Information processing program, information processing device, and information processing method

Abstract

The present invention makes it possible, in classification processing on input graph-structured data by machine learning models (110), to: acquire a contribution degree (T1) in the classification processing for each of a plurality of partial regions included in the graph-structured data; and select a machine learning model into which interpretability has been factored by determining an evaluation for the machine-learning models (110) on the basis of the similarity between the contribution degree (T1) and specified information (T2) about the partial region of the graph-structured data.

Description

Information processing program, information processing device, and information processing method

[0001] The present invention relates to an information processing program, an information processing device, and an information processing method.

[0002] Advances in computing have led to the use of machine learning to analyze large-scale protein 3D structure information. If 3D structure information can be used to distinguish (multi-value classification) types of virus strains (antigen clusters), it is expected that analysis of virus mutations, which is essential for vaccine development, can be improved.

[0003] In the classification prediction problem of antigen clusters of virus strains, it is becoming possible to perform classification prediction with higher accuracy by utilizing three-dimensional structural information.

[0004] Furthermore, it is expected that detailed analysis of the classification prediction results will lead to more advanced analysis of viral mutations.

[0005] Therefore, it is necessary to build a machine learning model that makes it easier to interpret the reasons behind the classification prediction results.

[0006] For example, protein three-dimensional structural information is converted into graph data, and a machine learning model is created based on this graph data.

[0007] Conventionally, multiple methods are prepared to represent input graph data as graph models by combining information such as three-dimensional structure information, and these are used as candidates. The candidates are compared using a classification prediction accuracy index, and the candidate with the highest accuracy is selected from these candidates, and this candidate is considered to be the graph model that is also easy to interpret.

[0008] Japanese Patent Application Publication No. 2019-159432 U.S. Patent No. 11176462 International Publication No. 2018 / 052131 JP 2022-536343 JP 2018-506963 JP U.S. Patent Application Publication No. 2022 / 0076783

[0009] However, a machine learning model with high classification accuracy is not necessarily an easy-to-interpret machine learning model, and in such cases, there is a problem that an easy-to-interpret machine learning model is not selected.

[0010] In one aspect, the present invention aims to enable the selection of a machine learning model that takes into account not only classification accuracy but also ease of interpretation.

[0011] Therefore, this information processing program causes a computer to execute a process of acquiring the contribution of each of a plurality of partial areas included in input graph structure data in a classification process using a machine learning model for the input graph structure data, and determining an evaluation of the machine learning model based on the similarity between the contribution and the specification information of the partial areas of the graph structure data.

[0012] According to one embodiment, a machine learning model can be selected that takes into account ease of interpretation.

[0013] FIG. 1 is a diagram schematically illustrating a configuration of an information processing device according to an embodiment; FIG. 2 is a block diagram illustrating an example hardware (HW) configuration of a computer that realizes the functions of an information processing device according to an embodiment; FIG. 3 is a diagram illustrating an example of explanation result information in an information processing device according to an embodiment; FIG. 4 is a diagram illustrating an example of protein position information in an information processing device according to an embodiment; FIG. 5 is a diagram illustrating processing by a coincidence index calculation unit of an information processing device according to an embodiment; FIG. 6 is a diagram illustrating processing by a graph selection unit of an information processing device according to an embodiment; and FIG. 7 is a flowchart illustrating processing by a coincidence index calculation unit and a graph selection unit of an information processing device according to an embodiment.

[0014] Hereinafter, embodiments of the present information processing program, information processing device, and information processing method will be described with reference to the drawings. However, the embodiments shown below are merely examples, and are not intended to exclude various modifications or application of techniques not explicitly stated in the embodiments. In other words, the present embodiment can be implemented in various modifications within the scope of its purpose. Furthermore, each figure does not intend to include only the components shown in the figure, but may include other functions, etc.

[0015] (A) Configuration FIG. 1 is a diagram schematically showing the configuration of an information processing device 1 according to an embodiment.

[0016] The information processing device 1 performs training (machine learning) of a machine learning model that predicts the type of virus strain (antigen cluster) from the amino acid sequence that constitutes the virus protein (training phase).

[0017] In the training phase, the amino acid sequence constituting the protein of a virus is input to the information processing device 1, and the antigen cluster name of the virus is used as the correct answer data.

[0018] Furthermore, the information processing device 1 uses the trained machine learning model to predict (infer) antigen clusters of viruses (prediction phase).

