Method for pre-training using graph information bottleneck based on subgraph and system thereof

Abstract

The present invention relates to a pre-learning method and system using a subgraph-based graph information bottleneck technique. According to the present invention, a pre-learning method using a subgraph-based graph information bottleneck technique comprises: a step of an input unit receiving a molecular graph represented as a graph; a step of a compression unit compressing the molecular graph to derive a graph core; and a step of a restoration unit identifying at least one subgraph in the molecular graph, extracting information of each subgraph, and restoring the molecular graph using the graph core and the information of each subgraph. As such, according to the present invention, by pre-training a graph neural network using a subgraph-based graph information bottleneck technique without prior knowledge of annotations or functional groups, functional groups of a molecular structure represented as a graph can be identified, and the graph of an input molecule can be restored using the identified subgraph and a compressed graph core.

Description

Technology Field

[0001] The present invention relates to a pre-training method and system using a subgraph-based graph information bottleneck technique, and more specifically, to a pre-training method and system using a subgraph-based graph information bottleneck technique that identifies the subgraphs of a molecule without annotations or prior knowledge, and trains a model to restore the graph of an input molecule using the identified subgraphs and compressed graph cores. Background Technology

[0002] Graph neural networks (GNNs) are artificial neural networks that use data with a graph structure as input and are recently being used to predict molecular properties. There are three representative methods for pre-training graph neural networks to predict molecular properties: node-level pre-training, node-level contrastive training, and pre-training of functional groups (or subgraphs) on the molecular structure.

[0003] First, when pre-training a graph neural network through node-level pre-training, there is a limitation in that the higher-order structure of functional groups cannot be fully utilized, making it impossible to reconstruct the graph at the graph level. In the case of node-level contrastive training, there is a disadvantage in that it is difficult to reconstruct a graph similar to the original graph because the structure and characteristics of the molecule cannot be considered. Furthermore, when performing pre-training on functional groups in molecular structure, there is a limitation in that although data for each functional group is required in advance, it is difficult to perform pre-training using data for all functional groups.

[0004] Meanwhile, the Conditioned Graph Information Bottleneck (CGB) is a model based on the principles of information bottleneck theory and is used to identify frequently used patterns in molecular structures represented as graph structures. The Conditioned Graph Information Bottleneck represents molecular structures as graphs and predicts conditional variables based on two input graphs.

[0005] There is a need for a technique that performs pre-training on graph neural networks without prior knowledge of annotations or operators by utilizing such graph information bottleneck techniques.

[0006] The technology forming the background of the present invention is described in Korean Published Patent No. 10-2024-0153883 (published on October 24, 2024). The problem to be solved

[0007] As such, the present invention aims to provide a pre-training method and system using a subgraph-based graph information bottleneck technique, which identifies a subgraph of a molecule without annotations or prior knowledge, and trains a model to restore the graph of an input molecule using the identified subgraph and a compressed graph core. means of solving the problem

[0008] According to an embodiment of the present invention for achieving such technical challenges, a pre-learning method using a subgraph-based graph information bottleneck technique may include: a step of receiving a molecular graph represented as a graph by an input unit; a step of deriving a graph core by a compression unit compressing the molecular graph; and a step of a restoration unit identifying at least one subgraph in the molecular graph, extracting information of each subgraph, and restoring the molecular graph using the graph core and the information of each subgraph.

[0009] The above derivation step may include: a step of adding a preset noise through a first encoder to each node embedding (H) of the molecular graph; a step of learning a probability for each node when noise is added to the node embedding; and a step of deriving a graph core by compressing the molecular graph through the node embedding to which the preset noise has been added.

[0010] The step of learning the above probability can learn the probability using a multilayer perceptron and a sigmoid function for each of the above nodes through the following mathematical formula:

[0011]

[0012] Here, is the i-th node( ) is the probability of the i-th node using a multilayer perceptron (MLP) and a sigmoid function, and H is a node embedding with noise added through the first encoder, and is the average value of the node embeddings, and is the variance of the node embedding, and ε is noise sampled from the node embedding based on a Gaussian distribution, and is a value sampled from a Bernoulli distribution parameterized by the probability of the i-th node, and is the node representation of the i-th node, and is the row vector of H and is the feature vector for the i-th node.

