Information processing method, learning model generation method, learning model, computer program, and information processing device

Abstract

Provided are an information processing method, a learning model generation method, a learning model, a computer program, and an information processing device.　In the present invention, a computer executes processing wherein: data of an adaptive immune receptor of a subject is acquired; a calculation based on the acquired data of the adaptive immune receptor of the subject is executed using a learning model trained to output information on a disease estimated from an adaptive immune receptor on the basis of data of the adaptive immune receptor; and information on a disease inferred from the adaptive immune receptor of the subject on the basis of the calculation result executed using the learning model is output.

Description

Information processing method, learning model generation method, learning model, computer program, and information processing device

[0001] The present disclosure relates to an information processing method, a learning model generation method, a learning model, a computer program, and an information processing device.

[0002] The adaptive immune system includes T cell receptors and B cell receptors as molecules that specifically recognize a wide variety of pathogens. Since living organisms cannot predict the pathogens that will invade from the outside world, they respond to a wide variety of pathogens by maintaining the diversity of T cell receptors and B cell receptors. T cell receptors and B cell receptors are composed of a combination of multiple gene regions. For example, human T cell receptors have 10 15 ~10 20 types, 10 for B cell receptors 16 ~10 18 It is believed that there are several types of receptors.

[0003] As a technique for determining a profile of recombinant DNA sequences in T cells and B cells, Patent Document 1 discloses a method for collecting a sample from a subject containing T cells and B cells, spatially isolating individual molecules of genomic DNA from the cells, determining the sequences of the spatially isolated individual molecules, and creating a profile of the recombinant DNA sequence.

[0004] Special Publication No. 2012-508011

[0005] The technique of Patent Document 1, which creates a profile of recombinant DNA sequences for individual T cells and B cells, has limitations in terms of generalizability to unknown data.

[0006] In one aspect, the present invention aims to provide an information processing method, a method for generating a learning model, a learning model, a computer program, and an information processing device that can predict diseases using a learning model that can be generalized even to unknown data on adaptive immune receptors.

[0007] The information processing method disclosed herein involves a computer-implemented process of acquiring adaptive immune receptor data of a subject, using a learning model trained to output information about diseases inferred from the adaptive immune receptors based on the adaptive immune receptor data, performing calculations based on the acquired adaptive immune receptor data of the subject, and outputting information about diseases inferred from the adaptive immune receptors of the subject based on the results of the calculations performed using the learning model.

[0008] In one aspect, diseases can be predicted using a learning model that can generalize to unknown data on adaptive immune receptors.

[0009] 1 is an explanatory diagram illustrating an overview of the present disclosure. FIG. 1 is a block diagram illustrating the internal configuration of an information processing device. FIG. 2 is a flowchart illustrating the procedure of processing executed by the information processing device in the learning phase. FIG. 3 is a diagram illustrating an example of a dataset used by the information processing device in the learning phase. FIG. 4 is an explanatory diagram illustrating a first classification method for AIR data. FIG. 5 is an explanatory diagram illustrating a second classification method for AIR data. FIG. 6 is an explanatory diagram illustrating an outline of a method for calculating a cluster frequency CF. FIG. 7 is an explanatory diagram illustrating an outline of a method for calculating a cluster occupancy CO. FIG. 8 is a flowchart illustrating the procedure of processing executed by the information processing device in the operation phase. FIG. 9 is a graph illustrating evaluation results of sparsity. FIG. 10 is a diagram illustrating comparison results of ROC AUC of a test set using four different feature amounts. FIG. 11 is a graph illustrating ROC curves of CCF and PCO.

