G-quadruplex formation sequence prediction device and machine learning method

Abstract

To provide a G4 formation sequence prediction device that can predict the formation state of G4 in a base sequence with high accuracy.SOLUTION: A G4 formation sequence prediction device according to the present invention includes: a determination unit that determines, from base sequence data, a feature vector indicating a feature of a base sequence represented by 5'-Ny0Gx1 Ny1Gx2 Ny2Gx3 Ny3Gx4 Ny4-3' (in the base sequence, Ny0 to Ny4 represent any sequence fragment with 0 or more bases, and Gx1 to Gx4 represent sequence fragments containing only G with two or more bases); and a prediction unit that predicts the formation state of a guanine quadruplex in the base sequence from the feature vector by using a prediction model. The prediction model is constructed by performing machine learning using a set of the feature vector indicating a feature of the base sequence and G4 formation state information indicating the formation state of the guanine quadruplex in the base sequence.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a G-quadruplex-forming sequence prediction device that predicts the formation state of a G-quadruplex in a base sequence, such as the presence or absence of G-quadruplex formation in a base sequence (whether a base sequence forms a G-quadruplex) and the stability of a G-quadruplex in a base sequence, and a machine learning method for constructing a prediction model for use in the prediction device, as well as a compound or a salt thereof used for detecting a G-quadruplex-forming sequence, and a G-quadruplex-binding ligand. [Background technology]

[0002] Conventionally, the G-quadruplex (G-quadruplex, hereafter sometimes abbreviated as "G4") is a nucleic acid higher-order structure formed by the driving force of π-π interactions and hydrogen bonds between guanine bases, as shown in Figure 9b). More than 20 types of proteins that bind to G4 have also been identified, and G4 is a dynamic structure that repeatedly forms and dissociates in vivo. As shown in Figure 9a), when a low-molecular-weight compound that stabilizes G4 (G4 ligand) is applied to a base sequence where G4 is formed (hereafter sometimes abbreviated as "G4 forming sequence") that exists in the untranslated region of mRNA, etc., ribosomes stagnate at the G4 formation point, dissociate from mRNA, and translation is inhibited. G4 formed in genomic DNA / RNA inhibits transcription and replication by inhibiting the polymerase elongation reaction. Therefore, G4 has attracted attention as a characteristic nucleic acid higher-order structure, and is expected to be a new drug discovery target involved in protein expression and replication.

[0003] Clinical development of anticancer drugs based on the cancer gene inhibitory activity targeting G4 is underway overseas. Meanwhile, in the field of antiviral drug discovery, G4-forming sequences have been found in the genomes of several types of RNA viruses, such as Corona, Retro, and Flavi, and it has been reported that treatment with G4 ligands inhibits viral proliferation and inhibits viral protein synthesis. In other words, the formula "stabilization of G4 = control of viral function" is being proven. G4 is also present in host cells, but in (+)ssRNA viruses, the genome and mRNA have the same sequence, so the G4-targeting drug discovery strategy can inhibit all stages of the viral life cycle, including transcription, translation, and replication, and can achieve high drug selectivity against viruses.

[0004] G4 is considered to be in equilibrium with single-stranded sequences and to be dynamically formed at the timing of biological events such as replication and transcription. There is a demand for the development of molecular tools for detecting G4 in vivo, which plays an important role in biological processes, as well as a prediction technique for the higher-order structure of DNA / RNA nucleic acids that predicts the formation state of G4. Currently, fluorescently labeled low molecular weight compounds have been reported as molecular tools for detecting G4. Specifically, for example, a G4-binding ligand that emits fluorescence only when interacting with G4 has been reported (Patent Document 1).

[0005] In contrast, as a prediction technique for the higher-order structure of DNA / RNA, many G4 formation sequence prediction devices have been reported that apply prediction models for the higher-order structure of RNA, but these prediction models are based on typical base pair formation (AT, GU). On the other hand, G4 formation sequence prediction devices that apply prediction models that predict the formation state of G4s consisting of atypical base pairs have also been reported in recent years. As such prediction models, for example, a model that extracts only G4s with three or more overlapping G-quartet planes has been reported (Non-Patent Documents 1 and 2). In addition, a model that extracts G4s with two or more G-quartet planes and evaluates the length and evenness of the loop bases to score the formation probability of G4s has been reported (Non-Patent Document 3). In addition, a model that scores the formation probability of G4s based on the content of G (guanine) and C (cytosine) has been reported (Non-Patent Document 4). In addition, a model that extracts incomplete G4s in which G (guanine) is missing has also been reported (Non-Patent Document 5). [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2020-176108 A [Non-patent literature]

[0007] [Non-Patent Document 1] Nucleic Acids Res. 2005, 33, 2908-2916. [Non-Patent Document 2] Nucleic Acids Res., 2005, 33, 2901-2907. [Non-Patent Document 3] Nucleic Acids Res., 2006, 34, 676-682. [Non-Patent Document 4] Nucleic Acids Res. 2016, 44, 1746-1759. [Non-Patent Document 5] Biochimie., 2017, 135, 54-62. Summary of the Invention [Problem to be solved by the invention]

[0008] However, there have been few reports of G4 formation sequence prediction devices to which a prediction model for predicting the G4 formation state has been applied, and it has been difficult to accurately predict what kind of base sequence will form a G4 or the stability of G4 in a base sequence. In the above-mentioned conventional G4 formation sequence prediction devices, a prediction model was constructed based on the empirical rules of experimental researchers using the number of layers of the G-quartet plane and the loop sequence composition, etc., whereas there was no prediction model that statistically and objectively evaluates the presence or absence of G4 formation in a base sequence and the stability of G4 by integrating the content of G (guanine), the tether sequence composition, the number of occurrences of G-quartet planes, etc., and predicts the G4 formation state. This is due to the fact that there is no experimental data set in which a sufficient number of G4 formation sequences have been collected to perform sufficiently reliable statistical processing and obtained under a single condition. In addition, there are no experimental data sets for G4 formation sequences (G3N 1-7 Although sequence rules such as 1G3G4 and 1G4G5 have been proposed, several G4-forming sequences that deviate from these sequence rules have been reported. Furthermore, previous prediction models for predicting the G4 formation state had low prediction accuracy, especially for base sequences with a low G (guanine) content. In recent years, cases have been reported in which G4s are formed even in base sequences in which the G (guanine) content is 50% or less of the total. For this reason, there is a demand for a G4 formation sequence prediction device that can predict the G4 formation state, such as the presence or absence of G4 formation and the stability of G4s, with high accuracy even for base sequences with a low G (guanine) content.

[0009] The present invention has been made in consideration of these points, and its purpose is to provide a G4 formation sequence prediction device that can predict the G4 formation state in a base sequence with high accuracy, a machine learning method for constructing a prediction model for use in the prediction device, and a compound or a salt thereof, and a G4-binding ligand for use in detecting a G4 formation sequence. [Means for solving the problem]

[0010] Therefore, in the present invention, the inventors have created a G4-forming sequence prediction device by subjecting the G4-forming sequences experimentally identified to machine learning. Candidates for G4-forming sequences are predicted and extracted using the prediction device, and a verification experiment of G4 formation using a G4-binding ligand is carried out to test the prediction results of the prediction device. The results are then re-learned, thereby improving the prediction accuracy of the G4-forming sequence prediction device.