[0019] In the prediction phase, the amino acid sequence that constitutes the protein of a virus is input to the information processing device 1, and the machine learning model predicts the antigen cluster name of the virus.

[0020] (A-1) Hardware Configuration Example The functions of the information processing device 1 according to an embodiment may be implemented by one computer or two or more computers. Furthermore, at least some of the functions of the information processing device 1 may be implemented using hardware (HW) resources and network (NW) resources provided by a cloud environment.

[0021] 2 is a block diagram showing an example of the hardware (HW) configuration of a computer 10 that realizes the functions of the information processing device 1 according to one embodiment. When multiple computers are used as HW resources that realize the functions of the information processing device 1, each computer may have the HW configuration shown in FIG.

[0022] As shown in FIG. 2, the computer 10 may, as a HW configuration, illustratively include a processor 10a, a graphics processing unit 10b, a memory 10c, a storage unit 10d, an IF (Interface) unit 10e, an IO (Input / Output) unit 10f, and a reading unit 10g.

[0023] The processor 10a is an example of a processing unit that performs various controls and calculations, and is a control unit that executes various processes. The processor 10a may be connected to each block in the computer 10 via a bus 10j so that they can communicate with each other. The processor 10a may be a multiprocessor including multiple processors, a multi-core processor having multiple processor cores, or a configuration having multiple multi-core processors.

[0024] Examples of the processor 10a include integrated circuits (ICs) such as a CPU, MPU, APU, DSP, ASIC, and FPGA. Note that the processor 10a may be a combination of two or more of these integrated circuits. CPU is an abbreviation for Central Processing Unit, MPU is an abbreviation for Micro Processing Unit, APU is an abbreviation for Accelerated Processing Unit, DSP is an abbreviation for Digital Signal Processor, ASIC is an abbreviation for Application Specific IC, and FPGA is an abbreviation for Field-Programmable Gate Array.

[0025] The graphics processing device 10b controls screen display for an output device such as a monitor in the IO unit 10f. The graphics processing device 10b may also be configured as an accelerator that executes machine learning processing and prediction processing using a machine learning model. Examples of the graphics processing device 10b include various types of arithmetic processing devices, such as a graphics processing unit (GPU), an APU, a DSP, an ASIC, an FPGA, or other integrated circuits (ICs).

[0026] The memory 10c is an example of HW that stores various types of data, programs, and other information. The memory 10c may be, for example, a volatile memory such as a dynamic random access memory (DRAM) or a non-volatile memory such as a persistent memory (PM), or both.

[0027] The storage unit 10d is an example of HW that stores various types of data, programs, and other information. Examples of the storage unit 10d include various storage devices such as a magnetic disk device such as a hard disk drive (HDD), a semiconductor drive device such as a solid state drive (SSD), and a nonvolatile memory. Examples of nonvolatile memory include a flash memory, a storage class memory (SCM), and a read-only memory (ROM).

[0028] The storage unit 10d may store a program 10h (information processing program) that realizes all or part of the various functions of the computer 10.

[0029] For example, the processor 10a of the information processing device 1 can implement functions in a training phase and a prediction phase, which will be described later, by expanding the program 10h stored in the storage unit 10d into the memory 10c and executing it.

[0030] The IF unit 10e is an example of a communication IF that controls the connection and communication between the computer 10 and other computers. For example, the IF unit 10e may include an adapter that complies with a LAN (Local Area Network) such as Ethernet (registered trademark) or optical communication such as FC (Fibre Channel). The adapter may support either or both wireless and wired communication methods.

[0031] For example, the computer 10 may be connected to other computers and databases (not shown) via the IF unit 10e and a network so as to be able to communicate with each other. The program 10h may be downloaded from the network to the computer 10 via the communication IF and stored in the storage unit 10d.

[0032] The IO unit 10f may include one or both of an input device and an output device. Examples of input devices include a keyboard, a mouse, and a touch panel. Examples of output devices include a monitor, a projector, and a printer. The IO unit 10f may also include a touch panel that combines an input device and an output device. The output device may be connected to the graphics processing device 10b.

[0033] The reading unit 10g is an example of a reader that reads data and program information recorded on the recording medium 10i. The reading unit 10g may include a connection terminal or device to which the recording medium 10i can be connected or inserted. Examples of the reading unit 10g include an adapter compliant with USB (Universal Serial Bus) or the like, a drive device that accesses a recording disk, and a card reader that accesses a flash memory such as an SD card. Note that the recording medium 10i may store the program 10h, and the reading unit 10g may read the program 10h from the recording medium 10i and store it in the memory unit 10d.