[0010] The present invention will be described in detail below with reference to the drawings illustrating embodiments thereof. (Embodiment 1) FIG. 1 is an explanatory diagram illustrating an overview of the present disclosure. This disclosure focuses on biomarker discovery using adaptive immune receptors (AIRs). Hereinafter, adaptive immune receptors will also be referred to as AIRs. In this embodiment, AIR data and disease data for each donor are collected from multiple donors. Donors include patients suffering from diseases such as infectious diseases (HIV, COVID-19, etc.), autoimmune diseases (autoimmune hepatitis, primary biliary cholangitis, etc.), and cancers (breast cancer, colon cancer, etc.).

[0011] In the learning phase, the information processing device 10 according to this embodiment uses a data set including AIR data of multiple donors and data related to the disease of each donor as training data, and when AIR data of a subject is input, generates a learning model LM that outputs disease information estimated from the AIR. Furthermore, in the operation phase after learning is completed, the information processing device 10 according to this embodiment acquires AIR data from a new subject, performs calculations using the learning model LM based on the acquired AIR data, and outputs disease information estimated by the learning model LM.

[0012] 2 is a block diagram illustrating the internal configuration of the information processing device 10. The information processing device 10 is a dedicated or general-purpose computer, and includes a control unit 11, a storage unit 12, a communication unit 13, an operation unit 14, and a display unit 15.

[0013] The control unit 11 includes, for example, a CPU (Central Processing Unit), a ROM (Read Only Memory), and a RAM (Random Access Memory). The ROM included in the control unit 11 stores a control program that controls the operation of each hardware unit included in the information processing device 10. The CPU in the control unit 11 reads and executes the control program stored in the ROM and a computer program (described below) stored in the storage unit 12, and controls the operation of each hardware unit, thereby causing the entire device to function as the information processing device 10 of the present disclosure. The RAM included in the control unit 11 temporarily stores data used during calculation and control.

[0014] In the embodiment, the control unit 11 is configured to include a CPU, a ROM, and a RAM, but the configuration of the control unit 11 is not limited to the above. The control unit 11 may be, for example, one or more control circuits or arithmetic circuits including a GPU (Graphics Processing Unit), an FPGA (Field Programmable Gate Array), a DSP (Digital Signal Processor), a quantum processor, volatile or non-volatile memory, etc. The control unit 11 may also have functions such as a clock that outputs date and time information, a timer that measures the elapsed time from when a measurement start instruction is given until when a measurement end instruction is given, and a counter that counts numbers.

[0015] The storage unit 12 includes a storage device such as a hard disk drive (HDD), a solid state drive (SSD), etc. The storage unit 12 stores various computer programs executed by the control unit 11 and various data used by the control unit 11.

[0016] The computer programs stored in the storage unit 12 include a model generation program PG1 that causes a computer to execute a process of acquiring a dataset including adaptive immune receptor data obtained for multiple donors and disease data for each donor, using the acquired dataset as training data to generate a learning model LM that outputs disease information estimated from the AIR when AIR data of a subject is input.The computer programs stored in the storage unit 12 also include an estimation processing program PG2 that acquires the AIR data of the subject, inputs the acquired AIR data of the subject into the learning model LM to perform calculations, and outputs disease information obtained from the learning model LM.

[0017] Each of the computer programs may be a single computer program, or may be a group of programs consisting of multiple computer programs. Furthermore, the computer programs may partially use an existing library. Each of the computer programs may be executed on a single computer, or may be executed cooperatively by multiple computers.

[0018] The computer program may be provided by a non-transitory recording medium RM (program product) on which the computer program is readably recorded. The recording medium RM is a portable memory such as a CD-ROM, a USB memory, or an SD card (SD: Secure Digital). The control unit 11 reads various computer programs from the recording medium RM using a reading device (not shown) and stores the read various computer programs in the memory unit 12. The computer program may also be provided by communication. The control unit 11 acquires the computer program by communication via the communication unit 13 and stores the acquired computer program in the memory unit 12.