[0011] In order to solve the above problems, the present invention provides a device for predicting G-quadruplex formation sequences, y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3' (in the base sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G with two or more bases.) from sequence data of the base sequence, and a prediction unit that predicts a G-quadruplex formation state in the base sequence from the feature vector by using a prediction model, wherein the prediction model is a prediction model constructed by machine learning using training data that is a pair of a feature vector indicating the characteristics of the base sequence and G4 formation state information indicating the G-quadruplex formation state in the base sequence. Effect of the Invention

[0012] According to the present invention, the state of G4 formation in a base sequence can be predicted with high accuracy. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram for explaining an outline of a G4 formation sequence prediction device according to a first embodiment. [Diagram 2] FIG. 1 is a diagram illustrating a schematic configuration of a computer that realizes a G4 formation sequence prediction device according to the first and second embodiments. [Diagram 3] FIG. 13 is a diagram for illustrating an outline of a G4 formation sequence prediction device according to a second embodiment. [Figure 4] FIG. 1 is a diagram showing an outline of a process flow for predicting whether or not a DNA sequence is a G4 forming sequence when evaluating the accuracy of the prediction device constructed in Example 1. [Diagram 5] 1 is a table showing the results of determining the prediction results of the prediction device constructed in the first embodiment through a verification experiment. [Figure 6] This is a table showing the dissociation constant Kd values ​​of the training data and the predicted stability values ​​of G4 when the stability of G4 in a DNA sequence is predicted using the prediction device of Example 2, together with the predicted value of the presence or absence of G4 formation obtained when the presence or absence of G4 formation in a DNA sequence is predicted using the prediction device of Example 1. [Figure 7] 13A is a graph showing the predicted value of G4 stability versus log(Kd[M]), which is the logarithm of the dissociation constant Kd value of the teacher data, when the stability of G4 in a DNA sequence is predicted using the prediction device of Example 2; and FIG. 13B is a graph showing the predicted value of the presence or absence of G4 formation versus log(Kd[M]), which is the logarithm of the dissociation constant Kd value of the teacher data, when the prediction device of Example 1 is used to predict the presence or absence of G4 formation in a DNA sequence. [Figure 8] FIG. 1 shows graphs in which the solid lines represent the UV absorption difference spectra of six types of G4-forming sequences (NCOR2, telo24, BIN1, OTP, WT1, and ZN318) before the addition of compound L2H2-6OTD, and the dashed lines represent the UV absorption difference spectra of three types of sequences in the high Kd value group (OTP, WT1, and ZN318) after the addition of compound L2H2-6OTD. [Figure 9] FIG. 1 is a diagram that illustrates the structure of a guanine quadruplex (G4), and a schematic diagram of the phenomenon that occurs when a G4-binding ligand, which is a low-molecular-weight compound that stabilizes G4, is applied to a G4-forming sequence. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, embodiments of the G4 formation sequence prediction device, the machine learning method, the compound or a salt thereof, and the G4 binding ligand of the present invention will be described. In the following, "G4" refers to a guanine quadruplex.

[0015] A. First embodiment DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described by taking as examples a G4 formation sequence prediction device, a machine learning method, and a G4 binding ligand according to a first embodiment.

[0016] <G4 formation sequence prediction device> Fig. 1 is a diagram for explaining the G4 formation sequence prediction device according to the first embodiment. Fig. 2 is a diagram for showing the configuration of a computer for realizing the G4 formation sequence prediction device according to the first embodiment.

[0017] As shown in Fig. 1, the G4 formation sequence prediction device 10 according to the first embodiment includes an input unit 2, a storage unit 4, a processing unit 6 including a determination unit 6a and a G4 stability prediction unit 6b, and an output unit 8. The G4 formation sequence prediction device 10 is realized, for example, by a computer 100 shown in Fig. 2. The computer 100 includes a processing device (computer) 110, a storage device 120, an input device 130, an output device 140, an input / output interface (I / F) 150, and the like.

[0018] The input part of the G4 formation sequence prediction system is 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3' DNA sequence (base sequence) (in the DNA sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4represents a sequence fragment containing only G with two or more bases.) is input. The memory unit stores a prediction model and other data. The determination unit of the processing unit determines a feature vector x indicating the characteristics of the DNA sequence from the sequence data of the DNA sequence. The G4 stability prediction unit of the processing unit predicts the stability of G4 in the DNA sequence from the feature vector x by using the prediction model. The output unit outputs the prediction result by the G4 stability prediction unit.

[0019] The prediction model stored in the memory is a feature vector x i and the dissociation constant K of the DNA sequence compound 6OTD-Py d The prediction model is constructed by performing machine learning using teacher data that is a pair of the value [M] of [M]. The compound 6OTD-Py is a G4 binding ligand according to the first embodiment. The prediction model is constructed by using the machine learning method according to the first embodiment described later.

[0020] In the prediction method for predicting the stability of G4 in a DNA sequence by using the G4 formation sequence prediction device according to the first embodiment, first, a 5'-N sequence that is the target for predicting the stability of G4 is selected. y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3' DNA sequence (in the DNA sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G bases of 2 or more.)

[0021] Next, sequence data of the DNA sequence is input via the input unit of the G4 formation sequence prediction device. The input sequence data is stored in the memory unit by the processing unit.

[0022] Next, a feature vector x indicating the characteristics of the DNA sequence is determined from the sequence data by a determination unit of the processing unit of the G4 formation sequence prediction device. The determined feature vector x is stored in the storage unit by the processing unit. Note that the method of determining the feature vector x from the sequence data is the same as that of the feature vector x in the machine learning method according to the first embodiment described later. i The method is similar to the method for determining

[0023] Next, the G4 stability prediction unit of the processing unit of the G4 formation sequence prediction device predicts the stability of G4 in the DNA sequence from the feature vector x by using the prediction model stored in the storage unit. Specifically, the value calculated by inputting the feature vector x into the regression equation f(x) of the prediction model represented by the equations (3-1) and (3-2) described later is used as the dissociation constant K of the compound 6OTD-Py in the DNA sequence. d The predicted value of [M] is obtained.

[0024] Next, the output unit of the G4 formation sequence prediction device outputs the dissociation constant K of the compound 6OTD-Py of the DNA sequence as the prediction result by the G4 stability prediction unit. d Output the predicted value of the dissociation constant K d According to the predicted value of, the lower the predicted value, the higher the possibility of G4 formation in the DNA sequence and the higher the stability of the G4 formed in the DNA sequence is predicted. Also, the higher the predicted value, the lower the possibility of G4 formation in the DNA sequence and the lower the stability of the G4 formed in the DNA sequence is predicted.

[0025] <Machine learning method> A description will be given of a machine learning method according to the first embodiment. The machine learning method according to the first embodiment includes a teacher data acquisition step and a machine learning step.

[0026] [Teacher data acquisition process] In the teacher data acquisition process according to the first embodiment, first, 5'-N y0 G x1 N y1 G x2 N y2 G x3 Ny3 G x4 N y4 -3' DNA sequence (base sequence) (in the DNA sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G with two or more bases.) and the dissociation constant K of the compound 6OTD-Py in the DNA sequence d The pair of values ​​[M] is obtained as training data.

