Computer system, model learning method, and information processing method

Abstract

This computer system generates a first model serving as a base by machine learning using a first training dataset, and generates a second model by transfer learning using the first model and a second training dataset. In the transfer learning, the second model is generated using at least one of a third loss function including a second loss function that, in accordance with a distribution characteristic of an item corresponding to the prediction object of the second model included in training data constituting the second training dataset, adjusts an influence on training of the training data, and a fifth loss function including a fourth loss function that evaluates the difference between the outputs of the first model and the second model.

Description

Computer system, model learning method, and information processing method Incorporation by Reference

[0001] This application claims priority from Japanese Patent Application No. 2023-220838, filed December 27, 2023, the contents of which are incorporated herein by reference.

[0002] The present invention relates to transfer learning, which utilizes a prediction model generated using a large-scale public database to improve the accuracy of a prediction model that handles data with a small amount of information (e.g., actual clinical data).

[0003] Advances in gene measurement technologies, such as Next Generation Sequencing (NGS), have enabled high-throughput screening, and in recent years, large-scale registries of gene sequence data and the like have been rapidly accumulated. Examples of such data include many clinical trial data, including master clinical trials sponsored by medical institutions such as The Cancer Genome Atlas (TCGA, National Cancer Institute), public cell line data (such as Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC)), and patient-derived xenograft (PDX) data.

[0004] Analytical research using these so-called medical big data is also very active, including research into drug susceptibility prediction models using publicly available cell line data. However, the accuracy of current prediction models, as well as the methods for interpreting the results obtained from them, do not provide sufficient clinical predictive power. This is believed to be due to the fact that commonly used cell lines and xenografts are models that cannot effectively mimic or predict human responses (see Non-Patent Document 1).

[0005] On the other hand, the cell line data is a huge amount of actual case data, with approximately 1,000 cell lines, approximately 300 drugs, and approximately 240,000 drug efficacy experiments (see Non-Patent Document 2), and there is a need for a method to utilize this data in an effective manner.

[0006] To address the issue of validation accuracy using public cell line data, drug response profiling using PDX is progressing, but there is a problem that creating PDX is costly and time-consuming.

[0007] In Non-Patent Document 2, a predictive model trained on public cell line data was transferred to in-vivo tumor and ex-vivo patient-derived xenograft (PDX) data to construct a model for predicting outcomes. It was reported that the AUROC was improved by 49.8% and 47.8% for drug-specific prediction and cancer-specific prediction, respectively, compared to when transfer learning was not used (on average).

[0008] Jennifer L Wilding, Walter F Bodmer, “Cancer cell lines for drug discovery and development”, Cancer Res. 2014 May 1;74(9):2377-84. doi: 10.1158 / 0008-5472.CAN-13-2971.Yi-Ching Tang, Reid T. Powell & Assaf Gottlieb, 「Molecular pathways enhance drug response prediction using transfer learning from cell lines to tumors and patient-derived xenografts」, Scientifc Reports(2022) 12:16109.Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., and He, Q. (2020) A Comprehensive Survey on Transfer Learning. Proceedings of the IEEE 109 (1), 43-76.Motiian, S., Piccirilli, M., Adjeroh, D. A., and Doretto, G.: Unified deep supervised domain adaptation and generalization, in Proceedings of the IEEE International Conference on Computer Vision, pp. 5715-5725 (2017)G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.Ji-Hoon Bae Junho Yim, Donggyu Joo and Junmo Kim.A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. CVPR, 2017. Maaten, LVD and Hinton, G. “Visualizing Data using t-SNE.” Journal of Machine Learning Research (2008), pp. 1-48.

[0009] For example, in clinical practice, medical data generally has an extremely low sample size (N) relative to the input parameters (explanatory variables, features) (P) (P>>N), meaning that machine learning prediction models are unlikely to achieve sufficient prediction accuracy or are likely to overfit. Therefore, there is a demand for clinical application of highly accurate prediction models that utilize large-scale databases, such as publicly available cell line data.

[0010] The transfer learning method used in Non-Patent Document 2 is a general transfer learning method that fine-tunes the final layer. In addition, the publicly available cell line data is not differentiated by target disease, and a prediction model is generated using all cell line data as training data.

[0011] The present invention aims to provide a system and method for generating a predictive model with high prediction performance through transfer learning.