[0019] The memory unit 12 also stores a learning model LM that has been trained to output information about the subject's disease based on the adaptive immune receptor data of the subject. More specifically, the learning model LM is trained to output information about the disease the subject is suffering from when a feature (such as the paratope cluster occupancy rate described below) extracted from the adaptive immune receptor data is input. The learning model LM uses classifiers such as XGBoost (eXtreme Gradient Boosting), SVM (Support Vector Machine), Bayesian network, regression tree, and LightGBM (Light Gradient Boosting Machine). The memory unit 12 stores information describing the learned learning model LM, such as the model's layer structure, the number and connection relationships of nodes constituting each layer, and weight coefficients and biases between nodes, as well as learned parameters.

[0020] Alternatively, the learning model LM may be stored in an external server accessible from the information processing device 10. In this case, the information processing device 10 may access the external server via communication and perform calculations using the learning model LM.

[0021] The communication unit 13 includes a communication interface for transmitting and receiving various data to and from an external device. The communication interface of the communication unit 13 may be a communication interface conforming to a communication standard such as Wi-Fi (registered trademark), LAN (Local Area Network), Bluetooth (registered trademark), ZigBee (registered trademark), 3G, 4G, 5G, or LTE (Long Term Evolution). When data to be transmitted is input from the control unit 11, the communication unit 13 transmits the data to the destination external device, and when data transmitted from the external device is received, the communication unit 13 outputs the received data to the control unit 11. In this embodiment, an example of an external device is a terminal device used by a medical professional such as a doctor or a subject.

[0022] The operation unit 14 includes operation devices such as a touch panel, a keyboard, and switches, and receives various operations and settings from a user, etc. The control unit 11 performs appropriate control based on various pieces of operation information provided by the operation unit 14, and stores setting information in the storage unit 12 as necessary.

[0023] The display unit 15 includes a display device such as a liquid crystal display or an organic EL (Electro-Luminescence) display, and displays information to be presented in response to an instruction from the control unit 11 .

[0024] In this embodiment, the information processing device 10 may be a single computer, or may be a computer system configured with multiple computers and peripheral devices, etc. Furthermore, the information processing device 10 may be a virtual machine whose entity is virtualized, or may be a cloud.

[0025] 3 is a flowchart illustrating the procedure of the process executed by the information processing device 10 in the learning phase. In the learning phase for generating a learning model LM, the control unit 11 of the information processing device 10 executes the model generation program PG1 to perform the following process.

[0026] The control unit 11 first acquires a dataset including AIR data obtained for multiple donors and disease data for each donor (step S101). Multiple AIR data are obtained for each donor. Furthermore, as the disease data, a correct answer label indicating whether each donor is suffering from any of the following diseases, such as COVID-19, CRC (colorectal cancer), AIH (autoimmune hepatitis), PBC (primary biliary cholangitis), BC (lung cancer), or HIV (human immunodeficiency virus), or whether each donor is suffering from none of the diseases, is acquired.

[0027] Next, the control unit 11 extracts features from the acquired AIR data (step S102). The method for calculating the features will be described in detail later; however, in this example, the features are extracted by clustering the AIR data and calculating the occupancy rate of AIR within the resulting clusters. In this embodiment, during clustering, the AIR data is clustered using a paratope pseudo-sequence, which is a pseudo-link of the three complementarity-determining regions (CDR1, CDR2, and CDR3) contained in AIR, as an index. In this embodiment, as a comparative example, a method for clustering AIR data using the sequences of the V gene, J gene, and CDR3 contained in AIR, and a method for extracting features by calculating the frequency of AIR within the resulting clusters will also be described.

[0028] Next, the control unit 11 selects a set of features and a correct label from a dataset including the features calculated for each donor and the disease data (correct label) of each donor (step S103), inputs the selected features into the learning model LM, and executes calculations using the learning model LM (step S104). It is assumed that initial values ​​are set for the internal parameters of the learning model LM.

[0029] The control unit 11 calculates the error from the correct label as the calculation using the learning model LM proceeds, and determines whether learning is complete or not based on the error finally obtained (step S105). If it determines that learning is not complete (S105: NO), the control unit 11 updates the internal parameters of the learning model LM (step S106), returns the process to step S103, and continues learning.

