Learning system, determination system, prediction system, learning method, determination method, and prediction method

Through the learning system and probability model processor, the error characteristics of DNA methylation determination are modeled and calibrated, which solves the measurement error problems in bisulfite conversion and PCR amplification, and improves the accuracy of DNA methylation determination. Reliability of sexual and disease diagnosis.

CN118974834APending Publication Date: 2024-11-15FUJIFILM CORP
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202380031812.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-03-30
Filing Date
2023-03-24
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art has problems with bisulfite conversion incompleteness, PCR deviation and measurement error in the multiplexing protocol in DNA methylation assay, which makes it difficult to quantify and learn the measurement error characteristics, affecting the accuracy of disease diagnosis.

Method used

Using learning systems and methods, the error characteristics in bisulfite conversion, PCR amplification and multiplexed panels are modeled and calibrated through a probability model processor, using prior parameter initialization and multinomial distribution modeling, combined with calibration data and metadata to determine and predict the determination error characteristics of biomarker sequences.

Benefits of technology

Effectively processing DNA methylation assays of multiple genes improves the accuracy and reliability of measurement errors, enhances the determination accuracy in liquid virology and cancer diagnosis, and reduces the underestimation of the actual methylation ratio.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118974834A_ABST
    Figure CN118974834A_ABST
Patent Text Reader

Abstract

One embodiment of the present invention provides a learning system, a determination system, a prediction system, a learning method, a determination method, and a prediction method. In DNA methylation assays, there have been a problem of incompleteness of bisulfite conversion (problem 1), a problem of variation occurring when several different biomarker sequences / genes are amplified together (problem 2), and a problem of the degree of excessive amplification of non-methylation signals depending on the gene sequences themselves and the chemical substances used in the assays (problem 3). In one embodiment of the present invention, provided are a system for learning measurement error characteristics when there are three problems and reflecting the learned error characteristics to a biomarker selection reference, and a method corresponding to the system. The main novelty of the present invention is achieved by solving the problem of evaluating the measurement error characteristics of DNA methylation when there is a combination of problems 1-3.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a technique for determining the value of a biomarker. Background Art

[0002] In DNA (deoxyribonucleic acid), a phenomenon called "methylation" is known to occur. Methylation refers to a modification based on the chemical bonding of a methyl molecule to cytosine. This cytosine (C: cytosine) together with guanine (G: guanine), adenine (A: adenine), and thymine (T: thymine) constitute the four essential nucleic acid bases that constitute DNA. Any sequence of nucleic acid bases is called a "nucleotide sequence", and the nucleotide sequence that encodes important information such as proteins is called a "genomic sequence" or "gene".

[0003] In the human body, methylation is particularly common at locations where cytosine is closely linked to guanine on the DNA chain (called "CpG sites"). The methylation state affects the activation or inhibition of genes, and the methylation state of CpG sites of certain genes forms an important biomarker for many diseases. In general, in order to make a quantitative model for disease diagnosis, data obtained from a combination of several biomarker candidate sequences are used. Therefore, it becomes important to measure the DNA methylation of biomarkers.

[0004] In the DNA assay process, errors in the data can affect the reliability of any inference / prediction. Previous studies to optimize the selection of biomarkers have assumed that there are only very few errors in the assay process and have focused only on the predictive value of the available data. As an example of this approach, a feature selection algorithm is known that relies on the output signal from a quantitative model (such as the performance of an Artificial Intelligence (AI) classifier) ​​to determine whether to use a biomarker sequence as a feature for classification.

[0005] Regarding this kind of conventional technology, for example, Patent Document 1 describes selecting a biomarker set from representative biomarker data for evaluation. Also, Non-Patent Document 1 describes measuring and relaxing PCR bias (PCR: polymerase chain reaction).

[0006] Previous technical literature

[0007] Patent Document 1: Japanese Patent Application No. 2017-523437

[0008] Non-patent literature

[0009] Non-patent document 1: “Measuring and Mitigating PCR Bias in Microbiome Data”, Justin D. Silverman et al., [retrieved on March 22, 2022], Internet (https: / / www.biorxiv.org / content / 10.1101 / 604025v1) Summary of the invention

[0010] Technical issues to be solved by the invention

[0011] In the next section, the existing studies on the measurement error characteristics of the biomarker sequence to be learned are discussed in detail. These existing studies and the problems associated with them are discussed, and each stage is explained in detail.

[0012] [DNA methylation determination]

[0013] The outline of the methylation assay is shown in Figure 1 In the measurement of methylation, the blood sample 10 is converted with bisulfite, the gene / signal is amplified by a PCR device, and the measurement is performed by a new generation sequencer, etc. These series of measurement sequences constitute a wet experiment protocol 20 (wet experiment protocol).

[0014] [STEP 1: Bisulfite conversion]

[0015] In order to distinguish Cm (methylated cytosine) from Cu (unmethylated cytosine), an additional step of bisulfite conversion is used. In bisulfite conversion, Cu is converted to uracil (U), and Cm remains Cm. If the converted sample is sequenced, Cm is read as C (cytosine), while uracil is read as thymine. In this way, the methylation state of cytosine can be distinguished.

[0016] [Question 1: Problems in bisulfite conversion]

[0017] The ideal result of this sequence is that Cu is converted to 100% uracil and Cm is not converted to uracil at all (conversion to 0%, Cm remains Cm). However, in the nature of chemical reactions, the degree of success (or failure) of the conversion is probabilistic, and quantitative studies are difficult. This incompleteness of bisulfite conversion is hereinafter referred to as "Problem 1".

[0018] [STEP 2: PCR amplification]

[0019] This stage can be understood as the signal amplification stage of the assay. Standardly (i.e., not for methylation but for bisulfite conversion), each "signal" is a gene or sequence of interest. In the original data, the number of such sequences is very small, so the derived signal is weak. Therefore, it is believed that by repeatedly replicating the original sequence, the number of sequences can be increased and the signal can be amplified. For example, the signal intensity of gene 1 before PCR is called G1_pre, and the signal intensity after PCR is called G1_post. In addition, in fact, the focus is on amplifying multiple genes / signals at the same time. Therefore, with respect to gene 2, G2_pre and G2_post are defined in the same manner as gene 1.