[0034] Examples of the recording medium 10i include non-transitory computer-readable recording media such as magnetic / optical disks and flash memories. Examples of magnetic / optical disks include flexible disks, CDs (Compact Discs), DVDs (Digital Versatile Discs), Blu-ray Discs, and HVDs (Holographic Versatile Discs). Examples of flash memories include semiconductor memories such as USB memories and SD cards.

[0035] The above-described HW configuration of the computer 10 is an example. Therefore, the HW in the computer 10 may be increased or decreased (for example, adding or deleting any block), divided, integrated in any combination, or the HW may be added or deleted as needed.

[0036] 1, the information processing device 1 may illustratively include functions as a three-dimensional structure calculation processing unit 101, a feature amount calculation unit 102, a graph generation unit 103, a graph AI calculation processing unit 104, a classification accuracy calculation unit 105, a correspondence processing unit 106, a match index value calculation unit 107, a graph selection unit 108, and a machine learning model 110. These functions may be realized by hardware of a computer 10 (see FIG. 2).

[0037] The 3D structure calculation processor 101 analyzes the 3D structure of a viral protein. When an amino acid sequence (amino acid sequence information) constituting a viral protein is input, the 3D structure calculation processor 101 performs 3D structure analysis of the protein. The 3D structure calculation processor 101 outputs protein 3D structure information as the analysis result. The protein 3D structure information may include, for example, coordinates of each amino acid atom as information representing the 3D structure, and may also include information representing molecular movement.

[0038] The function of the 3D structure calculation processor 101 may be realized using a known protein structure calculation tool. For example, a protein 3D structure prediction tool such as AlphaFold2 may be used as the protein structure calculation tool. The 3D structure calculation processor 101 may output the results of searching a database in which information on protein 3D structures experimentally determined by, for example, X-ray crystallography, NMR (Nuclear Magnetic Resonance), single particle analysis using a cryo-electron microscope, or electron diffraction has been registered in advance. Furthermore, the 3D structure calculation processor 101 may perform, for example, molecular dynamics simulation (Molecular Dynamics; MD) using the 3D structure information output from these methods as initial values.

[0039] In the three-dimensional structure information of a protein, for example, the coordinate values ​​of each amino acid and information representing the movement of the molecule may be associated with identification information that identifies the virus.

[0040] The protein three-dimensional structure information output by the three-dimensional structure calculation processing unit 101 may be stored in a predetermined storage area of ​​the memory 10c or the storage unit 10d.

[0041] The feature calculation unit 102 generates a three-dimensional structure feature based on the protein three-dimensional structure information created by the three-dimensional structure calculation processing unit 101. The three-dimensional structure feature represents a feature of the three-dimensional structure.

[0042] The feature calculation unit 102 may generate the three-dimensional structure feature using a known feature transformation method. For example, the feature calculation unit 102 may perform feature transformation using a statistical method such as support vector regression (SVR), neural network (NN), or principal component analysis (PCA). Alternatively, the feature calculation unit 102 may discretize categorical variables using the knowledge of experts familiar with the field of virology.

[0043] The three-dimensional structural features calculated by the feature calculation unit 102 may be stored in a predetermined storage area of ​​the memory 10c or the storage unit 10d.

[0044] The graph generation unit 103 generates graph information based on the protein 3D structure information generated by the 3D structure calculation processing unit 101 and the 3D structure feature amount generated by the feature amount calculation unit 102. The graph information may be called graph data. The graph information is an example of graph structure data.

[0045] The graph generation unit 103 may convert information about multiple viruses included in the graph information into data in a format that can be processed by the machine learning model 110.

[0046] The graph generating unit 103 may generate multiple types of graph information from one protein three-dimensional structure by, for example, appropriately changing parameters for graph generation, such as the distance between amino acids.

[0047] The graph AI calculation processing unit 104 creates a machine learning model 110 based on the graph information. In the training phase, the graph AI calculation processing unit 104 performs training (machine learning) of the machine learning model 110. The machine learning model 110 may be called a graph AI.

[0048] Here, the machine learning model 110 performs graph-based relationship learning and realizes graph classification (classification: classification processing).

[0049] A graph is a mathematical model that is composed of a set of nodes and a set of edges between the nodes.

[0050] When applying a graph to a virus, amino acids correspond to nodes, and bonds between amino acids correspond to edges. The bonds between amino acids may be, for example, peptide bonds, or other bonds such as bonds based on electrostatic forces.