[0013] The step of deriving a graph core by compressing the molecular graph through the above node embeddings can be performed by compressing the molecular graph through the above node embeddings with added noise to generate a node representation of the graph core, and deriving the graph core based on the node representation of the graph core.

[0014] The above restoration step may include: a step of extracting a plurality of subgraph candidates using neighbor nodes within a pre-designated k-hop centered on each node of the molecular graph; a step of sequentially applying a second encoder and a pooling function to each of the subgraph candidates to extract node embeddings of each of the subgraph candidates; a step of selecting subgraphs from among the subgraph candidates; and a step of restoring the molecular graph using the selected subgraphs and the graph core.

[0015] The step of extracting the node embeddings of the above subgraph candidates can extract the node embeddings of the above subgraph candidates of the above molecular graph through the following mathematical formula:

[0016]

[0017] Here, is a candidate subgraph connected to node (v), and is the second encoder, and is the pooling function of the subgraph candidates, and is the node embedding of a subgraph candidate.

[0018] The step of selecting the subgraph can be performed by applying a pooling function to each subgraph candidate to derive a node representation for each, calculating a normalized attention coefficient for each subgraph candidate, and identifying the correlation between the graph core and the subgraph candidate to select the subgraph from among the plurality of subgraph candidates.

[0019] The step of restoring the molecular graph using the above graph core can reflect the molecular-level characteristics of the selected subgraph by calculating the final node embedding of the subgraph connected to the i-th node through the following mathematical formula:

[0020]

[0021] Here, is the node representation of the i-th node, and is the normalized attention coefficient of the i-th node, and is the node embedding of the subgraph connected to the i-th node.

[0022] In a pre-learning system using a subgraph-based graph information bottleneck technique according to another embodiment of the present invention, the system may include: an input unit that receives a molecular graph represented as a graph; a compression unit that compresses the molecular graph to derive a graph core; and a restoration unit that identifies at least one subgraph in the molecular graph, extracts information of each subgraph, and restores the molecular graph using the graph core and the information of each subgraph. Effects of the invention

[0023] As such, according to the present invention, by pre-training a graph neural network using a subgraph-based graph information bottleneck technique without prior knowledge of annotations or functional groups, functional groups of a molecular structure represented as a graph can be identified, and the graph of an input molecule can be restored using the identified subgraph and a compressed graph core. Brief explanation of the drawing

[0024] Figure 1 is a structural diagram of a pre-learning system using a subgraph-based graph information bottleneck technique according to one embodiment of the present invention. FIG. 2 is a flowchart of a pre-learning method using a subgraph-based graph information bottleneck technique according to another embodiment of the present invention. FIG. 3 is a table comparing the graph classification performance of a first graph restoration model pre-trained according to another embodiment of the present invention and a second graph restoration model pre-trained using conventional technology. FIG. 4 is a table comparing the performance of a first graph restoration model pre-trained according to another embodiment of the present invention and a second graph restoration model pre-trained using conventional technology on a large molecular graph. FIG. 5 is a table comparing the efficiency of a first graph restoration model pre-trained according to another embodiment of the present invention. FIG. 6 is a table comparing the distinguishability using Jensen-Shannon divergence measurements of a first graph restoration model pre-trained according to another embodiment of the present invention and a second graph restoration model pre-trained using conventional technology. FIG. 7 is a table comparing the fidelity scores of a first graph restoration model pre-trained according to another embodiment of the present invention and a second graph restoration model pre-trained using conventional technology. FIG. 8 is a graph showing the results of qualitative verification of the graph interpretation of a first graph restoration model pre-trained according to another embodiment of the present invention. FIG. 9 is a table of the contributions of graph core and subgraph learning of a first graph restoration model pre-trained according to another embodiment of the present invention. Specific details for implementing the invention

[0025] Preferred embodiments according to the present invention will be described in detail below with reference to the attached drawings. In this process, the thickness of lines or the size of components shown in the drawings may be exaggerated for clarity and convenience of explanation.

[0026] Furthermore, the terms described below are defined in consideration of their functions within the present invention, and these may vary depending on the intent or practice of the user or operator. Therefore, the definitions of these terms should be based on the content throughout this specification.

[0027] In the embodiments described below, a pre-learning system (100) using a subgraph-based graph information bottleneck technique is described by specific examples as being performed by a computing device comprising one or more memories or one or more processors capable of performing the following process.