[0030] If it is determined that the learning is completed (S105: YES), a trained learning model LM is obtained, and the control unit 11 stores the trained learning model LM in the memory unit 12 (step S107).

[0031] The processing in the learning phase will be described in detail below. FIG. 4 is a diagram illustrating an example of a dataset used by the information processing device 10 in the learning phase. In this embodiment, a dataset including 106,640 pieces of AIR data acquired from 44 donors infected with COVID-19 and 174,139 pieces of AIR data acquired from 58 donors not infected with COVID-19 was used. Similarly, as shown in FIG. 4 , a dataset including AIR data acquired from donors infected with and not infected with CRC (colorectal cancer), AIH (autoimmune hepatitis), PBC (primary biliary cholangitis), BC (lung cancer), and HIV (human immunodeficiency virus) was used. This AIR data includes data on BCR (B cell receptor) and TCR (T cell receptor). The dataset used in the learning phase may be collected in advance and already published as a benchmark dataset.

[0032] In the example of Fig. 4, a data set is shown in which AIR data is classified into a class belonging to "disease" and a class belonging to "health" for each type of disease, but if the disease has subtypes, a data set in which AIR data is classified by subtype may be used. Also, instead of the data set shown in Fig. 4, a data set in which AIR data is classified by the severity of the disease may be used, or a data set in which AIR data is classified by the disease that the user is likely to contract in the future may be used.

[0033] In this study, we employed the XGBoost algorithm for the binary classification task. Training was performed on 70% of randomly selected donors, with all parameters set to default values ​​except for weighting the positive class according to the ratio of true to false training donors.

[0034] AIR data can be classified according to AIR expression. FIG. 5 is an explanatory diagram illustrating a first classification method for AIR data. AIR consists of the amino acid sequence of a B cell receptor or T cell receptor and is described by the V gene name, J gene name, and the position of the complementarity-determining region (CDR) within the amino acid sequence. The CDR includes three regions (CDR1, CDR2, and CDR3). In the first classification method, AIR data is represented by the V gene name, J gene name, and CDR3 sequence. The representation using the V gene name, J gene name, and CDR3 sequence is referred to as a clonotype representation. Furthermore, a cluster of AIRs having the same V gene and J gene and obtained by clustering based on the similarity of CDR3 sequences is referred to as a clonotype cluster in this specification.

[0035] To cluster by clonotype, AIR data with identical V gene, J gene, and CDR3 lengths were combined. Next, the AIR data within these groups was clustered by CDR3 amino acid sequence. In this embodiment, 80% sequence identity and 100% alignment coverage were used, as implemented in CD-HIT (Fu, L., B. Niu, Z. Zhu, S. Wu, and W. Li (2012). "CD-HIT: accelerated for clustering the next-generation sequencing data." Bioinformatics 28(23): 3150-3152).

[0036] FIG. 6 is an explanatory diagram illustrating a second classification method for AIR data. In the second classification method, AIR data is represented by CDR sequences. Representation using CDR sequences is called a paratope representation. In this embodiment, three segments, CDR1, CDR2, and CDR3, are pseudo-linked to construct one pseudo-sequence. A cluster of AIRs obtained by clustering from the perspective of similarity of this pseudo-sequence (hereinafter referred to as a paratope pseudo-sequence) is called a paratope cluster in this specification.

[0037] Clustering of paratope pseudosequences was performed using the MMseqs2 search method with a sequence identity threshold of 0.8 and a coverage threshold of 0.9. Alignments of paratope pseudosequences with fixed sequence identity and coverage are statistically more significant than CDR3 alignments using the same criteria due to the longer sequence lengths.

