A material property prediction method, terminal and storage medium

Abstract

The present invention discloses a method, terminal and storage medium for predicting the properties of a substance, the method comprising: obtaining a target graph corresponding to the molecular structure of a target substance, obtaining target information according to the target graph; inputting the target information into a trained neural network, processing the target information, and outputting a property prediction result of the target substance. The present invention can provide auxiliary decision-making for the study of the properties of a substance and shorten the time for the study of drug properties by converting the molecular structure of the substance into a corresponding graph and then predicting the properties of the target substance through a trained neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of drug property detection, and in particular to a material property prediction method, a terminal and a storage medium. Background Art

[0002] The purpose of drug property research is to ensure that drugs are safe and effective. It relies on a series of laboratory and clinical tests on drugs manually. Due to the wide range of drug properties, a new drug needs to undergo a large number of tests to test the possibility of various drug properties, which makes drug property research time-consuming.

[0003] Therefore, the prior art still needs to be improved and enhanced. Summary of the invention

[0004] The present invention provides a substance property prediction method, a terminal and a storage medium, aiming to solve the problem in the prior art that the research on drug properties is limited to laboratories and clinical trials, resulting in a long time-consuming research on drug properties.

[0005] A first aspect of the present invention provides a method for predicting material properties, comprising:

[0006] Obtain a target graph corresponding to the molecular structure of the target substance, and obtain target information according to the target graph, wherein the target information includes first codes corresponding to nodes in the target graph, second codes corresponding to connecting edges in the target graph, and connection relationships between the nodes;

[0007] Inputting the target information into a trained neural network, processing the target information, and outputting a property prediction result of the target substance;

[0008] Among them, each node in the target graph corresponds one-to-one to each atom in the molecular structure, and each connecting edge in the target graph corresponds one-to-one to each connecting chemical bond in the molecular structure; the neural network is trained based on multiple groups of training data, and each group of training data includes sample target information and characteristic information of the substance corresponding to the sample target information.

[0009] The material property prediction method, wherein the step of obtaining target information according to the target map comprises:

[0010] Acquire the categories of each atom in the molecular structure, and acquire the first codes corresponding to each node in the target graph according to the categories of each atom;

[0011] The categories of the connecting chemical bonds in the molecular structure are obtained, and the second codes corresponding to the connecting edges in the target graph are obtained according to the categories of the connecting chemical bonds.

[0012] In the material property prediction method, the property prediction result of the target material output by the neural network is a property prediction vector, and the property prediction vector includes the probability that the target material has each preset property.

[0013] In the material property prediction method, the property information of the material corresponding to the sample target information includes information on whether the material corresponding to the sample target information has each of the preset properties.

[0014] The material property prediction method, wherein the neural network includes multiple network layers, the input of each network layer is the output of the previous layer, the multiple network layers include at least one first network layer and at least one second network layer, the first network layer is a feature extraction layer, the second network layer is a graph pooling layer, and the processing of the target information includes:

[0015] In the first network layer, extracting features of each first node according to the features of each first node output by the previous layer and the connection relationship between the first nodes, and outputting the extracted features of each first node;

[0016] In the second network layer, each second node is merged according to the features of each second node output by the previous layer, and each third node and the features corresponding to each third node are output, wherein the number of the third nodes is less than the number of the second nodes;

[0017] Among them, the first network layer in the neural network is the first network layer, and the initial features corresponding to the respective nodes in the target graph are the respective first codes.

[0018] The material property prediction method, wherein, in the first network layer before the first second network layer in the neural network, extracting features of each first node according to the features of each first node output by the previous layer and the connection relationship between each first node comprises:

[0019] For a target first node among the first nodes, features of the target first node are extracted based on features of the target first node output by a previous layer, features of a node connected to the target first node, and a second code corresponding to a connection edge connected to the target first node.

[0020] The material property prediction method, wherein the second network layer includes a global pooling layer and a hierarchical pooling layer.