[0028] In the embodiment described below, a pre-learning system (100) using a subgraph-based graph information bottleneck technique receives a molecular graph (G) representing a molecular structure as a graph, identifies a subgraph representing at least one functional group, and compresses the graph to form a graph core ( It derives ) and explains with a specific example how to restore the input molecular graph using the identified subgraph and graph core.

[0029] Figure 1 is a structural diagram of a pre-learning system using a subgraph-based graph information bottleneck technique according to one embodiment of the present invention.

[0030] As illustrated in FIG. 1, the pre-learning system (100) includes an input unit (110), a compression unit (120), and a restoration unit (130).

[0031] First, the input unit (110) can receive a molecular graph that represents a single molecular structure as a graph.

[0032] Specifically, the input unit (110) may receive a molecular graph represented as a single graph, excluding annotations or prior knowledge regarding the molecular structure. In this case, the molecular graph consists of a set of atoms (V) containing N atoms and a set of bonds (E).

[0033] Furthermore, the input unit (110) can receive a molecular structure expressed as a string and convert it into a graph.

[0034] Next, the compression unit (120) can compress the input molecular graph to derive a graph core.

[0035] Specifically, the compression unit (120) generates a node representation (Z) of a graph core by adding a preset noise (ε) through a first encoder to each node embedding (H) of a molecular graph, and can derive a graph core based on the node representation of the graph core. At this time, the first encoder is an encoder built based on a graph neural network (GNN) of the graph information bottleneck (GIB) technique, in which a preset noise is input to important nodes with a low probability and a preset noise is input to unimportant nodes with a high probability.

[0036] Next, the restoration unit (130) can identify at least one subgraph in the input molecular graph, extract information of each subgraph, and restore the molecular graph using the derived graph core and the information of each subgraph. At this time, the information of the subgraph includes the structural pattern and features of the subgraph.

[0037] Specifically, the restoration unit (130) can extract multiple subgraph candidates (Functional group candidates, S) using neighbor nodes within a pre-designated K-hop centered on each node of the input molecular graph.

[0038] Additionally, the restoration unit (130) can learn node embeddings for each subgraph candidate through a second encoder.

[0039] Through this, the restoration unit (130) can effectively extract the structural patterns and features of the subgraph.

[0040] That is, the restoration unit (130) can extract information of each subgraph candidate through a second encoder for each subgraph candidate.

[0041] Additionally, the restoration unit (130) can identify a subgraph among the subgraph candidates through attention-based interaction with the information of the derived graph core and the extracted subgraph candidate, and restore the molecular graph using the identified subgraph and graph core.

[0042] Hereinafter, a pre-learning method using a subgraph-based graph information bottleneck technique, performed by a pre-learning system (100) using a subgraph-based graph information bottleneck technique, will be explained in more detail using FIGS. 2 to 9.

[0043] FIG. 2 is a flowchart of a pre-learning method using a subgraph-based graph information bottleneck technique according to another embodiment of the present invention.

[0044] As shown in FIG. 2, the input unit (110) can receive a molecular graph representing a molecular structure as a graph (S210).

[0045] Specifically, the input unit (110) can receive a molecular graph (G=(V, E)) represented as a single graph excluding annotations or prior knowledge regarding the molecular structure. At this time, the molecular graph is an atomic set containing N atoms ( It consists of a combination set (E) and a molecular graph, and the molecular graph is associated with an adjacency matrix (A) and a feature matrix (X).

[0046] Next, the compression unit (120) can compress the input molecular graph to derive a graph core (S220).

[0047] Specifically, the compression unit (120) can add a preset noise (ε) to each node embedding (H) of the molecular graph through a first encoder (S221). Here, the first encoder is an encoder constructed as a graph neural network based on a graph information bottleneck technique, in which the preset noise is input to important nodes with a low probability and the preset noise is input to unimportant nodes with a high probability.

[0048] At this time, noise that has been set in the node embedding can be added using [Equation 1] below the compression unit (120).

[0049]

[0050] Here, is a node embedding (or feature matrix for all nodes) with noise added through the first encoder, and is a set of real numbers, and d is the dimension size of the node feature vector, and is the first encoder, X is the characteristic matrix of the molecular graph, and A is the adjacency matrix of the molecular graph.