[0020] Now, if the above-mentioned STEP 1 is performed first, two signals can be obtained even for only one gene. For example, gene 1 has several sequences that are converted into other sequences containing uracil and have CpG as non-methylation. Similarly, sequences in which CpG is methylated will not be converted. This is common and is found in a mixture of DNA from the liver and stomach. In this mixture, it is possible that genes important to the liver are not methylated in hepatocytes but are methylated (and therefore inhibited) in gastric cells. Therefore, with respect to gene 1, the intensity of the pre-PCR signal and the intensity of the post-PCR signal are set to G1_U_Pre and G1_U_post (in the case of non-methylation), set to G1_M_Pre and G1_M_post (in the case of methylation), and the decrypted sequence is set to G1_M_Pre and G1_M_post.

[0021] [Issue 2: PCR bias in the monobisulfite protocol]

[0022] Even if the signal of the same gene is amplified, bisulfite conversion will result in two signal types. Therefore, G1_U_post / G1_U_pre=G1_M_post / G1_M_pre does not hold. It is known that even in the case of G1_U_pre=G1_M_pre, after amplification it becomes G1_U_post / G1_U_pre>G1_M_post / G1_M_pre (i.e., the non-methylated gene is over-amplified relative to the methylated gene). However, the degree of over-amplification of this non-methylated signal depends on the gene sequence itself and the chemical substances used in the measurement. Hereinafter, this problem is referred to as "Problem 2".

[0023] [Question 3: Problems in PCR amplification]

[0024] The ideal result of PCR is G1_post / G1_pre=G2_post / G2_pre. However, in reality, some gene sequences are easier to measure than other gene sequences and this equivalence does not hold. This deviation that occurs when several different biomarker sequences / genes are amplified together is called "PCR deviation" in the multiplex protocol (hereinafter referred to as "problem 3").

[0025] [Correspondence in the prior art]

[0026] The corresponding conventional techniques for the above-mentioned problems 1 to 3 are explained. In the conventional techniques, regarding problem 1, since extreme accuracy is usually not required in quantitative research, the degree of success of bisulfite conversion is not considered. Also, regarding problem 3, in microbiological research to date, the effect of PCR is considered in a multiplicative manner. That is, in the conventional techniques, it is considered that if the signal intensity of gene 1 after one PCR cycle is j, then the signal intensity after two cycles is j. 2 , the signal intensity after x cycles is also j x Using this assumption, PCR was modeled as a log-linear process using a multinomial logistic-general linear model. Other covariates such as "batch effects" (PCR of samples showing slightly different bias characteristics for each batch) were also included in the probabilistic approach. The model was used to calibrate for PCR bias after "training" the generated calibration data.

[0027] In addition, regarding question 2, it is easier to characterize the measurement error and bias in a single protocol setting, so a linear regression is performed in a portion of the PCR data to find the degree of bias. After calculating the linear regression estimator, the equation can be used to correct for this bias.

[0028] The importance of accurate measurement of DNA methylation has been described. In application areas such as disease diagnosis, it is not uncommon to use data from multiple biomarker sequences and input them into a quantitative model. When designing a measurement process for simultaneously measuring the index values ​​of multiple biomarkers, Problem 1, Problem 2, and Problem 3 are combined together, and all of these become problems. Therefore, quantification of errors and learning of error characteristics become very difficult. In the present invention, a system is studied that learns measurement error characteristics in the presence of three problems and reflects the learned error characteristics on a biomarker selection criterion. In the presence of this combination of problems, the problem of evaluating measurement error characteristics for DNA methylation is solved, forming the main novelty of the present invention.

[0029] The present invention is particularly important in situations such as liquid biopsy where it is necessary to measure DNA methylation from multiple genes simultaneously and very accurately. In particular, it is known that in order to accurately identify diseases such as cancer, certain cancer cell genes show high methylation compared to the same genes in healthy cells. In this case, problem 2 refers to the underestimation of the true methylation ratio (negative bias) from the mixture of cancer and normal DNA during measurement. Problems 1 and 3 further exacerbate the degree of underestimation.

[0030] The present invention has been made in view of the above circumstances, and one embodiment thereof provides a learning system and a learning method for learning the measurement error characteristics of a biomarker sequence. In addition, one embodiment of the present invention provides a determination system and a determination method for determining a sequence set by reflecting the learned error characteristics, and a prediction system and a prediction method for predicting the measurement error characteristics of a gene sequence using data obtained by the learning system or the learning method.

[0031] Means for solving technical problems

[0032] A learning system according to a first aspect of the present invention is a learning system for learning the relationship between an assay protocol variable and an error characteristic generated as a result of a biomarker sequence, the learning system comprising a processor for performing the following processing: inputting calibration data designed so as to obtain appropriate data for a variable of importance; and learning the characteristics of the error distribution across each assay protocol for the variable of importance using a probability model, the probability model including: a first parameter initialized with a priori parameters appropriately selected to model the error of bisulfite conversion; a second parameter initialized with a priori parameters appropriately selected to model the interdependence of amplification of the biomarker sequence; and a third parameter initialized with a priori parameters appropriately selected to model the overall PCR bias. The learning system according to the first aspect is a system for learning the relationship between an assay protocol variable and an error characteristic generated as a result of a biomarker sequence (defined as a template to product ratio).

[0033] In the first embodiment and the following embodiments, the "important variable" is a variable known by laboratory experts to have an impact on signal amplification performance, and the PCR device is adjusted for such a variable. For example, as described below Figure 2The PCR temperature or PCR cycle number shown is an example of an "important variable". If the temperature is too high, the DNA will be decomposed and the reaction required for replicating the target gene sequence will not occur. In addition, regarding the first to third parameters, the same parameters can be used as "appropriately selected prior parameters". In addition, regarding the "input of calibration data", for example, in the case of PCR temperature, it is required to be able to display appropriately within the range of temperatures used in normal PCR.

[0034] The learning system involved in the second mode is a parameter in the first mode in which the second parameter is the following parameter: the counts of the methylated sequence and the unmethylated sequence of the gene after bisulfite conversion are obtained respectively, and for each sequence of the methylated sequence and the unmethylated sequence, the obtained counts are modeled with a multinomial distribution that can respectively determine the prior variable. The second mode stipulates a specific mode for the second parameter corresponding to the above-mentioned problem 2, and the error of bisulfite conversion is modeled and corrected so that the methylation of the biomarker sequence can be correctly evaluated. In the second mode, a better prior variable can be selected based on empirical data analysis. In addition, the counts obtained in the second mode can be modeled based on factors such as the GC ratio (ratio of guanine to cytosine) of the base sequence.