[0012] A representative example of the invention disclosed in the present application is as follows: That is, a computer system including a computer having a processor, a storage device connected to the processor, and a connection interface connected to the processor, which generates a first base model by machine learning using a first training dataset stored in the storage device, and generates a second model by transfer learning using the first model and a second training dataset stored in the storage device, and in the transfer learning, the second model is generated using at least one of a third loss function including a second loss function that adjusts the influence of training data constituting the second training dataset on training in accordance with distribution characteristics of an item corresponding to a target to be predicted by the second model, and a fifth loss function that includes a fourth loss function that evaluates a difference between outputs of the first model and the second model.

[0013] According to the present invention, effective transfer learning using a base model (first model) can be realized. This can improve the prediction accuracy of a prediction model. Problems, configurations, and effects other than those described above will become clear from the following description of the embodiments.

[0014] 1 is a diagram illustrating an example of the configuration of a computer system according to Example 1. FIG. 2 is a diagram illustrating an example of the software configuration in the computer system according to Example 1. FIG. 3 is a diagram illustrating an example of input / output of a base model and a prediction model according to Example 1. FIG. 4 is a diagram illustrating an example of the structure of a base model and a prediction model according to Example 1. FIG. 5 is a diagram illustrating an example of the structure of a base model and a prediction model according to Example 1. FIG. 6 is a diagram illustrating the distribution of dependent variables in the prediction of the prediction model according to Example 1. FIG. 7 is a diagram illustrating an example of clinical trial data (Beat AML data) according to Example 1. FIG. 8 is a diagram illustrating experimental results for the prediction model according to Example 1. FIG. 9 is a diagram for illustrating transfer learning according to Example 2. FIG. 10 is an example of parameters during learning of the prediction model according to Example 2. FIG. 11 is a diagram illustrating an example of experimental results for Example 2. FIG. 12 is a diagram illustrating a transfer learning framework according to Example 3. FIG. 13 is a diagram illustrating a technique for switching learning parameters according to Example 2. FIG. 14 is a diagram illustrating a GUI display of the results of Example 1.

[0015] Hereinafter, embodiments of the present invention will be described with reference to the drawings. However, the present invention should not be construed as being limited to the description of the embodiments shown below. Those skilled in the art will readily understand that the specific configuration can be changed without departing from the spirit or intent of the present invention.

[0016] In the configuration of the invention described below, the same or similar configurations or functions are denoted by the same reference numerals, and redundant explanations will be omitted.

[0017] In this specification, the terms "first," "second," "third," etc. are used to identify components and do not necessarily limit the number or order.

[0018] To facilitate understanding of the invention, the position, size, shape, range, etc. of each component shown in the drawings etc. may not represent the actual position, size, shape, range, etc. Therefore, the present invention is not limited to the position, size, shape, range, etc. disclosed in the drawings etc.

[0019] FIG. 1 is a diagram illustrating an example of the configuration of a computer system according to a first embodiment.

[0020] The computer system is composed of a computer 100, an information terminal 110, and an external storage device 111. The computer 100, the information terminal 110, and the external storage device 111 are connected to one another via a network 109. The network 109 is, for example, a local area network (LAN) or a wide area network (WAN), and the connection method may be either wired or wireless.

[0021] The computer 100 performs machine learning to generate a prediction model that makes predictions (such as diagnosis prediction, optimal treatment prediction, and prognosis prediction) regarding a patient's disease / pathological condition, and also uses the model to present prediction results for actual clinical data (input data 220). The computer 100 includes a CPU 101, a main memory device 102, a secondary memory device 103, a network adapter 104, an input device 105, and an output device 106. The hardware elements are connected to each other via an internal bus 108.

[0022] The CPU 101 executes a program stored in the main memory device 102. The CPU 101 executes processing in accordance with the program, thereby operating as a functional unit (module) that realizes a specific function. In the following description, when a processing is described using a functional unit as the subject, this indicates that the CPU 101 is executing a program that realizes the functional unit.

[0023] The main memory device 102 is a dynamic random access memory (DRAM) that stores programs executed by the CPU 101 and data used by the programs. The main memory device 102 is also used as a work area.