[0051] In addition, when noise is added to the node embedding, the compression unit (120) [Equation 2] below [Equation 2] each node ( Probability using a multi-layer perceptron (MLP) and a sigmoid function in ) ) can be trained (S222).

[0052]

[0053] Here, is the i-th node( is the probability of the i-th node using a multilayer perceptron (MLP) and a sigmoid function in ), and H is the node embedding (or feature matrix for all nodes) with noise added through the first encoder, and is the average value of the node embeddings, and is the variance of the node embedding, and ε is noise sampled from the node embedding based on a Gaussian distribution, and is a value sampled from a Bernoulli distribution parameterized by the probability of the i-th node, and is the node representation of the i-th node, and is the row vector of H and is the feature vector for the i-th node.

[0054] In addition, the compression unit (120) compresses the input molecular graph through a node embedding with added noise and a graph core ( A node representation (Z) of ) is generated, and a graph core can be derived based on the node representation of the graph core (S223).

[0055] At this time, the compression unit (120) can remove unnecessary information from the input molecular graph and derive a graph core containing only essential information by using a compression loss function such as [Equation 3] below.

[0056]

[0057] Here, and N is the number of atoms, and is a value sampled from a Bernoulli distribution parameterized by the probability of the j-th node, and am.

[0058] Next, the restoration unit (130) can identify at least one subgraph in the input molecular graph, extract information of each subgraph, and restore the molecular graph using the derived graph core and the information of each subgraph (S230). At this time, the information of the subgraph includes the structural pattern and features of the subgraph.

[0059] Specifically, the restoration unit (130) can extract multiple subgraph candidates (S) using neighbor nodes within a pre-designated k-hop centered on each node of the input molecular graph (S231).

[0060] Additionally, the restoration unit (130) can extract node embeddings of subgraph candidates by sequentially applying a second encoder and a pooling function (e.g., Sum) to each subgraph candidate (S232). Here, the second encoder is an encoder constructed as a graph neural network based on a graph information bottleneck technique.

[0061] At this time, the restoration unit (130) can extract the node embeddings of the subgraph candidates of the input molecular graph using the [Equation 4] below.

[0062]

[0063] Here, is a candidate subgraph connected to node (v), and is the second encoder, and is the pooling function of the subgraph candidates, and is the node embedding of the subgraph candidate (or the feature matrix for all nodes of the subgraph candidate).

[0064] Additionally, the restoration unit (130) can select a subgraph from among the subgraph candidates through attention-based interaction with the information of the derived graph core and the extracted subgraph candidate (S233).

[0065] To elaborate, the restoration unit (130) applies a pooling function to each subgraph candidate to the node representation ( ) can be derived. In this case, the derived node representation is a single vector representation of the graph core.

[0066] At this time, the restoration unit (130) can ensure that the graph core derived using the contrast loss function calculated through [Equation 5] below maintains a high correlation with the input molecular graph and that the graph-level representation is well distinguished.

[0067]

[0068] Here, B is the number of graphs in the mini-batch, and s is the cosine similarity between the derived graph core and the molecular graph, and And, am.

[0069] Additionally, the restoration unit (130) can calculate a normalized attention coefficient (α) for each subgraph candidate using the following [Equation 6].

[0070]

[0071] Here, is the normalized attention coefficient of the i-th node, and is a node representation, and is the node embedding of the subgraph connected to the i-th node, and is a learnable projection.

[0072] Through this, the correlation between the graph core and the subgraph candidates can be identified to select a subgraph from among multiple subgraph candidates.

[0073] Additionally, the restoration unit (130) can restore the molecular graph using the identified subgraph and graph core (S234).

[0074] At this time, the restoration unit (130) can derive the final node embedding of the subgraph connected to the i-th node through the [Equation 7] below.

[0075]

[0076] Through this, the restoration unit (130) can effectively reflect molecular-level characteristics in the identified subgraph.

[0077] At this time, the restoration unit (130) can evaluate whether it can restore the molecular graph while maintaining important structural information based on the graph core derived and the identified subgraph using the restoration loss function calculated through [Equation 8] below.

[0078]

[0079] Here, V is the set of atoms of the input molecular graph, and A is the adjacency matrix of the input molecular graph, and is the adjacency matrix of the restored molecular graph, and is the Frovenius norm for the difference between the adjacency matrix of the input molecular graph and the adjacency matrix of the restored molecular graph.

