained in the quadrilateral, and the interaction message between each quadrilateral is the chirality of the volume between the quadrilaterals.

[0017] Optionally, the step of passing initialization messages at each level through the hierarchical geometric graph neural network to obtain lateral messages for each substructure includes:

[0018] Based on the initialization messages of each substructure, lateral message passing is performed on the adjacent substructures to obtain the lateral messages of each substructure. The initialization messages of the four bodies do not need to be laterally passed. The initialization messages of the four bodies are used to directly pass to the substructures of other levels except the fourth level.

[0019] Optionally, the lateral messages of each of the three bodies in the third level are:

[0020] Among them, AGG-H p This is a three-body horizontal aggregation function, including the Hadamard operation and the cumulative operation, p h For each of the aforementioned tripartites, there is a tripartite adjacent to it. For the interactive geometry embedding corresponding to the third level, N p A set representing three bodies;

[0021] The horizontal messages of each two bodies in the second level are:

[0022] e AH =AGG-H e ({e h ,e CBF (θ)}|e h ∈N e}), where AGG-H e For two-body lateral aggregation functions, e h For each of the two bodies adjacent to the aforementioned two bodies, e CBF (θ) represents the interactive geometry embedding corresponding to the second level, N e A set representing two bodies;

[0023] The horizontal messages of each entity in the first level are:

[0024] v AH =AGG-H v ({v h ,e RBF (d)}|v h ∈N v}), where AGG-H v As a unified horizontal aggregation function, v h For an entity adjacent to each of the aforementioned entities, e RBF (d) represents the interactive geometry embedding corresponding to the first level, N v A collection that represents a single entity.

[0025] Optionally, the step of passing the horizontally aggregated message obtained by aggregating multiple horizontal messages to the various substructures contained in other levels through the hierarchical geometric graph neural network in a manner from high-level to low-level, so that the substructures of each level receive vertical messages, includes:

[0026] The vertical messages passed from the three bodies in the third level to the two bodies in the second level are as follows:

[0027] e AUP =AGG-UP e ({p n |p n ∈N e}), among which, AGG-UP e p is a vertical aggregation function that includes an addition operation. n For a message of a tribody containing the chemical bond e between two atoms in the two-body, N e A set representing two bodies;

[0028] The vertical messages passed from the four entities of the fourth level and the two entities of the second level to each entity in the first level are as follows:

[0029] v AUP =AGG-UP v ({e n ,b n |e n ,b n ∈N v}), among which, AGG-UP v For vertical aggregation functions, e n For a message containing the atom v in the unified entity, b h For a message containing the atom v in the aforementioned body, N v A collection that represents a single entity.

[0030] Optionally, the step of updating and reading out the initialization messages of each substructure according to the vertical and horizontal messages through the hierarchical geometric graph neural network to obtain the target molecular representation structure of the molecule to be processed includes:

[0031] The four bodies retain the initialization message, and the substructures in the first level to the third level are updated according to their respective horizontal messages, vertical messages and initialization messages to obtain the update messages of each substructure.

[0032] The method further includes:

[0033] The target molecular representation structure of the molecule to be processed is represented as follows: u = Readout(v′)

[0034] Wherein, u is the molecular graph representation of the target molecule representation structure, and Readout is the message reading function, which is implemented through additive pooling.

[0035] Optionally, the update message of the three bodies is: p′ = COM p (p,p AH );

[0036] The update message for the two bodies is: e′ = COM e (e,e AH ,e AUP );

[0037] The update message for the unified system is: v′=COM v (v,v AH ,v AUP );

[0038] Wherein, COM is the message update function including addition and residual implementation, p is the initialization message of the three bodies, e is the initialization message of the two bodies, and v is the initialization message of the one body.

[0039] Optionally, the step of inputting the spatial coordinate information of each atom contained in the multi-body molecule to be processed into a pre-trained hierarchical geometric graph neural network, and converting the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information through the hierarchical geometric graph neural network, includes:

[0040] Obtain the spatial coordinate information of the molecule to be processed, wherein the spatial coordinate information includes the atomic attribute information, chemical bond information, and coordinate file representing the molecular geometry of the molecule to be processed;

[0041] The spatial coordinate information is encoded through the hierarchical geometric graph neural network to obtain the node feature matrix and interior coordinate matrix of the relatively geometrically structured 3D molecule;

[0042] The relative geometric information of the molecule to be processed is obtained based on the node feature matrix and internal coordinate matrix of the relatively geometrically structured 3D molecule.

[0043] Optionally, the method further includes:

[0044] The hierarchical geometric graph neural network to be trained is trained using a training dataset containing multiple multibody molecules, and the gradient of energy with respect to the three-dimensional coordinates of the molecules is calculated, with the gradient as the force.

[0045] The loss value is calculated by comparing the output of the hierarchical geometric graph neural network to be trained with the real labels in the dataset, and then backpropagation is performed. The real labels are the energy of the multibody molecule and the force situation of each atom contained in the multibody molecule.

[0046] After a preset number of training iterations, a trained hierarchical geometric neural network is obtained.

[0047] A second aspect of this application provides a hierarchical geometric graph neural network construction apparatus for molecular representation learning, the apparatus comprising:

[0048] The input module is used to input the spatial coordinate information of each atom contained in the multibody molecule to be processed into a pre-trained hierarchical geometric graph neural network. The hierarchical geometric graph neural network converts the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information.

[0049] The decomposition module is used to decompose the molecule to be processed according to different numbers of atoms based on the relative geometric information through the hierarchical geometric graph neural network, to obtain multiple substructures belonging to different levels, wherein each level includes multiple substructures, and the number of atoms contained in the substructures of each level is the same.

[0050] The calculation module is used to calculate the interaction messages between substructures contained in each level and the internal messages of the substructures through the hierarchical geometric graph neural network, so as to obtain the initialization message of each substructure.

[0051] The horizontal message passing module is used to pass initialization messages of each level through the hierarchical geometric graph neural network to obtain the horizontal messages of each of the substructures.

[0052] The horizontal aggregation and vertical message passing module is used to pass the horizontally aggregated message obtained by aggregating multiple horizontal messages to each substructure contained in other layers in a manner from high-level to low-level through the hierarchical geometric graph neural network, so that each substructure of the layer receives the vertical message.

[0053] The update and readout module is used to update and read out the initialization messages of each of the substructures according to the vertical messages and the horizontal messages through the hierarchical geometric graph neural network, so as to obtain the target molecular representation structure of the molecule to be processed;

[0054] The prediction module is used to predict the molecular properties and dynamics of the target molecule representation structure through the hierarchical geometric graph neural network, so as to obtain the state of the molecule to be processed and the force situation of each atom contained in the molecule to be processed.

