Method and device for training model for estimating polygenic risk score

The method trains a deep learning model to estimate polygenic risk scores by constructing a learning dataset with genetic information and using backpropagation to update parameters, addressing the challenge of accurately assigning SNP weights and improving prediction accuracy by considering SNP interactions, thus enhancing the model's ability to predict disease risk.

WO2024228610A1PCT designated stage expired Publication Date: 2024-11-07GENOPLAN INC
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
PCT/KR2024/095732
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-23
Filing Date
2024-04-15
Publication Date
2024-11-07

AI Technical Summary

Technical Problem

Current methods for calculating polygenic risk scores using deep learning models face challenges in accurately assigning weights to Single Nucleotide Polymorphisms (SNPs) due to overfitting and the complexity of interactions between SNPs, which affects the model's accuracy in predicting disease risk.

Method used

A method and apparatus for training a deep learning model that constructs a learning dataset with genetic information and corresponding polygenic risk scores, using backpropagation to update parameters based on a loss function that considers the error between estimated and calculated scores, while also accounting for the correlation between SNPs to assign appropriate weights.

Benefits of technology

This approach prevents overfitting and improves the model's accuracy in estimating polygenic risk scores by effectively learning the interactions between SNPs, leading to a more reliable prediction of disease risk.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure KR2024095732_07112024_PF_FP_ABST
    Figure KR2024095732_07112024_PF_FP_ABST
Patent Text Reader

Abstract

A method for training a model for estimating a polygenic risk score (PRS) according to one embodiment of the present disclosure may configure a training data set including at least one training data having genetic information about a single nucleotide polymorphism (SNP) as an input and a PRS calculation value corresponding to the genetic information as a label. In addition, the method may train the model for estimating the polygenic risk score on the basis of the training data set.
Need to check novelty before this filing date? Find Prior Art

Description

Method and device for training a model for estimating a polygenic risk score

[0001] The present disclosure relates to a method and apparatus for training a model for estimating a polygenic risk score.

[0002] The polygenic risk score (PRS) method measures the risk of a specific disease due to inherited factors. Recently, research is being conducted on methods that utilize deep learning models to calculate polygenic risk scores, reflecting the interactions between individual SNPs within a SNP list.

[0003] Accordingly, to improve the accuracy of the model that calculates the polygenic risk score, there is a need to train a deep learning model to assign weights to each SNP by considering the association between SNPs.

[0004] The background technology described above is technical information that the inventor possessed for the purpose of deriving the present invention or acquired in the process of deriving the present invention, and cannot necessarily be considered as publicly known technology disclosed to the general public prior to the application for the present invention.

[0005] Some embodiments of the present disclosure provide a method and device for training a model to estimate a polygenic risk score. The problems addressed by the present invention are not limited to those mentioned above. Other problems and advantages of the present invention, not mentioned above, can be understood through the following description and will be more clearly understood through embodiments of the present invention. Furthermore, it will be appreciated that the problems and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims.

[0006] As a technical means for achieving the above-described technical problem, a first aspect of the present disclosure may provide a method for training a model for estimating a polygenic risk score (PRS), the method comprising: a step of configuring a training data set including at least one training data having as input genetic information about a single nucleotide polymorphism (SNP) and a PRS calculated value corresponding to the genetic information as a label; and a step of training the model for estimating the polygenic risk score based on the training data set; wherein the training step comprises: a step of computing the genetic information using the model; a step of obtaining a PRS estimated value corresponding to the genetic information from the model; and a step of updating parameters for the model by performing backpropagation based on a loss function regarding an error between the PRS estimated value and the PRS calculated value, which is a label for the genetic information.

[0007] A second aspect of the present disclosure provides a device for training a model for estimating a polygenic risk score, comprising: a memory in which at least one program is stored; and a processor for executing the at least one program, wherein the processor is configured to input genetic information about a single nucleotide polymorphism (SNP), construct a training data set including at least one training data labeled with a PRS calculation value corresponding to the genetic information, and train the model for estimating the polygenic risk score based on the training data set, wherein the processor calculates the genetic information using the model, obtains a PRS estimate value corresponding to the genetic information from the model, and performs backpropagation on the model based on a loss function relating to an error between the PRS estimate value and the PRS calculation value, which is a label for the genetic information, thereby updating parameters for the model.

[0008] A third aspect of the present disclosure can provide a computer-readable recording medium having recorded thereon a program for executing the method according to the first aspect on a computer.