[0027] [Selection of G4-forming and non-G4-forming sequences for training data] Specifically, in the teacher data acquisition step, first, for a large number of types of DNA sequences arranged on a DNA microarray previously acquired, a method for detecting the presence of a base sequence forming G4 in a sample using a G4-binding ligand according to the first embodiment described later is used to detect 2000 types of sequences to which the G4-binding ligand binds as G4-forming sequences (sequences that form G4), and 2000 types of sequences to which the G4-binding ligand does not bind as G4 non-forming sequences (sequences that do not form G4). Next, for these detected sequences, whether or not they form G4 is experimentally verified using methods such as DMS footprinting, EMSA, and CD spectroscopy. As a result, 130 types of sequences that have been experimentally verified to form G4 are selected from the 2000 types of sequences detected as G4-forming sequences, and 130 types that have been experimentally verified not to form G4 are selected from the 2000 types of sequences detected as G4 non-forming sequences. At this time, among the sequences experimentally verified to form G4, 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 Although sequences not represented by -3' cannot be included, sequences that are experimentally verified not to form G4 include those that are not represented by 5'-N y0 G x1 N y1 G x2 Ny2 G x3 N y3 G x4 N y4 Therefore, in the selection of 130 sequences that were experimentally verified not to form G4, the 5'-N sequence was selected from the sequences that were verified not to form G4. y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 - Exclude sequences not represented in 3' and 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 Only sequences represented by -3' were selected. Next, the dissociation constant K of the compound 6OTD-Py was determined from 130 sequences that were experimentally verified to form G4. d The value of K d ≦1×10 -5.2 The 83 sequences that are M are selected as G4-forming sequences for training data. Then, from the 130 sequences that were experimentally verified as not forming G4, the dissociation constant K d The value of K d >1×10 -5.2 89 types of sequences that are M are selected as G4 non-forming sequences for training data.

[0028] [Acquisition of training data] Next, 83 types of teacher data are obtained for 83 types of G4-forming sequences for teacher data, and 89 types of teacher data are obtained for 89 types of G4 non-forming sequences for teacher data, to obtain a total of 172 types of teacher data. At this time, as the teacher data for various G4-forming sequences, a feature vector x i and the dissociation constant K of the compound 6OTD-Py of the G4-forming sequence dAs training data for various non-G4 sequences, a feature vector x i and the dissociation constant K of the compound 6OTD-Py of the G4 non-forming sequence d In this case, the dissociation constant K of the compound 6OTD-Py of the G4-forming sequence in the training data for various G4-forming sequences is obtained. d The value of can be set to, for example, a value actually obtained by a general method for the oligonucleotide of the G4-forming sequence. In addition, the dissociation constant K of the compound 6OTD-Py of the G4-non-forming sequence in the training data for various G4-non-forming sequences can be set to, for example, d The value of can be set to a value actually determined by a general method for the oligonucleotide of the non-G4-forming sequence. d Considering the possibility that the value of may not be accurately obtained, the dissociation constant K of the compound 6OTD-Py of the G4 non-forming sequence in the training data for all types of G4 non-forming sequences for the training data was d The dissociation constant K d The value is 1×10 -5.2 A constant value larger than M may be set.

[0029] Here, the feature vector x , which shows the characteristics of various DNA sequences of 83 types of G4 forming sequences and 89 types of G4 non-forming sequences, is i A method for determining the above will be described.

[0030] A feature vector x that indicates the characteristics of various DNA sequences i In the method of determining the feature vector x from the sequence data of the DNA sequence based on certain rules, i Ingredient x in Specifically, we obtain the stability of G4 in DNA sequences by comparing the N y0 ~N y4 Along with the homology of the sequence fragments, x1 ~G x4Since the number of G bases in is assumed to be an important influencing factor, the N for any 5-base sequence can be calculated from the DNA sequence data. y0 ~N y4 The homology score of the sequence fragment and G x1 ~G x4 The length of the sequence fragment and the feature vector x i Ingredient x in Obtain as.

[0031] N for any 5-base sequence y0 ~N y4 The homology score of the sequence fragment and the method of obtaining it will be explained. The type of base sequence of any 5 bases is the type of arrangement when 5 arbitrary letters are selected from the four types of letters A, T, C, and G that represent bases. 5 = 1024 types. Therefore, N for any 5-base sequence is y0 ~N y4 When obtaining the homology score for each sequence fragment, the homology score for each sequence fragment for 1024 types of arbitrary 5-base base sequences is obtained, so 1024 scores are obtained. y0 ~N y4 The homology scores for the sequence fragments are 1024 × 5 = 5120 scores, and these 5120 scores are expressed as a feature vector x i 5120 components x in It becomes.

[0032] In addition, the N y0 ~N y4To obtain the homology score for each sequence fragment, the Smith-Waterman method is used to align each sequence fragment with an arbitrary 5-base base sequence, and then a score of +2 is given if each character in the arbitrary 5-base base sequence matches a character in the same position in each sequence fragment (match(m)), a score of -1 is given if the character in the same position in each sequence fragment does not match (mismatch(i)), a score of -1 is given if a character deletion occurs in the arbitrary 5-base base sequence or each sequence fragment (gapopen(o)), and a score of -0.1 is given if a character deletion occurs consecutively in the arbitrary 5-base base sequence or each sequence fragment (gapextension(x)). The alignment pattern that gives the maximum total score among all alignment patterns is then obtained. The maximum total score is then calculated as the Nth alignment pattern for the arbitrary 5-base base sequence. y0 ~N y4 The homology score for each sequence fragment is obtained.

[0033] Here is an example of obtaining a homology score. For an example of a 5-base sequence, "ATGCA", y0 ~N y4 When an example of each sequence fragment, "ATCCGCA", is aligned as shown in Table 1 below, the total score of the relationship between characters at the same position is 2.9.

[0034] [Table 1]

[0035] Then, for an arbitrary 5-base sequence, such as "ATGCA", y0 ~N y4 Among all the alignment patterns when aligning an example of each sequence fragment, "ATCCGCA", the alignment pattern that maximizes the total score of the relationship of characters at the same position is as shown in Table 2 below, and the total score of the relationship of characters at the same position in that alignment pattern is 8.9. Therefore, the N for an example of a base sequence of any 5 bases, "ATGCA", y0 ~N y4The homology score for each sequence fragment, "ATCCGCA", is 8.9.

[0036] [Table 2]

[0037] On the other hand, G x1 ~G x4 The length of each sequence fragment is the number of G bases contained in the fragment. x1 ~G x4 To obtain the length of a sequence fragment, use G x1 ~G x4 The base lengths of the four sequence fragments are then obtained, and the base lengths of these four sequence fragments are expressed as a feature vector x i 4 components of x in It becomes.

[0038] Feature vector x i Ingredient x in In the acquisition of the above, the N for all the arbitrary 5-base sequences acquired as above is y0 ~N y4 The homology score of the sequence fragment (5120 components x in ) and G x1 ~G x4 The length of the sequence fragment (four components x in ) to obtain the feature vector x i 5124 components x i1 ~x i5124 This gives us a feature vector x i 5124 components x i1 ~x i5124 A 5124-dimensional feature vector x i Determine.

[0039] [Machine learning process] Next, in the machine learning process, the above 5124-dimensional feature vector x i A total of 172 types of training data were used to perform machine learning, and 5'-N y0 G x1 Ny1 G x2 N y2 G x3 N y3 G x4 N y4 -3' DNA sequence (in the DNA sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G with 2 or more bases.) A predictive model for predicting the stability of G4 in DNA sequences is constructed (machine learning process). Specifically, SVR (support vector regression) to which an RBF kernel is applied is used to perform machine learning using a total of 172 types of training data to construct a predictive model that constitutes a regression equation for predicting the stability of G4 in a DNA sequence. Below, the process flow for constructing a predictive model using SVR to which an RBF kernel is applied is specifically described.