[0024] The secondary storage device 103 is a hard disk drive (HDD) or a solid state drive (SSD), and stores data permanently. The programs and data stored in the main storage device 102 may be stored in the secondary storage device 103. In this case, the CPU 101 reads the programs and information from the secondary storage device 103 and loads them into the main storage device 102.

[0025] The network adapter 104 is an interface for connecting to an external device via a network 109 .

[0026] The input device 105 is a device such as a keyboard, a mouse, or a touch panel, and is used to input data to the computer 100 .

[0027] The output device 106 is a display, a printer, or the like, and is a device for outputting the processing results of the computer 100, etc.

[0028] The hardware configuration of the computer 100 is merely an example and is not limiting. For example, the computer 100 does not need to include the input device 105 and the output device 106.

[0029] The information terminal 110 is a terminal that performs various operations on the computer 100. For example, the information terminal 110 registers / deletes learning data, registers models, and inputs user data. The hardware configuration of the information terminal 110 is the same as that of the computer 100.

[0030] The external storage device 111 stores various types of information and is, for example, an external HDD or a storage system.

[0031] FIG. 2 is a diagram illustrating an example of the software configuration of the computer 100 according to the first embodiment.

[0032] The computer 100 includes a first learning unit 200, a second learning unit 201, and a prediction unit 202, as well as a first learning data DB 210, a second learning data DB 211, and a model DB 212. The first learning data DB 210, the second learning data DB 211, and the model DB 212 may be stored in the external storage device 111.

[0033] The first training data DB 210 is a database that stores training data used for machine learning of the base model. The first training data DB 210 is a large-scale public database that stores public clinical data, clinical trial data, cell line data, etc.

[0034] The second learning data DB 211 is a database that stores learning data used in transfer learning. The learning data is data obtained in actual clinical practice, and includes items such as the patient's age, gender, blood pressure, test results, genetic information, treatment results, and prognosis information.

[0035] The model DB 212 is a database that stores information about models, such as the structure and hyperparameters of the model, such as weights.

[0036] The first learning unit 200 performs machine learning using the learning data stored in the first learning data DB 210 and the base model stored in the model DB 212 , and stores the learned base model in the model DB 212 .

[0037] The second learning unit 201 performs transfer learning using the base model stored in the second learning data DB 211 and the model DB 212 , and stores the generated prediction model in the model DB 212 .

[0038] The prediction unit 202 receives input data 220 using a prediction model stored in the model DB 212 , predicts the medical effects of the patient, and outputs the prediction result 221 .

[0039] Next, a method for learning the prediction model of the first embodiment will be described.

[0040] FIG. 3 illustrates an example of input and output of the base model and the prediction model according to the first embodiment.

[0041] Input / output 301 indicates the input and output of a base model (drug efficacy prediction model) trained using publicly available cell line data. Input / output 302 indicates the input and output of a prediction model (drug efficacy prediction model) obtained by transfer learning of the base model using clinical trial data (Beat AML data published by the Leukemia and Lymphoma Society).

[0042] The input data includes gene mutations, expression levels, gene copy number variations, and drug targets, and the output data is the IC50 value for the drug target (the drug concentration required to halve the number of cells). The inputs and outputs are not limited to these, and an appropriate input / output model can be set in the problem configuration. For example, genetic structural polymorphisms, drug-related genes, the chemical structure of the drug, and treatment history information can be used as inputs, and RECIST, test values, or prognostic information can be used as outputs.

[0043] Transfer learning is generally a technique for solving a problem in a target domain using a solution method (pre-trained model) in a source domain that is different from but related to the target domain (see, for example, Non-Patent Document 1). In the case of the input / output shown in Figure 3, the task solved using Beat AML data is the target domain, and the task solved using the public cell line database is the source domain.

[0044] 4A and 4B are diagrams illustrating an example of the structure of a base model and a prediction model according to the first embodiment.

[0045] Network 401 is an example of a base model to be trained in the source domain, and network 402 is an example of a predictive model to be trained in the target domain.

[0046] Note that there is no guarantee that the data handled in the source domain and the target domain are of the same type and characteristics. In this example, gene copy number variation data (CNV) present in public cell line data is used to train the base model. This is based on medical knowledge that CNV has a high impact on drug efficacy prediction, and is incorporated as a feature of the base model. CNV related to the actual target tumor is not input into the prediction model.