[0009] In addition, other methods for implementing the present invention, other systems, and computer-readable recording media storing a computer program for executing the method may be further provided.

[0010] Other aspects, features and advantages other than those described above will become apparent from the following drawings, claims and detailed description of the invention.

[0011] According to one embodiment of the present disclosure, a model for estimating a polygenic risk score can be prevented from overfitting to training data during the learning process.

[0012] More specifically, according to one embodiment of the present disclosure, a model for estimating a polygenic risk score can be trained to assign a weight to each single nucleotide polymorphism (SNP) by taking into account the association between the SNPs.

[0013] FIG. 1 is a diagram illustrating an example of a model for estimating a polygenic risk score (PRS) according to one embodiment.

[0014] FIG. 2 is a diagram illustrating an example of a learning data set used to train a model according to one embodiment.

[0015] FIG. 3 is a diagram illustrating an example of a method for training a model based on a learning data set according to one embodiment.

[0016] FIG. 4 is a diagram illustrating another example of a method for training a model based on a learning data set according to one embodiment.

[0017] FIG. 5 is a flowchart illustrating an example of a method for training a model for estimating a polygenic risk score (PRS) according to one embodiment.

[0018] FIG. 6 is a schematic diagram illustrating an example of the internal configuration of a device for training a model for estimating a polygenic risk score (PRS) according to one embodiment.

[0019] According to one embodiment of the present disclosure, a method for training a model for estimating a polygenic risk score (PRS) may include the steps of: configuring a training data set including one or more training data having as input genetic information regarding a single nucleotide polymorphism (SNP) and a PRS calculated value corresponding to the genetic information as a label; and training the model for estimating the polygenic risk score based on the training data set. In addition, the training step may include the steps of: calculating the genetic information using the model; obtaining a PRS estimated value corresponding to the genetic information from the model; and updating parameters for the model by performing backpropagation based on a loss function regarding an error between the PRS estimated value and the PRS calculated value, which is a label for the genetic information.

[0020] The advantages and features of the present invention, and the methods for achieving them, will become clearer with reference to the embodiments described in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments presented below, but may be implemented in various different forms, and it should be understood that it includes all transformations, equivalents, and substitutes included in the spirit and technical scope of the present invention. The embodiments presented below are provided to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention. In describing the present invention, if a detailed description of a related known technology is judged to obscure the gist of the present invention, the detailed description thereof will be omitted.

[0021] The terminology used in this application is only used to describe specific embodiments and is not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly indicates otherwise. In this application, it should be understood that the terms "comprise" or "have" indicate the presence of a feature, number, step, operation, component, part, or combination thereof described in the specification, but do not preclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

[0022] Some embodiments of the present disclosure may be represented by functional block configurations and various processing steps. Some or all of these functional blocks may be implemented by various hardware and / or software configurations that perform specific functions. For example, the functional blocks of the present disclosure may be implemented by one or more microprocessors or by circuit configurations for a given function. Furthermore, for example, the functional blocks of the present disclosure may be implemented in various programming or scripting languages. The functional blocks may be implemented by algorithms that execute on one or more processors. Furthermore, the present disclosure may employ conventional techniques for electronic configuration, signal processing, and / or data processing. Terms such as "mechanism," "element," "means," and "configuration" may be used broadly and are not limited to mechanical and physical configurations.

[0023] Additionally, the connecting lines or connecting members between components depicted in the drawings are merely exemplary representations of functional connections and / or physical or circuit connections. In an actual device, connections between components may be represented by various functional connections, physical connections, or circuit connections that may be replaced or added.

[0024] The term "gene" as used herein means a segment of a nucleic acid sequence (also referred to herein as a "coding sequence" or "coding region") that encodes a protein or RNA, which may be accompanied by a regulatory region, such as a promoter, operator, terminator, etc., which may be located upstream or downstream of the coding sequence.

[0025] The term "genetic information" as used herein encompasses information obtained through genetic analysis of a subject, including, for example, information on genetic traits or genetic mutations associated with the occurrence of a specific disease. Such genetic mutations may be, but are not limited to, missense mutations, frameshift mutations, nonsense mutations, splice mutations, nucleotide substitutions, insertions, or deletions. In a specific example, the genetic information may include single nucleotide polymorphisms (SNPs). The disease occurrence risk calculated based on such genetic information includes the meaning of an innate risk of occurrence for the disease.