[0080] In addition, the restoration unit (130) can improve restoration performance by using the final loss function calculated through [Equation 9] below.

[0081]

[0082] Here, β is a hyperparameter for balancing compression and structure preservation for the molecular graph.

[0083] According to another embodiment of the present invention, a pre-trained graph restoration model can perform domain-oriented generalization on a downstream dataset using unsupervised domain adaptation and restore node feature information for each molecule graph by reconstructing it using a loss function such as [Equation 10] below.

[0084]

[0085] Here, V is the set of atoms in the molecular graph, and is the i-th node (or atom) of the molecular graph, and is the initial feature of the i-th node, and am.

[0086] FIG. 3 is a table comparing the graph classification performance of a first graph restoration model pre-trained according to another embodiment of the present invention and a second graph restoration model pre-trained using conventional technology.

[0087] In Figure 3, ContesxtPred, AttrMasking, and EdgePred are examples of node-level pre-trained second graph reconstruction models, Infomax, JOAO, JOAOv2, GraphCL, and GraphLoG are examples of node-level contrastive pre-trained second graph reconstruction models, and GraphFP, MICRO-Graph, MGSSL, GROVE, So,SGT, and MoAMa are examples of molecular structural functional groups (or subgraphs) pre-trained second graph reconstruction models.

[0088] Referring to Figure 3(a), using the area under curve (AUC) of the receiver operating characteristic (ROC), it can be seen that the graph classification performance of the first and second graph restoration models for multiple datasets is best with respect to the graph classification performance of the first graph restoration model (Ours) pre-trained according to another embodiment of the present invention.

[0089] Referring to FIG. 3(b), it can be seen that the graph classification accuracy of the first graph restoration model, which is pre-trained according to another embodiment of the present invention, is much higher than that of the second graph restoration model, which is pre-trained by the prior art.

[0090] FIG. 4 is a table comparing the performance of a first graph restoration model pre-trained according to another embodiment of the present invention and a second graph restoration model pre-trained using conventional technology on a large molecular graph.

[0091] In Figure 4, the large molecular graphs are Peptides-func and Peptides-struct, ContesxtPred, AttrMasking, and EdgePred are examples of node-level pre-trained second graph reconstruction models, Infomax, JOAO, JOAOv2, GraphCL, and GraphLoG are examples of node-level contrastive pre-trained second graph reconstruction models, and GraphFP, MICRO-Graph, MGSSL, GROVE, SimSGT, and MoAMa are examples of pre-trained second graph reconstruction models of functional groups (or subgraphs) on the molecular structure.

[0092] Referring to FIG. 4, it can be seen that a first graph reconstruction model (S-CGIB(Ours)) pre-trained using domain adaptation according to another embodiment of the present invention has excellent ability to identify subclasses and to reconstruct using graph core and subclass information.

[0093] FIG. 5 is a table comparing the efficiency of a first graph restoration model pre-trained according to another embodiment of the present invention.

[0094] Referring to FIG. 5, it can be seen that a first graph reconstruction model (S-CGIB(Ours)) pre-trained (PT) using domain adaptation (DA) according to another embodiment of the present invention has excellent ability to identify subclasses and to reconstruct using graph core and subclass information.

[0095] In Figure 5, the solid line is a graph using training data, and the dotted line is a graph using validation data.

[0096] FIG. 6 is a table comparing the distinguishability using Jensen-Shannon divergence measurements of a first graph restoration model pre-trained according to another embodiment of the present invention and a second graph restoration model pre-trained using conventional technology.

[0097] Referring to Figure 6, when comparing the distinguishability between learned embeddings using Jensen-Shannon divergence (JSD) measures, it can be seen that for three datasets (benzene, alkane carbonyl, and fluoride carbonyl), the first graph reconstruction model distinguishes subclasses better than the second graph reconstruction model.

[0098] FIG. 7 is a table comparing the fidelity scores of a first graph restoration model pre-trained according to another embodiment of the present invention and a second graph restoration model pre-trained using conventional technology, and FIG. 8 is a graph showing the results of qualitative verification of the graph interpretation of the first graph restoration model pre-trained according to another embodiment of the present invention.