[0035] The learning system involved in the third mode is a parameter in the first or second mode, wherein the third parameter is a parameter that imposes a structural data constraint that the sum of each count calculated by multinomial distribution follows Gaussian distribution when multiple sequences are amplified simultaneously using common primers. The third mode specifies a specific mode for the second parameter corresponding to the above-mentioned problem 3, and when the number of multiple biomarkers is large and the modeling parameters are simplified in a manner that can calculate the structural data constraint, the sum of each count in multiple dispersed counts follows Gaussian distribution. Furthermore, when multiple sequences are amplified simultaneously, an amplification method is performed such that the count value of each marker is not independent and the sum value becomes approximately constant, so the modeling based on multinomial distribution as described above is suitable. In addition, in the methylation measurement accompanied by bisulfite conversion, each marker has two states, methylated and unmethylated, so it becomes a modeling of the count value relative to the number of markers × 2.

[0036] The determination system according to the fourth aspect of the present invention is a determination system having a processor, the processor performing the following processing: inputting the nucleotide sequence and measurement protocol information of the biomarker sequence of interest used in the multiplex panel; inputting the learned error characteristics and metadata associated with the error characteristics from the learning system according to any one of the first to third aspects; outputting a first score for grouping possible biomarker sequences using the nucleotide sequence, measurement protocol information, learned error characteristics, and metadata input using a predetermined reference; and determining the biomarker sequence set in consideration of the value of the first score for each group. In the determination system according to the fourth aspect, the output from the system according to the first aspect is used to determine whether the biomarker sequence is used in the multiplex panel. The first score is a score derived from the measurement accuracy, and is a "low error score" having a higher value as the measurement error is smaller.

[0037] The determination system involved in the fifth embodiment is a system in which the processor performs the following processing in the fourth embodiment: inputting a second score for each biomarker sequence to be determined; and considering the first score for each biomarker sequence in the biomarker sequence set, optimizing the balance between the first score and the second score, thereby selecting the best subset of the multiplex panel. In the determination system involved in the fifth embodiment, the fourth embodiment is enhanced by considering the ultimate goal of the multiplex panel in order to enable a more balanced trade-off selection of biomarker sequences. For example, the greater the correlation with the disease to be predicted, the higher the second score (correlation score). And, for example, by calculating a third score defined by the additive average or multiplicative average of the first score and the second score and maximizing the third score, the "balance between the first score and the second score" can be optimized.

[0038] The prediction system according to the sixth aspect of the present invention is a prediction system for predicting measurement error characteristics of a gene sequence and includes a processor that performs the following processing: inputting a nucleotide sequence of a biomarker sequence of interest used in a multiplexed panel and measurement protocol information; inputting a learned error characteristic and metadata associated with the error characteristic from the learning system according to any one of the first to third aspects; calculating the similarity between a biomarker sequence previously included in calibration data and a new biomarker sequence using a measurement benchmark for calculating a scale of similarity between two gene sequences; and predicting the error characteristic when measuring a biomarker sequence not included in the calibration data by combining the calculated similarity with other related inputs and the learned error characteristic. The prediction system according to the sixth aspect can use the learning system according to the first to third aspects for a biomarker sequence not included in the calibration data.

[0039] In the sixth aspect, "other related inputs" refer to metadata corresponding to the biomarker sequence, for example. For example, if the gene type is "promoter or enhancer", the CpG type is "island, island shore, island shelf", and the CG abundance is "high, low", then the combination of this information (an example of metadata) for a certain biomarker sequence G1 can be expressed as a vector of "promoter, island, low".

[0040] In the prediction system according to the sixth aspect, the processor performs the following processing: using the predicted error characteristics to obtain a biomarker sequence that is most similar to a biomarker sequence not included in the calibration data and can be used in the calibration data; and reflecting information on the obtained biomarker sequence in determining the biomarker sequence set in the determination system according to the fourth or fifth aspect. In the seventh aspect, using the determination system according to the fourth or fifth aspect, a biomarker sequence not included in the calibration data can be used in the biomarker sequence set selection.

[0041] The learning method according to an eighth aspect of the present invention is a learning method performed by a learning system having a processor and learning the relationship between a measurement protocol variable and an error characteristic generated as a result of a biomarker sequence, wherein the processor performs the following processing: inputting calibration data designed so as to obtain appropriate data for important variables (calibration data input step); and

[0042] The characteristics of the error distribution across each assay protocol for a variable of significance are learned using a probabilistic model (learning step), the probabilistic model comprising: a first parameter initialized with a priori parameters appropriately selected to model the error of bisulfite conversion; a second parameter initialized with a priori parameters appropriately selected to model the interdependency of amplification of biomarker sequences; and a third parameter initialized with a priori parameters appropriately selected to model the overall PCR bias. The eighth aspect provides a learning method corresponding to the first aspect described above.

[0043] The learning method according to the ninth aspect is the learning method according to the eighth aspect, wherein the second parameter is a parameter that includes obtaining counts of methylated sequences and unmethylated sequences of the gene after bisulfite conversion, respectively, and modeling the obtained counts with a multinomial distribution capable of respectively determining a priori variables for each of the methylated sequences and the unmethylated sequences. The ninth aspect provides a learning method corresponding to the second aspect described above.

[0044] The learning method according to the tenth aspect is a learning method according to the eighth or ninth aspect, wherein the third parameter is a parameter that imposes a structural data constraint that the sum of counts calculated using a multinomial distribution follows a Gaussian distribution when a plurality of sequences are simultaneously amplified using a common primer. The tenth aspect specifies a learning method corresponding to the third aspect described above.

[0045] The determination method according to the eleventh aspect of the present invention is a determination method executed by a determination system having a processor, wherein the processor performs the following processing: inputting a nucleotide sequence and measurement protocol information of a biomarker sequence of interest used in a multiplexed panel (sequence information input step); inputting a learned error characteristic and metadata associated with the error characteristic obtained as a result of the learning method according to any one of the eighth to tenth aspects (learning result input step); outputting a first score for grouping possible biomarker sequences using the nucleotide sequence, measurement protocol information, learned error characteristic, and metadata input using a predetermined reference (score output step); and determining a biomarker sequence set taking into account the value of the first score for each set (sequence set determination step). The eleventh aspect provides a determination method corresponding to the fourth aspect described above.