[0055] The beneficial effects of this application are:

[0056] This application provides a method and apparatus for constructing a hierarchical geometric graph neural network (HLM) for molecular representation learning. The HLM, trained using deep learning, converts the spatial coordinates of each atom in the molecule to be processed into relative geometric information. The molecule is then decomposed according to different numbers of atoms, resulting in substructures at different levels. Initialization messages for each substructure are obtained, and horizontal and vertical message passing of molecular geometric information occurs between different levels, leading to a more complete representation of the target molecule. This target molecular representation helps the HLM understand molecular geometric features, thereby enabling downstream tasks such as predicting molecular properties and dynamics. This achieves an effective and complete representation of molecular geometry, facilitating the modeling and learning of molecular geometry by the HLM, and enabling various computational chemistry downstream tasks. Attached Figure Description

[0057] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0058] To more clearly illustrate the technical solution of this application, the drawings used in the description of this application 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.

[0059] Figure 1 This is a flowchart of a hierarchical geometric graph neural network construction method for molecular representation learning provided in one embodiment of this application;

[0060] Figure 2 This is a schematic diagram of the internal messages of the alanine molecule from one to four forms provided in the embodiments of this application;

[0061] Figure 3 This is a schematic diagram of the interaction messages between the monoclonal and tetraclonal forms of the alanine molecule provided in the embodiments of this application;

[0062] Figure 4 This application provides a schematic diagram of the horizontal and vertical message passing of the alanine molecule from the first to the fourth level in an embodiment;

[0063] Figure 5 This is a schematic diagram of the molecular geometry information preprocessing of the molecule to be processed according to the embodiments of this application;

[0064] Figure 6 This is a schematic diagram of a hierarchical geometric graph neural network in an embodiment of this application;

[0065] Figure 7 This is an architecture diagram of each layer of the hierarchical geometric graph neural network in the embodiments of this application;

[0066] Figure 8 This is a functional block diagram of a hierarchical graph neural network for molecular dynamics prediction, provided in an embodiment of this application.

[0067] Figure 9 This is a schematic diagram of the framework of a hierarchical geometric graph neural network construction device for molecular representation learning provided in one embodiment of this application. Detailed Implementation

[0068] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0069] 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, 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.

[0070] Graph Neural Networks (GNNs) are a general term for algorithms that use neural networks to learn graph-structured data, extract and discover features and patterns in graph-structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation.

[0071] Molecular dynamics simulation: Molecular dynamics simulation (MD) uses computers to directly simulate the motion of atoms. That is, by calculating the interactions between microscopic particles, each particle is regarded as moving under the force field provided by all other particles and charges, following Newton's laws. The structure and properties of the system are obtained by statistical averaging of the particles.

[0072] The relevant technologies mainly suffer from the following defects:

[0073] 1) The ability of 2D molecular graph neural network and 3D molecular graph invariant neural network models to model geometric structure and geometric properties is limited, and it is difficult to learn an effective and complete representation of molecular geometry.

[0074] 2) Although equivariant neural networks based on 3D molecular graphs model molecular geometry more completely, such as GemNet, this method has low computational efficiency, high cost, and limited scalability.

[0075] Therefore, to address the aforementioned shortcomings, this application aims to achieve efficient modeling and learning of molecular geometry, and to use the obtained molecular geometry to complete various downstream computational chemistry tasks, thereby meeting the following requirements:

[0076] 1) Users can use hierarchical geometric molecular diagram modeling methods to characterize molecular structures at multiple levels;

[0077] 2) Users can use hierarchical geometric graph neural network design methods to construct efficient hierarchical geometric graph neural network structures;

[0078] 3) Users can use hierarchical geometric graph neural networks to efficiently and accurately predict molecular properties and perform dynamic simulations.

[0079] Figure 1 This is a flowchart of a hierarchical geometric graph neural network construction method for molecular representation learning provided in one embodiment of this application, as follows: Figure 1 As shown in this embodiment, a hierarchical geometric graph neural network construction method for molecular representation learning is provided. The method includes:

[0080] In step S101, the spatial coordinate information of each atom contained in the multibody molecule to be processed is input into a pre-trained hierarchical geometric graph neural network. The hierarchical geometric graph neural network converts the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information.

[0081] In step S102, the molecule to be processed is decomposed according to the relative geometric information based on the hierarchical geometric graph neural network, according to different numbers of atoms, to obtain multiple substructures belonging to different levels. Each level includes multiple substructures, and the number of atoms contained in the substructures of each level is the same.

[0082] In step S103, the hierarchical geometric graph neural network is used to calculate the interaction messages between substructures contained in each level and the internal messages of the substructures to obtain the initialization messages of each substructure.

[0083] In step S104, the initialization messages of each level are transmitted through the hierarchical geometric graph neural network to obtain the lateral messages of each substructure.

[0084] In step S105, the hierarchical geometric graph neural network transmits the horizontally aggregated message obtained by aggregating multiple horizontal messages to the substructures contained in other levels in a manner from high-level to low-level, so that the substructures of each level can obtain vertical messages.

[0085] In step S106, the initialization messages of each substructure are updated and read out according to the vertical messages and the horizontal messages through the hierarchical geometric graph neural network to obtain the target molecular representation structure of the molecule to be processed.

[0086] In step S107, the hierarchical geometric graph neural network is used to predict the molecular properties and dynamics of the target molecule representation structure, thereby obtaining the state of the molecule to be processed and the force situation of each atom contained in the molecule to be processed.

[0087] First, in step S101, the spatial coordinate information of each atom contained in the multibody molecule to be processed is input into a pre-trained hierarchical geometric graph neural network. The hierarchical geometric graph neural network converts the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information.

[0088] In this embodiment, the target molecular representation structure of the molecule to be processed is obtained through a pre-trained hierarchical geometric neural network. In particular, for multi-body molecules, the spatial coordinate information of each atom contained in the molecule to be processed needs to be input into the pre-trained hierarchical geometric neural network to convert the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information. The spatial coordinate information is the structural and unstructured information of the molecule to be processed in 3D form. By processing the spatial coordinate information of the molecule to be processed into relative geometric information, the properties of the molecule to be processed can be kept unchanged after translation or rotation.

[0089] Furthermore, in step S102, the molecule to be processed is decomposed according to the relative geometric information based on the hierarchical geometric graph neural network to obtain multiple substructures belonging to different levels, wherein each level includes multiple substructures and the number of atoms contained in the substructures of each level is the same.

[0090] In this embodiment of the application, after obtaining the relative geometric information of the molecule to be processed, the molecule to be processed is decomposed based on the obtained relative geometric information. The decomposition process is carried out sequentially according to the principle of different atomic numbers, resulting in multiple substructures belonging to different levels. That is, each level includes multiple substructures, and the number of atoms contained in the substructures of each level is the same, while the number of atoms contained in the substructures of different levels is different.