[0047] The computer 100 trains a base model (network 401) using training data (public cell line data) stored in a first training data DB 210. The computer 100 sets the parameters of each layer of the base model as the initial values ​​of the parameters of each layer of a prediction model (network 402), and trains the prediction model (network 402) using training data (clinical trial data) stored in a second training data DB 211.

[0048] Note that a fixed value may be set as the initial value of the parameter of at least one layer. Alternatively, a fixed value may be set as the initial value of the parameter of all layers. When the initial values ​​of the parameters of all layers are set to fixed values, a method may be used in which only the last layer (discrimination layer) 404 is trained. Furthermore, when using a domain adaptation technique (see, for example, Non-Patent Document 4), only the initial value of the first input layer 403 of the network 402 may be set randomly.

[0049] Fig. 5A is a diagram illustrating the distribution of the dependent variable of the clinical trial data (Beat AML data) of Example 1. Fig. 5B is a diagram illustrating the distribution of the dependent variable in the prediction of the prediction model of Example 1. Fig. 6 is a diagram illustrating an example of the clinical trial data (Beat AML data) of Example 1.

[0050] In actual clinical case databases, there are many cases where the names of diseases, drugs, and methods of expressing gene mutations are not standardized. For example, clinical trial data such as Beat AML is collected according to the actual clinical trial conditions, which can make it difficult for those who later use the database to understand the intentions behind the data.

[0051] Figure 6 shows an example of Beat AML data. In this efficacy test, each drug was administered to a sample in vitro, and the amount of drug required to inhibit 10%, 25%, 50%, 75%, and 90% of the drug was examined. However, the amount of drug examined was limited to a certain range. This is thought to be due to the fact that in actual cases, the amount of sample that can actually be used is small compared to cell lines that can be cultured. In Figure 6, the minimum (min_conc) and maximum (max_conc) values ​​of the drug amounts used in the efficacy test are set, but the maximum experimental value (max_conc) is exceeded at the IC50 point. In other words, there is data in which the experiment was terminated when IC50 = max_conc. Of the 17,000 cases in the data used in this example, approximately 7,000 cases contained data such as those described above.

[0052] The aforementioned data is considered to be not useful for training a predictive model. Therefore, in this embodiment, a loss function (second loss function) is introduced to suppress the influence of the aforementioned data on training (parameter update). For example, the loss function shown in Equation (1) can be considered. This makes it possible to adjust the influence of the training data on training according to the distribution characteristics of the objective variable included in the training data.

[0053]

[0054] The graph in FIG. 5A shows the −log １０ This graph shows the distribution of IC50 cases. In this graph, α is a value corresponding to the maximum IC50 value, and β is a value corresponding to the minimum IC50 value. The loss function L shown in equation (1) acts to prevent an increase in the loss of the predicted value for training data whose values ​​are not within the range 501 defined by α and β. In other words, training is performed taking into account only the loss for training data whose values ​​are within the range 501.

[0055] Transfer learning may be performed using a loss function (third loss function) defined by combining the loss function (second loss function) shown in Equation (1) with a loss function (first loss function) that evaluates the prediction accuracy of the prediction model. The loss function that evaluates the prediction accuracy of the prediction model is, for example, cross entropy.

[0056] The graph in Figure 5B shows the predicted distribution after training using loss functions including the loss function in Equation (1). It can be seen that a natural data distribution is obtained even near the boundaries (α and β).

[0057] 7 and 8 are diagrams showing experimental results of the prediction model of Example 1.

[0058] Figure 7 shows a comparison between a prediction model trained using only Beat AML data and a prediction model generated by transfer learning using a model trained using public cell line data. Note that fine tuning was used as the transfer learning method.

[0059] As in Non-Patent Document 1, the evaluation indices used were the RMSE shown in Equation (2) and the Spearman coefficient shown in Equation (3), where D represents the difference in rank between the corresponding estimated value and the correct value, and N represents the number of pairs of values.

[0060]

[0061]

[0062] As shown in FIG. 7, it can be seen that the prediction model of this embodiment has improved accuracy in both RMSE and Spearman coefficient compared to the conventional prediction model.

[0063] Figure 8 shows a performance comparison between the method described in Non-Patent Document 1 and the method of Example 1. The comparison is based on the Spearman coefficient, and the numerical superiority is evident. This is thought to be due to the addition of a loss function according to the characteristics of the data distribution and the incorporation of CNV data features, as described in Example 1.