[0046] The determination method according to the twelfth aspect is the determination method according to the eleventh aspect, wherein the processor performs the following processing: inputting a second score for each biomarker sequence to be determined (a score input step); and optimizing a balance between the first score and the second score by considering the first score for each biomarker sequence in the biomarker sequence set, thereby selecting an optimal subset of the multiplexed panel (a subset selection step). The twelfth aspect provides a determination method corresponding to the fifth aspect described above.

[0047] The prediction method according to the 13th aspect of the present invention is a prediction method performed by a prediction system that predicts measurement error characteristics of a gene sequence and includes a processor, wherein the processor performs the following processing: inputting a nucleotide sequence and measurement protocol information of a biomarker sequence of interest used in a multiplexed panel (sequence information input step); inputting a learned error characteristic and metadata associated with the error characteristic obtained by the learning method according to any one of the 8th to 10th aspects (learning result input step); calculating the similarity between a biomarker sequence previously included in calibration data and a new biomarker sequence using a measurement standard that is a scale for calculating similarity between two gene sequences (similarity calculation step); and predicting the error characteristic when measuring a biomarker sequence not included in the calibration data by combining the calculated similarity with other related inputs and the learned error characteristic (error characteristic prediction step). The 13th aspect provides a prediction method corresponding to the above-mentioned 6th aspect.

[0048] The prediction method according to the 14th aspect is the prediction method according to the 13th aspect, wherein the processor performs the following processing: using the predicted error characteristics to obtain a biomarker sequence that is most similar to the biomarker sequence not included in the calibration data and can be used in the calibration data (sequence acquisition step); and reflecting information of the obtained biomarker sequence in the determination of the biomarker sequence set in the determination method according to the 11th or 12th aspect (information reflection step). The 14th aspect provides a prediction method corresponding to the above-mentioned 7th aspect.

[0049] In addition, programs (learning program, determination program, prediction program) for executing the learning method, determination method, and prediction method in the above-mentioned manner by a processor and non-transitory recording media having computer-readable codes of these programs recorded thereon are also included in the scope of the present invention.

[0050] Effects of the Invention

[0051] As described above, the learning system, determination system, prediction system, and learning method, determination method, and prediction method according to the present invention have the following effects.

[0052] (1) It is possible to handle a multiplexed panel in which a plurality of gene sequences are measured simultaneously.

[0053] (2) Capable of processing samples converted with bisulfite.

[0054] (3) Ability to predict assay errors using sequence parameters and protocol parameters as input.

[0055] (4) Ability to determine whether a sequence is to be used for analytical / classification purposes. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a diagram showing the state of measuring DNA methylation.

[0057] Figure 2 It is a figure which shows the state of creating calibration data.

[0058] Figure 3 A graph representing a learning system and data related to the learning system.

[0059] Figure 4 is a diagram showing the structure of a learning system.

[0060] Figure 5 It is a diagram showing an implementation of a probability model.

[0061] Figure 6 It is a diagram showing the relationship between the determination system and the prediction system.

[0062] Figure 7 It is a diagram showing the structure of a determination system.

[0063] Figure 8 is a diagram showing the structure of a prediction system. DETAILED DESCRIPTION

[0064] Hereinafter, embodiments of the present invention will be described. In the description, the drawings are referred to as needed. In addition, in the drawings, for the sake of convenience of description, the description of some constituent elements may be omitted.

[0065] [Creation of calibration data]

[0066] In the present invention, if Figure 2 As shown, first, calibration data consisting of important measurement protocol variables such as PCR temperature or PCR cycle number needs to be prepared from the blood sample 10 through the wet experiment protocol 20. The calibration data is preferably designed to be able to obtain appropriate data for important variables. Finally, the measurement results of the sequence are stored together with the protocol information in the calibration data DB30 (DB: database) (hereinafter, the database is sometimes recorded as "DB"). In addition, Figure 2 In the figure, part of the calibration data creation procedure is omitted for clarity.

[0067] The calibration data is used by a learning algorithm (learning system, learning method) in order to learn the relationship between such protocol variables and their assay characteristics. Next, for a given set of assay protocol variables (not included in the calibration data), the system (prediction system, prediction method) can predict the assay error characteristics of a given biomarker sequence. Using this prediction, the system (determination system, determination method) can determine whether the biomarker sequence is suitable for use in certain quantitative studies. Finally, even for biomarker sequences that are not present in the calibration data, the system (determination system, determination method) of the present invention finds the most similar sequence for which the assay error characteristics are known, and can use this sequence to make similar determinations for new sequences.

[0068] Specifically, the present invention details the role of the probabilistic model to characterize the measurement error of the biomarker sequence by estimating the "template to product" ratio. "Template" refers to the initial amount of the biomarker sequence (the amount before PCR amplification), and "product" refers to the final amount of the same biomarker sequence after PCR amplification (the amount after PCR amplification).

[0069] [Structure of the learning system]

[0070] Figure 3The learning system 100 and data related thereto according to one embodiment of the present invention are shown. Application of such a learning system for learning the DNA methylation measurement error characteristics for the multiple bisulfite PCR protocol is a minimum requirement for ensuring the novelty of the present invention. In addition, as described later, the learning system 100 may be supplemented with a determination system 200 (determination system) and a prediction system 300 (prediction system) that utilize the results (learned error distribution DB50).

[0071] Figure 4 is a diagram showing a configuration example of the learning system 100. Figure 4 As shown, the learning system 100 includes a processor 110 (processor, computer), a probability model 120 (probability model), a storage unit 130, a ROM 140 (ROM: Read Only Memory), and a RAM 150 (RAM: Random Access Memory). The processor 110 performs overall control of the processing performed by each unit of the learning system 100, and therefore includes a calibration data input unit 112 and a learning unit 114.

[0072] The processor 110 may include Figure 4 In addition to the elements shown, a display control unit, a communication control unit, an output control unit, etc., which are not shown in the figures, may be included.