[0021] In the material property prediction method, when the first second network layer in the neural network is a hierarchical pooling layer, in the first second network layer in the neural network, merging each second node according to the features of each second node output by the previous layer includes:

[0022] The second nodes are merged according to the characteristics of the second nodes and the second codes of the connection edges corresponding to the second nodes.

[0023] The second aspect of the present invention provides a terminal, comprising: a processor, a storage medium communicatively connected to the processor, the storage medium being suitable for storing a plurality of instructions, and the processor being suitable for calling the instructions in the storage medium to execute the steps of implementing any of the above-mentioned material property prediction methods.

[0024] According to a third aspect of the present invention, a storage medium is provided, wherein the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of any of the above-mentioned material property prediction methods.

[0025] Beneficial effects: Compared with the prior art, the present invention provides a material property prediction method, terminal and storage medium. The material property prediction method provided by the present invention converts the molecular structure of the substance into a corresponding graph, and then predicts the properties of the target substance through a trained neural network. It can provide auxiliary decision-making for material property research and shorten the time for drug property research. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flow chart of an embodiment of a material property prediction method provided by the present invention;

[0027] Figure 2 An example diagram of the molecular structure of a target substance in an embodiment of the method for predicting a substance property provided by the present invention;

[0028] Figure 3 A schematic diagram of a target graph corresponding to the molecular structure of a target substance in an embodiment of the material property prediction method provided by the present invention;

[0029] Figure 4 A schematic diagram of a neural network training process in an embodiment of a material property prediction method provided by the present invention;

[0030] Figure 5 Schematic diagram of the structure of the neural network in the embodiment of the material property prediction method provided by the present invention Figure 1 ;

[0031] Figure 6 Schematic diagram of the structure of the neural network in the embodiment of the material property prediction method provided by the present invention Figure 2 ;

[0032] Figure 7 A schematic diagram of the processing process of the first and second network layers of a neural network in an embodiment of the material property prediction method provided by the present invention;

[0033] Figure 8 This is an experimental verification effect diagram of an embodiment of the material property prediction method provided by the present invention;

[0034] Fig. 9 A structural principle diagram of an embodiment of a terminal provided by the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0036] The present invention provides a material property prediction method that can be applied in a terminal, which can be, but is not limited to, various personal computers, laptops, mobile phones, tablet computers, car computers, and portable wearable devices. After obtaining the molecular structure of the target substance, the terminal can predict the properties of the target substance using the material property prediction method provided by the present invention.

[0037] Embodiment 1

[0038] like Figure 1 As shown, the material property prediction method provided by the present invention comprises the steps of:

[0039] S100, obtaining a target graph corresponding to the molecular structure of a target substance, and obtaining target information according to the target graph.

[0040] Specifically, the target substance is a substance that needs to be predicted for properties. A graph is a graphic composed of nodes and edges connecting the nodes. The molecular structure of a substance is composed of atoms and chemical bonds connecting atoms, which is similar to a graph. In the present invention, the molecular structure of the target substance is converted into a graph, and then processed by a graph neural network to achieve prediction of the properties of the substance. Specifically, after obtaining the molecular structure of the target substance, a target graph corresponding to the molecular structure of the target substance is obtained. Specifically, each node in the target graph corresponds to each atom in the molecular structure, and each connecting edge in the target graph corresponds to each connecting chemical bond in the molecular structure. Taking hydroxyurea (Hydroxyurea) as an example, its molecular formula is CH4N2O2, and its molecular structure is as follows: Figure 2 The corresponding target graph is shown in Figure 3 shown.

[0041] The target information includes first codes corresponding to nodes in the target graph, second codes corresponding to edges in the target graph, and connection relationships between nodes. Acquiring target information according to the target graph includes:

[0042] Acquire the categories of each atom in the molecular structure, and acquire the first codes corresponding to each node in the target graph according to the categories of each atom;

[0043] The categories of the connecting chemical bonds in the molecular structure are obtained, and the second codes corresponding to the connecting edges in the target graph are obtained according to the categories of the connecting chemical bonds.