[0064] FIG. 15 is a diagram illustrating an example of a GUI presented by the computer 100 according to the first embodiment.

[0065] The GUI 1500 includes display fields 1501 , 1502 , 1503 , 1504 , and 1505 .

[0066] A display field 1501 displays information related to the base model, including the version of the base model used, the training data, the number of data used for training, and the performance index (AUC).

[0067] Display field 1502 is a field that displays information about the input data to be input to the prediction model, and shows the items and numbers included in the input data, such as gene mutations and expression levels, and information about target drugs.

[0068] A display field 1503 displays information about the prediction model. A display field 1504 displays the number of data points ultimately used from the input data (the number of data points after preprocessing by data cleaning due to missing data / missing measurements, etc.), the transfer learning method used, and performance indices (AUC, sensitivity / specificity).

[0069] Display field 1504 is a field that displays information about the performance of the predictive model. The user can select from various transfer learning methods, and display field 1504 displays the performance of the predictive model generated by each method.

[0070] Gene mutations corresponding to drugs that are expected to be effective are displayed in a display field 1506. In addition, as an interpretation thereof, for example, a pathway showing the interaction relationship between the gene mutations and the drug is displayed.

[0071] According to Example 1, the prediction accuracy for actual clinical data can be improved by performing transfer learning using a drug efficacy prediction model using public cell line data as a base model. Furthermore, since a base model incorporating clinically influential data and its features is used for the prediction target (output), which cannot be extracted by a prediction model using only actual clinical data, an improvement in the prediction accuracy of the prediction model can be expected.

[0072] In Example 2, the loss function used in transfer learning is different from that in Example 1. Example 2 will be described below, focusing on the differences from Example 1.

[0073] The configuration of the computer system of the second embodiment is the same as that of the first embodiment. The hardware configuration and software configuration of the computer 100 of the second embodiment are the same as those of the first embodiment.

[0074] FIG. 9 is a diagram illustrating transfer learning according to the second embodiment.

[0075] In Example 1, a domain adaptation method was used to generate a prediction model. In Example 2, we propose a knowledge distillation method (see, for example, Non-Patent Document 5) that transfers learning features (tacit knowledge) so that the model can be applied to a clinical data usage environment with further limited input data.

[0076] 9, a network 901 shows an example of a prediction model generated by the method of Example 1. Hereinafter, the prediction model generated by the method of Example 1 will be referred to as a teacher model.

[0077] Here, a CNN (Convolutional Neural Network) is used as the prediction model (VGG16 model in this example), but the prediction model is not limited to this.

[0078] The input data 903 includes, for example, gene mutation, expression level, copy number, patient information, and the like.

[0079] Network 902 is an example of a model that handles input data 904, which has a more limited amount of information than input data 903. The input data 904 is, for example, data that includes only genetic mutations. Hereinafter, the prediction model corresponding to network 902 will be referred to as a student model.

[0080] Here, limiting the amount of information in the input data means that the number of items contained in the input data is less than the number of items contained in the input data to be compared, or that the amount of data in the values ​​of the items contained in the input data is less than the amount of data in the values ​​of the items contained in the input data to be compared.

[0081] In the method of this embodiment, in addition to the output of the teacher model, knowledge regarding the features of the intermediate layer is also distilled into the teacher model. The distillation of the features of the intermediate layer uses a well-known method, the FSP matrix (see Non-Patent Document 6). By converting the outputs of multiple intermediate layers into a Gram matrix called an FSP matrix, an FSP matrix is ​​generated for each of the teacher model and the student model, and the L2-loss between the FSP matrices is calculated, thereby realizing learning that takes into account the relationship (flow of process) between the intermediate layers.

[0082] In FIG. 9, the computer 100 calculates the Gram Matrix (G １ Ａ , G ３ Ａ , G ５ Ａ , G １ Ｂ , G ３ Ｂ , G ５ Ｂ In equation (4), l represents the channel number, k represents the block number, and h and w represent the convolution layer C １ , C ２ The width and depth of

[0083]

[0084] The value of the Gram Matrix represents the correlation between the feature amounts output from each Block. The feature amounts output from each Block are normalized using Equation (5).