[0091] Furthermore, in step S103, the interaction messages between substructures contained in each level and the internal messages of the substructures are calculated through the hierarchical geometric graph neural network to obtain the initialization message of each substructure.

[0092] In this embodiment of the application, the internal messages of each substructure contained in each level and the interaction messages between substructures in each level are calculated by using a hierarchical geometric graph neural network to obtain the initialization message of each substructure. The internal message is the atomic information and structural characteristics of the substructure itself, and the interaction information is the relationship between the substructure and its multiple adjacent substructures. Each substructure has an initialization message.

[0093] Furthermore, in step S104, the initialization messages of each level are transmitted through the hierarchical geometric graph neural network to obtain the lateral messages of each substructure.

[0094] In this embodiment of the application, the initialization messages of each substructure contained in each layer are transmitted at each layer through a hierarchical geometric graph neural network, so that the initialization message of each substructure is transmitted to its adjacent substructure. After obtaining the initialization message of the adjacent substructure, each substructure obtains the horizontal message of each substructure.

[0095] Furthermore, in step S105, the hierarchical geometric graph neural network transmits the horizontally aggregated message obtained by aggregating multiple horizontal messages to the various substructures contained in other levels in a manner from high-level to low-level, so that the substructures of each level receive vertical messages.

[0096] In this embodiment of the application, a hierarchical geometric graph neural network is used to aggregate the horizontal messages obtained by each substructure in step S104 through horizontal message passing at each level to obtain horizontal aggregated messages. The horizontal aggregated messages at each level are used for vertical message passing to each lower level relative to that level, so that each substructure contained in each lower level relative to that level obtains the vertical messages of the higher level. Here, lower level or higher level are relative concepts. The number of atoms in each substructure in the higher level is more than the number of atoms in each substructure in the lower level. Furthermore, the direction of vertical message passing is unidirectional. Among them, the transmission effect is best when the message is passed from the higher level to the lower level.

[0097] Furthermore, in step S106, the initialization messages of each substructure are updated and read out according to the vertical messages and the horizontal messages through the hierarchical geometric graph neural network to obtain the target molecular representation structure of the molecule to be processed.

[0098] In this embodiment of the application, a hierarchical geometric graph neural network is used to aggregate the horizontal messages obtained by horizontal message passing and the vertical messages obtained by vertical message passing for each substructure in each level. The aggregated messages are then used to update the initialization messages of each substructure, so that each substructure receives the updated messages and reads the message representation of the molecule to be processed, thereby obtaining the target molecule representation structure of the molecule to be processed.

[0099] Finally, in step S107, the molecular properties and dynamics of the target molecule representation structure are predicted using the hierarchical geometric graph neural network to obtain the state of the molecule to be processed and the force situation of each atom contained in the molecule to be processed.

[0100] In this embodiment, the hierarchical geometric graph neural network predicts the chemical and dynamic properties of the molecule to be processed using the aforementioned atomic spatial coordinates, atomic information (including type, quantity, etc.), and the target molecule representation structure. This is used for subsequent task prediction of molecular properties and dynamics. All pooling methods are incorporated to obtain the target molecule representation structure, which is used to predict the molecular properties and dynamics of the molecule to be processed, thereby obtaining the state of the molecule and the forces acting on each atom within it.

[0101] Through the above embodiments, the spatial coordinate information of each atom in the molecule to be processed is converted into relative geometric information. The molecule is then decomposed according to different numbers of atoms to obtain substructures at different levels. Initialization messages for each substructure are then obtained, and horizontal and vertical message passing of molecular geometry messages is performed between different levels to obtain a more complete representation of the target molecule. This target molecule representation helps the hierarchical geometric graph neural network understand molecular geometric features, thereby completing downstream tasks such as predicting molecular properties and dynamics. Achieving an effective and complete representation of molecular geometry facilitates the modeling and learning of molecular geometry by the hierarchical geometric graph neural network, enabling the completion of various downstream tasks in computational chemistry.

[0102] Optionally, the hierarchy includes a first level, a second level, a third level, and a fourth level, and the substructure includes one-body, two-body, three-body, and four-body structures. The process involves using the hierarchical geometric graph neural network to decompose the molecule to be processed according to the selected number of atoms based on the relative geometric information, resulting in multiple substructures at different levels, including:

[0103] The hierarchical geometric graph neural network decomposes the molecule to be processed into a first level based on the relative geometric information, according to the number of atoms. The first level includes multiple units, the internal message of each unit is the atomic number of the unit, and the interaction message between each unit is the distance between atoms.

[0104] The hierarchical geometric graph neural network decomposes the molecule to be processed into two atoms according to the relative geometric information to obtain a second level. The second level includes multiple binary bodies. The internal message of each binary body is the bond length between the two atoms contained in the binary body. The interaction message between each binary body is the bond angle between the bond length of the binary body.

[0105] The hierarchical geometric graph neural network decomposes the molecule to be processed into three atoms according to the relative geometric information to obtain a third level. The third level includes multiple three-body structures. The internal message of each three-body structure is the area formed between the three atoms contained in the three-body structure. The interaction message between each three-body structure is the torsion angle of the area between the three-body structures.

[0106] The hierarchical geometric graph neural network decomposes the molecule to be processed into four atoms according to the relative geometric information to obtain a fourth level. The fourth level includes multiple quadrilaterals. The internal message of each quadrilateral is the volume formed between the four atoms contained in the quadrilateral, and the interaction message between each quadrilateral is the chirality of the volume between the quadrilaterals.

[0107] Specifically, in the embodiments of this application, such as Figure 2 , Figure 3 and Figure 4 As shown, as an example, the molecule to be treated is alanine. Figure 2 This is a schematic diagram of the internal messages of the alanine molecule from one to four forms provided in the embodiments of this application. Figure 3 This is a schematic diagram of the interaction messages between the monoclonal and tetraclonal forms of alanine molecule provided in an embodiment of this application. Figure 4 This application provides a schematic diagram of the horizontal and vertical message passing of the alanine molecule from the first to the fourth level.

[0108] Based on relative geometric information, the molecule to be processed is decomposed according to different selected numbers of atoms to obtain multiple substructures at different levels. The levels of the molecule to be processed can be divided into four levels, such as... Figure 4 As shown, these represent the first, second, third, and fourth levels, respectively. The first level includes multiple substructures, each containing one atom, and each substructure is a single unit (in...). Figure 2 and Figure 3 The second level includes multiple substructures, each with 2 atoms, and each substructure is a two-body structure (corresponding to 1-body). Figure 2 and Figure 3 The third level includes multiple substructures, each with 3 atoms, making each substructure a three-body structure (corresponding to 2-body). Figure 2 and Figure 3 The fourth level includes multiple substructures, each with 4 atoms, making each substructure a four-body structure (in the context of 3-body structures). Figure 2 and Figure 3 (Corresponding to 4-body).