[0051] The machine learning model 110 performs graph classification based on the graph and edge information. In this case, the machine learning model 110 may use, as explanatory variables, a set of nodes in the graph information created by the graph generation unit 103, a set of edges between nodes included in the node set, and a set of attributes of the nodes included in the node set. In addition, the machine learning model 110 may use, as a target variable, a cluster number corresponding to the antigen cluster.

[0052] In graph classification, parameters for each node may be used as node attributes to help with classification. The machine learning model 110 may output, as a result of graph classification, cluster numbers and probabilities into which the protein 3-dimensional structures are classified.

[0053] To have the machine learning model 110 perform graph classification, it is necessary to explicitly assign edges to the machine learning model 110. Therefore, the graph AI calculation processing unit 104 determines that adjacent amino acids have an edge based on the amino acid sequence. In addition, amino acids within a certain distance due to electrostatic force or the like may also be determined to have an edge.

[0054] The function of the machine learning model 110 can be realized using a known method. For example, the function of the machine learning model 110 may be realized by Deep Tensor (registered trademark).

[0055] Based on the graph information, the graph AI calculation processing unit 104 may create information (graph AI input information) to be input to the machine learning model 110 by arranging the attributes of the two amino acids connected by each edge in the amino acid sequence that constitutes the virus, in units of bonds. The two amino acids connected by an edge may be referred to as an amino acid pair. The amino acid that is the starting point of the edge in the amino acid pair may be referred to as the start node, and the amino acid that is the end point of the edge may be referred to as the end node.

[0056] The machine learning model 110 may be a deep neural network (DNN) that includes multiple hidden layers between an input layer and an output layer.

[0057] For example, a NN inputs input data to an input layer and sequentially executes predetermined calculations in a hidden layer composed of a convolutional layer, a pooling layer, and the like, thereby executing forward processing (forward propagation processing) to sequentially transmit information obtained by the calculations from the input side to the output side. After the forward processing is executed, backward processing (backpropagation processing) is executed to determine parameters to be used in the forward processing in order to reduce the value of the error function obtained from the output data (graph classification result) output from the output layer and the correct data (cluster name). Then, an update processing is executed to update variables such as weights based on the results of the backpropagation processing. For example, gradient descent may be used as an algorithm to determine the update width of the weights used in the backpropagation calculation.

[0058] In addition, in the prediction phase, the graph AI calculation processing unit 104 uses the graph AI input information as input data to cause the machine learning model 110 to perform graph classification, for example, predicting (inferring) cluster numbers and predicting the probability of cluster numbers.

[0059] The graph AI calculation processing unit 104 may input features (three-dimensional structure features) related to the three-dimensional structure of a viral protein into the machine learning model 110 to predict a cluster number.

[0060] In addition, in each of the training phase and the prediction phase, the graph AI calculation processing unit 104 inputs graph AI input information to the machine learning model 110, performs graph classification (class classification), and then creates prediction basis information.

[0061] The prediction basis information may be, for example, a contribution (prediction contribution) for obtaining a prediction result when the machine learning model 110 performs graph classification. The graph AI calculation processing unit 104 obtains prediction basis information for each amino acid contained in the virus.

[0062] The graph AI calculation processing unit 104 may calculate the prediction basis information using explainable AI technology such as Deep Tensor. In the information processing device 1, the explainable AI technology has a function of identifying important parts that are the reason (basis) for the prediction, particularly for the components (subregions) of the input data that express the information to be predicted, and outputting values ​​according to their importance.

[0063] The components (subregions) of the input data that represent the information to be predicted may be nodes or edges of graph data, and in this embodiment, an example in which the components are nodes will be shown.

[0064] The graph AI calculation processing unit 104 receives graph data representing protein information as input data and outputs predicted contributions for nodes that are components (subregions) of the graph data. Here, the predicted contributions are calculated by constructing a learning process that creates a machine learning model 110 that better classifies antigen clusters and by using the machine learning model 110 to perform inference on the prediction target.

[0065] In addition, the graph AI calculation processing unit 104 creates explanation result information using the prediction basis information.

[0066] FIG. 3 is a diagram illustrating an example of explanation result information in the information processing device 1 according to an embodiment.

[0067] The explanation result information shown in FIG. 3 associates a node ID, which indicates the entity of a node, with a predicted contribution. FIG. 3 shows an example in which the full length is 9. Also, in FIG. 3, sorting is performed based on the node ID. The node ID may be referred to as a position in the protein three-dimensional structure, or simply as a position. The node ID is information that identifies multiple partial regions included in the graph structure data.