[0085]

[0086] Loss function L ｋｄｆｔ The fifth loss function is defined as a combination of the fourth loss function, which evaluates the difference between the outputs of each layer. Specifically, the binary-cross entropy loss (L ｂｃ ) and the L2-loss (L Ｇｋ ) and the MSE error (L ｆｔ ) and are combined to define as in equation (9).

[0087]

[0088]

[0089]

[0090]

[0091] Here, y represents the output (predicted value) of the teacher model, and y^ represents the output (predicted value) of the student model. １ , λ ３ , λ ５ , λ ｆｔ , λ ｂｃ represents the weight.

[0092] L Ｇｋ The purpose of introducing the above is to more accurately distill the features of the teacher model 901 into the student model 902.

[0093] In equation (9), each weight is, for example, λ １ , λ ３ , λ ５ is 0.4, λ ｆｔ is 0.04, λ ｂｃ is set to 0.01. An example of the parameters during learning is shown in FIG.

[0094] In addition, λ １ , λ ３ , λ ５ , λ ｆｔ , λ ｂｃ may be changed depending on the progress of the study.

[0095] Specifically, as shown in FIG. 14, in the learning process, when the learning error shown on the vertical axis is equal to or greater than threshold A (initial stage of learning), λ is set to match the feature quantities of the previous stages of the network. １ , λ ３ , λ ５ Also, when the learning error becomes equal to or smaller than the threshold B (later learning stage), the value of λ is set to be large in order to further improve the accuracy of the prediction results near the final layer. ｆｔ , λ ｂｃ The value of λ is set to a large value. Alternatively, the weight may be set based on the number of epochs. For example, when the value is equal to or less than the threshold C (initial stage of learning), １ , λ ３ , λ ５ When the value of is set large and is above the threshold C (late learning stage), λ ｆｔ , λ ｂｃ Set the value to a larger value.

[0096] Furthermore, a weight setting method that combines both of the above may also be used.

[0097] FIG. 11 is a diagram showing an example of predictions of the teacher model 901 and the student model 902.

[0098] Here, it is assumed that pathological tissue image data is input to the teacher model 901, and MRI data is input to the student model 902. It is also assumed that the teacher model 901 and the student model 902 predict ACU.

[0099] The resolution of pathological tissue image data is two or more orders of magnitude higher than that of MRI data, and it contains a large amount of information. Here, a model was generated that uses only MRI data, which contains little information, to estimate disease risk equivalent to pathological data, which contains a large amount of information. The student model (Model_kdft) using the method of this example exhibited higher discrimination performance than the student model (Model_mri) that predicts using only MRI data, and achieved performance close to that of the teacher model (Modek_hst).

[0100] FIG. 12 shows two-dimensional space mapping after dimensionality reduction of the feature vectors (dense2 output) of Model_mri, Model_hist, and Model_kdft using t-SNE (t-distributed Stochastic Neighbor Embedding) (see Non-Patent Document 7).

[0101] In Model_mri, the data of the two classes overlap at the boundary, but Model_hist shows that a distribution that allows the two classes to be classified is obtained. Although there is still overlap in Model_kdft, the degree of overlap is smaller than in Model_mri and the range is narrower, suggesting that the classification performance is closer to that of Model_hst.

[0102] This embodiment makes it possible to maintain the performance of a highly accurate prediction model and effectively inherit the features of a teacher model both in clinical practice where input data is more limited and in model construction. In other words, a prediction model with high prediction accuracy can be generated even for input data with a limited amount of information.

[0103] In the third embodiment, a framework for realizing the transfer learning described in the first and second embodiments will be described. The configuration of the computer system in the third embodiment is the same as that in the first embodiment. In addition, the hardware configuration and software configuration of the computer 100 in the third embodiment are the same as those in the first embodiment.

[0104] FIG. 13 is a diagram illustrating a framework of transfer learning in the third embodiment.

[0105] The computer 100 performs preprocessing 1302 on the data stored in the database 1301 (public database and proprietary database) to refine the data and extract the data characteristics and relationships between the data, and generates a base model 1303 by performing machine learning using the preprocessed data.

[0106] The computer 100 executes the transfer learning of the first embodiment using actual clinical data (in vivo) 1304 starting from the base model, and generates a highly accurate clinical prediction model (in-vivo model) 1305.