[0044] Specifically, different codes are set for different categories of atomic categories to obtain first codes corresponding to each atomic category, and different codes are set for different categories of chemical bonds to obtain second codes corresponding to each chemical bond category. In a possible implementation, the first code and the second code can be one hot codes. Of course, it can be understood that the first code and the second code can also be other forms of unique codes.

[0045] One-hot encoding is a commonly used data preprocessing method. In this embodiment, each atom of the same category corresponds to the same first encoding, and each chemical bond of the same category corresponds to the same second encoding. When the first encoding and the second encoding are one-hot encoding, the number of elements in the first encoding and the second encoding can be preset. Specifically, the atom category library and the chemical bond category library can be preset. The number of elements in the first encoding can be equal to or greater than the number of atom categories in the atom category library, and the number of elements in the second encoding can be equal to or greater than the number of chemical bond categories in the chemical bond category library. For example, the one-hot encoding corresponding to C atoms and N atoms can be h. C =[1,0,0...0], h N =[0,1,0...0], h C and h N It is a vector containing m elements, where the elements are only 0 or 1. The unique encoding corresponding to single covalent bonds and double covalent bonds can be l Ⅰ =[1,0...0], l Ⅱ =[0,1...0], l Ⅰ and l Ⅱ is a vector containing n elements, where the elements are either 0 or 1.

[0046] Please refer again Figure 1 The material property prediction method provided in this embodiment further includes the steps of:

[0047] S200, inputting the target information into a trained neural network, processing the target information, and outputting a property prediction result of the target substance.

[0048] The neural network is completed based on multiple sets of training data, each set of training data includes sample target information and property information of the material corresponding to the sample target information, so that the neural network has the ability to output the corresponding material properties according to the input target information and realize the ability to predict the material properties. The training process of the neural network can be as follows: Figure 4 The model optimization method in the neural network training process can be selected from SGD (Stochastic Gradient Descent), Adam (Adaptive moment estimation), etc.

[0049] Specifically, the characteristic prediction result of the target substance is a characteristic prediction vector, which includes the probability that the target substance has each preset characteristic. In this embodiment, a characteristic library can be pre-constructed, and the characteristic library contains multiple characteristics, such as whether it will cause abdominal pain, whether it can treat tumor diseases, etc. The characteristic prediction vector is an M-element vector, where M is the number of categories of the characteristics in the characteristic library. The M element in the characteristic prediction vector is a number in the range of [0,1], which indicates the probability that the target substance has this characteristic. For example, assuming that the first element indicates whether it causes abdominal pain, and the second element indicates whether it can treat tumor diseases, then the characteristic prediction vector p=[0.9,0.1,...0.7] indicates that the graph neural network predicts that the drug is more likely to cause abdominal pain and is unlikely to be effective in treating tumor diseases. In the sample data used to train the neural network, the characteristic information of the substance corresponding to the sample target information is also input in the form of a vector, and each element in the vector is "1" or "0", indicating whether the substance corresponding to the sample target information has a certain preset characteristic. It is not difficult to see that the arrangement order of the preset characteristics represented by each element in the characteristic information of the substance corresponding to the sample target information should be consistent.

[0050] The neural network includes multiple network layers, the input of each network layer is the output of the previous layer, the multiple network layers include at least one first network layer and at least one second network layer, the first network layer is a feature extraction layer, the second network layer is a graph pooling layer, and the processing of the target information includes:

[0051] S210. In the first network layer, extract features of each first node according to the features of each first node output by the previous layer and the connection relationship between the first nodes, and output the extracted features of each first node.

[0052] Specifically, the first network layer in the neural network is the first network layer, and the initial features corresponding to the respective nodes in the target graph are the respective first codes. That is to say, in the first network layer in the neural network, features of the respective nodes are extracted according to the respective first codes corresponding to the respective nodes and the connection relationship between the respective nodes.

[0053] S220. In the second network layer, the second nodes are merged according to the features of the second nodes output by the previous layer, and the third nodes and the features corresponding to the third nodes are output, wherein the number of the third nodes is less than the number of the second nodes.