[0026] In this specification, "polymorphism" refers to the presence of two or more alleles at a single genetic locus, and among the polymorphic sites, a polymorphism in which only a single base differs from person to person is called a single nucleotide polymorphism (SNP). A desirable polymorphic marker has two or more alleles that exhibit an occurrence frequency of 1% or more, more preferably 5% or 10% or more, in a selected population.

[0027] In this specification, the "Polygenic Risk Score" is a genetic assessment index that assesses the likelihood of an individual or entity developing a specific disease by synthesizing the influence of multiple genetic variations found in the genome of each individual or entity. Generally, a PRS can be calculated as a combination of single nucleotide polymorphisms (SNPs). In other words, a single nucleotide polymorphism (SNP) is a replacement of one base in a DNA sequence with another base. When such a base replacement occurs at a location within a gene, a trait or disease associated with the gene can be expressed. Therefore, a PRS can be calculated as a combination of SNPs associated with a specific disease.

[0028] In this specification, "model" may refer to a deep learning model based on an artificial neural network (ANN). A deep learning model has a deep structure with multiple hidden layers in an ANN, which allows it to learn complex patterns. For example, a deep learning model may be composed of an input layer, a hidden layer, and an output layer, and each layer may be composed of multiple neurons. Furthermore, each neuron may calculate an output value by applying a value obtained by multiplying an input value by a weight to an activation function, and each layer may transmit a signal to the next layer, which may be adjusted through weights and biases. In other words, a deep learning model uses training data to optimize weights and biases, thereby learning complex relationships between inputs and outputs. Meanwhile, deep learning models can be implemented using various types of models, including multi-layer perceptrons (MLPs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). Furthermore, deep learning models can be implemented in a variety of ways depending on how they configure activation functions, loss functions, and optimization algorithms, and are not limited to a specific type of model.

[0029] In this specification, "training data" and "training data set" refer to data used to train an artificial intelligence model. Each data set may consist of an input and a label. "Input" refers to data included in the training data that is input to the model, and "label" refers to the correct answer for the input data, which is the correct answer that the model should predict among the data included in the training data.

[0030] Below, embodiments are described in detail with reference to the attached drawings. However, the embodiments may be implemented in various different forms and are not limited to the examples described herein.

[0031] FIG. 1 is a diagram illustrating an example of a model for estimating a polygenic risk score (PRS) according to one embodiment.

[0032] Referring to FIG. 1, a device for calculating a polygenic risk score (PRS) according to one embodiment of the present disclosure (hereinafter, “PRS calculating device”) can calculate a PRS estimate (130) from a genome-wide association study (GWAS) result (110) using a model (hereinafter, “model”) (120) for estimating a polygenic risk score (PRS). For example, input data of the model (120) may be a GWAS result (110), and output data of the model (120) may be a PRS estimate (130).

[0033] Here, genome-wide association study (GWAS) is an exploratory method for identifying traits (e.g., height, hair color, eye color, and disease risk) associated with genetic variants. Genome-wide association study typically compares the genetic information of cases (a group with the trait of interest, such as patients) and controls (a group without the trait, such as normal individuals) across the entire genome, and selects genetic variants with a higher frequency in cases as those associated with the trait.

[0034] In other words, when the aforementioned trait of interest is limited to the onset of a specific disease, the genome-wide association analysis result (110) may include information regarding the association between a genetic mutation and the onset of a specific disease. That is, the genome-wide association analysis result (110) may include one or more single nucleotide polymorphisms (SNPs) that increase the likelihood of developing a specific disease.

[0035] Accordingly, disease risk can be predicted based on the results of genome-wide association analysis (110). For example, disease risk can be predicted by considering the presence or absence of mutations in specific genes known to play a significant role in disease development, including numerous genetic mutations identified through genome-wide association analysis.

[0036] An example of a method for predicting disease risk may include calculating a polygenic risk score (PRS). Furthermore, the PRS can be estimated using a deep learning model. For example, the model (120) may be a deep learning model consisting of an input layer, an output layer, and multiple hidden layers between the input and output layers.

[0037] For example, the model (120) can receive a SNP list including one or more SNPs as input and output a PRS estimate based on the SNP list. Here, each SNP included in the SNP list can constitute a node of the input layer. In addition, each hidden layer can transmit the output to the next layer through a series of linear operations or activation functions using the input data. The output layer can finally perform the above-described operations to output a PRS estimate.