[0109] Preferably, it can also contain more levels, so the number of atoms in the substructures in each level can also be greater (in Figure 2 and Figure 3 (corresponding to n-body).

[0110] Using a hierarchical geometric graph neural network, the molecule to be processed is decomposed into a single atom based on its relative geometric information to obtain the first level. The first level includes multiple units (i.e., the substructures in the first level are units). The internal message of a unit is the atomic number of the unit, and the interaction message between units is the distance between atoms.

[0111] Using a hierarchical geometric graph neural network, the molecule to be processed is decomposed into two atoms based on the relative geometric information of the molecule to be processed, resulting in a second level. The second level includes multiple diatoms (i.e., the substructures in the second level are diatoms). The internal message of a diatom is the bond length between the two atoms contained in the diatom, and the interaction message between each diatom is the bond angle between the bond lengths of the diatoms.

[0112] Using a hierarchical geometric graph neural network, the molecule to be processed is decomposed into three atoms based on the relative geometric information of the molecule to be processed, resulting in a third level. The third level includes multiple tribodies (i.e., the substructures in the third level are tribodies). The internal message of a tribody is the area formed between the three atoms contained in the tribody (as shown in the figure, it is the area formed by the line connecting the midpoints of the three atoms), and the interaction message between each tribody is the torsion angle of the area between the tribodies.

[0113] Using a hierarchical geometric graph neural network, the molecule to be processed is decomposed into four atoms based on the relative geometric information of the molecule to be processed, resulting in a fourth level. The fourth level includes multiple tetrabodies (i.e., the substructures in the fourth level are tetrabodies). The internal message of a tetrabody is the volume formed between the four atoms contained in the tetrabody (as shown in the figure, it is the volume formed by the line connecting the midpoints of the four atoms). The interaction message between each tetrabody is the chirality (direction) of the volume between the tetrabodies, which is used to distinguish chiral enantiomers.

[0114] Optionally, the step of passing initialization messages at each level through the hierarchical geometric graph neural network to obtain lateral messages for each substructure includes:

[0115] Based on the initialization messages of each substructure, lateral message passing is performed on the adjacent substructures to obtain the lateral messages of each substructure. The initialization messages of the four bodies do not need to be laterally passed. The initialization messages of the four bodies are used to directly pass to the substructures of other levels except the fourth level.

[0116] Specifically, such as Figure 4 As shown in this embodiment, the initialization message of each substructure at each level is the interaction message and internal message of that substructure. Horizontal message passing involves each substructure passing its own initialization message to other adjacent substructures at the same level, and receiving initialization messages from other adjacent substructures. This achieves internal messages between substructures in the aggregation domain and interaction messages between two substructures. It should be noted that the initialization message of the fourth body does not require horizontal message passing, but is directly passed to substructures at other levels except the fourth level through vertical message passing.

[0117] Optionally, the lateral messages of each of the three bodies in the third level are:

[0118] Among them, AGG-H p This is a three-body horizontal aggregation function, including the Hadamard operation and the cumulative operation, p h For each of the aforementioned tripartites, there is a tripartite adjacent to it. For the interactive geometry embedding corresponding to the third level, N p A set representing three bodies;

[0119] The horizontal messages of each two bodies in the second level are:

[0120] e AH =AGG-H e ({e h ,e CBF (θ)}|e h ∈N e}), where AGG-H e For two-body lateral aggregation functions, e h For each of the two bodies adjacent to the aforementioned two bodies, e CBF (θ) represents the interactive geometry embedding corresponding to the second level, N e A set representing two bodies;

[0121] The horizontal messages of each entity in the first level are:

[0122] v AH =AGG-H v ({v h ,e RBF (d)}|v h ∈Nv}), where AGG-H v As a unified horizontal aggregation function, v h For an entity adjacent to each of the aforementioned entities, e RBF (d) represents the interactive geometry embedding corresponding to the first level, N v A collection that represents a single entity.

[0123] Specifically, in this embodiment, the horizontal messages obtained by each tripartite in the third level through horizontal message passing are: Among them, AGG-H p This is a three-body horizontal aggregation function, including the Hadamard operation and the cumulative operation, p h For each tribosome adjacent to another tribosome, For the interactive geometry embedding corresponding to the third level, N p This represents a set of three bodies; the horizontal message obtained by each two bodies in the second level through horizontal message passing is e. AH =AGG-H e ({e h ,e CBF (θ)}|e h ∈N e}), where AGG-H e For two-body lateral aggregation functions, e h For each two-body adjacent to another two-body, e CBF (θ) represents the interactive geometric embedding corresponding to the second level, N e This represents a set of two entities; the horizontal messages obtained by each entity in the first level through horizontal message passing are: v AH =AGG-H v ({v h ,e RBF (d)}|v h ∈N v}), where AGG-H v As a unified horizontal aggregation function, v h For each entity adjacent to another entity, e RBF (d) represents the interactive geometry embedding corresponding to the first level, N v A collection that represents a single entity.

[0124] Optionally, the step of passing the horizontally aggregated message obtained by aggregating multiple horizontal messages to the various substructures contained in other levels through the hierarchical geometric graph neural network in a manner from high-level to low-level, so that the substructures of each level receive vertical messages, includes:

[0125] The vertical messages passed from the three bodies in the third level to the two bodies in the second level are as follows:

[0126] e AUP =AGG-UP e ({p n |p n ∈N e}), among which, AGG-UP e p is a vertical aggregation function that includes an addition operation. n For a message of a tribody containing the chemical bond e between two atoms in the two-body, N e A set representing two bodies;

[0127] The vertical messages passed from the four entities of the fourth level and the two entities of the second level to each entity in the first level are as follows:

[0128] v AUP =AGG-UP v ({e n ,b n |e n ,b n ∈N v}), among which, AGG-UP v For vertical aggregation functions, e n For a message containing the atom v in the unified entity, b h For a message containing the atom v in the aforementioned body, N v A collection that represents a single entity.

[0129] Specifically, in this embodiment, vertical message passing proceeds from a higher level to a lower level. A horizontally aggregated message, obtained by aggregating multiple horizontal messages, is passed to the various substructures contained in other levels (relatively lower levels), so that the substructures at each level (relatively lower level) receive the vertical message. The vertical message passed from the three bodies of the third level to the two bodies in the second level is as follows:

[0130] e AUP =AGG-UP e ({p n |p n ∈N e}), among which, AGG-UP e This is a two-body vertical aggregation function, which includes an addition operation, p n For a message of a tribody containing the chemical bond e between two atoms in the two-body, N e The set of two entities; the vertical message passed from the four entities of the fourth level and the two entities of the second level to each entity in the first level is v. AUP =AGG-UP v({e n ,b n |e n ,b n ∈N v}), among which, AGG-UP v A vertical aggregation function that integrates all functions, e n For a message containing the atom v in the unified entity, b h For a message containing the atom v in the aforementioned body, N v A collection that represents a single entity.