[0054] Graph pooling is the process of integrating nodes in a neighborhood in a graph to obtain new features. During the graph pooling process, the nodes are integrated, so the number of nodes is reduced. Specifically, the second network layer includes a global pooling layer and a hierarchical pooling layer, that is, the second network layer can be a global pooling layer or a hierarchical pooling layer, and the multiple second network layers included in the neural network can all be global pooling layers or all be hierarchical pooling layers, or some can be global pooling layers and some can be hierarchical pooling layers.

[0055] The neural network may further include a graph normalization layer, an activation function, a fully connected layer, etc. The graph normalization layer may use an existing graph normalization network such as GraphNorm, and the activation function may use an existing activation function such as ReLU, Mish, ELU, PReLu, etc. The various layers in the neural network may be combined in a certain manner, such as Figure 5-6 As shown, Figure 5-6 2 is an example of two combinations, where x is the input of the neural network, p is the output of the neural network, concat refers to feature concatenation, FC is a fully connected layer, GMaxP is a global maximum pooling, and GMeanP is a global average pooling. Of course, those skilled in the art can use other combinations.

[0056] It is not difficult to see from the above description that in the first network layer, when extracting features from each first node outputted by the previous layer, the number of each first node is not changed, and only the features of each first node are updated, while in the second network layer, the number of nodes is changed. Without changing the number of nodes, the connection relationship of each node can maintain the connection relationship in the target information unchanged. In this embodiment, when processing the features of the nodes before the connection relationship of the nodes is changed, the features of the connecting edges between the nodes are combined for processing.

[0057] Specifically, in the first network layer before the first second network layer in the neural network, extracting features of each first node according to the features of each first node output by the previous layer and the connection relationship between the first nodes includes:

[0058] For a target first node among the first nodes, features of the target first node are extracted based on features of the target first node output by a previous layer, features of a node connected to the target first node, and a second code corresponding to a connection edge connected to the target first node.

[0059] The first network layer can adopt the structure of the existing graph feature extraction network, such as GCN (graph convolutional network), GAT (graph attention network), GIN (graph isomorphism network), GraphSAGE (SAmple and aggreGatE), etc. In the prior art, there is no difference in the connection edges in the graph. Therefore, when extracting the features of the nodes in the graph feature extraction network, only the features of the nodes are considered, and the features of the connection edges are not considered. In this embodiment, before the connection relationship of the nodes changes, in the first network layer, the features of the connection edges of the target first node are also combined to extract the features of the target first node. Specifically, taking the structure of the first network layer using GAT (graph attention network) as an example, the feature extraction calculation process of the target first node i in the first network layer is:

[0060]

[0061]

[0062] Among them, N i is the neighborhood of node i, that is, the set of nodes connected to node i, h i is the feature of node i outputted by the previous layer, l ijis the feature of the edge connecting node i and node j, that is, the second encoding corresponding to the edge connecting node i and node j. || represents the concatenation of vectors or tensors. a is a single-layer forward neural network. W is an updateable weight matrix, and W is the same for each node. w is an updateable weight matrix applied to the edge feature, and w is the same for each edge. W and w are continuously optimized during the training of the neural network. f is an activation function. h i ' is the calculation output of node i in the first network layer. It can be seen that in the first network layer before the first second network layer, when extracting the characteristics of the nodes, the characteristics of the connecting edges of the nodes are also combined, that is, in the process of predicting the characteristics of the target substance, not only the influence of each atom on the characteristics is considered, but also the influence of the chemical bonds between atoms on the characteristics is considered.

[0063] Of course, those skilled in the art can, based on the above description, combine the features of the connection edges accordingly when using other feature extraction network structures (such as GCN, GIN) to construct the first network layer.

[0064] In the first network layer after the second network layer, the nodes are merged and the original connection relationship between the nodes has been lost. Therefore, the calculation method for extracting the features of the nodes in the first network layer after the second network layer can be consistent with the calculation method in the existing graph feature extraction network.