[0073] The processor 110 is composed of various processors or circuits such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an FPGA (Field Programmable Gate Array), and a PLD (Programmable Logic Device). When these processors or circuits execute software (programs), the code that can be read by the computer executing the software (for example, various processors or circuits and / or their combination that constitute the processor) is stored in a non-temporary and tangible recording medium such as ROM140, and the computer refers to the software. The software stored in the non-temporary and tangible recording medium includes a program (learning program, prediction program, determination program) for executing the learning method, prediction method, and determination method involved in the present invention and data used during execution. The code can be recorded in various non-temporary and tangible recording media such as optical magnetic recording devices and semiconductor memories instead of in ROM140. When software processing is used, for example, RAM 150 is used as a temporary storage area, and data stored in a non-temporary and tangible recording medium such as an EEPROM (Electronically Erasable and Programmable Read Only Memory) or a flash memory (not shown) can also be referenced. The storage unit 130 can also be used as a "non-temporary and tangible recording medium".

[0074] The storage unit 130 is composed of various storage devices such as a hard disk and a semiconductor memory and a control unit thereof, and can store the above-mentioned calibration data or execution conditions and execution results (learned error distribution data) of the learning method.

[0075] Learning System 100 Figure 4 In addition to the elements shown, a display device (e.g., a liquid crystal monitor) or an operating device (e.g., a mouse or a keyboard) not shown may also be included. The display device can display calibration data, error distribution data, etc., and the user can perform operations required for executing the learning method (learning program) involved in the present invention through the operating unit.

[0076] The above Figure 3 Blood sample data 11 is shown, but this is any biological data including tissue samples. Figure 1As shown, the blood sample data 11 is measured by the above-mentioned STEP 1, STEP 2 and the measurement order determined by the DNA sequence, and has several variables (important variables) such as the number of PCR cycles that affect its effectiveness. Since it is necessary to obtain data based on several values ​​of such variables, the relevant variables are first identified and the measurement is performed within the range of these values. For example, if the number of PCR cycles is the only important variable, data of the same blood sample with 5, 10 and 15 PCR cycles can be generated. This is so-called calibration data.

[0077] [Probability model]

[0078] The learning system 100 trains the probability model (learning step) using the calibration data (training data) stored in the calibration data DB 30 (calibration data input step). Figure 5 , a probability model 120 as an example of such a probability model is shown by a Bayesian hierarchical model. The important novelty of the present invention is that (i) prior information (prior parameters; the same below) of bisulfite conversion errors, (ii) prior information of covariates of bisulfite conversion, and (iii) prior information of the interdependence of amplification of biomarker sequences are used. These prior information (i) to (iii) correspond to the first to third parameters of the present invention, and therefore correspond to the above-mentioned problems 1 to 3. If the above three requirements are combined, the present invention is different from the previous model such as the above-mentioned patent document 1 or non-patent document 1.

[0079] Furthermore, through these three factors, the above-mentioned problem 1+problem 2+problem 3 can be solved as one. The learning system 100 is adjusted according to an optimization method (loss function for minimization, etc.) and through a series of hyperparameters (hyperparameters 40). This adjustment is performed by confirming the final performance of the system and selecting the hyperparameters that maximize it.

[0080] In addition, the first to third parameters are part of the probability model 120 (and therefore part of the learning system 100), and the values ​​of these parameters are updated during the training process. Figure 3 Not shown in .

[0081] On the other hand, if hyperparameters are used, it is possible to control a certain aspect of the probability model 120. However, the value of the hyperparameter is set by the user and the value is not updated during the training process. Figure 6 ), the hyperparameters are different.

[0082] More specifically, for biomarkers with a binomial distribution, it is possible to choose to model the error of bisulfite conversion. Therefore, the prior probability of the bisulfite conversion error (an example of a prior parameter) can be selected as a value between [0,1]. When the prior probability is 0, complete conversion of the biomarker is assumed (100% conversion of Cu and uracil and 0% conversion of Cm), and when the prior probability is greater than 0, incomplete conversion of the biomarker is assumed (only a portion of Cu is converted to uracil, and a portion of Cm is also converted to uracil). Ideally, the prior variables should be set based on empirical data analysis. The bisulfite covariate includes the amount of sulfite added to the sample measured in nanograms and the initial DNA amount. In this way, the bisulfite conversion error initialized with the prior probability is the first parameter.

[0083] Similarly, the PCR error distribution can be modeled as a multinomial distribution, and appropriate prior probabilities can be set. At this stage, the sequence counts (number of sequences) after PCR can be expressed as N1, N2, ..., Nx when the number of selected biomarkers is x, and the sequence counts can be modeled as a multinomial distribution. Here, "Ni" is the sequence count of the i-th biomarker. PCR covariates sometimes include factors such as PCR temperature or PCR cycle number that are selected to create calibration data.

[0084] The novelty of the present invention lies in the ability to consider the possibility of two different counts from the same sequence. One is the count of the bases of a certain sequence after bisulfite conversion (the count of the methylated sequence), and the other is the count of the non-base type of the sequence (the count of the non-methylated sequence). As a result, the number of possibilities considered, such as N1_M, N1_U, N2_M, N2_U, etc., is doubled. Here, "Ni_M" represents the count of the methylated sequence for the i-th biomarker, and "Ni_U" represents the count of the non-methylated sequence for the i-th biomarker. In order to impose natural constraints on each other (a high average number of times for one means a low average number for the other), the use of such constraints (interdependence) can simplify the modeling problem. In this way, the interdependence of the amplification of the biomarker sequence initialized with the prior probability is the second parameter.

[0085] Finally, the overall distribution model (a model that represents the deviation of the PCR as a whole) quantifies the sequence of the total number of all biomarkers, namely N1+N2+....+Nx, and can even be used to impose interdependence restrictions (structural data restrictions) through biomarker counts (for example, when N1 is too high, N3 is too low). N1, N2, etc. are each (each count) multinomially distributed, so it is believed that under the condition that the number of selected biomarkers is large (for example, more than 30), their total (the total of the counts calculated with the multinomial distribution) satisfies the central limit theorem and therefore follows a Gaussian distribution. This interdependence (structural data restriction) between sequence types may not be immediately clear, but it is known that this interdependence exists when multiple sequences (multiple biomarker sequences) are amplified simultaneously using universal primers. In this way, the deviation of the PCR as a whole initialized with a priori probability is the third parameter.