[0107] Even when the target is actual clinical data, if the amount of information in the input data is limited compared to the amount of information in the input data used in training the base model, or if it is desired to generate a model smaller in size than the base model, the computer 100 executes the transfer learning of Example 2 to generate a highly accurate clinical prediction model (Tiny Model) 1307.

[0108] A case where a small-sized model is required is assumed to be a case where there are limitations on the hardware implementation environment of the computer or computer system that uses the prediction model.

[0109] Data refinement in preprocessing 1302 refers to data cleaning processes such as data cleansing, missing value interpolation, noise removal of outliers, and normalization. Feature extraction refers to the process of identifying important biological pathways using pathway enrichment analysis. Pathways with high enrichment scores are analyzed from a group of genes included in a gene list. Extraction of substructures using fingerprinting is also possible. The most commonly used fingerprinting method is Morgan Fingerprint (ECFP). It is a method of representing molecular structure based on two-dimensional structure and is primarily used for molecular "similarity searches" and for building quantitative structure-activity / property relationship (QSAR, QSPR) models.

[0110] Other examples include the extraction of specific features through filtering, dimensionality compression, etc. Extracting relationships between data refers to the process of integrating multiple data sets by taking into account the relationships between each piece of input data, and adding new parameters. For example, one approach is to include the correlation coefficient between data sets as input data. Note that feature extraction and data relationship extraction are not necessarily required processes.

[0111] The framework of this embodiment makes it possible to effectively transfer knowledge and perform distillation processing from a base model that utilizes a large-scale public database or existing archive samples to a small number of clinical samples (with few features), thereby improving the sophistication and efficiency of predictive model generation that is in line with the on-site environment and requirements of actual clinical practice.

[0112] The present invention is not limited to the above-described embodiments, but includes various modifications. For example, the above-described embodiments are provided to explain the present invention in detail, and the present invention is not necessarily limited to those including all of the described configurations. Furthermore, some of the configurations of each embodiment can be added to, deleted from, or replaced with other configurations.

[0113] Furthermore, some or all of the above-described configurations, functions, processing units, processing means, etc. may be implemented in hardware, for example, by designing them as integrated circuits. The present invention can also be realized by software program code that implements the functions of the embodiments. In this case, a storage medium on which the program code is recorded is provided to a computer, and a processor included in the computer reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium implements the functions of the above-described embodiments, and the program code itself and the storage medium on which it is stored constitute the present invention. Examples of storage media for providing such program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, solid-state drives (SSDs), optical disks, magneto-optical disks, CD-Rs, magnetic tape, non-volatile memory cards, and ROMs.

[0114] Furthermore, the program code that realizes the functions described in this embodiment can be implemented in a wide range of programming or scripting languages, such as assembler, C / C++, perl, Shell, PHP, Python, and Java.

[0115] Furthermore, the program code of the software that realizes the functions of the embodiments may be distributed via a network and stored in a storage means such as a computer's hard disk or memory, or in a storage medium such as a CD-RW or CD-R, and the processor of the computer may read and execute the program code stored in the storage means or the storage medium.

[0116] In the above-described embodiment, the control lines and information lines are those that are considered necessary for the explanation, and not all control lines and information lines are necessarily shown in the product. All components may be interconnected.

Claims

1. A computer system comprising a computer having a processor, a storage device connected to the processor, and a connection interface connected to the processor, generating a first base model by machine learning using a first learning dataset stored in the storage device, generating a second model by transfer learning using the first model and a second learning dataset stored in the storage device, wherein in the transfer learning, a third loss function including a second loss function for adjusting the influence of the learning data on the learning according to the distribution characteristics of the items corresponding to the prediction target of the second model included in the learning data constituting the second learning dataset, and a fifth loss function including a fourth loss function for evaluating the difference between the outputs of the first model and the second model, at least one of which is used to generate the second model.

2. The computer system according to claim 1, wherein the second loss function is a function for suppressing the update of the second model by the learning data in which the values of the items corresponding to the prediction target of the second model included in the learning data constituting the second learning dataset are not included in a predetermined range.

3. The computer system according to claim 1, wherein the first model and the second model are networks including an input layer, a plurality of intermediate layers, and an output layer, wherein the fourth loss function includes a loss function for evaluating the difference between the outputs of at least one of the input layer and the plurality of intermediate layers of the first model and the second model, and a loss function for evaluating the difference between the outputs of the output layers of the first model and the second model.