[0065] The second network layer can adopt the structure of the existing graph pooling network. For example, the global pooling layer can use global average pooling, global maximum pooling, global sum pooling, etc. The hierarchical pooling layer can use existing graph pooling algorithms such as DIFF pool, TopKpool, Eigen pool, SAG pool, Edge pool, etc. to construct. When the second network layer is a global pooling layer, the number of the third node is 1, that is, no matter how many nodes are input, the second network layer only outputs one node feature. When the second network layer is a hierarchical pooling layer, the number of the third node is greater than or equal to 1. The number of the third node can be set, and the pooling ratio can also be set. Assuming that the number of the second nodes input is N_in and the pooling ratio is r, the number of the third nodes output N_out is INT(N_in×r), where INT() represents a rounding operation.

[0066] When the first second network layer in the neural network is a hierarchical pooling layer, in the first second network layer in the neural network, merging the second nodes according to the features of the second nodes output by the previous layer includes:

[0067] The second nodes are merged according to the characteristics of the second nodes and the second codes of the connection edges corresponding to the second nodes.

[0068] In the existing hierarchical pooling layer, the characteristics of the connection edges of the nodes are not considered when the nodes are merged. In this embodiment, in the first and second network layers in the neural network, since the nodes have not been merged before, the connection relationship between the original nodes is still retained in the first and second network layers. When the first and second network layers are hierarchical pooling layers, the characteristics of the nodes can be processed in combination with the characteristics of the connection points corresponding to each node, so that the neural network further considers the influence of chemical bonds when processing the target information to predict the properties of the target substance, and the prediction result is more accurate.

[0069] Take the first and second network layers as Edge pool as an example. Figure 7 As shown, in the network layer, the process of merging each second node input to the network layer is as follows:

[0070] First, the weight matrix v is used to calculate the contraction score for each connection edge. For example, the contraction score r of the connection edge connecting node i and node j is ij yes:

[0071] r ij =r ji =v[h i ||h j ||l ij ];

[0072] Then, the shrinkage score is normalized using the softmax function in the neighborhood to obtain the normalized shrinkage score, namely:

[0073]

[0074] Among them, N i is the neighborhood of node i, that is, the set of nodes connected to node i, h i is the feature of node i outputted by the previous layer, l ij is the feature of the edge connecting node i and node j, that is, the second code corresponding to the edge connecting node i and node j. || represents the concatenation of vectors or tensors. v is continuously optimized during the training of the neural network. s ij Not necessarily equal to s ji, that is, each connecting edge has two normalized shrinkage scores. After obtaining the normalized shrinkage score of each connecting edge, the two unselected nodes with the highest normalized shrinkage score are selected in turn, and the two nodes are merged into one node. The merged node feature can be the average, maximum value or concatenation vector of the original two node features.

[0075] It can be seen that when the first second network layer is a hierarchical pooling layer, the characteristics of the connection edges of the nodes are combined in the first second network layer. That is to say, in the process of predicting the properties of the target substance, not only the influence of each atom on the properties is taken into account, but also the influence of the chemical bonds between atoms on the properties.

[0076] Of course, those skilled in the art can, according to the above description, combine the features of the connection edges accordingly when adopting other hierarchical pooling network structures (such as DIFFpool, TopK pool) to construct the second network layer.

[0077] In the second network layer after the first second network layer, the nodes are merged and the connection relationship between the original nodes has been lost. Therefore, the calculation method for extracting node features in the second network layer after the first second network layer can be consistent with the calculation method in the existing graph pooling network.