[0038] In this embodiment, an AIR adjacency matrix is ​​constructed from this paratope pseudosequence. The AIR adjacency matrix can be calculated from the paratope pseudosequence using the MMseqs2 search method, for example, with a sequence identity threshold of 0.8 and a coverage threshold of 0.9. That is, the sequence identity is calculated by ij , coverage q ij Then, for a given AIR pair a i and a j On the other hand, the AIR adjacency matrix A ij is expressed by the following equation 1.

[0039]

[0040] In this embodiment, two types of feature amounts, namely, cluster frequency CF and cluster occupancy CO, are introduced as feature amounts that characterize the AIR data of each donor.

[0041] Figure 7 is an explanatory diagram outlining the method for calculating the cluster frequency CF. In this embodiment, to construct a feature vector for each donor, all AIR data in a given dataset is clustered, and each donor is described by the frequency of AIR within the resulting cluster. Donor D has M AIRs, each of which is {S1 ... SN}. Formally, the set of different AIRs of donor D is defined as AIR D = {a1 ... a M} D where: 1. For all a i ∈AIR D Whereas, a i ∈S k 2. There exists k∈[1,N] such that for all a i , a j ∈AIR D For i=j only, a i = a j is.

[0042] C M×N Let be an M × N matrix containing the one-hot encoding of the cluster assignments of each AIR, then each element C ik ∈C M×N is defined as follows:

[0043]

[0044] That is, AIRa i is cluster S k If it belongs to C ik is equal to 1 and is 0 for all other values ​​of k. k The cluster frequency CF of C is expressed as M×N It is calculated by summing over all rows of

[0045]

[0046] In this embodiment, the cluster frequency CF is used as a vector representation of each donor. The feature vector of a certain donor D is {CF1, ..., CF N} D Hereinafter, the features based on the clonotype clusters will be referred to as clonotype cluster frequencies (CCFs), and the features based on the paratope clusters will be referred to as paratope cluster frequencies (PCFs).

[0047] FIG. 8 is an explanatory diagram outlining the method for calculating the cluster occupancy rate CO. Since the distribution of the AIR sequence is a long-tail distribution, the cluster size also follows the long-tail distribution. Furthermore, the matrix element C ik Since one-hot encoding is used to define O, the features are very sparse. Therefore, the inventors of the present application hypothesized that reducing sparsity is effective in reducing the size of the features, and used an adjacency matrix to identify that AIR occupies two or more clusters. The occupancy matrix O M×N is the adjacency matrix A M×M and a cluster matrix C containing one-hot cluster encodings M×N It is defined as multiplication with A. M×M is an M×M symmetric matrix. Formally, O M×N =A M×M C M×N Here, each element O ik ∈O M×N is defined as follows:

[0048]

[0049] Here, A ij represents the similarity between the i-th AIR and the j-th AIR. If the neighbors connect different clusters, O M×N is C M×N Furthermore, similar to Equation 3, by summing the AIRs of all donors, the cluster S of a given donor can be k The overall cluster occupancy of CO k can be calculated.

[0050]

[0051] Hereinafter, the cluster occupancy derived from the clonotype clusters will be referred to as the clonotype cluster occupancy (CCO), and the cluster occupancy derived from the paratope clusters will be referred to as the paratope cluster occupancy (PCO).

[0052] Using all cluster information to create a feature vector often results in a very high-dimensional feature space. This is mainly because cluster sizes follow a long-tail distribution, resulting in a large number of clusters with a very small number of data points. Furthermore, when there are a large number of AIRs, it becomes difficult to calculate the complete matrix. To address these issues, we propose an algorithm to reduce the number of clusters and the number of AIRs. Furthermore, reducing the number of clusters allows us to utilize meaningful clusters for feature generation. To find meaningful clusters, we identify clusters that meet a given population criterion, as described below.