[0068] 3, the explanation result information is denoted by the symbol T1, and may be referred to as explanation result information T1 hereinafter.

[0069] The graph AI calculation processing unit 104 acquires the node ID and the predicted contribution based on the prediction basis information and associates them to create explanation result information T1. The explanation result information T1 indicates positions that are indicated as having a high contribution by the graph machine learning explanation function of the graph AI calculation processing unit 104.

[0070] The graph AI calculation processing unit 104 generates explanation result information T1 for each sample (input protein 3-dimensional structure) and for each of multiple types of machine learning models 110.

[0071] The graph AI calculation processing unit 104 creates contributions (predicted contributions, explanation result information T1) in the classification process for each of the multiple nodes and edges (subregions) contained in the input graph data (graph structure data) in the classification process using the machine learning model 110.

[0072] The explanation result information T1 created by the graph AI calculation processing unit 104 may be stored in a predetermined storage area of ​​the memory 10c or the storage unit 10d.

[0073] In the machine learning model 110 (Graph AI), the contribution is obtained for each three-dimensional structure or each amino acid. Therefore, the graph AI calculation processing unit 104 may calculate a sample average of the contribution in a predetermined unit such as a cluster, year, or amino acid, and use the average as prediction basis information.

[0074] The prediction results performed by the machine learning model 110 by the graph AI calculation processing unit 104, and the values ​​of the prediction basis information calculated by the machine learning model 110, may be stored in a specified memory area of ​​the memory 10c or the storage unit 10d.

[0075] In the prediction phase, the classification accuracy calculation unit 105 calculates the classification accuracy (prediction accuracy) by comparing the classification result (prediction result) of the cluster numbers obtained by the machine learning model 110 with the correct answer data. The classification accuracy calculated by the classification accuracy calculation unit 105 may be stored in a predetermined storage area of ​​the memory 10c or the storage unit 10d.

[0076] The association processing unit 106 creates protein position information. The protein position information is information indicating a position (region) in an amino acid sequence (protein three-dimensional structure), and indicates a position (region) where features for classifying antigen clusters are particularly prominent. The protein position information may be said to indicate a position (region) known by experts in the field of virology as being particularly relevant (having a large influence) on the classification of antigen clusters. The position particularly relevant to the classification of antigen clusters may be a set of node IDs indicating nodes that the above experts recognize as having a large influence on the classification of antigen clusters in graph data created based on the protein three-dimensional structure. The protein position information is an example of information specifying a partial region of graph structure data.

[0077] In the protein position information, positions (regions) particularly related to the classification of antigen clusters may be represented by a set of residue number positions. Fig. 4 is a diagram illustrating an example of protein position information in the information processing device 1 according to an embodiment.

[0078] 4, the protein location information is denoted by the symbol T2, and may hereinafter be referred to as protein location information T2.

[0079] The protein position information T2 shown in Fig. 4 indicates a plurality of numerical values ​​indicating positions in the protein three-dimensional structure that have a large influence on the classification of antigen clusters. The numerical values ​​indicating each position included in the protein position information T2 correspond to a node ID. A process of matching the positions in the protein position information with the node IDs in the explanation result information may be performed.

[0080] The association processing unit 106 creates protein position information based on information about the amino acid sequence already known by experts in the field of virology and information about the position (region) of the protein in the three-dimensional structure. Alternatively, the association processing unit 106 may obtain pre-created protein position information T2 by downloading or other methods.

[0081] Based on the explanation result information T1 and the protein location information T2, the match index value calculation unit 107 calculates a match index value indicating the degree of match between a node with a high predicted contribution in the explanation result information T1 and a node indicated in the protein location information T2.

[0082] The match index value represents the degree of agreement between the explanation result information T1 (contribution, predicted contribution) and the protein location information T2, which includes information specifying partial regions such as nodes and edges in the graph structure data, and indicates the similarity between them. The match index value can also be said to represent the ease of interpretation (classification interpretation) of the mechanism of prediction (classification) performed by the machine learning model 110.

[0083] The coincidence index value calculation unit 107 acquires the explanation result information T1 created by the graph AI calculation processing unit 104 and the protein position information T2 created by the correspondence processing unit 106, and calculates the coincidence index value using this explanation result information T1 and protein position information T2.