[0040] In the process of constructing a predictive model, first, a 5124-dimensional feature vector x i A total of 172 types of training data were used. i (i=1~N, N=172) and use the Lagrange multiplier α i and α i * The objective function L expressed by the following equations (1-1) and (1-2) D The Lagrange multiplier α that is the maximum within the range where (α) satisfies the constraints expressed by the following equations (2-1) to (2-3) is i and α i * In this machine learning, the RBF kernel parameter γ is set to an optimal value within the range of 0.001 to 1000, and the cost parameter C is set to an optimal value within the range of 0.001 to 1000, and the objective function L D The Lagrange multiplier α that maximizes (α) within the range where the constraint is satisfied. i and α i * The combination of 172 types of training data Tr i The corresponding difference of Lagrange multipliers αi -α i * is α i -α i * Those that are ≠ 0 are called support vectors.

[0041]

number

[0042]

number

[0043] N: Number of training data (172) α i , α i * :Lagrange multiplier for each training data x i : 5124-dimensional feature vector for each training data y i : Dissociation constant K of each training data d Value of ε: Error deadband parameter γ: RBF kernel hyperparameter C: Cost parameters

[0044] Next, the Lagrange multiplier α determined by the above machine learning is i and α i * By using a combination of the above, f(x) is obtained, which is expressed by the following formulas (3-1) and (3-2), as a regression equation for predicting the stability of G4 in a DNA sequence to be predicted that is different from the sequence for the training data. The value calculated by inputting the feature vector x, which indicates the characteristics of the DNA sequence to be predicted, into the regression equation f(x) is the dissociation constant K d This is the predicted value of the dissociation constant K d A low value of K is sufficient for the G4 to be stable in the DNA sequence, so the dissociation constant K dAccording to the predicted value of, the lower the predicted value, the higher the possibility of G4 formation in the DNA sequence and the higher the stability of the G4 formed in the DNA sequence is predicted. Also, the higher the predicted value, the lower the possibility of G4 formation in the DNA sequence and the lower the stability of the G4 formed in the DNA sequence is predicted.

[0045]

number

[0046] x: A feature vector that indicates the characteristics of the DNA sequence to be predicted N: Number of training data (172) α i , α i * :Lagrange multiplier for each training data x i : 5124-dimensional feature vector for each training data γ: RBF kernel hyperparameter b: constant term C: Cost parameters

[0047] The optimum value of the error dead zone parameter ε, the optimum value of the RBF kernel hyperparameter γ, and the optimum value of the cost parameter C are examined using a grid search method. In this case, the candidates for the error dead zone parameter ε, the RBF kernel hyperparameter γ, and the cost parameter C are as follows:

[0048] Candidates for ε: 6 values ​​from 0.001 to 1000 in 10-fold increments Candidates for γ: 6 values ​​from 0.001 to 1000 in 10-fold increments Candidates for C: 6 values ​​from 0.001 to 1000 in 10-fold increments

[0049] Then, under each condition of all combinations of ε, γ, and C candidates (6 × 6 × 6 = 216 combinations), a 5-fold cross-validation was applied to a total of 172 types of training data to obtain the coefficient of determination r2 By calculating these values, the optimal values ​​of ε, γ, and C are examined. Specifically, a total of 172 types of training data are divided into five, and machine learning is performed using four subsets to obtain a regression equation for validation in the same manner as the regression equation f(x) was obtained. The dissociation constant K of the compound 6OTD-Py of the DNA sequence of the remaining one subset for validation is calculated using the regression equation for validation. d The process of calculating the predicted value of is repeated five times by changing one subset for validation. The predicted value calculated in this way (the dissociation constant K of the compound 6OTD-Py of the DNA sequence) is d The predicted value of the dissociation constant K of the compound 6OTD-Py of the DNA sequence in the training data, and the actual value ( d Using the value of , the coefficient of determination r 2 , correlation coefficient, or RMSE (Root Mean Squared Error) is calculated. 2 In the calculation, the coefficient of determination r is calculated from the predicted value itself and the actual value itself. 2 Alternatively, the coefficient of determination r can be calculated by using the logarithm of the predicted value and the logarithm of the actual value instead of the predicted value and the actual value, respectively. 2 In the calculation of RMSE, the RMSE may be calculated directly from the predicted values ​​themselves and the actual measured values ​​themselves, or the RMSE may be calculated using the logarithmic values ​​of the predicted values ​​and the logarithmic values ​​of the actual measured values ​​instead of the predicted values ​​themselves and the actual measured values ​​themselves. In the calculation of the correlation coefficient, the correlation coefficient representing the relationship between the predicted values ​​themselves and the actual measured values ​​themselves may be calculated, or the correlation coefficient representing the relationship between the logarithmic values ​​of the predicted values ​​and the logarithmic values ​​of the actual measured values ​​may be calculated. Then, the coefficient of determination r 2 The values ​​of ε, γ, and C that maximize the correlation coefficient or RMSE are set as the optimal values ​​of ε, γ, and C, respectively.

[0050] In this way, a prediction model is constructed that forms a regression equation for predicting the stability of G4 in a DNA sequence.

[0051] G4-binding ligands The G4 binding ligand according to the first embodiment will be described. The G4 binding ligand according to the first embodiment has the following structure: [ka] The G4-binding ligand according to the first embodiment is a compound (compound 6OTD-Py) or a salt thereof having the formula: When a DNA sequence forming G4 is present in a sample to be measured, the G4-binding ligand according to the first embodiment selectively binds to the G4, and can detect the presence and amount of such a DNA sequence using a change in its fluorescent property as an indicator.

[0052] B. Second embodiment Furthermore, the G4-forming sequence prediction device, the machine learning method, and the G4-binding ligand according to the second embodiment will be described as examples.

[0053] <G4 formation sequence prediction device> FIG. 3 is a diagram for explaining the outline of the G4 formation sequence prediction device according to the second embodiment.

[0054] As shown in Fig. 3, the G4 formation sequence prediction device 20 according to the second embodiment includes an input unit 12, a storage unit 14, a processing unit 16 including a determination unit 16a and a G4 formation prediction unit 16b, and an output unit 18. The G4 formation sequence prediction device 20 is realized, for example, by the computer 100 shown in Fig. 2 described above.

[0055] The input part of the G4 formation sequence prediction system is 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3' DNA sequence (base sequence) (in the DNA sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4represents a sequence fragment containing only G with two or more bases.) is input. The memory unit stores a prediction model and other data. The determination unit of the processing unit determines a feature vector x indicating the characteristics of the DNA sequence from the sequence data of the DNA sequence. The G4 formation prediction unit of the processing unit predicts whether or not G4 will be formed in the DNA sequence (whether or not the DNA sequence is a G4 forming sequence) from the feature vector x by using the prediction model. The output unit outputs the prediction result by the G4 formation prediction unit.

[0056] The prediction model stored in the memory is a feature vector x i and a classification label for classifying the presence or absence of G4 formation in a DNA sequence (a classification label for classifying whether a DNA sequence is a G4 forming sequence or not) t i The prediction model is constructed by performing machine learning using training data that is a pair of the prediction model and the prediction model is constructed by using a machine learning method according to a second embodiment described later.