[0053] First, define a set of binary classes as Y = {y1, y2}, where y1 represents the class of healthy donors and y2 represents the class of diseased donors. k Among them, class y p The number of AIRs belonging to donors is n p (S k ) and the cluster S k The cluster ratio tuple of (r k y1 , r k y2 ) where r k yp is expressed as follows:

[0054]

[0055] That is, r k yp is class y p The total number of AIRs belonging to cluster S k Within class y p represents the ratio of the number of AIRs belonging to

[0056] Next, the threshold parameter d min This threshold parameter d min defines the minimum relative parameter below which all clusters are removed and their respective AIRs are ignored. For a class to be selected, at least one value of the cluster ratio tuple must be greater than d minSpecifically, cluster S k For max(r k y1 , r k y2 ) <d min If so, then the cluster S k This downsizing is min This preprocessing is applied to all datasets during training using as a hyperparameter. For very large datasets (e.g., the HIV dataset shown in Figure 4), this preprocessing is applied because the neighborhood calculations take a significant amount of time across the entire dataset.

[0057] For example, in the HIV data set shown in FIG. 4, a total of 13,136,538 pieces of AIR data are obtained. For the clusters obtained from this AIR data, the threshold parameter d min When the value of d is set to 0.05 and downsizing is performed as preprocessing, the number of AIRs can be reduced to 4,424,328. min Increasing the value of can further reduce the number of AIRs, but resulted in the removal of all AIRs with more than one donor. The reduced HIV dataset was used in all subsequent calculations.

[0058] 9 is a flowchart illustrating the procedure of processing executed by the information processing device 10 in the operation phase. In the operation phase after generating the trained learning model LM, the control unit 11 of the information processing device 10 executes the estimation processing program PG2 to perform the following processing.

[0059] The control unit 11 acquires air data obtained for the subject (step S121), and calculates feature quantities from the acquired air data (step S122). The feature quantity calculation method is the same as the calculation method in the learning phase.

[0060] The control unit 11 inputs the feature values ​​calculated in step S122 into the trained learning model LM and executes calculations using the learning model LM (step S123).The control unit 11 acquires information on diseases estimated from the AIR data as calculation results using the learning model LM (step S124).

[0061] The control unit 11 outputs the estimation result of step S123 (step S125). The control unit 11 may display the estimation result on the display unit 15, or may notify the doctor or the subject's terminal of the estimation result via the communication unit 13.

[0062] For example, in the data set of FIG. 4, the AIR data is classified into a class belonging to "disease" and a class belonging to "health" for each disease type, and the learning model LM trained using such a data set can obtain information on whether or not the subject has each disease (presence or absence of disease) for each disease type. Furthermore, the control unit 11 may calculate the probability of each disease in step S123 and output the calculated probability information as an estimation result. Furthermore, the control unit 11 may output information on two or more types of diseases (presence or absence, probability, etc.).

[0063] The learning model LM used in the operation phase may be a learning model trained using a data set in which AIR data is classified by disease subtype. In this case, information about the disease subtype is obtained as a calculation result by the learning model LM. The control unit 11 may output the information about the disease subtype obtained from the learning model LM.

[0064] Furthermore, the learning model LM used in the operation phase may be a learning model trained using a data set in which AIR data is classified by the severity of the disease. In this case, information regarding the severity of the disease is obtained as a calculation result by the learning model LM. The control unit 11 may output information regarding the importance of the disease obtained from the learning model LM.

[0065] Furthermore, the learning model LM used in the operation phase may be a learning model trained using a data set in which AIR data is classified into classes according to diseases that are likely to be contracted in the future. In this case, information about diseases that are likely to be contracted in the future is obtained as a result of calculation by the learning model LM. The control unit 11 may output information about diseases that are likely to be contracted in the future that is obtained from the learning model LM.

[0066] The evaluation of the learning model LM will be described below. Classification performance was evaluated using the AIR data collected from the remaining 30% of donors using the receiver operating characteristic curve (ROC AUC). In this embodiment, four workflows were constructed with different clustering methods (clonotype or paratope) and feature construction methods (cluster frequency or cluster occupancy), and each workflow was evaluated.