[0084] FIG. 5 is a diagram for explaining the processing of the match index value calculation unit 107 of the information processing device 1 according to an embodiment.

[0085] The coincidence index value calculation unit 107 calculates the coincidence index value as the ratio of the "sum of predicted contributions of node IDs (positions) included in the explanation result information T1 that correspond to all values ​​included in the protein position information T2" to the "sum of predicted contributions of all node IDs (positions) included in the explanation result information T1."

[0086] 5, values ​​corresponding to node IDs 3, 4, 5, and 8 are set as positions in protein position information T2. ​​In this case, coincidence index calculation unit 107 calculates the coincidence index by dividing (the sum of the predicted contributions of node IDs 3, 4, 5, and 8) by (the sum of the predicted contributions of node IDs 1 to 9).

[0087] In the example shown in FIG. 5, (total of predicted contributions of nodes IDs 3, 4, 5, and 8)=0.8+0.9+0.8+0.6=3.1.

[0088] In addition, (total of predicted contributions of nodes ID 1 to 9) = 0.2 + 0.1 + 0.8 + 0.9 + 0.8 + 0.1 + 0.2 + 0.6 + 0.1 = 3.8.

[0089] Therefore, the match index value calculation unit 107 calculates 3.1 / 3.8=0.816 to obtain a match index value of 0.816.

[0090] The match index value represents the degree of match between a position (node ​​ID) in the amino acid 3D structure indicated by the protein position information T2 and a node ID with a high predicted contribution calculated by the graph AI calculation processing unit 104. In other words, the match index value represents whether a position (node ​​ID) in the amino acid 3D structure indicated by the protein position information T2 is similar to a node ID with a high predicted contribution calculated by the graph AI calculation processing unit 104. It can also be said that the match index value represents whether a high predicted contribution is concentrated at a position (node ​​ID) indicated in the protein position information T2.

[0091] In this embodiment, the larger the agreement index value (closer to 1), the higher the degree of agreement between the tendency indicated by the explanation result information T1 and the tendency indicated by the protein location information T2.

[0092] If the match index value is greater than a predetermined threshold (for example, 0.5), it may be determined that the trend indicated by the explanation result information T1 and the trend indicated by the protein location information T2 match.

[0093] In the example shown in Figure 5, each of the nodes with node IDs 3, 4, 5, and 8 shown in the protein position information T2 has a predicted contribution of 0.6 or more, and it can be said that there is a high degree of agreement between the positions (node ​​IDs) in the amino acid three-dimensional structure shown in the protein position information T2 and the node IDs with high predicted contributions calculated by the graph AI calculation processing unit 104.

[0094] The graph selection unit 108 calculates an evaluation value for the machine learning model 110 based on the match index value calculated by the match index value calculation unit 107, the classification accuracy calculated by the classification accuracy calculation unit 105, and the balance index.

[0095] The balance index indicates the degree to which the match index and classification accuracy are reflected in the calculation of the evaluation value, and is a weight applied to each of the match index and classification accuracy. The weight applied to the match index is represented by symbol w1, and the weight applied to the classification accuracy is represented by symbol w2.

[0096] The graph selection unit 108 may calculate the evaluation value of the machine learning model 110 by, for example, calculating the following formula (1): Evaluation value = w1 × classification accuracy + w2 × match index value (1) The balance index (w1, w2) may be w1 + w2 = 1.0. The balance index may be set in advance by a user or the like. When determining the balance index, the values ​​of w1 and w2 may be changed as appropriate, taking into consideration the balance between the match index value (ease of interpretation) and the classification accuracy (prediction accuracy).

[0097] For example, if the match index value and classification accuracy are treated equally, w1 = 0.5 and w2 = 0.5 may be used. If a machine learning model 110 with a classification mechanism that is easy to interpret is to be selected, the value of the weight w1 applied to the match index value may be increased, and if a machine learning model 110 with high classification accuracy is to be selected, the value of the weight w2 applied to classification accuracy may be increased.

[0098] The graph selection unit 108 calculates an evaluation value for each of the multiple machine learning models 110. Then, the graph selection unit 108 evaluates the machine learning models 110 based on the evaluation value. The graph selection unit 108 may determine, for example, the machine learning model 110 with the highest evaluation value as the optimal machine learning model 110.

[0099] FIG. 6 is a diagram for explaining the processing of the graph selection unit 108 of the information processing device 1 according to an embodiment.

[0100] The example shown in Figure 6 illustrates the process in which the graph selection unit 108 compares machine learning model #1 and machine learning model #2 and selects machine learning model #1 as the machine learning model to present to the user.