[0038] In this process, the model (120) can assign a weight to each SNP included in the SNP list and output a PRS estimate (130) based on the assigned weight. For example, the model (120) can receive an SNP list including three SNPs with SNP IDs of rs1234, rs5678, and rs9101, respectively. For example, if the model (120) has learned to output a high PRS in response to the inclusion of the rs5678 SNP in the input data during the learning phase, the model (120) can assign a high weight to the rs5678 SNP. As another example, if the model (120) has learned to output a low PRS in response to the inclusion of the rs1234 SNP in the input data during the learning phase, the model (120) can assign a low weight to the rs1234 SNP. As in the example described above, the model (120) can assign weights to SNPs included in the input SNP list according to what has been learned in the learning step, and can output a PRS estimate (130) based on the weights.

[0039] In this way, the PRS calculation device according to the embodiment of the present disclosure can calculate the PRS by reflecting the interaction between each SNP included in the SNP list by calculating the PRS estimate value (130) using a deep learning model.

[0040] Referring to FIGS. 2 and 3 below, an example of a method for training a model to output a PRS estimate that appropriately reflects the interaction between SNPs is described.

[0041] FIG. 2 is a diagram illustrating an example of a learning data set used to train a model according to one embodiment.

[0042] Referring to FIG. 2, a device (hereinafter referred to as “device”) for training a model for estimating a polygenic risk score (PRS) may input genetic information about a single nucleotide polymorphism (SNP) as input (220) and configure a learning data set (210) including one or more learning data having a PRS calculation value corresponding to the genetic information as a label (230).

[0043] For example, the genetic information about SNPs may be a SNP list including one or more SNPs. More specifically, the device may generate a SNP list based on a predetermined P-value in data about genome-wide association analysis (GWAS), and use the generated SNP list as an input (220) of the training data set (210). For example, the device may determine a predetermined P-value for selecting SNPs, and may select one or more SNPs having data significance from data about GWAS based on the determined P-value, thereby generating a SNP list.

[0044] For example, the training data set (210) may include training data having as input (220) a first SNP list having a P-value less than 0.0003 and a first PRS calculation value corresponding to the first SNP list as a label (230). As another example, the training data set (210) may include training data having as input (220) a second SNP list having a P-value less than 0.0002 and a second PRS calculation value corresponding to the second SNP list as a label (230).

[0045] Meanwhile, according to one embodiment, the label (230) included in the learning data set (210) may be a disease determination result corresponding to genetic information. For example, a PRS value may be determined based on a list of SNPs included in the input (220) of the learning data set (210), and the result of determining whether or not a disease is present based on the determined PRS value may be a label corresponding to the corresponding input.

[0046] According to another embodiment, the label (230) included in the learning data set (210) may be a trait value corresponding to genetic information. For example, a PRS value may be determined based on a list of SNPs included in the input (220) of the learning data set (210), and a trait value calculated based on the determined PRS value may be a label corresponding to the corresponding input. Here, the trait value is a value related to a feature of an individual or an organism, and may include values ​​that can be expressed numerically, such as height and weight, as well as hair color, blood type, and disease resistance values.

[0047] FIG. 3 is a diagram illustrating an example of a method for training a model based on a learning data set according to one embodiment.

[0048] Referring to FIG. 3, the device can train a model (320) that estimates a polygenic risk score (PRS) based on a learning data set.

[0049] For example, the device can calculate genetic information (310) using a model (320). In addition, the device can obtain a PRS estimate (330) corresponding to the genetic information (310) from the model (320). For example, the PRS estimate (330) may be an estimate of the PRS output by the model (320) based on the weight assigned to each SNP included in the SNP list, as described in FIG. 1.

[0050] In addition, the device can update the parameters of the model (320) by performing backpropagation on the model (320) based on a loss function regarding the error between the PRS estimated value (330) and the PRS calculated value (340), which is a label for the genetic information (310). In other words, the device can train the model (320) in a direction that reduces the error between the PRS calculated value (340) included in the training data set and the PRS estimated value (330) output from the model (320). That is, the model (320) can update its own parameters (e.g., weights, biases, etc.) so that the model outputs an accurate PRS estimated value for the input SNP list as the training is repeated. Meanwhile, the loss function here refers to a function that measures the difference between the model's predicted value for PRS and the actual value, that is, the PRS estimated value (330) and the PRS calculated value (340), and the type of function, such as mean squared error (MSE) and cross entropy, is not limited.