[0067] Figure 10 is a graph showing the sparsity evaluation results. Figure 10 shows a graph evaluating sparsity for CCF and PCO using AIR data for COVID-19, CRC, AIH, PBC, BC, and HIV. The horizontal axis of each graph represents the threshold parameter d min where the vertical axis represents sparsity. Sparsity can be defined as the fraction of zeros in the feature matrix (i.e., the set of all feature vectors for a given set of donors). In Figure 10, all d min The results show that the PCO features are less sparsity than the CCF features in terms of the value of . These results are independent of the clustering method; the sparsity of the PCF features is comparable to that of the CCF features, and the sparsity of the CCO features is comparable to that of the PCO features. This evaluation suggests that the main factor in reducing sparsity is the use of paratope adjacency in the computation of the CO features.

[0068] Next, we show the ROC AUC of test data obtained from datasets of six diseases. Figure 11 is a chart showing the comparison results of ROC AUC of test sets using four different feature amounts. Figure 11 shows the results of calculating ROC AUC for six diseases: COVID-19, CRC, AIH, PBC, BC, and HIV, when using CCF features, PCF features, CCO features, and PCO features. For reference, Figure 11 also shows the evaluation results for Ave (average).

[0069] Of the six diseases for which CCF features failed to generalize, CCF generalized to subjects with HIV, but failed to generalize to new subjects for the remaining five diseases. The use of PCF features yielded relatively good results for three of the five diseases for which CCF features failed to generalize (COVID-19, AIH, and BC), but poor results for the other two (CRC and PBC). Similarly, the use of CCO features yielded good results for the same three diseases (COVID-19, AIH, and BC), but poor results for the remaining diseases (CRC and PBC). In contrast, the use of PCO features yielded good results for all disease datasets.

[0070] Figure 12 shows the ROC curves for CCF and PCO. The horizontal axis of each graph represents the false positive rate, and the vertical axis represents the positive rate (sensitivity). Figure 12 shows that the benefits outweigh the losses across all different diseases. While the learning model LM cannot perfectly classify diseases, the benchmark demonstrates that it was able to fairly test the generalizability of the features within the context of the data used in this embodiment.

[0071] As described above, in this embodiment, by using a learning model LM that has been trained using a dataset containing adaptive immune receptor data obtained for multiple donors and data related to the disease of each donor as training data, it is possible to estimate the disease that a subject is suffering from based on the adaptive immune receptor data of the subject.

[0072] Of particular note is that the use of a paratopic representation of AIR proved beneficial not only for BCR data but also for TCR data. This observation, likely due to the sparseness of the exact cluster-encoding features, was addressed by converting cluster frequencies to cluster occupancies using an adjacency matrix. The inclusion of paratopic features representing triplets of CDR sequence fragments implies the potential existence of a cohesive network of similarity relationships. The cluster occupancy formulation effectively captures these relationships, allowing for flexible connections to other clusters' AIRs while adjusting the exact cluster identity. The improved generalization of the PCO classification method demonstrates the robustness and effectiveness of this approach.

[0073] The embodiments disclosed herein are illustrative in all respects and should not be considered limiting. The scope of the present invention is defined by the claims, not by the above meaning, and is intended to include all modifications within the meaning and scope of the claims.

[0074] REFERENCE SIGNS LIST 10 Information processing device 11 Control unit 12 Storage unit 13 Communication unit 14 Operation unit 15 Display unit PG1 Model generation program PG2 Estimation processing program LM Learning model

Claims

1. An information processing method in which a computer executes the following processes: acquiring adaptive immune receptor data of a subject; performing a calculation based on the acquired adaptive immune receptor data of the subject using a learning model trained to output information on diseases inferred from the adaptive immune receptor based on the adaptive immune receptor data; and outputting information on diseases inferred from the adaptive immune receptor of the subject based on the results of the calculation performed using the learning model.