[0101] For machine learning model #1, the graph selection unit 108 receives as input the match index (see symbol P1) calculated by the match index calculation unit 107 based on the explanation result information T1 and the protein position information T2, and the classification accuracy (see symbol P2) calculated by the classification accuracy calculation unit 105. The graph selection unit 108 also acquires balance indices (w1, w2).

[0102] The graph selection unit 108 calculates the evaluation value by calculating the above formula (1) based on the match index value, classification accuracy, and balance index.

[0103] In the example shown in FIG. 6, the evaluation value of machine learning model #1 is 0.5×0.0784+0.5×0.816=0.80.

[0104] The graph selection unit 108 similarly performs the same operation on machine learning model #2, calculating the formula (0.5 x 0.800 + 0.5 x 0.6) based on the match index value (0.6), classification accuracy (0.800), and balance index (0.5, 0.5) to obtain an evaluation value of 0.70.

[0105] The graph selection unit 108 may compare the calculated evaluation values ​​of each machine learning model 110 and present the machine learning model #1 with the highest evaluation value to the user. In the example shown in Figure 6, the graph selection unit 108 selects the machine learning model #1 with the highest evaluation value as the preferred machine learning model 110. Note that the selection of the machine learning model 110 by the graph selection unit 108 can be considered to be equivalent to selecting the graph information that is the basis of the machine learning model 110.

[0106] The graph AI calculation processing unit 104 presents the selected machine learning model #1 to the user. The graph AI calculation processing unit 104 may present only the selected machine learning model #1 to the user.

[0107] Note that the information presented to the user by the graph selection unit 108 is not limited to this. For example, the graph selection unit 108 may present the user with the evaluation values ​​of each of the multiple machine learning models 110, and this can be changed as appropriate.

[0108] (B) Operation The processing of the coincidence index calculation unit 107 and the graph selection unit 108 of the information processing device 1 according to the embodiment configured as described above will be described with reference to the flowchart (steps S1 to S4) shown in FIG.

[0109] The following processing is performed in the training phase in the information processing device 1.

[0110] In step S1 , the match index calculation unit 107 acquires protein position information T2 and explanation result information T1 created for each sample based on the machine learning model 110 .

[0111] In step S2, the match index calculation unit 107 calculates a match index based on the explanation result information T1 and the protein location information T2.

[0112] In step S3, the graph selection unit 108 calculates an evaluation value for the machine learning model 110 based on the match index value calculated by the match index value calculation unit 107, the classification accuracy calculated by the classification accuracy calculation unit 105, and the balance index. The graph selection unit 108 calculates an evaluation value for each of the multiple machine learning models 110.

[0113] In step S4, the graph selection unit 108 evaluates the machine learning model 110 based on the multiple evaluation values. For example, the graph selection unit 108 may select the machine learning model 110 with the highest evaluation value and present it to the user. Alternatively, the graph selection unit 108 may present the machine learning model 110 with a high evaluation value (e.g., equal to or greater than a threshold) to the user. Then, the process ends.

[0114] (C) Effect As described above, according to the information processing device 1 of one embodiment, the coincidence index value calculation unit 107 calculates, based on the explanation result information T1 and the protein location information T2, a coincidence index value indicating the degree of coincidence between a node with a high predicted contribution in the explanation result information T1 and a node indicated in the protein location information T2.

[0115] In addition, the graph selection unit 108 calculates an evaluation value for the machine learning model 110 based on the match index value calculated by the match index value calculation unit 107, the classification accuracy calculated by the classification accuracy calculation unit 105, and the balance index.

[0116] Then, the graph selection unit 108 evaluates the machine learning model 110 based on the calculated evaluation value.

[0117] This allows the graph selection unit 108 to select a machine learning model 110 that takes into account classification accuracy and ease of interpretation (good interpretability).

[0118] By setting the balance index (w1, w2) taking into account the balance between the match index value (ease of interpretation) and classification accuracy (prediction accuracy), it is possible to select a machine learning model 110 that meets the user's needs in terms of the relationship between ease of interpretation and classification accuracy.

[0119] The coincidence index value calculation unit 107 calculates the coincidence index value as the ratio of the "sum of predicted contributions of node IDs (positions) included in the explanation result information T1 that correspond to all values ​​included in the protein position information T2" to the "sum of predicted contributions of all node IDs (positions) included in the explanation result information T1."