[0131] Optionally, step S106 includes:

[0132] The four bodies retain the initialization message, and the substructures in the first level to the third level are updated according to their respective horizontal messages, vertical messages and initialization messages to obtain the update messages of each substructure.

[0133] The method further includes:

[0134] The target molecular representation structure of the molecule to be processed is represented as follows: u = Readout(v′)

[0135] Wherein, u is the molecular graph representation of the target molecule representation structure, and Readout is the message reading function, which is implemented through additive pooling.

[0136] Specifically, in the embodiments of this application, after horizontal and vertical message passing are performed in the substructures at each level, each substructure has horizontal messages, vertical messages and initialization messages. For each substructure, the horizontal messages, vertical messages and initialization messages are used to update each substructure so that each substructure has update messages. The update messages of each substructure are obtained based on the horizontal messages, vertical messages and initialization messages.

[0137] It should be noted that since the four bodies are not updated, each of the four bodies at the fourth level only has the initialization message, that is, the initialization message is retained as the message of the four bodies.

[0138] Optionally, the update message of the three bodies is: p′ = COM p (p,p AH );

[0139] The update message for the two bodies is: e′ = COM e (e,e AH ,e AUP );

[0140] The update message for the unified system is: v′=COM v (v,v AH ,v AUP )

[0141] Wherein, COM is the message update function including addition and residual implementation, p is the initialization message of the three bodies, e is the initialization message of the two bodies, and v is the initialization message of the one body.

[0142] Specifically, in this embodiment, the update message for the three bodies is: p′ = COM p (p,p AH The update message for the two-body system is: e′ = COM e (e,e AH ,e AUP The update message for the whole system is: v′=COM v (v,v AH ,v AUP In the above formula, COM is the message update function including addition and residual implementation, p is the initialization message of the three-body system, e is the initialization message of the two-body system, and v is the initialization message of the one-body system.

[0143] It is understandable that the update message of the three-body system is obtained based on the initialization message, the vertical message, and the horizontal message of the three-body system; the update message of the two-body system is obtained based on the initialization message, the vertical message, and the horizontal message of the two-body system; and the update message of the unified system is obtained based on the initialization message, the vertical message, and the horizontal message of the unified system.

[0144] Optionally, the step of inputting the spatial coordinate information of each atom contained in the multi-body molecule to be processed into a pre-trained hierarchical geometric graph neural network, and converting the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information through the hierarchical geometric graph neural network, includes:

[0145] Obtain the spatial coordinate information of the molecule to be processed, wherein the spatial coordinate information includes the atomic attribute information, chemical bond information, and coordinate file representing the molecular geometry of the molecule to be processed;

[0146] The spatial coordinate information is encoded through the hierarchical geometric graph neural network to obtain the node feature matrix and interior coordinate matrix of the relatively geometrically structured 3D molecule;

[0147] The relative geometric information of the molecule to be processed is obtained based on the node feature matrix and internal coordinate matrix of the relatively geometrically structured 3D molecule.

[0148] Specifically, Figure 5This is a schematic diagram illustrating the preprocessing of molecular geometry information of the molecule to be processed according to an embodiment of this application, as shown below. Figure 5 As shown, the molecules to be processed are a natural graph structure data, requiring preprocessing of molecular geometry information from multiple molecules. However, currently, the data used in chemistry to represent 3D molecules is usually an SDF file, which contains atomic attribute information, chemical bond information, and coordinates representing the molecular geometry. Through preprocessing including encoding, the node feature matrix and interior coordinate matrix of the structured 3D molecule are obtained.

[0149] By incorporating the atomic property information of the molecule to be processed into the spatial coordinate information (i.e., Figure 5 The system uses nodes, chemical bond information, and coordinate files representing the molecular geometry to encode spatial coordinate information through a hierarchical geometric graph neural network, obtaining a node feature matrix of a relatively geometrically structured 3D molecule. It then performs coordinate transformation on the spatial position information to obtain an inner coordinate matrix. Finally, based on the feature matrix and inner coordinate matrix of the relatively geometrically structured 3D molecule, the system obtains the relative geometric information (i.e., internal messages and interactive messages) of the molecule to be processed.

[0150] In the embodiments of this application, Figure 6 This is a schematic diagram of a hierarchical geometric graph neural network in an embodiment of this application. Figure 7 This is an architecture diagram of each layer of the hierarchical geometric graph neural network in the embodiments of this application, such as... Figure 6 As shown, the design of the hierarchical geometric graph neural network decouples the transmission of several types of geometric information using a hierarchical update strategy. First, two update methods for the message update function in the message passing strategy are proposed: horizontal message passing and vertical message passing. Second, the settings for message aggregation, update, and readout functions are proposed to obtain a complete representation of the molecule to be processed, containing the 3D structure.

[0151] like Figure 7 As shown, the hierarchical geometric graph neural network consists of four initialization modules (1-body initialization, 2-body initialization, 3-body initialization, and 4-body initialization), three interaction modules (1-body interaction layer, 2-body interaction layer, and 3-body interaction layer), residual layers, and the final output. It predicts the chemical and dynamic properties of molecules based on the input atomic numbers and atomic 3D coordinates. Here, Embedding is the embedding representation after substructure initialization, σ(w□) is the linear layer with added activation functions, || represents concatenation, and the output of the final interaction layer is the message representation of each substructure. Since the task objective is to predict molecular properties and dynamics, a pooling method is added for all substructures to obtain the molecular graph representation u of the molecule to be processed, which is used for subsequent prediction tasks of the molecule.

[0152] Optionally, the method further includes:

[0153] The hierarchical geometric graph neural network to be trained is trained using a training dataset containing multiple multibody molecules, and the gradient of energy with respect to the three-dimensional coordinates of the molecules is calculated, with the gradient as the force.

[0154] The loss value is calculated by comparing the output of the hierarchical geometric graph neural network to be trained with the real labels in the dataset, and then backpropagation is performed. The real labels are the energy of the multibody molecule and the force situation of each atom contained in the multibody molecule.

[0155] After a preset number of training iterations, a trained hierarchical geometric neural network is obtained.

[0156] The following section provides a detailed explanation of the process of using hierarchical geometric graph neural networks for molecular dynamics prediction. Figure 8 This is a functional block diagram of a hierarchical graph neural network for molecular dynamics prediction, provided in one embodiment of this application.

[0157] 1) Through the hierarchical geometric molecular graph modeling module, the geometric information preprocessing technology of the molecule to be processed, the internal geometry and interaction geometry modeling technology, and the geometric structure information representation technology for quantum systems are integrated. The molecular information is input and the hierarchical geometric molecular graph representation of the molecular structure is output as effective and complete.