[0086] The use of such universal primers can only be performed after appropriate adapter sequences are configured at both ends of the target biomarker sequence. The limited amount of universal primers added at this stage creates compositional dependencies between biomarker sequences, affecting pure signal amplification. The second novelty of the present invention is to impose structural data constraints on the multiplexed panel in PCR amplification and use universal primers to model the relative abundance of biomarker sequences.

[0087] [Determine the positioning of the system and predict the system]

[0088] like Figure 6 As shown, the learning system 100 described above may be provided with a determination system 200 (determination system) and a prediction system 300 (prediction system). It is recommended that these determination system and prediction system be added to the learning system 100 as an option. By adding the determination system 200 and the prediction system 300, for example, the error characteristics learned by the learning system 100 can be used to find the best subset of candidate biomarkers by the determination system 200 (including the learning result input step or the score input step, the subset selection step, etc., based on the execution of the determination method involved in the present invention), thereby providing information on the selection criteria of the biomarker sequence (including the information reflection step, etc., based on the execution of the prediction method involved in the present invention; the prediction system 300), which can help to effectively use the learning system 100.

[0089] The learning system 100 described above has a probability model 120 that is learned by maximizing or minimizing an optimization benchmark by statistical methods. Such learning broadly covers the meaning of "training" the algorithm. On the other hand, the determination system 200 takes effect after the learning system 100 has finished training. The determination system 200 itself does not have a defined optimization benchmark to be maximized or minimized, so it is not "trained" and the system does not learn. However, the determination system 200 is composed of hyperparameters that make the system "adjustable".

[0090] [Structure of determination system and prediction system]

[0091] Figure 7 2 is a diagram showing the structure of the determination system 200. Figure 7 As shown, the determination system 200 includes a processor 210 (processor), a ROM 230 (non-transitory and tangible recording medium), and a RAM 240. The processor 210 includes a sequence information input unit 212, a learning result input unit 214, a score output unit 216, and a sequence set determination unit 218. In addition to these elements, the determination system 200 may also include a display control unit or a display device, a storage device, an operation unit, etc., which are not shown.

[0092] Figure 8 3 is a diagram showing the structure of the prediction system 300. Figure 8 As shown, the prediction system 300 includes a processor 310 (processor), a ROM 330 (non-transitory and tangible recording medium), and a RAM 340. The processor 310 includes a sequence information input unit 312, a learning result input unit 314, a similarity calculation unit 316, an error characteristic prediction unit 318, and a sequence information reflection unit 320. In addition to these elements, the prediction system 300 may also include a display control unit or a display device, a storage device, an operation unit, etc., which are not shown.

[0093] Similar to the learning system 100, these elements of the determination system 200 and the prediction system 300 are composed of various processors or circuits such as CPU, GPU, FPGA, PLD, etc. When these processors or circuits execute software (program), the code that can be read by the computer of the executed software is stored in a non-temporary and tangible recording medium such as ROM230 or ROM330, and the computer refers to the software. The software stored in the non-temporary and tangible recording medium includes a program (prediction program, determination program) for executing the prediction method and determination method involved in the present invention and data used during execution. The code can be recorded in various non-temporary and tangible recording media such as optical magnetic recording devices and semiconductor memories instead of ROM230 or ROM330. When the software is used for processing, for example, RAM240, RAM340 is used as a temporary storage area, and it is also possible to refer to data stored in a non-temporary and tangible recording medium such as an EEPROM or flash memory not shown.

[0094] [Determination of score-based biomarker sequence sets]

[0095] Hereinafter, the binary-based method and the combinatorial-based method, which are two major categories of “score optimization processes (optimization methods)”, will be described.

[0096] In the binary-based optimization benchmark, the sequence information input unit 212 (processor) of the determination system 200 inputs the nucleotide sequence of the biomarker sequence of interest and the measurement protocol information (sequence information input step), and the learning result input unit 214 (processor) inputs the learned error characteristics and metadata associated with the error characteristics from the learning system 100 (learning result input step). The score output unit 216 (processor) can independently consider the learned measurement error characteristics and assign a score (measurement error score; an example of a first score) of {+1, 0, -1} to each biomarker sequence according to the slope of the measurement error graph generated as a result (score output step). The sequence set determination unit 218 can sum up the scores (first score) according to the order of each biomarker and determine whether to use the combination (biomarker sequence set) (sequence set determination step).

[0097] If they are installed more durably, it is possible to design an optimization benchmark for the combination basis by the same method as the feature selection algorithm. The feature selection algorithm relies on the output of the quantitative model and updates these benchmarks in order to optimize the performance of the quantitative model. The viewpoint of this previous feature selection algorithm considers the score (first score) generated by the measurement error characteristics, and selects the best information from the subset of the assigned biomarker sequence set, so it is corrected to use the same score (first score) as the signal, thereby being able to design the optimization benchmark for the combination basis used in the present invention. In the case of the optimization benchmark for the combination basis, as in the case of the optimization benchmark for the binary basis, the biomarker sequence set can be determined using each element of the determination system 200 (execution of the sequence information input step to the sequence set determination step).

[0098] In addition, in the above-mentioned binary-based optimization criteria, scores are assigned independently to each biomarker sequence, while in the case of the combination-based optimization criteria, scores are assigned to the combination of biomarker sequences. Therefore, when the size of the measurement error can be handled independently in each marker sequence, the binary-based optimization criteria are suitable, and when the mutual dependence is particularly large, the combination-based optimization criteria are suitable. For example, mutual dependence refers to the situation that "the measurement error is small when biomarker sequence 1 and biomarker sequence 2 are measured simultaneously, but the measurement error is large when it is measured simultaneously with biomarker sequence 3."

[0099] In the present invention, not only the measurement error score (first score) described above can be considered, but also "the greater the correlation with the disease to be predicted, the higher the score" (correlation score; an example of the second score) can be considered, and the biomarker sequence set can be determined by optimizing the balance of these scores. When such a correlation score (second score) is used simultaneously, the balance between the measurement error score (first score) and the correlation score (second score) described above is optimized (for example, the arithmetic mean or geometric mean of the measurement error score and the correlation score is maximized), thereby determining the biomarker sequence set. In this case, the correlation score can also be independently assigned to each biomarker sequence as in the case of the measurement error score, and can also be assigned to a combination of biomarker sequences. For example, when markers 1, 2, and 3 are all related to the disease, and the correlation between markers 1 and 2 is small and the correlation between markers 1 and 3 is large, the combination of markers 1 and 2 is more effective in disease prediction, and the correlation score becomes higher.