[0078] The experimental verification effect diagram of the material property prediction method provided in this embodiment is as follows Figure 8 As shown in the figure, in the verification, the drug data of Drugbank was used to construct a data set of "drug-whether it has the characteristics of treating viral infection diseases", including 998 samples. The ratio of the number of samples in the training set and the test set is 9:1, and the ratio of positive and negative samples in the training set and the test set is 1:1. Among them, the positive sample means that the drug has the characteristics of treating viral infection diseases, and the negative sample means that the drug does not have the characteristics of treating viral infection diseases. The structure of the neural network used is as follows Figure 6 As shown in the figure, the Adam optimization method was used to train 750 epochs. The initial learning rate was 0.0001, and the learning rate was attenuated by half every 150 epochs. Figure 8 The two curves in the figure are training curves on the training set and the test set, respectively. It can be seen from the figure that the accuracy of the training set can reach 96.67%, and the accuracy of the test set can reach 84.04%. The small size of the data set limits the further increase of the accuracy of the test set. However, even when trained on such a small-scale data set, the accuracy of the test set is significantly higher than the random prediction result (50%), which proves that the material property prediction method provided in this embodiment can provide auxiliary decision-making for the property research of drugs.

[0079] In summary, this embodiment provides a method for predicting material properties, which converts the molecular structure of the substance into a corresponding graph and then predicts the properties of the target substance through a trained neural network, thereby providing auxiliary decision-making for material property research and shortening the time for drug property research.

[0080] It should be understood that, although the various steps in the flowcharts given in the accompanying drawings of the present invention are displayed in sequence according to the indications of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless there is a clear description in this article, the execution of the steps in the present invention is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps of the present invention may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0081] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0082] Embodiment 2

[0083] Based on the above embodiment, the present invention further provides a terminal, whose principle block diagram can be shown as follows: Fig. 9The terminal includes a memory 10 and a processor 20. The memory 10 stores a computer program. When the processor 10 executes the computer program, at least the following steps can be implemented:

[0084] Obtain a target graph corresponding to the molecular structure of the target substance, and obtain target information according to the target graph, wherein the target information includes first codes corresponding to nodes in the target graph, second codes corresponding to connecting edges in the target graph, and connection relationships between the nodes;

[0085] Inputting the target information into a trained neural network, processing the target information, and outputting a property prediction result of the target substance;

[0086] Among them, each node in the target graph corresponds one-to-one to each atom in the molecular structure, and each connecting edge in the target graph corresponds one-to-one to each connecting chemical bond in the molecular structure; the neural network is trained based on multiple groups of training data, and each group of training data includes sample target information and characteristic information of the substance corresponding to the sample target information.

[0087] Wherein, acquiring target information according to the target map comprises:

[0088] Acquire the categories of each atom in the molecular structure, and acquire the first codes corresponding to each node in the target graph according to the categories of each atom;

[0089] The categories of the connecting chemical bonds in the molecular structure are obtained, and the second codes corresponding to the connecting edges in the target graph are obtained according to the categories of the connecting chemical bonds.

[0090] The characteristic prediction result of the target substance output by the neural network is a characteristic prediction vector, and the characteristic prediction vector includes the probability that the target substance has each preset characteristic.

[0091] The characteristic information of the substance corresponding to the sample target information includes information on whether the substance corresponding to the sample target information has the respective preset characteristics.

[0092] The neural network includes multiple network layers, the input of each network layer is the output of the previous layer, the multiple network layers include at least one first network layer and at least one second network layer, the first network layer is a feature extraction layer, the second network layer is a graph pooling layer, and the processing of the target information includes:

[0093] In the first network layer, extracting features of each first node according to the features of each first node output by the previous layer and the connection relationship between the first nodes, and outputting the extracted features of each first node;

[0094] Merge the second nodes according to the features of the second nodes output by the previous layer, and output the third nodes and the features corresponding to the third nodes, wherein the number of the third nodes is less than the number of the second nodes;

[0095] Among them, the first network layer in the neural network is the first network layer, and the initial features corresponding to the respective nodes in the target graph are the respective first codes.

[0096] Wherein, in the first network layer before the first second network layer in the neural network, extracting features of each first node according to the features of each first node output by the previous layer and the connection relationship between each first node includes:

[0097] For a target first node among the first nodes, features of the target first node are extracted based on features of the target first node output by a previous layer, features of a node connected to the target first node, and a second code corresponding to a connection edge connected to the target first node.