[0051] Meanwhile, the model (320) can learn the weights assigned to each SNP included in the SNP list during the learning process. For example, as in the example described in FIG. 1, if the model (320) receives a SNP list including three SNPs with SNP IDs of rs1234, rs5678, and rs9101, the model (320) can learn to assign a weight of 0.4 to rs1234, a weight of 0.2 to rs5678, and a weight of 0.4 to rs9101 through learning.

[0052] Here, the model (320) needs to be trained to assign appropriate weights to each SNP included in the SNP list, taking into account the associations between SNPs. For example, the rs9101 SNP, which should be given a low weight due to its low association with the risk of developing a specific disease, may be given a relatively high weight due to its association with the rs1234 SNP.

[0053] For example, due to the nature of genetic information, SNPs located around the same gene are likely to be highly related to each other because they are derived from the same ancestor on the same chromosome. For example, Linkage Disequilibrium (LD) can be an indicator that measures the genetic association between one SNP and another. For example, LD can be expressed as a value from 0 to 1, and the closer the LD value is to 1, the stronger the association between the two SNPs. In the example described above, the rs1234 SNP and the rs9101 SNP may have an LD value close to 1 because they are located close to each other within the gene.

[0054] As another example, the model (320) may be input with a list of SNPs including four SNPs, each with SNP IDs of rs1234, rs5678, rs9101, and rs1121. In this case, in order to calculate an appropriate PRS corresponding to the SNP list, the influence of rs1234 and rs1121 should be evaluated highly. However, the model (320) may evaluate the influence of rs1234 and rs9101 as high due to the high correlation between rs1234 and rs9101 based on what it has learned. In other words, the model (320) may assign a high weight to the rs9101 SNP instead of the rs1121 SNP due to overfitting to the training data, which may ultimately reduce the accuracy of the output of the model (320).

[0055] Meanwhile, as an example of a method for enabling the model (320) to estimate an accurate PRS value, a method using an LD panel that includes information on LD values ​​between SNPs included in the SNP list can be used. For example, the model (320) can estimate a PRS value by reflecting the association between SNPs by constructing a prior distribution based on the LD panel and adjusting (shrinking) the distribution of the GWAS results. However, there may be a problem that using the LD panel for this purpose unnecessarily requires excessive data.

[0056] Accordingly, with reference to FIG. 4 below, another example of a method for training a model based on a training data set is described. More specifically, the model training method described with reference to FIG. 4 relates to a method for training a model to assign weights to each SNP by considering the association between SNPs.

[0057] FIG. 4 is a diagram illustrating another example of a method for training a model based on a learning data set according to one embodiment.

[0058] Referring to FIG. 4, the computation process between the first layer (410) and the second layer (420), which are layers of the hidden layer that constitutes the model, is illustrated. For example, the value corresponding to the 2.1 node (421), which is the 1st node of the second layer (420), may be calculated based on a plurality of nodes of the first layer (410). Here, w(1,1) may mean a weight assigned to the value of the 1.1 node (411), which is the 1st node of the first layer (410), in calculating the value of the 2.1 node (421) of the second layer (420). Similarly, w(1,2) may refer to a weight assigned to the value of the 1.2 node (412), which is the 2nd node of the 1st layer (410), and w(1,3) may refer to a weight assigned to the value of the 1.3 node (413), which is the 3rd node of the 1st layer (410). For convenience of explanation, when explaining excluding the bias value, the value of the 2.1 node (421) may be calculated as the sum of the value of the 1.1 node (411) * w(1, 1), the value of the 1.2 node (412) * w(1,2), and the value of the 1.3 node (413) * w(1,3).

[0059] In the example, if a higher weight is given to the values ​​of the 1.1 node (411) and the 1.3 node (413) due to the high correlation between the 1.1 node (411) and the 1.3 node (413), than the influence, the device can adjust to decrease the weight given to the values ​​of the 1.1 node (411) and the 1.3 node (410). Conversely, if a lower weight is given to the values ​​of the 1.1 node (411) and the 1.3 node (413), than the influence, the device can adjust to increase the weight given to the values ​​of the 1.1 node (411) and the 1.3 node (410).

[0060] To this end, according to one embodiment of the present disclosure, the device can redefine a loss function that guides the learning direction of the model. In other words, the device can update model parameters (e.g., weights and biases assigned to each node) by performing backpropagation on the model based on the gradient of the newly derived loss function.

[0061] In one embodiment, the loss function may include a first term that computes the error between the PRS estimate and the PRS calculation, and a second term that computes the correlation between the weights assigned to SNPs input to the model.