[0100] Any optimization basis (e.g., binary basis, feature selection algorithm, or combinatorial basis) is preferably dependent on the application domain, user, and time constraints, and can be appropriately selected based on these conditions. The output is a set of biomarker sequences that are considered by the system together, with minimal error in the protocol of assigned measurement errors for multiplex PCR sequence determination. As with the fifth and twelfth aspects of the present invention, the embodiments can be varied depending on the implementation of the determination system (whether or not balanced sequence selection is considered).

[0101] In addition, the determination system 200 relies on the error distribution obtained in the learning system 100 (in Figure 6 The learned error distribution database 50 is not able to calculate the score of the biomarker sequence not included in the original calibration data. Figure 6 As shown and described below, the prediction system 300 according to the sixth and seventh aspects of the present invention (and the prediction methods according to the thirteenth and fourteenth aspects of the present invention) is required. The prediction system 300 is an addition to the learning system 100 (learning system) according to the present invention, similar to the above-mentioned content of the determination system 200, and relies on the use case, existence, and importance of such new biomarker sequences.

[0102] [Prediction of measurement error characteristics of biomarker sequences not included in calibration data]

[0103] When the biomarker sequence of interest contained in the sequence of interest database 60 is determined to be a biomarker sequence not included in the calibration data ( Figure 6 In the case where the judgment "Is it a sequence included in the training data?" is "Yes", a method for predicting the measurement error characteristics of the biomarker sequence of interest is described. In this case, the input is first transmitted to the prediction system 300. Specifically, the sequence information input unit 312 (processor) of the prediction system 300 inputs the nucleotide sequence and measurement protocol information of the biomarker sequence of interest (sequence information input step), and the learning result input unit 314 (processor) inputs, for example, the learned error characteristics and metadata associated with the error characteristics from the learning system 100 (learning result input step). Here, "metadata" is, for example, the type of gene (promoter or enhancer) and the region of the gene (transcription start site, etc.), but is not limited to this.

[0104] Furthermore, the similarity calculation unit 316 (processor) calculates the similarity between the biomarker sequence contained in the previous calibration data (the biomarker sequence that can be used in the calibration data) and the new biomarker sequence (the biomarker sequence of interest) using a measurement standard (similarity scale) for the similarity between two gene sequences such as the Levenshtein distance or GC content (GC-content; the ratio of guanine to cytosine in the nitrogen bases in the DNA molecule) (similarity calculation step). The similarity calculation unit 316 finds the biomarker sequence of interest and the "most similar" biomarker sequence from the biomarker sequences existing in the learned error distribution database 50 (similarity calculation step). The prediction system 300 can use the information of the detected "most similar sequence" to obtain the learned error characteristics corresponding to the "closest sequence" (from the learned error distribution database 50) (error characteristic prediction step, sequence acquisition step), thereby being able to fully implement the sixth and thirteenth aspects of the present invention. The sequence information reflection unit 320 can also use the information simultaneously with the determination system 200 implementing the 4th and 5th methods of the present invention (and the determination method involved in the 11th and 12th methods of the present invention), and reflect it in the determination of the biomarker sequence set in the determination system 200 (information reflection step).

[0105] [Example]

[0106] The candidate set of gene sequences is considered to first show sufficient variation in measurement-related factors such as GC content. For example, assuming that only the sequence GC content is important, 3 gene sequences with "high" and 3 gene sequences with "low" GC content are determined because they can be measured simultaneously. Next, a series of important measurement protocol-related variables are determined and ranges are considered. Then, a wet experiment sequence is performed for the entire range of all values ​​and all variables. For example, considering PCR cycles of {5, 10, 15}, and considering the methylation ratio as {5%, 10%}, 3*2=6 independent measurements of 6 genes are performed from aliquots of the same biological specimen. Next, the above-mentioned probability model used in the learning system is trained, and the hyperparameters of the determination system are adjusted (tuned). Now, as in the case of cancer diagnosis, it is possible to use the determination model to evaluate the goodness of the measurement characteristics of the gene while searching for more gene biomarkers.

[0107] If we consider that the Artificial Intelligence cancer classification model trained on 100 gene sequence measurements performs at 70% sensitivity, part of the reason for the low performance may be that the measurement noise is high in some genes. If we use the above system to reconsider the 100 genes and remove genes that are difficult to measure, it is possible that the performance of Artificial Intelligence can be increased to 80% by avoiding measurement errors, which has better robustness.

[0108] The above-described embodiment has the following effects.

[0109] (1) It is possible to handle a multiplexed panel in which a plurality of gene sequences are measured simultaneously.

[0110] (2) Capable of processing samples converted with bisulfite.

[0111] (3) Ability to predict assay errors using sequence parameters and protocol parameters as input.

[0112] (4) Ability to determine whether a sequence is to be used for analytical / classification purposes.

[0113] (5) Based on appropriately learned error characteristics, appropriately selected biomarker sequence sets (and their subsets), and accurately predicted measurement errors of the biomarker sequence sets, it is possible to perform analysis or diagnosis using biomarker sequences (for example, the above-mentioned AI-based cancer classification), etc. with high accuracy.

[0114] As mentioned above, although embodiment of this invention was described, this invention is not limited to the said form, Various deformation|transformation is possible.

[0115] Explanation of symbols

[0116] 10-blood sample, 11-blood sample data, 20-wet experiment protocol, 30-calibration data DB, 40-hyperparameter, 50-learned error distribution database, 60-focus sequence database, 100-learning system, 110-processor, 112-calibration data input unit, 114-learning unit, 120-probability model, 130-storage unit, 200-determination system, 210-processor, 212-sequence information input unit, 214-learning result input unit, 216-score output unit, 218-sequence set determination unit, 300-prediction system, 310-processor, 312-sequence information input unit, 314-learning result input unit, 316-similarity calculation unit, 318-error characteristic prediction unit, 320-sequence information reflection unit.