[0057] In the method for predicting the presence or absence of G4 formation in a DNA sequence by using the G4 formation sequence prediction device according to the second embodiment, first, a 5'-N target for predicting the presence or absence of G4 formation is selected. y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3' DNA sequence (in the DNA sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G bases of 2 or more.)

[0058] Next, sequence data of the DNA sequence is input via the input unit of the G4 formation sequence prediction device. The input sequence data is stored in the memory unit by the processing unit.

[0059] Next, a feature vector x indicating the characteristics of the DNA sequence is determined from the sequence data by a determination unit of the processing unit of the G4 formation sequence prediction device. The determined feature vector x is stored in the storage unit by the processing unit. Note that the method of determining the feature vector x from the sequence data is the same as that of the feature vector x in the machine learning method according to the second embodiment described later. i The method is similar to the method for determining

[0060] Next, the G4 formation prediction unit of the processing unit of the G4 formation sequence prediction device predicts the presence or absence of G4 formation in the DNA sequence from the feature vector x by using the prediction model stored in the storage unit. Specifically, the feature vector x is input to the discrimination function f(x) of the prediction model expressed by the formulas (7-1) to (7-3) described later, and the calculated value is obtained as the predicted value of the presence or absence of G4 formation.

[0061] Next, the output unit of the G4 formation sequence prediction device outputs a predicted value of the presence or absence of G4 formation as a prediction result by the G4 formation prediction unit. According to the predicted value of the presence or absence of G4 formation, with 0.5 as the boundary, the closer the predicted value is to 1, the more likely the DNA sequence is to be a G4 forming sequence, and the closer the predicted value is to 0, the more likely the DNA sequence is to be a non-G4 forming sequence.

[0062] <Machine learning method> A description will now be given of a machine learning method according to the second embodiment. The machine learning method according to the second embodiment includes a teacher data acquisition step and a machine learning step.

[0063] [Teacher data acquisition process] In the teacher data acquisition process according to the second embodiment, first, 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3' DNA sequence (base sequence) (in the DNA sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and Gx1 ~G x4 represents a sequence fragment containing only G with two or more bases.) and a classification label that classifies whether or not G4 is formed in the DNA sequence (classification label that classifies whether or not the DNA sequence is a G4-forming sequence) are obtained as training data.

[0064] [Selection of G4-forming and non-G4-forming sequences for training data] Specifically, in the teacher data acquisition step, first, for a large number of types of DNA sequences arranged on a DNA microarray previously acquired, a method for detecting the presence of base sequences forming G4 in a sample using a G4-binding ligand according to the second embodiment described later is used to detect 2000 types of sequences to which the G4-binding ligand binds as G4-forming sequences (sequences that form G4), and 2000 types of sequences to which the G4-binding ligand does not bind as G4 non-forming sequences (sequences that do not form G4). Next, for these detected sequences, whether or not they form G4 is experimentally verified using methods such as DMS footprinting, EMSA, and CD spectroscopy. As a result, 130 types that have been experimentally verified to form G4 are selected as G4-forming sequences for teacher data from the 2000 types of sequences detected as G4-forming sequences, and 130 types that have been experimentally verified not to form G4 are selected as G4 non-forming sequences for teacher data from the 2000 types of sequences detected as G4 non-forming sequences. Note that among the sequences verified to form G4, 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 Although sequences not represented by -3' cannot be included, sequences verified not to form G4 include those containing 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4Therefore, the 130 types of G4 non-forming sequences for the training data were selected from the sequences verified not to form G4, and the 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 - Exclude sequences not represented in 3' and 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 Only sequences represented in -3' are selected.

[0065] [Acquisition of training data] Next, 130 types of teacher data are obtained for each of the 130 types of G4-forming sequences for teacher data, and 130 types of teacher data are obtained for each of the 130 types of G4 non-forming sequences for teacher data, thereby obtaining a total of 260 types of teacher data. At this time, as the teacher data for each type of G4-forming sequence, a feature vector x i and a classification label t that classifies the DNA sequence as a G4-forming sequence. i As training data for various types of G4 non-forming sequences, a feature vector x i and a classification label t that classifies the DNA sequence as a non-G4 sequence. i Get a pair with "-1".

[0066] Here, the feature vector x , which shows the features of various DNA sequences of 130 types of G4 forming sequences and 130 types of G4 non-forming sequences, is i A method for determining the above will be described.

[0067] (Obtaining components of feature vectors from DNA sequences) A feature vector x that indicates the characteristics of various DNA sequences iIn the method of determining the initial feature vector x from the sequence data of the DNA sequence based on a certain rule, p Ingredient x pn Specifically, the N for any of the 5-base base sequences is obtained for whether the DNA sequence forms G4 or not. y0 ~N y4 Along with the homology of the sequence fragments, x1 ~G x4 Since the number of G bases in is assumed to be an important influencing factor, the N for any 5-base sequence can be calculated from the DNA sequence data. y0 ~N y4 The homology score of the sequence fragment and G x1 ~G x4 The length of the sequence fragment and the initial feature vector x p Ingredient x pn Obtain as.

[0068] N for any 5-base sequence y0 ~N y4 The homology score of the sequence fragment and the method of obtaining it will be explained. The type of base sequence of any 5 bases is the type of arrangement when 5 arbitrary letters are selected from the four types of letters A, T, C, and G that represent bases. 5 = 1024 types. Therefore, N for any 5-base sequence is y0 ~N y4 When obtaining the homology score for each sequence fragment, the homology score for each sequence fragment for 1024 types of arbitrary 5-base base sequences is obtained, so 1024 scores are obtained. y0 ~N y4 The homology scores for the sequence fragments are 1024 × 5 = 5120 scores, and these 5120 scores are calculated by the initial feature vector x p 5120 components x pn It becomes.

[0069] In addition, the N y0 ~N y4To obtain the homology score for each sequence fragment, the Smith-Waterman method is used to align each sequence fragment with an arbitrary 5-base base sequence, and then a score of +2 is given if each character in the arbitrary 5-base base sequence matches a character in the same position in each sequence fragment (match(m)), a score of -1 is given if the character in the same position in each sequence fragment does not match (mismatch(i)), a score of -1 is given if a character deletion occurs in the arbitrary 5-base base sequence or each sequence fragment (gapopen(o)), and a score of -0.1 is given if a character deletion occurs consecutively in the arbitrary 5-base base sequence or each sequence fragment (gapextension(x)). The alignment pattern that gives the maximum total score among all alignment patterns is then obtained. The maximum total score is then calculated as the Nth alignment pattern for the arbitrary 5-base base sequence. y0 ~N y4 The homology score for each sequence fragment is obtained.

[0070] Here is an example of obtaining a homology score. For an example of a 5-base sequence, "ATGCA", y0 ~N y4 When an example of each sequence fragment, "ATCCGCA", is aligned as shown in Table 3 below, the total score of the relationship between characters at the same position is 2.9.

[0071] [Table 3]

[0072] Then, for an arbitrary 5-base sequence, such as "ATGCA", y0 ~N y4 Among all the alignment patterns when aligning an example of each sequence fragment, "ATCCGCA", the alignment pattern that maximizes the total score of the relationship of characters at the same position is as shown in Table 4 below, and the total score of the relationship of characters at the same position in that alignment pattern is 8.9. Therefore, the N for an example of a base sequence of any 5 bases, "ATGCA", y0 ~N y4The homology score for each sequence fragment, "ATCCGCA", is 8.9.