[0062] For example, the correlation between weights can be calculated by computing the vector similarity for the weight vectors. In other words, a high degree of similarity between weight vectors may indicate a high degree of association between the corresponding SNPs. For example, the second term described above may refer to a term related to the correlation between weights that must be considered to ensure that the upper bound of the generalization error remains below a certain level, based on the PAC-Bayesian (Probably Approximately Correct-Bayesian) framework.

[0063] More specifically, the device can compute a prior distribution matrix based on the correlation between the weights assigned to SNPs input to the model. For example, the device can compute the average similarity between all weight vectors by calculating the similarity between the weight vectors assigned to SNPs through cosine similarity, and can compute the prior distribution matrix using the average similarity.

[0064] Additionally, the device can compute a posterior distribution matrix based on the covariance of the prior distribution matrix and the weight vector. For example, the device can compute the posterior distribution matrix through a Kronecker product operation on the prior distribution matrix and the covariance matrix of the weight vector.

[0065] Additionally, the device can derive a second term that computes the difference in information between the prior distribution matrix and the posterior distribution matrix based on the Kullback-Leibler Divergence. In other words, the second term of the loss function may be a term related to the average correlation of the weight vector distribution assigned to each node corresponding to the SNP.

[0066] Meanwhile, the device can update parameters for the model (e.g., weights, biases, etc. assigned to each node) by performing backpropagation for the model based on the gradient of the loss function as described above.

[0067] For example, the device can compute the gradient of the loss function by computing the gradient for the first and second terms in the loss function, and update the parameters for the SNPs input to the model based on the computed gradient.

[0068] If the gradient of the loss function is negative, the weights for the current model parameters can be increased using gradient descent. Conversely, if the gradient of the loss function is positive, the weights for the current model parameters can be decreased using gradient descent.

[0069] For example, if the slope values ​​calculated by the device for the first and second terms are negative, the device may increase the weight for the SNP input into the model. Alternatively, if the slope values ​​calculated by the device for the first and second terms are positive, the device may decrease the weight for the SNP input into the model. In this example, the device may update the parameters of the model in a way that reduces the correlation between the weights assigned to the SNPs by adjusting the weights for the SNPs based on the slope of the second term.

[0070] Meanwhile, even if the gradient value calculated for the first term is positive, if the gradient value calculated for the second term is a negative number with a larger absolute value, the gradient calculated for the overall loss function may be negative. Furthermore, even if the gradient value calculated for the first term is negative, if the gradient value calculated for the second term is a positive number with a larger absolute value, the gradient calculated for the overall loss function may be positive. In this case, the device can update the model's parameters based on the gradient of the overall loss function including the first and second terms.

[0071] For example, the device may decrease the weight for a SNP input into the model if the slope of the overall loss function is positive. Alternatively, the device may increase the weight for a SNP input into the model if the slope of the overall loss function is negative. In other words, according to one embodiment of the present disclosure, the device may adjust the weights for current model parameters using a descent method that takes into account weight correlation.

[0072] FIG. 5 is a flowchart illustrating an example of a method for training a model for estimating a polygenic risk score (PRS) according to one embodiment.

[0073] Referring to FIG. 5, a method for training a model for estimating a polygenic risk score (PRS) may include steps 510 to 530. However, even if omitted below, the contents described with reference to FIGS. 1 to 4 may also be applied to the method for training a model for estimating a polygenic risk score (PRS) of FIG. 5.

[0074] In step 510, the device may input genetic information about a single nucleotide polymorphism (SNP) and construct a learning data set including one or more learning data having a PRS calculation value corresponding to the genetic information as a label.

[0075] According to one embodiment, the device can generate a list of SNPs based on a predetermined P-value from data on genome-wide association study (GWAS), and configure a learning data set by using the generated list of SNPs as input to the learning data set.

[0076] At step 530, the device can train a model to estimate a polygenic risk score (PRS) based on a training data set.

[0077] For example, the device can update parameters for the model by computing genetic information using a model, obtaining a PRS estimate corresponding to the genetic information from the model, and performing backpropagation for the model based on a loss function regarding the error between the PRS estimate and the PRS calculated value, which is a label for the genetic information.

[0078] In one embodiment, the loss function may include a first term that computes the error between the PRS estimate and the PRS calculation, and a second term that computes the correlation between the weights assigned to the SNPs input to the model.

[0079] In more detail, the device can derive a second term that calculates a prior distribution matrix based on a correlation between weights assigned to SNPs input to the model, calculates a posterior distribution matrix based on a covariance matrix of the prior distribution matrix and the weight vector, and calculates an information difference between the prior distribution matrix and the posterior distribution matrix based on Kullback-Leibler divergence.