[0098] The second network layer includes a global pooling layer and a hierarchical pooling layer.

[0099] Wherein, when the first second network layer in the neural network is a hierarchical pooling layer, in the first second network layer in the neural network, merging each second node according to the features of each second node output by the previous layer includes:

[0100] The second nodes are merged according to the characteristics of the second nodes and the second codes of the connection edges corresponding to the second nodes.

[0101] Embodiment 3

[0102] The present invention also provides a storage medium storing one or more programs, which can be executed by one or more processors to implement the steps of the material property prediction method described in the above embodiment.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting material properties, characterized in that: include: Obtain a target graph corresponding to the molecular structure of the target substance, and obtain target information according to the target graph, wherein the target information includes first codes corresponding to nodes in the target graph, second codes corresponding to connecting edges in the target graph, and connection relationships between the nodes; Inputting the target information into a trained neural network, processing the target information, and outputting a property prediction result of the target substance; Wherein, each node in the target graph corresponds one-to-one to each atom in the molecular structure, and each connecting edge in the target graph corresponds one-to-one to each connecting chemical bond in the molecular structure; the neural network is trained based on multiple sets of training data, each set of training data includes sample target information and characteristic information of the substance corresponding to the sample target information; The neural network includes multiple network layers, the input of each network layer is the output of the previous layer, the multiple network layers include at least one first network layer and at least one second network layer, the first network layer is a feature extraction layer, the second network layer is a graph pooling layer, and the processing of the target information includes: In the first network layer, extracting features of each first node according to the features of each first node output by the previous layer and the connection relationship between the first nodes, and outputting the extracted features of each first node; In the second network layer, each second node is merged according to the features of each second node output by the previous layer, and each third node and the features corresponding to each third node are output, wherein the number of the third nodes is less than the number of the second nodes; and the number of the third nodes is greater than or equal to 1; Among them, the first network layer in the neural network is the first network layer, and the initial features corresponding to the respective nodes in the target graph are the respective first codes.

2. The material property prediction method according to claim 1, characterized in that: The acquiring target information according to the target map comprises: Acquire the categories of each atom in the molecular structure, and acquire the first codes corresponding to each node in the target graph according to the categories of each atom; The categories of the connecting chemical bonds in the molecular structure are obtained, and the second codes corresponding to the connecting edges in the target graph are obtained according to the categories of the connecting chemical bonds.

3. The material property prediction method according to claim 1, characterized in that: The characteristic prediction result of the target substance output by the neural network is a characteristic prediction vector, and the characteristic prediction vector includes the probability that the target substance has each preset characteristic.

4. The material property prediction method according to claim 3, characterized in that: The object corresponding to the sample target information The characteristic information of the substance includes information on whether the substance corresponding to the sample target information has the respective preset characteristics.

5. The material property prediction method according to claim 1, characterized in that: In the first network layer before the first second network layer in the neural network, extracting features of each first node according to the features of each first node output by the previous layer and the connection relationship between the first nodes includes: For a target first node among the first nodes, features of the target first node are extracted based on features of the target first node output by a previous layer, features of a node connected to the target first node, and a second code corresponding to a connection edge connected to the target first node.

6. The material property prediction method according to claim 4, characterized in that: The second network layer includes a global pooling layer and a hierarchical pooling layer.

7. The material property prediction method according to claim 6, characterized in that: When the first second network layer in the neural network is a hierarchical pooling layer, in the first second network layer in the neural network, merging the second nodes according to the features of the second nodes output by the previous layer includes: The second nodes are merged according to the characteristics of the second nodes and the second codes of the connection edges corresponding to the second nodes.

8. A terminal, characterized in that: The terminal includes: a processor, a storage medium communicatively connected to the processor, the storage medium being suitable for storing a plurality of instructions, and the processor being suitable for calling the instructions in the storage medium to execute the steps of the material property prediction method described in any one of claims 1 to 7.

9. A storage medium, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the material property prediction method as described in any one of claims 1-7.

Images