[0158] 2) Through the hierarchical geometric graph neural network module, the hierarchical geometric graph representation is instantiated through message passing, message aggregation, message updating and reading at each level, and the hierarchical geometric graph neural network is modeled based on the hierarchical geometric graph representation;

[0159] 3) Input the hierarchical geometric graph neural network into the molecular dynamics training module, and train the hierarchical geometric graph neural network on the molecular dynamics simulation dataset. The simulation dataset is a training dataset containing multiple multibody molecules. The gradient of the energy with respect to the three-dimensional coordinates of the molecules in the training dataset is calculated. The gradient is used as the force to calculate the loss value with the real label in the training dataset and backpropagation is performed. The real label is the energy of the multibody molecule and the force situation of each atom contained in the multibody molecule. After several rounds of training, a hierarchical geometric graph neural network specifically for molecular dynamics simulation tasks is obtained.

[0160] 4) Input the finely tuned hierarchical geometric neural network into the molecular dynamics simulation module, and use the hierarchical geometric neural network to simulate molecular dynamics motion on a dedicated simulation library.

[0161] Through the above embodiments, the technical effects of this application are as follows:

[0162] 1) This application proposes the concept of hierarchical geometric decomposition of the molecule to be treated, and conducts subsequent design around the molecule to be treated after hierarchical geometric decomposition;

[0163] 2) This application proposes a modeling method for hierarchical geometric decomposition, which takes substructures (one to four bodies) and geometric quantities at different levels as the core, and extends the modeling of the structure of the molecules to be treated to a higher level, achieving efficient and complete modeling results.

[0164] 3) This application proposes the concept of hierarchical geometric graph neural network, which enables the geometric information of the molecule to be processed at each level to be transmitted interactively, and can effectively learn the molecular geometric structure.

[0165] 4) This application proposes using a hierarchical geometric graph neural network for molecular dynamics prediction. Molecular dynamics simulation methods directly simulate the motion of atoms using computers, that is, by calculating the interactions between microscopic particles. In the simulation process, each particle is considered to move according to Newton's laws under the force field provided by all other particles and charges, and the structure and properties of the system are obtained by statistically averaging the particles. However, traditional molecular dynamics simulations are very time-consuming, while machine learning methods can maintain similar simulation results while reducing the time consumption by several degrees of complexity. The hierarchical geometric graph neural network proposed in this application can improve the efficiency and accuracy of the above prediction task.

[0166] Figure 9 This is a schematic diagram of the framework of a hierarchical geometric graph neural network construction device for molecular representation learning provided in one embodiment of this application, as shown below. Figure 9 As shown in the figure, this embodiment provides a hierarchical geometric graph neural network construction device for molecular representation learning, the device comprising:

[0167] The input module 11 is used to input the spatial coordinate information of each atom contained in the multibody molecule to be processed into a pre-trained hierarchical geometric graph neural network. The hierarchical geometric graph neural network converts the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information.

[0168] The decomposition module 12 is used to decompose the molecule to be processed according to different numbers of atoms based on the relative geometric information through the hierarchical geometric graph neural network, to obtain multiple substructures belonging to different levels, wherein each level includes multiple substructures, and the number of atoms contained in the substructures of each level is the same.

[0169] The calculation module 13 is used to calculate the interaction messages between substructures contained in each level and the internal messages of the substructures through the hierarchical geometric graph neural network, so as to obtain the initialization message of each substructure.

[0170] The horizontal message passing module 14 is used to pass initialization messages of each level through the hierarchical geometric graph neural network to obtain the horizontal messages of each substructure.

[0171] The horizontal aggregation and vertical message passing module 15 is used to pass the horizontally aggregated message obtained by aggregating multiple horizontal messages to each substructure contained in other layers in a manner from high-level to low-level through the hierarchical geometric graph neural network, so that each substructure of the layer receives the vertical message.

[0172] The update and readout module 16 is used to update and read out the initialization messages of each of the substructures according to the vertical messages and the horizontal messages through the hierarchical geometric graph neural network, so as to obtain the target molecular representation structure of the molecule to be processed.

[0173] The prediction module 17 is used to predict the molecular properties and dynamics of the target molecule representation structure through the hierarchical geometric graph neural network, so as to obtain the state of the molecule to be processed and the force situation of each atom contained in the molecule to be processed.

[0174] Optionally, the hierarchy includes a first level, a second level, a third level, and a fourth level; the substructure includes one body, two bodies, three bodies, and four bodies; and the decomposition module 12, through the hierarchical geometric graph neural network, includes:

[0175] The first-level decomposition module is used to decompose the molecule to be processed according to the relative geometric information through the hierarchical geometric graph neural network, according to the number of atoms, to obtain the first level. The first level includes multiple units, the internal message of each unit is the atomic number of the unit, and the interaction message between each unit is the distance between atoms.

[0176] The second-level decomposition module is used to decompose the molecule to be processed according to the relative geometric information through the hierarchical geometric graph neural network, according to the number of two atoms, to obtain the second level. The second level includes multiple two-body structures. The internal message of each two-body structure is the bond length formed between the two atoms contained in the two-body structure, and the interaction message between each two-body structure is the bond angle of the bond length between the two-body structures.

[0177] The third-level decomposition module is used to decompose the molecule to be processed into three atoms according to the relative geometric information through the hierarchical geometric graph neural network to obtain the third level. The third level includes multiple three-body structures. The internal message of each three-body structure is the area formed between the three atoms contained in the three-body structure. The interaction message between each three-body structure is the torsion angle of the area between the three-body structures.

[0178] The fourth-level decomposition module is used to decompose the molecule to be processed according to the relative geometric information through the hierarchical geometric graph neural network, according to the number of four atoms, to obtain the fourth level. The fourth level includes multiple quad bodies. The internal message of the quad body is the volume formed between the four atoms contained in the quad body, and the interaction message between each quad body is the chirality of the volume between the quad bodies.

[0179] Optionally, the lateral messaging module 14 includes:

[0180] The horizontal message passing submodule is used to perform horizontal message passing on the adjacent substructures according to the initialization messages of each substructure, so as to obtain the horizontal messages of each substructure. The initialization messages of the four bodies do not need to be horizontally passed. The initialization messages of the four bodies are used to be directly passed to the substructures of other levels except the fourth level.