[0120] The graph selection unit 108 selects the machine learning model 110 with the high match index value thus obtained, thereby selecting the machine learning model 110 that calculates a high predicted contribution to a partial region of the protein that an expert in the field of virology believes is particularly relevant to the classification of antigen clusters. Therefore, it is possible to select a machine learning model 110 that is easy to interpret.

[0121] (D) Others The disclosed technology is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present embodiment. The configurations and processes of the present embodiment can be selected or combined as needed.

[0122] In the above-described embodiment, the predicted contributions and protein location information are set for nodes in graph data, but the present invention is not limited to this. Instead of nodes, the predicted contributions and protein location information may be set for edges.

[0123] In the above-described embodiment, the match index calculation unit 107 calculates the match index for each sample, but this is not limiting. For example, the match index may be calculated based on the average or maximum value of multiple samples (e.g., all samples), and the calculation may be carried out with appropriate modifications.

[0124] Alternatively, the coincidence index calculation unit 107 may calculate the rate at which the magnitude of the predicted contribution values ​​is concentrated at the positions set in the protein position information T2, and use this as the coincidence index (degree of coincidence in tendency).

[0125] Furthermore, the above disclosure enables those skilled in the art to implement and manufacture the present embodiment.

[0126] 1 Information processing device 10 Computer 10a Processor 10b Graphics processing device 10c Memory 10d Storage unit 10e IF unit 10f IO unit 10g Reading unit 10h Program 10i Recording medium 10j Bus 101 Three-dimensional structure feature processing unit 102 Feature calculation unit 103 Graph generation unit 104 Graph AI calculation processing unit 105 Classification accuracy calculation unit 106 Correspondence processing unit 107 Matching index value calculation unit 108 Graph selection unit 110 Machine learning model T1 Explanation result information T2 Protein position information

Claims

1. An information processing program that causes a computer to execute a process of obtaining a contribution level in a classification process using a machine learning model for input graph structure data for each of a plurality of partial regions contained in the graph structure data, and determining an evaluation of the machine learning model based on a similarity between the contribution levels and information specifying the partial regions of the graph structure data.

2. The information processing program according to claim 1, characterized in that the computer is caused to execute a process of calculating a matching index value representing the similarity by calculating a ratio of the sum of the contribution degrees corresponding to the partial regions identified by the designation information among all the partial regions included in the graph structure data to the sum of the contribution degrees corresponding to all the partial regions included in the graph structure data.

3. The information processing program according to claim 2, characterized in that the computer is caused to execute a process of calculating an evaluation value for the machine learning model based on the sum of the match index value and the classification accuracy of the machine learning model.

4. The information processing program according to claim 3, characterized in that weights are set for the match index value and the classification accuracy.

5. An information processing device comprising: a control unit that executes a process of acquiring a contribution level in a classification process using a machine learning model for input graph structure data for each of a plurality of partial regions contained in the graph structure data; and determining an evaluation of the machine learning model based on a similarity between the contribution levels and designation information for the partial regions of the graph structure data.

6. The information processing device according to claim 5, characterized in that the control unit executes a process of calculating a matching index value representing the similarity by calculating a ratio of the sum of the contribution degrees corresponding to the partial regions identified by the designation information among all the partial regions included in the graph structure data to the sum of the contribution degrees corresponding to all the partial regions included in the graph structure data.

7. The information processing device according to claim 6, characterized in that the control unit executes a process of calculating an evaluation value for the machine learning model based on the sum of the match index value and the classification accuracy of the machine learning model.

8. The information processing device according to claim 7, characterized in that weights are set for the match index value and the classification accuracy.

9. An information processing method, characterized in that a computer executes a process of obtaining a contribution level in a classification process using a machine learning model for input graph structure data, for each of a plurality of partial regions contained in the graph structure data, and determining an evaluation of the machine learning model based on a similarity between the contribution levels and information specifying the partial regions of the graph structure data.

10. The information processing method according to claim 9, characterized in that the computer executes a process of calculating a match index value representing the similarity by calculating a ratio of the sum of the contribution degrees corresponding to the partial regions identified by the designation information among all the partial regions included in the graph structure data to the sum of the contribution degrees corresponding to all the partial regions included in the graph structure data.

11. The information processing method according to claim 10, characterized in that the computer executes a process of calculating an evaluation value for the machine learning model based on the sum of the match index value and the classification accuracy of the machine learning model.

12. The information processing method according to claim 11, characterized in that weights are set for the match index value and the classification accuracy.

Images