[0073] [Table 4]

[0074] On the other hand, G x1 ~G x4 The length of each sequence fragment is the number of G bases contained in the fragment. x1 ~G x4 To obtain the length of a sequence fragment, use G x1 ~G x4 The base lengths of the four sequence fragments are obtained, and the base lengths of these four sequence fragments are expressed as the initial feature vector x p 4 components of x pn It becomes.

[0075] Initial feature vector x p Ingredient x pn In the acquisition of the above, the N for all the arbitrary 5-base sequences acquired as above is y0 ~N y4 The homology score of the sequence fragment (5120 components x pn ) and G x1 ~G x4 The length of the sequence fragment (four components x pn ) to obtain the initial feature vector x p 5124 components x p1 ~x p5124 Get the.

[0076] (Feature Selection) The above 5124 components x p1 ~x p5124 The initial feature vector x p The number of dimensions is larger than the total number of training data, 260. Therefore, the initial feature vector x p If a G4 formation sequence prediction device is constructed by performing machine learning using training data that is directly adopted from the above, the generalization performance may be poor. Taking this into consideration, the 5124-dimensional initial feature vector xp By performing feature selection on 5124 components x p1 ~x p5124 From the above, we select the components that contribute most to the prediction accuracy of whether a DNA sequence is a G4 non-forming sequence, and then convert the feature vector consisting of only the selected components into a feature vector x i It is determined as follows.

[0077] 5124-dimensional initial feature vector x p When selecting features, first, we set constraints appropriately using a linear SVM (support vector machine) and select only the selected components from the 260 types of training data. i Instead of the 5124-dimensional initial feature vector x p By using 260 types of provisional training data that adopts the above data as is, we construct a feature selection prediction model that configures a discrimination function that separates DNA sequences into G4-forming sequences and non-G4-forming sequences. The discrimination function of the feature selection prediction model is expressed by the following formula (4).

[0078]

number

[0079] D(x p ): Discriminant function of predictive model for feature selection w: 5124-dimensional weight vector x p : 5124-dimensional initial feature vector w1~w 5124 :5124-dimensional weight vector components x p1 ~x p5124 :5124-dimensional initial feature vector components b: constant term

[0080] In the classification function of the feature selection prediction model expressed by the above formula (4), each component w of the 5124-dimensional weight vector w n The magnitude of the absolute value of (n = 1 to 5124) is the 5124-dimensional initial feature vector x pThe component w of the weight vector w in n Component x corresponding to pn This indicates the degree to which the prediction accuracy is affected.

[0081] So, the initial feature vector x of 5124 dimensions p When selecting features, the initial feature vector x p 5124 components x p1 ~x p5124 From the weight vector w, 5124 components w1 to w 5124 The 84 components that correspond to the 84 components with absolute values ​​of 0.0175 or more in are selected as components that contribute highly to the prediction accuracy. i The 84-dimensional feature vector x consisting of only selected components is i Determine.

[0082] [Machine learning process] Next, in the machine learning process, the above 84-dimensional feature vector x i By performing machine learning using a total of 260 types of training data, y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3' DNA sequence (in the DNA sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G with two or more bases.) A predictive model is constructed to predict the presence or absence of G4 formation (whether the DNA sequence is a G4-forming sequence or not) in the DNA sequence (machine learning process). Specifically, an SVM to which an RBF kernel is applied is used to perform machine learning using a total of 260 types of training data to construct a predictive model that constitutes a discrimination function that separates the DNA sequence into a G4-forming sequence and a non-G4-forming sequence. Below, a specific description is given of the process flow for constructing a predictive model using an SVM to which an RBF kernel is applied.

[0083] In the process of constructing a predictive model, first, an 84-dimensional feature vector x i A total of 260 types of training data were used. i (i=1~N, N=260) and use the Lagrange multiplier α i The objective function L expressed by the following equations (5-1) and (5-2) D The Lagrange multiplier α that is the maximum within the range where (α) satisfies the constraints expressed by the following equations (6-1) and (6-2) is i In this machine learning, the RBF kernel parameter γ is set to an optimal value within the range of 0.001 to 1000, and the cost parameter C is set to an optimal value within the range of 0.001 to 1000, and the objective function L D The Lagrange multiplier α that maximizes (α) within the range where the constraint is satisfied. i The combination of Tr is determined. i The corresponding Lagrange multiplier α i A vector that is not 0 is specifically called a support vector.

[0084]

number

[0085]

number

[0086] α i :Lagrange multiplier for each training data x i : 84-dimensional feature vector for each training data t i : Classification label of each training data ("1" or "-1") γ: RBF kernel hyperparameter C: Cost parameters

[0087] Next, the Lagrange multiplier α determined by the above machine learning is i By using a combination of the above, f(x) is obtained as a discrimination function for separating a DNA sequence to be predicted, other than the sequence for teacher data, into a G4-forming sequence and a G4 non-forming sequence, as expressed by the following formulas (7-1) to (7-3). The value calculated by inputting a feature vector x indicating the characteristics of the DNA sequence to be predicted into the discrimination function f(x) is normalized to the range of 0 to 1 by a sigmoid function, and becomes a predicted value for predicting the presence or absence of G4 formation in the DNA sequence to be predicted. According to the predicted value for the presence or absence of G4 formation, with 0.5 as the boundary, the closer the predicted value is to 1, the more likely it is that the DNA sequence is a G4-forming sequence, and the closer the predicted value is to 0, the more likely it is that the DNA sequence is a non-G4-forming sequence.

[0088]

number

[0089] x: A feature vector that indicates the characteristics of the DNA sequence to be predicted α i :Lagrange multiplier for each training data x i : 84-dimensional feature vector for each training data t i : Classification label of each training data ("1" or "-1") γ: RBF kernel hyperparameter C: Cost parameters ρ: Threshold constant

[0090] The optimal value of the RBF kernel hyperparameter γ and the optimal value of the cost parameter C are examined using a grid search method. In this case, under each condition where γ and C are set separately to each value in 10-fold increments from 0.001 to 1000, 5-fold cross validation is applied to a total of 260 types of teacher data to calculate the accuracy rate (5-fold cross validation accuracy), and the optimal values ​​of γ and C are examined. Specifically, a total of 260 types of teacher data are divided into 5, and machine learning is performed using 4 subsets to obtain a verification discrimination function in the same manner as the method of obtaining the discrimination function f(x), and the process of separating the DNA sequence of the remaining one verification subset into G4-forming sequences and G4 non-forming sequences using the verification discrimination function is performed 5 times by changing one verification subset. From the results of such examination, the values ​​of γ and C that give the highest accuracy rate are set as the optimal values ​​of γ and C, respectively.

[0091] In this manner, a prediction model is constructed that constitutes a discrimination function for separating DNA sequences into G4-forming sequences and non-G4-forming sequences.

[0092] G4-binding ligands The G4-binding ligand according to the second embodiment will be described. The G4-binding ligand according to the second embodiment is the same as the G4-binding ligand according to the first embodiment, and when a DNA sequence forming G4 is present in a sample to be measured, it selectively binds to the G4 and can detect the presence and amount of such a DNA sequence using a change in its fluorescence property as an indicator.