[0181] Optionally, the lateral messages of each of the three bodies in the third level are:

[0182] Among them, AGG-H p This is a three-body horizontal aggregation function, including the Hadamard operation and the cumulative operation, p h For each of the aforementioned tripartites, there is a tripartite adjacent to it. For the interactive geometry embedding corresponding to the third level, N p A set representing three bodies;

[0183] The horizontal messages of each two bodies in the second level are:

[0184] e AH =AGG-H e ({e h ,e CBF (θ)}|e h ∈N e}), where AGG-H e For two-body lateral aggregation functions, e h For each of the two bodies adjacent to the aforementioned two bodies, e CBF (θ) represents the interactive geometry embedding corresponding to the second level, N e A set representing two bodies;

[0185] The horizontal messages of each entity in the first level are:

[0186] v AH =AGG-H v ({v h ,e RBF (d)}|v h ∈N v}), where AGG-Hv As a unified horizontal aggregation function, v h For an entity adjacent to each of the aforementioned entities, e RBF (d) represents the interactive geometry embedding corresponding to the first level, N v A collection that represents a single entity.

[0187] Optionally, the horizontal aggregation and vertical messaging module 15 includes:

[0188] The second-level vertical message passing module is used to pass vertical messages from the three bodies of the third level to each two bodies in the second level as follows:

[0189] e AUP =AGG-UP e ({p n |p n ∈N e}), among which, AGG-UP e p is a vertical aggregation function that includes an addition operation. n For a message of a tribody containing the chemical bond e between two atoms in the two-body, N e A set representing two bodies;

[0190] The first-level vertical message passing module is used to pass vertical messages from the four entities of the fourth level and the two entities of the second level to each entity in the first level as follows:

[0191] v AUP =AGG-UP v ({e n ,b n |e n ,b n ∈N v}), among which, AGG-UP v For vertical aggregation functions, e n For a message containing the atom v in the unified entity, b h For a message containing the atom v in the aforementioned body, N v A collection that represents a single entity.

[0192] Optionally, the update and read module 16 includes:

[0193] The update module is used to retain the initialization message of the four bodies, and the substructures in the first level to the third level are updated according to their respective horizontal messages, vertical messages and initialization messages to obtain the update messages of each substructure;

[0194] The device further includes:

[0195] The molecular graph representation module is used to represent the target molecular representation structure of the molecule to be processed. The molecular graph representation of the target molecular representation structure of the molecule to be processed is: u = Readout(v′)

[0196] Wherein, u is the molecular graph representation of the target molecule representation structure, and Readout is the message reading function, which is implemented through additive pooling.

[0197] Optionally, the update message of the three bodies is: p′ = COM p (p,p AH );

[0198] The update message for the two bodies is: e′ = COM e (e,e AH ,e AUP );

[0199] The update message for the unified system is: v′=COM v (v,v AH ,v AUP )

[0200] Wherein, COM is the message update function including addition and residual implementation, p is the initialization message of the three bodies, e is the initialization message of the two bodies, and v is the initialization message of the one body.

[0201] Optionally, the input module 11 includes:

[0202] The acquisition module is used to acquire the spatial coordinate information of the molecule to be processed. The spatial coordinate information includes the atomic attribute information, chemical bond information and coordinate file representing the molecular geometry of the molecule to be processed.

[0203] The encoding module is used to encode the spatial coordinate information through the hierarchical geometric graph neural network to obtain the node feature matrix and interior coordinate matrix of the relatively geometrically structured 3D molecule;

[0204] The relative geometric information acquisition module is used to obtain the relative geometric information of the molecule to be processed based on the node feature matrix and internal coordinate matrix of the 3D molecule with relative geometric structure.

[0205] Optionally, the device further includes:

[0206] The training module is used to train the hierarchical geometric graph neural network to be trained using a training dataset containing multiple multibody molecules, and to calculate the gradient of energy with respect to the three-dimensional coordinates of the molecules, using this gradient as the force.

[0207] The loss value calculation module is used to calculate the loss value by comparing the output of the hierarchical geometric graph neural network to be trained with the real labels in the dataset, and to perform backpropagation. The real labels are the energy of the multibody molecule and the force situation of each atom contained in the multibody molecule.

[0208] The multiple training module is used to obtain a trained hierarchical geometric neural network after a preset number of training iterations.

[0209] As the apparatus is basically similar to the method embodiment, it is described in a relatively simple way. For relevant details, please refer to the description of the method embodiment.

[0210] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0211] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0212] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0213] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0214] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0215] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0216] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0217] The above provides a detailed description of the hierarchical geometric graph neural network construction method and apparatus for molecular representation learning provided in this application. Specific examples have been used 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 method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for constructing a hierarchical geometric graph neural network for molecular representation learning, characterized in that, The method includes: The spatial coordinate information of each atom contained in the multibody molecule to be processed is input into a pre-trained hierarchical geometric graph neural network. The hierarchical geometric graph neural network converts the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information. The hierarchical geometric graph neural network decomposes the molecule to be processed according to different numbers of atoms based on the relative geometric information, resulting in multiple substructures belonging to different levels. Each level includes multiple substructures, and the number of atoms contained in the substructures of each level is the same. The hierarchical geometric graph neural network is used to calculate the interaction messages between substructures in each level and the internal messages of the substructures to obtain the initialization messages of each substructure. The hierarchical geometric graph neural network is used to transmit the initialization messages of each level to obtain the horizontal messages of each substructure. Through the hierarchical geometric graph neural network, the horizontally aggregated message obtained by aggregating multiple horizontal messages is passed to the various substructures contained in other levels in a way from high level to low level, so that the substructures of each level can obtain vertical messages. By using the hierarchical geometric graph neural network, the initialization messages of each substructure are updated and read out according to the vertical messages and the horizontal messages, so as to obtain the target molecular representation structure of the molecule to be processed. The hierarchical geometric graph neural network is used to predict the molecular properties and dynamics of the target molecule representation structure, thereby obtaining the state of the molecule to be processed and the force situation of each atom contained in the molecule to be processed.

2. The method according to claim 1, characterized in that, The hierarchy includes a first level, a second level, a third level, and a fourth level. The substructures include one-body, two-body, three-body, and four-body structures. The process involves using the hierarchical geometric graph neural network to decompose the molecule to be processed according to different selected atomic numbers based on the relative geometric information, resulting in multiple substructures at different levels, including: The hierarchical geometric graph neural network decomposes the molecule to be processed into a first level based on the relative geometric information, according to the number of atoms. The first level includes multiple units, the internal message of each unit is the atomic number of the unit, and the interaction message between each unit is the distance between atoms. The hierarchical geometric graph neural network decomposes the molecule to be processed into two atoms according to the relative geometric information to obtain a second level. The second level includes multiple binary bodies. The internal message of each binary body is the bond length between the two atoms contained in the binary body. The interaction message between each binary body is the bond angle between the bond length of the binary body. The hierarchical geometric graph neural network decomposes the molecule to be processed into three atoms according to the relative geometric information to obtain a third level. The third level includes multiple three-body structures. The internal message of each three-body structure is the area formed between the three atoms contained in the three-body structure. The interaction message between each three-body structure is the torsion angle of the area between the three-body structures. The hierarchical geometric graph neural network decomposes the molecule to be processed into four atoms according to the relative geometric information to obtain a fourth level. The fourth level includes multiple quadrilaterals. The internal message of each quadrilateral is the volume formed between the four atoms contained in the quadrilateral, and the interaction message between each quadrilateral is the chirality of the volume between the quadrilaterals.