Claims

1. A learning system for learning a relationship between an assay protocol variable and an error characteristic resulting from a biomarker sequence, the learning system comprising a processor, The processor performs the following processing: Input calibration data designed in such a way as to obtain appropriate data for variables of importance; and Using a probabilistic model, learn the characteristics of the error distribution across various assay protocols for the variables of interest, The probability model includes: The first parameters, in order to model the errors in the bisulfite conversion, were initialized with appropriately chosen a priori parameters; The second parameters are initialized with appropriately selected a priori parameters in order to model the interdependence of the amplification of the biomarker sequences; and The third parameter is initialized with an appropriately chosen a priori parameter in order to model the overall PCR bias.

2. The learning system according to claim 1, wherein: The second parameter is the following parameter: respectively obtaining counts of methylated sequences and unmethylated sequences of the gene after bisulfite conversion, and for each of the methylated sequences and the unmethylated sequences, modeling the obtained counts with a multinomial distribution that can respectively determine a priori variables.

3. The learning system according to claim 1 or 2, wherein: The third parameter is a parameter that imposes a structural data restriction that the sum of each count calculated using a multinomial distribution follows a Gaussian distribution when a plurality of sequences are simultaneously amplified using common primers.

4. A determination system comprising a processor, wherein: The processor performs the following processing: inputting nucleotide sequence and assay protocol information for biomarker sequences of interest used in the multiplexed panel; inputting the learned error characteristic and metadata associated with the error characteristic from the learning system according to any one of claims 1 to 3; outputting a first score for a set of possible biomarker sequences using the nucleotide sequence inputted using a predetermined benchmark, the assay protocol information, the learned error characteristics, and the metadata; and The biomarker sequence set is determined by considering the value of the first score for each set.

5. The determination system according to claim 4, wherein: The processor performs the following processing: Entering a second score for each biomarker sequence to be determined; and The best subset of the multiplexed panel is selected by optimizing the balance between the first score and the second score considering the first score for each biomarker sequence in the set of biomarker sequences.

6. A prediction system for predicting measurement error characteristics of a gene sequence, the prediction system comprising a processor, The processor performs the following processing: inputting nucleotide sequence and assay protocol information for biomarker sequences of interest used in the multiplexed panel; inputting the learned error characteristic and metadata associated with the error characteristic from the learning system according to any one of claims 1 to 3; Calculating the similarity of a biomarker sequence previously included in the calibration data and a new biomarker sequence using a measure of the metric used to calculate the similarity between two gene sequences; and The calculated similarity is used in combination with other relevant inputs and the learned error characteristic to predict the error characteristic when measuring a biomarker sequence not included in the calibration data.

7. The prediction system according to claim 6, wherein: The processor performs the following processing: using the predicted error characteristic to obtain a biomarker sequence that is most similar to a biomarker sequence not included in the calibration data and that can be utilized in the calibration data; and The acquired information of the biomarker sequence is reflected in the determination of the biomarker sequence set in the determination system according to claim 4 or 5.

8. A learning method performed by a learning system having a processor and learning a relationship between an assay protocol variable and an error characteristic generated as a result of a biomarker sequence, wherein: The processor performs the following processing: Input calibration data designed in such a way as to obtain appropriate data for variables of importance; and Using a probabilistic model, learn the characteristics of the error distribution across various assay protocols for the variables of interest, The probability model includes: The first parameters, in order to model the errors in the bisulfite conversion, were initialized with appropriately chosen a priori parameters; The second parameters are initialized with appropriately selected a priori parameters in order to model the interdependence of the amplification of the biomarker sequences; and The third parameter is initialized with an appropriately chosen a priori parameter in order to model the overall PCR bias.

9. The learning method according to claim 8, wherein: The second parameter is the following parameter: respectively obtaining counts of methylated sequences and unmethylated sequences of the gene after bisulfite conversion, and for each of the methylated sequences and the unmethylated sequences, modeling the obtained counts with a multinomial distribution that can respectively determine a priori variables.

10. The learning method according to claim 8 or 9, wherein: The third parameter is a parameter that imposes a structural data restriction that the sum of each count calculated using a multinomial distribution follows a Gaussian distribution when a plurality of sequences are simultaneously amplified using common primers.

11. A determination method, which is executed by a determination system having a processor, wherein: The processor performs the following processing: inputting nucleotide sequence and assay protocol information for biomarker sequences of interest used in the multiplexed panel; inputting the error characteristic learned and metadata associated with the error characteristic obtained as a result of the learning method according to any one of claims 8 to 10; outputting a first score for a set of possible biomarker sequences using the nucleotide sequence inputted using a predetermined benchmark, the assay protocol information, the learned error characteristics, and the metadata; and The biomarker sequence set is determined by considering the value of the first score for each set.

12. The determination method according to claim 11, wherein: The processor performs the following processing: Entering a second score for each biomarker sequence to be determined; and The best subset of the multiplexed panel is selected by optimizing the balance between the first score and the second score considering the first score for each biomarker sequence in the set of biomarker sequences.

13. A prediction method, which is performed by a prediction system having a processor and predicting measurement error characteristics of a gene sequence, wherein: The processor performs the following processing: inputting nucleotide sequence and assay protocol information for biomarker sequences of interest used in the multiplexed panel; inputting the error characteristic learned and metadata associated with the error characteristic obtained by the learning method according to any one of claims 8 to 10; Calculating the similarity of a biomarker sequence previously included in the calibration data and a new biomarker sequence using a measure of the metric used to calculate the similarity between two gene sequences; and The calculated similarity is used in combination with other relevant inputs and the learned error characteristic to predict the error characteristic when measuring biomarker sequences not included in the calibration data.

14. The prediction method according to claim 13, wherein: The processor performs the following processing: using the predicted error characteristic to obtain a biomarker sequence that is most similar to a biomarker sequence not included in the calibration data and that can be utilized in the calibration data; and The acquired information of the biomarker sequence is reflected in the determination of the biomarker sequence set in the determination method according to claim 11 or 12.

Citation Information

Patent Citations

  • Machine learning implementation for multi-analyte assays of biological samples

    JP2021521536A

  • Systems and methods for methylation prediction

    US20080172183A1

  • Methylation-based diagnostic method

    WO2005038046A1