[0093] C. Effects Viral RNA or DNA is known to form a G4 structure, and the virus-derived G4 structure is involved in transcription, translation, replication, genome stability, and the like. Therefore, the G4 structure is a drug discovery target for antiviral drugs. According to the G4 formation sequence prediction device and machine learning method of the embodiment, as in the first and second embodiments, the presence or absence of G4 formation in the base sequence of nucleic acid such as DNA, RNA, etc., and the G4 formation state such as the stability of G4 can be predicted with high accuracy from genome information, independent of the biological species. Therefore, it is possible to quickly present drug discovery targets for emerging and re-emerging infectious diseases.

[0094] In particular, according to the embodiment, it is possible to predict the formation state of G4 such as the presence or absence of G4 formation in a base sequence and the stability of G4 based on a statistical method, not on rules based on empirical rules as in the first and second embodiments. In addition, since it is possible to predict the formation state of G4 such as the presence or absence of G4 formation and the stability of G4 for a base sequence containing a tether sequence, it is possible to make a prediction that reflects the actual formation process of G4 in a living body. Furthermore, since it is possible to predict the formation state of G4 such as the presence or absence of G4 formation and the stability of G4 for base sequences other than base sequences that have been assumed to have a high content of G (guanine) and have a uniform length, it is also possible to obtain an advantage that it is easy to obtain target specificity for G4-forming sequences that can be uniquely predicted by the method according to the embodiment.

[0095] Therefore, in the embodiment, it becomes possible to predict higher-order structures of nucleic acid structures, particularly those due to atypical base pairs, and therefore it is possible to present highly unique drug discovery targets.

[0096] Furthermore, according to the compound or its salt and the G4-binding ligand according to the embodiment, as in the first and second embodiments, when a base sequence forming G4 is present in a sample to be measured, the compound or its salt and the G4-binding ligand selectively bind to the G4, and the presence and amount of such a base sequence can be detected using a change in its fluorescent properties as an indicator. Therefore, the formation / non-formation of G4 in the base sequence in the sample to be measured can be detected quickly and sensitively, and a foundation for ligand-based disease marker detection can be constructed. The compound or its salt and the G4-binding ligand according to the embodiment do not require a separation step of the ligand not bound to the base sequence, and therefore can be suitably used for imaging in living cells.

[0097] D. Detailed Description of the Embodiments Next, the configurations of the G4-forming sequence prediction device and machine learning method, as well as the compound or a salt thereof, and the G4-binding ligand according to the embodiments will be described in detail.

[0098] 1. G4 formation sequence prediction device The G4 formation sequence prediction device according to the embodiment is characterized by including a determination unit and a prediction unit.

[0099] The determining part of the G4 formation sequence prediction system is the 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3' (in the base sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G with two or more bases.) A feature vector showing the characteristics of the above-mentioned base sequence is determined from the sequence data of the base sequence.

[0100] The base sequence is 5'-N y0 G x1 N y1 G x2 Ny2 G x3 N y3 G x4 N y4 The base sequence is not particularly limited as long as it is a base sequence represented by -3', and may be a DNA sequence or an RNA sequence. "Sequence data of a base sequence" means data that represents the order of letters that represent bases in a base sequence. For example, when the base sequence is a DNA sequence, the sequence data of a base sequence is data that represents the order of four types of letters, A, T, C, and G, that represent bases in a DNA sequence, and when the base sequence is an RNA sequence, the sequence data is data that represents the order of four types of letters, A, U, C, and G, that represent bases in an RNA sequence.

[0101] The method of determining a feature vector showing the characteristics of a base sequence is not particularly limited. For example, as in the first embodiment, N y0 ~N y4 The homology score of the sequence fragment and G x1 ~G x4 The base lengths of the sequence fragments of the sequence fragments are obtained, and an initial feature vector is obtained by integrating them. Then, the initial feature vector is subjected to feature selection to select components that contribute to the prediction accuracy to a high extent, and the feature vector consisting of the selected components is determined as the feature vector showing the characteristics of the base sequence. Alternatively, N for all arbitrary 5-base base sequences is y0 ~N y4 The homology score of the sequence fragment and G x1 ~G x4 and the base lengths of the sequence fragments, and a feature vector is obtained by integrating these, and the feature vector is determined as the feature vector indicating the features of the base sequence.

[0102] The prediction unit of the G4-forming sequence prediction device predicts the G4 formation state in the base sequence from the feature vector by using a prediction model. Note that the "G4-forming sequence" refers to a base sequence that forms a G4, and the "G4 non-forming sequence" refers to a base sequence that does not form a G4.

[0103] The prediction model is constructed by performing machine learning using training data that is a set of a feature vector indicating the characteristics of the base sequence and G4 formation state information indicating the formation state of G4 in the base sequence. The prediction model is constructed, for example, by using the machine learning method described in the section "2. Machine learning method" described later.

[0104] The prediction unit of the G4 formation sequence prediction device is not particularly limited, but for example, as in the first embodiment, a G4 stability prediction unit that predicts the stability of G4 in the base sequence as the formation state of G4 in the base sequence, and the G4 formation state information is the value of the dissociation constant of the G4 binding ligand of the base sequence, is preferable. This is because it is possible to predict not only the presence or absence of G4 formation in the base sequence, but also the stability of G4. In addition, as the prediction unit of the G4 formation sequence prediction device, for example, as in the second embodiment, a G4 formation prediction unit that predicts the presence or absence of G4 formation in the base sequence (whether the base sequence is a G4 forming sequence or not) as the formation state of G4 in the base sequence, and the G4 formation state information may be a classification label that classifies the presence or absence of G4 formation in the base sequence (classification label that classifies whether the base sequence is a G4 forming sequence or not).

[0105] The G4 formation sequence prediction device is not particularly limited as long as it comprises a determination unit and a prediction unit, but as in the first embodiment, it may further comprise an input unit for inputting sequence data of a base sequence, a memory unit for storing a prediction model, an output unit for outputting the prediction result by the prediction unit, etc.

[0106] 2. Machine Learning Methods The machine learning method according to the embodiment is characterized by including a teacher data acquisition step and a machine learning step.

[0107] (1) Training data acquisition process In the training data acquisition process, 5'-N y0 G x1 N y1 G x2 N y2 Gx3 N y3 G x4 N y4 -3' (in the base sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G with two or more bases.) and G4 formation state information indicating the formation state of the G-quadruplex in the base sequence are obtained as training data.

[0108] The base sequence is similar to that explained in the above section "1. G4 formation sequence prediction device," and therefore will not be explained here.

[0109] The G4 formation state information is not particularly limited, but for example, as in the first embodiment, the value of the dissociation constant of the G-quadruplex binding ligand of the base sequence is preferable. This is because a prediction model capable of predicting not only the presence or absence of G4 formation in the base sequence but also the stability of G4 can be constructed. In addition, as in the second embodiment, the G4 formation state information may be a classification label that classifies the presence or absence of the G-quadruplex formation in the base sequence (a classification label that classifies whether the base sequence is a G4-forming sequence or not).