[0099] In Figure 7, ContesxtPred, AttrMasking, and EdgePred are examples of node-level pre-trained second graph reconstruction models, Infomax, JOAO, JOAOv2, GraphCL, and GraphLoG are examples of node-level contrastive pre-trained second graph reconstruction models, and GraphFP, MICRO-Graph, MGSSL, GROVE, and MoAMa are examples of molecular structural functional groups (or subgraphs) pre-trained second graph reconstruction models.

[0100] Referring to FIG. 7, it can be seen that a first graph restoration model (S-CGIB(Ours)) pre-trained using domain adaptation according to another embodiment of the present invention on downstream datasets of mutagenicity, benzene, alkane carbonyl, and fluorine carbonyl achieves the highest fidelity score in two indicators.

[0101] Referring to FIG. 8, it can be seen that for two types of datasets (downstream datasets of alkane carbonyls and fluorine carbonyls), graph interpretation through visualization is interpreted more accurately than the standard by the first graph reconstruction model (S-CGIB) pre-trained according to another embodiment of the present invention.

[0102] FIG. 9 is a table of the contributions of graph core and subgraph learning of a first graph restoration model pre-trained according to another embodiment of the present invention.

[0103] Referring to Fig. 9, Graph Core ( It can be seen that the first graph reconstruction model, which derives both the attention coefficient (α) and the attention coefficient (α), has the best performance.

[0104] In addition, while deriving graph cores is important, it can be seen that identifying subgraphs through attention coefficients is more important in predicting the attributes of the molecule.

[0105] According to the embodiment of the present invention described above, by pre-training a graph neural network using a subgraph-based graph information bottleneck technique without prior knowledge of annotations or functional groups, functional groups of a molecular structure represented as a graph can be identified, and the graph of an input molecule can be restored using the identified subgraph and a compressed graph core.

[0106] The present invention has been described with reference to the embodiments illustrated in the drawings, but this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the present invention should be determined by the technical spirit of the following claims. Explanation of the symbols

[0107] 100: Pre-training system using subgraph-based graph information bottleneck techniques 110: Input section 120: Compression section 130: Restoration section