2. The information processing method according to claim 1, wherein the disease information includes at least one of information regarding whether the subject is affected by the disease, the subtype of the disease, the severity of the disease, and information regarding diseases that the subject is likely to develop in the future.

3. The information processing method according to claim 1, wherein the disease information includes information on the probability of suffering from the disease.

4. The information processing method according to claim 1, wherein the disease information includes information on two or more types of diseases.

5. The information processing method according to claim 1, wherein the learning model is a trained learning model that has been trained to output information on diseases estimated from the adaptive immune receptor based on the adaptive immune receptor data of a subject, using a dataset including adaptive immune receptor data obtained for multiple donors and data related to the disease of each donor as training data.

6. The information processing method according to claim 5, wherein the learning model is a trained learning model that has learned the relationship between features extracted from adaptive immune receptor data included in the dataset and disease-related data included in the dataset, and the features are features that characterize a paratope cluster obtained when the adaptive immune receptor data is clustered using a paratope pseudo-sequence as an index, the paratope pseudo-sequence being constructed by pseudo-linking segments representing three complementarity determining regions included in the adaptive immune receptor.

7. A method for generating a learning model, which comprises the steps of: acquiring a dataset including adaptive immune receptor data obtained for a plurality of donors and data relating to the disease of each donor; and using the acquired dataset as training data to generate a learning model that outputs information on diseases inferred from the adaptive immune receptor based on the adaptive immune receptor data of a subject.

8. The method for generating a learning model according to claim 7, wherein the computer executes a process of extracting features from adaptive immune receptor data included in the dataset, and generating the learning model by learning the relationship between the features extracted from the data and disease-related data included in the dataset.

9. The method for generating a learning model according to claim 8, wherein the feature is a cluster occupancy derived from a paratope cluster.

10. The method for generating a learning model described in claim 8, wherein the feature is a feature that characterizes a paratope cluster generated by clustering the adaptive immune receptor data with a plurality of types of paratope pseudo sequences.

11. The method for generating a learning model according to claim 10, wherein the paratope pseudo-sequence is constructed by pseudo-linking segments representing three complementarity determining regions contained in the adaptive immune receptor.

12. The method for generating a learning model according to claim 9, further comprising the steps of: selecting, from all paratope clusters, paratope clusters that satisfy a given population criterion; and extracting the feature quantities using the selected paratope clusters.

13. A method for generating a learning model described in any one of claims 9 to 12, wherein the computer executes a process of deriving a cluster occupancy rate in the paratope cluster using an adjacency matrix constructed from the similarity of adaptive immune receptors contained in the data.

14. A learning model that is trained using a dataset containing adaptive immune receptor data obtained for multiple donors and data related to the disease of each donor as training data, which causes a computer to function such that, when adaptive immune receptor data of a subject is input, the computer performs calculations based on the input adaptive immune receptor data of the subject and outputs information on the disease estimated from the adaptive immune receptor of the subject.

15. A computer program for causing a computer to execute a process of acquiring a dataset including adaptive immune receptor data obtained for a plurality of donors and data relating to the disease of each donor, and using the acquired dataset as training data to generate a learning model that outputs information on diseases inferred from the adaptive immune receptor based on the adaptive immune receptor data of a subject.

16. A computer program for causing a computer to execute the following processes: acquiring adaptive immune receptor data of a subject; using a learning model trained to output information on a disease predicted from the adaptive immune receptor based on the adaptive immune receptor data, performing a calculation based on the acquired adaptive immune receptor data of the subject; and outputting information on the disease obtained from the learning model.

17. An information processing device comprising at least one processor, which acquires adaptive immune receptor data of a subject, performs calculations based on the acquired adaptive immune receptor data of the subject using a learning model trained to output information on a disease predicted from the adaptive immune receptor based on the adaptive immune receptor data, and outputs disease information obtained from the learning model.

Images