4. The computer system according to claim 1, wherein the third loss function and the fifth loss function include coefficients, and the computer system executes the transfer learning while changing the coefficients included in the third loss function or the fifth loss function according to the progress of the transfer learning.

5. The computer system according to claim 1, wherein the second model is a model for predicting a patient's disease or medical condition.

6. A computer system according to claim 1, wherein the learning data included in the first learning dataset includes gene information of a cell line and drug response, the first model outputs the drug response from the gene information of the cell line, the learning data included in the second learning dataset is clinical data indicating the drug response of a patient, and the second model is a model that receives the gene mutation information of the patient as an input and makes a prediction regarding the drug response in the patient.

7. A method for learning a model executed by a computer system, the computer system having a computer including a processor, a storage device connected to the processor, and a connection interface connected to the processor, the method for learning the model including: a first step in which the processor generates a first model by performing machine learning using a first learning dataset stored in the storage device and stores the first model in the storage device; and a second step in which the processor generates a second model by performing transfer learning using the first model and a second learning dataset stored in the storage device and stores the second model in the storage device, the second step including a step of generating the second model using at least one of a third loss function including a second loss function for adjusting the influence of the learning data on the learning of the second model according to the distribution characteristics of the items corresponding to the prediction target of the second model included in the learning data constituting the second learning dataset, and a fifth loss function including a fourth loss function for evaluating the difference between the outputs of the first model and the second model.

8. A method for learning a model according to claim 7, wherein the second loss function is a function for suppressing the update of the second model by the learning data included in the learning data constituting the second learning dataset, in which the value of the item corresponding to the prediction target of the second model is not included in a predetermined range.

9. The method for training a model according to claim 7, wherein the first model and the second model are networks including an input layer, a plurality of intermediate layers, and an output layer, and the fourth loss function includes a loss function for evaluating a difference in outputs of at least one of the input layer and the plurality of intermediate layers of the first model and the second model, and a loss function for evaluating a difference in outputs of the output layers of the first model and the second model.

10. The method for training a model according to claim 7, wherein the third loss function and the fifth loss function include coefficients, and the second step includes a step in which the processor performs the transfer learning while changing the coefficients included in the third loss function or the fifth loss function according to the progress of the transfer learning.

11. The method for training a model according to claim 7, wherein the second model is a model for predicting a patient's disease or pathological condition.

12. The method for training a model according to claim 7, wherein the training data included in the first training dataset includes gene information of cell lines and drug responses, the first model outputs the drug responses from the gene information of the cell lines, the training data included in the second training dataset is clinical data indicating a patient's drug response, and the second model is a model that receives the patient's gene mutation information as an input and makes a prediction regarding the drug response in the patient.

13. A computer system comprising a computer having a processor, a storage device connected to the processor, and a connection interface connected to the processor, wherein the storage device stores a second model generated by transfer learning of a base first model, the first model being generated by machine learning using a first learning dataset, the second model being generated by the transfer learning using the first model and a second learning dataset, and in the transfer learning, at least one of a third loss function including a second loss function for adjusting the influence of the learning data on the learning according to the distribution characteristics of an item corresponding to the prediction target of the second model included in the learning data constituting the second learning dataset, and a fifth loss function including a fourth loss function for evaluating the difference between the outputs of the first model and the second model is used, and the computer system receives an input of input data and performs a prediction using the second model and the input data.

14. An information processing method executed by a computer system, wherein the computer system has a computer having a processor, a storage device connected to the processor, and a connection interface connected to the processor, the storage device stores a second model generated by transfer learning of a base first model, the first model being generated by machine learning using a first learning dataset, the second model being generated by the transfer learning using the first model and a second learning dataset, and in the transfer learning, at least one of a third loss function including a second loss function for adjusting the influence of the learning data on the learning according to the distribution characteristics of an item corresponding to the prediction target of the second model included in the learning data constituting the second learning dataset, and a fifth loss function including a fourth loss function for evaluating the difference between the outputs of the first model and the second model is used, and the information processing method includes a step of the processor receiving an input of input data and a step of the processor performing a prediction using the second model and the input data.

Images