3. The method according to claim 2, characterized in that, The process of transmitting initialization messages at each level through the hierarchical geometric graph neural network to obtain lateral messages for each substructure includes: Based on the initialization messages of each substructure, lateral message passing is performed on the adjacent substructures to obtain the lateral messages of each substructure. The initialization messages of the four bodies do not need to be laterally passed. The initialization messages of the four bodies are used to directly pass to the substructures of other levels except the fourth level.

4. The method according to claim 3, characterized in that, The lateral messages of each of the three bodies in the third level are: Among them, AGG-H p This is a three-body horizontal aggregation function, including the Hadamard operation and the cumulative operation, p h For each of the aforementioned triplets, there is a triplet adjacent to it. For the interactive geometry embedding corresponding to the third level, N p A set representing three bodies; The horizontal messages of each two bodies in the second level are: e AH =AGG-H e ({e h ,e CBF (θ)}|e h ∈N e }), where AGG-H e For two-body lateral aggregation functions, e h For each of the two bodies adjacent to the aforementioned two bodies, e CBF (θ) represents the interactive geometry embedding corresponding to the second level, N e A set representing two bodies; The horizontal messages of each entity in the first level are: v AH =AGG-H v ({v h ,e RBF (d)}|v h ∈N v }), where AGG-H v As a unified horizontal aggregation function, v h For an entity adjacent to each of the aforementioned entities, e RBF (d) represents the interactive geometry embedding corresponding to the first level, N v A collection that represents a single entity.

5. The method according to claim 2, characterized in that, The method of using the hierarchical geometric graph neural network, from higher to lower levels, to pass the horizontally aggregated message obtained by aggregating multiple horizontal messages to the various substructures contained in other levels, so that the substructures of each level receive vertical messages, includes: The vertical messages passed from the three bodies in the third level to the two bodies in the second level are as follows: e AUP =AGG-UP e ({p n |p n ∈N e }), among which, AGG-UP e p is a vertical aggregation function that includes an addition operation. n For a message of a tribody containing the chemical bond e between two atoms in the two-body, N e A set representing two bodies; The vertical messages passed from the four entities of the fourth level and the two entities of the second level to each entity in the first level are as follows: v AUP =AGG-UP v ({e n ,b n |e n ,b n ∈N v }), among which, AGG-UP v For vertical aggregation functions, e n For a message containing the atom v in the unified entity, b h For a message containing the atom v in the aforementioned body, N v A collection that represents a single entity.

6. The method according to claim 2, characterized in that, The step of updating and reading the initialization messages of each substructure according to the vertical and horizontal messages through the hierarchical geometric graph neural network to obtain the target molecular representation structure of the molecule to be processed includes: The four bodies retain the initialization message, and the substructures in the first level to the third level are updated according to their respective horizontal messages, vertical messages and initialization messages to obtain the update messages of each substructure. The method further includes: The target molecular representation structure of the molecule to be processed is represented as follows: u = Readout(v′) Wherein, u is the molecular graph representation of the target molecule representation structure, and Readout is the message reading function, which is implemented through additive pooling.

7. The method according to claim 6, characterized in that, The update message for the three bodies is: p′ = COM p (p,p AH ); The update message for the two bodies is: e′ = COM e (e,e AH ,e AUP ); The update message for the unified system is: v′=COM v (v,v AH ,v AUP ) Wherein, COM is the message update function including addition and residual implementation, p is the initialization message of the three bodies, e is the initialization message of the two bodies, and v is the initialization message of the one body.

8. The method according to claim 1, characterized in that, The step of inputting the spatial coordinate information of each atom in the multi-body molecule to be processed into a pre-trained hierarchical geometric graph neural network, and converting the spatial coordinate information of each atom in the molecule to be processed into relative geometric information through the hierarchical geometric graph neural network, includes: Obtain the spatial coordinate information of the molecule to be processed, wherein the spatial coordinate information includes the atomic attribute information, chemical bond information, and coordinate file representing the molecular geometry of the molecule to be processed; The spatial coordinate information is encoded through the hierarchical geometric graph neural network to obtain the node feature matrix and interior coordinate matrix of a relatively geometrically structured 3D molecule. The relative geometric information of the molecule to be processed is obtained based on the node feature matrix and internal coordinate matrix of the relatively geometrically structured 3D molecule.

9. The method according to claim 1, characterized in that, The method further includes: The hierarchical geometric graph neural network to be trained is trained using a training dataset containing multiple multibody molecules, and the gradient of energy with respect to the three-dimensional coordinates of the molecules is calculated, with the gradient as the force. The loss value is calculated by comparing the output of the hierarchical geometric graph neural network to be trained with the real labels in the dataset, and then backpropagation is performed. The real labels are the energy of the multibody molecule and the force situation of each atom contained in the multibody molecule. After a preset number of training iterations, a trained hierarchical geometric neural network is obtained.

10. A hierarchical geometric graph neural network construction device for molecular representation learning, characterized in that, The device includes: The input module is used to input the spatial coordinate information of each atom contained in the multibody molecule to be processed into a pre-trained hierarchical geometric graph neural network. The hierarchical geometric graph neural network converts the spatial coordinate information of each atom contained in the molecule to be processed into relative geometric information. The decomposition module is used to decompose the molecule to be processed according to different numbers of atoms based on the relative geometric information through the hierarchical geometric graph neural network, to obtain multiple substructures belonging to different levels, wherein each level includes multiple substructures, and the number of atoms contained in the substructures of each level is the same. The calculation module is used to calculate the interaction messages between substructures contained in each level and the internal messages of the substructures through the hierarchical geometric graph neural network, so as to obtain the initialization message of each substructure. The horizontal message passing module is used to pass initialization messages of each level through the hierarchical geometric graph neural network to obtain the horizontal messages of each of the substructures. The horizontal aggregation and vertical message passing module is used to pass the horizontally aggregated message obtained by aggregating multiple horizontal messages to each substructure contained in other layers in a manner from high-level to low-level through the hierarchical geometric graph neural network, so that each substructure of the layer receives the vertical message. The update and readout module is used to update and read out the initialization messages of each of the substructures according to the vertical messages and the horizontal messages through the hierarchical geometric graph neural network, so as to obtain the target molecular representation structure of the molecule to be processed; The prediction module is used to predict the molecular properties and dynamics of the target molecule representation structure through the hierarchical geometric graph neural network, so as to obtain the state of the molecule to be processed and the force situation of each atom contained in the molecule to be processed.

Images