Claims

Claim 1 A pre-learning method using a subgraph-based graph information bottleneck technique comprising: a step of receiving a molecular graph represented as a graph by an input unit; a step of deriving a graph core by a compression unit compressing the molecular graph; and a step of a restoration unit identifying at least one subgraph in the molecular graph, extracting information of each subgraph, and restoring the molecular graph using the graph core and the information of each subgraph. Claim 2 A pre-training method using a subgraph-based graph information bottleneck technique according to claim 1, wherein the deriving step comprises: a step of adding a preset noise through a first encoder to each node embedding (H) of the molecular graph; a step of learning a probability for each node when the noise is added to the node embedding; and a step of deriving a graph core by compressing the molecular graph through the node embedding to which the preset noise has been added. Claim 3 In paragraph 2, the step of learning the probability is a pre-training method using a subgraph-based graph information bottleneck technique that learns the probability using a multilayer perceptron and a sigmoid function at each node through the following mathematical formula: Here, is the i-th node( is the probability of the i-th node using a multilayer perceptron (MLP) and a sigmoid function in ), and H is the node embedding (or feature matrix for all nodes) with noise added through the first encoder, and is the average value of the node embeddings, and is the variance of the node embedding, and ε is noise sampled from the node embedding based on a Gaussian distribution, and is a value sampled from a Bernoulli distribution parameterized by the probability of the i-th node, and is the node representation of the i-th node, and is the row vector of H and is the feature vector for the i-th node. Claim 4 In paragraph 3, the step of deriving a graph core by compressing the molecular graph through the node embedding is a pre-learning method using a subgraph-based graph information bottleneck technique, wherein the molecular graph is compressed through the node embedding with added noise to the preset noise to generate a node representation of the graph core, and the graph core is derived based on the node representation of the graph core. Claim 5 In claim 1, the restoration step comprises: a step of extracting a plurality of subgraph candidates using neighbor nodes within a pre-designated k-hop centered on each node of the molecular graph; a step of sequentially applying a second encoder and a pooling function to each of the subgraph candidates to extract a node embedding of each of the subgraph candidates; a step of selecting a subgraph from among the subgraph candidates; and a step of restoring the molecular graph using the selected subgraph and the graph core. A pre-learning method using a subgraph-based graph information bottleneck technique. Claim 6 In claim 5, the step of extracting the node embeddings of the subgraph candidates comprises a pre-training method using a subgraph-based graph information bottleneck technique that extracts the node embeddings of the subgraph candidates of the molecular graph through the following mathematical formula: Here, is a candidate subgraph connected to node (v), and is the second encoder, and is the pooling function of the subgraph candidates, and is the node embedding of a subgraph candidate. Claim 7 In claim 5, the step of selecting the subgraph is a pre-learning method using a subgraph-based graph information bottleneck technique, wherein a pooling function is applied to each subgraph candidate to derive a node representation, a normalized attention coefficient is calculated for each subgraph candidate, and the correlation between the graph core and the subgraph candidate is identified to select the subgraph from among the plurality of subgraph candidates. Claim 8 In claim 5, the step of restoring the molecular graph using the graph core comprises a pre-learning method using a subgraph-based graph information bottleneck technique that reflects molecular-level characteristics of the selected subgraph by calculating the final node embedding of the subgraph connected to the i-th node through the following mathematical formula: Here, is the node representation of the i-th node, and is the normalized attention coefficient of the i-th node, and is the node embedding of the subgraph connected to the i-th node. Claim 9 A pre-learning system using a subgraph-based graph information bottleneck technique comprising: an input unit receiving a molecular graph represented as a graph; a compression unit deriving a graph core by compressing the molecular graph; and a restoration unit identifying at least one subgraph in the molecular graph to extract information of each subgraph and restoring the molecular graph using the graph core and the information of each subgraph. Claim 10 In claim 9, the compression unit adds a preset noise to each node embedding (H) of the molecular graph through a first encoder, learns a probability for each node when noise is added to the node embedding, and compresses the molecular graph through the node embedding with the preset noise added to derive a graph core, thereby pre-learning system using a subgraph-based graph information bottleneck technique. Claim 11 In claim 10, the above compression unit is a pre-training system using a subgraph-based graph information bottleneck technique that learns probabilities using a multilayer perceptron and a sigmoid function at each of the above nodes through the following mathematical formula: Here, is the i-th node( is the probability of the i-th node using a multilayer perceptron (MLP) and a sigmoid function in ), and H is the node embedding (or feature matrix for all nodes) with noise added through the first encoder, and is the average value of the node embeddings, and is the variance of the node embedding, and ε is noise sampled from the node embedding based on a Gaussian distribution, and is a value sampled from a Bernoulli distribution parameterized by the probability of the i-th node, and is the node representation of the i-th node, and is the row vector of H and is the feature vector for the i-th node. Claim 12 In claim 11, the compression unit compresses the molecular graph through the node embedding with added noise to generate a node representation of the graph core, and derives the graph core based on the node representation of the graph core, thereby creating a pre-learning system using a subgraph-based graph information bottleneck technique. Claim 13 In claim 9, the restoration unit extracts a plurality of subgraph candidates using neighbor nodes within a pre-designated k-hop centered on each node of the molecular graph, sequentially applies a second encoder and a pooling function to each of the subgraph candidates to extract node embeddings of each of the subgraph candidates, selects a subgraph from among the subgraph candidates, and restores the molecular graph using the selected subgraph and the graph core, thereby pre-learning system using a subgraph-based graph information bottleneck technique. Claim 14 In claim 13, the above restoration unit is a pre-learning system using a subgraph-based graph information bottleneck technique that extracts node embeddings of the above subgraph candidates of the above molecular graph through the following mathematical formula: Here, is a candidate subgraph connected to node (v), and is the second encoder, and is the pooling function of the subgraph candidates, and is the node embedding of a subgraph candidate. Claim 15 In claim 13, the restoration unit applies a pooling function to each of the subgraph candidates to derive a node representation, calculates a normalized attention coefficient for each of the subgraph candidates, and identifies the correlation between the graph core and the subgraph candidates to select the subgraph from among the plurality of subgraph candidates, thereby pre-learning a system using a subgraph-based graph information bottleneck technique. Claim 16 In claim 13, the above restoration unit is a pre-learning system using a subgraph-based graph information bottleneck technique that reflects molecular-level characteristics of the selected subgraph by calculating the final node embedding of the subgraph connected to the i-th node through the following mathematical formula: Here, is the node representation of the i-th node, and is the normalized attention coefficient of the i-th node, and is the node embedding of the subgraph connected to the i-th node.

Images