[0080] Meanwhile, the device can calculate the gradient of the loss function by calculating the gradient for the first and second terms in the loss function, and update the parameters regarding the SNPs input to the model based on the calculated gradient.

[0081] For example, the device may decrease the weight for an SNP input into the model if the slope of the loss function is positive, and may increase the weight for an SNP input into the model if the slope of the loss function is negative.

[0082] FIG. 6 is a schematic diagram illustrating an example of the internal configuration of a device for training a model for estimating a polygenic risk score (PRS) according to one embodiment.

[0083] Referring to FIG. 6, the device (600) includes a processor (610), a memory (620), an input / output interface (630), and a communication module (640). For convenience of explanation, only components related to the present invention are illustrated in FIG. 6. Therefore, in addition to the components illustrated in FIG. 6, other general-purpose components may be further included in the device (600). In addition, it will be apparent to a person skilled in the art related to the present invention that the processor (610), memory (620), input / output interface (630), and communication module (640) illustrated in FIG. 6 may be implemented as independent devices.

[0084] The processor (610) can process computer program instructions by performing basic arithmetic, logic, and input / output operations. Here, the instructions can be provided from memory (620) or an external device. In addition, the processor (610) can generally control the operation of other components included in the device (600).

[0085] The processor (610) may be implemented as an array of a plurality of logic gates, or may be implemented as a combination of a general-purpose microprocessor and a memory storing a program executable on the microprocessor. For example, the processor (610) may include a general-purpose processor, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a controller, a microcontroller, a state machine, etc. In some environments, the processor (610) may include an application-specific integrated circuit (ASIC), a programmable logic device (PLD), a field programmable gate array (FPGA), etc. For example, the processor (610) may refer to a combination of processing devices, such as a combination of a digital signal processor (DSP) and a microprocessor, a combination of a plurality of microprocessors, a combination of one or more microprocessors coupled with a digital signal processor (DSP) core, or any other such combination of configurations.

[0086] The memory (620) may include any non-transitory computer-readable recording medium. As an example, the memory (620) may include a non-volatile mass storage device such as a random access memory (RAM), a read only memory (ROM), a disk drive, a solid state drive (SSD), a flash memory, etc. As another example, the non-volatile mass storage device such as a ROM, an SSD, a flash memory, a disk drive, etc. may be a separate permanent storage device distinct from the memory. In addition, the memory (620) may store an operating system (OS) and at least one program code (e.g., code for the processor (610) to perform the operations described with reference to FIGS. 1 to 5).

[0087] These software components may be loaded from a computer-readable recording medium separate from the memory (620). This separate computer-readable recording medium may be a recording medium that can be directly connected to the device (600), and may include, for example, a computer-readable recording medium such as a floppy drive, a disk, a tape, a DVD / CD-ROM drive, a memory card, etc. Alternatively, the software components may be loaded into the memory (620) through a communication module (640) other than a computer-readable recording medium. For example, at least one program may be loaded into the memory (620) based on a computer program (e.g., a computer program for the processor (610) to perform the operations described with reference to FIGS. 1 to 5) that is installed by files provided by developers or a file distribution system that distributes installation files of applications through the communication module (640).

[0088] The input / output interface (630) may be a means for interfacing with a device (e.g., a keyboard, a mouse, etc.) for input or output that may be connected to or included in the device (600). In FIG. 6, the input / output interface (630) is illustrated as an element configured separately from the processor (610), but is not limited thereto, and the input / output interface (630) may be configured to be included in the processor (610).

[0089] The communication module (640) may provide a configuration or function for the server (60) and the device (600) to communicate with each other via a network. In addition, the communication module (640) may provide a configuration or function for the device (600) to communicate with other external devices. For example, control signals, commands, data, etc. provided under the control of the processor (610) may be transmitted to the server (60) and / or the external device via the communication module (640) and the network.

[0090] Embodiments of the present invention may be implemented in the form of a computer program that can be executed through various components on a computer, and such a computer program may be recorded on a computer-readable medium. In this case, the medium may include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specifically configured to store and execute program instructions, such as ROMs, RAMs, and flash memories.

[0091] Meanwhile, the computer program may be specifically designed and constructed for the present invention, or may be one known and available to those skilled in the computer software field. Examples of computer programs may include not only machine language code, such as that generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like.

[0092] According to one embodiment, the method according to various embodiments of the present disclosure may be provided as included in a computer program product. The computer program product may be traded as a product between a seller and a buyer. The computer program product may be distributed in the form of a machine-readable storage medium (e.g., compact disc read-only memory (CD-ROM)), or may be distributed online (e.g., downloaded or uploaded) through an application store (e.g., Play Store™) or directly between two user devices. In the case of online distribution, at least a portion of the computer program product may be temporarily stored or temporarily generated in a machine-readable storage medium, such as the memory of a manufacturer's server, an application store's server, or a relay server.

[0093] Unless the steps constituting the method according to the present invention are explicitly described in a specific order or are not described to the contrary, the steps may be performed in any appropriate order. The present invention is not necessarily limited to the order in which the steps are described. The use of all examples or exemplary terms (e.g., "for example," etc.) in the present invention is merely intended to illustrate the present invention in detail, and the scope of the present invention is not limited by the examples or exemplary terms unless otherwise defined by the claims. Furthermore, those skilled in the art will appreciate that various modifications, combinations, and variations can be configured according to design conditions and factors within the scope of the appended claims or their equivalents.

[0094] Therefore, the idea of ​​the present invention should not be limited to the embodiments described above, and not only the scope of the patent claims described below but also all scopes equivalent to or equivalently modified from the scope of the patent claims are considered to fall within the scope of the idea of ​​the present invention.

Claims

1. A method for training a model to estimate a polygenic risk score (PRS), A step of constructing a learning data set including at least one learning data set having as input genetic information about a single nucleotide polymorphism (SNP) and having as a label a PRS calculation value corresponding to the genetic information; and A step of training the model for estimating the polygenic risk score based on the learning data set; The above learning steps are: A step of calculating the genetic information using the model; A step of obtaining a PRS estimate corresponding to the genetic information from the model; and A method comprising: a step of updating parameters for the model by performing backpropagation for the model based on a loss function regarding the error between the PRS estimated value and the PRS calculated value, which is a label for the genetic information; 2. In paragraph 1, The steps for constructing the above learning data set are: A method comprising: generating a list of SNPs based on a predetermined p-value from data on a genome-wide association study (GWAS), and using the generated list of SNPs as the input of the learning data; 3. In paragraph 1, The above loss function is, A method comprising a first term for calculating an error between the PRS estimate and the PRS calculation, and a second term for calculating a correlation between a weight assigned to a SNP input to the model.

4. In paragraph 3, The above learning steps are: A step of calculating a prior distribution matrix based on the correlation between the weights assigned to SNPs input to the above model; A step of calculating a posterior distribution matrix based on the covariance matrix of the above prior distribution matrix and the weight vector; and A method further comprising: deriving the second term for calculating the difference in information between the prior distribution matrix and the posterior distribution matrix based on Kullback-Leibler divergence; 5. In paragraph 3, The steps to update the above parameters are: A step of calculating the slope of the loss function by calculating the slope for the first term and the second term in the loss function; and A step of updating the parameters based on the calculated slope; A method comprising:

6. In paragraph 5, The steps to update the above parameters are: A method comprising: a step of decreasing a weight for an SNP input into the model when the slope of the loss function is positive, and a step of increasing a weight for the SNP input into the model when the slope of the loss function is negative; 7. In a device for training a model for estimating a multi-gene risk score, memory in which at least one program is stored; and comprising a processor for executing at least one program; The above processor, A learning data set is constructed that includes at least one learning data set that inputs genetic information about a single nucleotide polymorphism (SNP) and labels the PRS calculation value corresponding to the genetic information, training the model for estimating the polygenic risk score based on the training data set, The above processor, The above genetic information is calculated using the above model, Obtain a PRS estimate corresponding to the genetic information from the above model, A device for updating parameters for the model by performing backpropagation for the model based on a loss function regarding the error between the PRS estimated value and the PRS calculated value, which is a label for the genetic information.

8. A computer-readable recording medium recording a program for executing the method of paragraph 1 on a computer.

Citation Information

Patent Citations

  • Metal bonded gasket having sub seailing structure

    KR1020240122994A

  • Notification service server capable of providing access notification service to harmful sites and operating method thereof

    KR102421572B1

  • Deep learning based method for estimating multi-output value of power plant and system for permoing the same

    KR102441135B1

  • System and method for calculating degree of messiness of space based on characterstics of product in the space

    KR102451757B1

  • Genomic sequence dataset generation

    WO2022056438A1