[0110] An embodiment of a method for acquiring a pair of a feature vector showing the characteristics of a base sequence and a classification label for classifying the presence or absence of G4 formation in the base sequence (a classification label for classifying whether the base sequence is a G4-forming sequence or not) as teacher data will be described. In this embodiment, first, a plurality of types of G4-forming sequences that have been experimentally verified to form G4 are selected as G4-forming sequences for teacher data, and a plurality of types of non-G4-forming sequences that have been experimentally verified not to form G4 are selected as non-G4-forming sequences for teacher data. Next, a plurality of types of teacher data are acquired for a plurality of types of G4-forming sequences for teacher data, and a plurality of types of teacher data are acquired for a plurality of types of non-G4-forming sequences for teacher data. At this time, as teacher data for various G4-forming sequences, a pair of a feature vector showing the characteristics of the G4-forming sequence and a classification label "1" that classifies the base sequence as a G4-forming sequence (a sequence that forms G4) is acquired. As training data for various types of non-G4-forming sequences, a pair of a feature vector showing the characteristics of the non-G4-forming sequence and a classification label "-1" that classifies the base sequence as a non-G4-forming sequence (a sequence that does not form a G4) is obtained. There are no particular limitations on the method for determining the feature vector showing the characteristics of the G4-forming sequence and the base sequence including the non-G4-forming sequence, but since it is the same as the method for determining the feature vector described above in "1. G4-forming sequence prediction device", a description thereof will be omitted here.

[0111] (2) Machine learning process In the machine learning step, machine learning is performed using the training data to construct a prediction model for predicting the formation state of the G-quadruplex in the base sequence.

[0112] The prediction model is not particularly limited, but for example, as in the first embodiment, a prediction model for predicting the stability of the G-quadruplex in the base sequence as the formation state of the G-quadruplex in the base sequence is preferable. This is because a prediction model capable of predicting not only the presence or absence of G4 formation in the base sequence but also the stability of G4 can be constructed. In addition, as in the second embodiment, the prediction model may be a prediction model for predicting the presence or absence of the formation of the G-quadruplex in the base sequence (whether the base sequence is a G4-forming sequence) as the formation state of the G-quadruplex in the base sequence.

[0113] The method of constructing a prediction model by performing machine learning using teacher data is not particularly limited, but a method capable of handling nonlinear relationships in teacher data is preferable, and examples of the method capable of handling nonlinear relationships include SVM (support vector machine), neural network, decision tree regression, random forest, gradient boosting, etc. Among them, for example, SVM to which a kernel function such as RBF kernel is applied, such as the SVM used in the first embodiment, is preferable.

[0114] (3) Machine learning methods The machine learning method may use the following method. That is, first, a verification experiment of G4 formation is performed for the candidates of G4-forming sequences and candidates of G4-non-forming sequences predicted and extracted using the G4-forming sequence prediction device to find new G4-forming sequences and non-G4 sequences. Next, new teacher data is obtained from the newly found G4-forming sequences and non-G4 sequences, and machine learning is performed using the new teacher data together with the existing teacher data. This can improve the prediction accuracy of the G4-forming sequence prediction device.

[0115] 3. Compound or its salt The compound according to the embodiment, or a salt thereof, has the following general formula (I): [ka] [In the formula, R 2 is H, or [ka] In this case, R 2 R in 1 is an aromatic fused ring consisting of 2 to 4 conjugated planar rings which may be substituted, R 3 and R 4 are each independently [ka] [ka] It is.] or a salt thereof.

[0116] The above compound or its salt selectively binds to G4 and shows an increase in fluorescence depending on the concentration of G4, and therefore can be suitably used to detect the presence or absence of G4 in a base sequence such as a DNA sequence in a sample by fluorescence. Those skilled in the art can understand that the compound having the above structure can be suitably used as a ligand for detecting G4, similar to the compounds specifically shown in the examples, based on the description of this specification and the technical common sense in the art. The above compound or its salt is further superior in detecting G4 in that it has a large Stokes shift (the difference between the maximum excitation wavelength and the maximum fluorescence wavelength), high brightness, and a large difference in fluorescence brightness between the presence and absence of binding to G4.

[0117] R in general formula (I) 2 R in 1may be an aromatic fused ring group selected from the group consisting of naphthyl, anthracenyl, pyrenyl, quinolyl, isoquinolyl, 1-methylquinolinium, acridinyl, indolyl, isoindolyl, benzothiophenyl, benzofuranyl, isobenzofuranyl, benzimidazolyl, benzoxazolyl, benzisoxazolyl, and benzothiazolyl, which may be unsubstituted or substituted with one substituent. These aromatic fused rings are all composed of 2 to 4 conjugated planar rings, and each ring forming the fused ring has a 5- or 6-membered ring structure having an atom selected from carbon (C), oxygen (O), nitrogen (N), and sulfur (S) as a ring atom. These aromatic fused rings have a conjugated ring structure, and it has been found that the planar structure acts advantageously in the interaction with G4.

[0118] R 1 When has a substituent, the substituent is preferably selected from the group consisting of a hydroxyl group (-OH), a methoxy group (-OMe), an acetoxy group (-OAc), a mesyloxy group (-OMs), a trifluoromethylsulfonyloxy group (-OTf), an amino group (-NH2), a dimethylamino group (-NMe2), and a diethylamino group (-NEt2). The number of the substituent may be two or more, but it is particularly preferable that the substituent is substituted with one substituent.

[0119] More specifically, R in general formula (I) 2 R in 1 For example, but not limited to, may have the structure shown in Table 5 below.

[0120] [Table 5]

[0121] The salt of the above compound is not particularly limited, but examples thereof include hydrochloride, trifluoroacetate, etc. Furthermore, when the above compound is contacted with a sample for detecting G4, for example, it may form a salt or solvate with a substance derived from the living body or buffer contained in the sample.

[0122] Compounds according to the invention have the following structure: [ka] In the present specification, this compound is referred to as "compound 6OTD-Py."

[0123] The compound according to the embodiment or a salt thereof can be suitably synthesized according to the description in this specification including the following examples, and the common general technical knowledge in the art.

[0124] 4. G4-binding ligands The G4-binding ligand according to the embodiment is a G4-binding ligand consisting of a compound according to the embodiment or a salt thereof.

[0125] The G4-binding ligand according to the embodiment selectively binds to G4 when a base sequence such as a DNA sequence forming G4 is present in a sample, and can detect the presence and amount of such a base sequence using a change in its fluorescence property as an indicator. The G4-binding ligand according to the embodiment is a "light-up type ligand" consisting of a compound or a salt thereof that does not emit fluorescence alone (i.e., when not bound to G4).

[0126] Thus, the embodiments also provide methods for detecting the presence of a base sequence forming a G4 in a sample using a G4-binding ligand according to the embodiments. More specifically, the method of the present invention comprises the following steps: (i) measuring the fluorescence generated when the G4-binding ligand of the present invention is contacted with a sample; (n) Compare with the fluorescence in the absence of sample The present invention may be a method for detecting a base sequence that has formed G4 in a sample, which comprises the steps of:

[0127] Since the G4-binding ligand according to the embodiment is a "light-up ligand", the fluorescence in the absence of a sample in the above step (n) is hardly or not detected at all, and can be ignored in the above method, that is, the method can be such that it does not include step (n) and only performs step (i). Therefore, in this case, the detection method according to the embodiment can indicate the presence of a base sequence forming G4 in the sample if fluorescence occurs in the presence of a sample.

[0128] The detection method according to the embodiment can be a method for visually confirming an increase or change in fluorescence. In the detection method according to the embodiment, a "light-up ligand" is used as a G4-binding ligand, and therefore the generation of fluorescence indicates the presence of a base sequence forming G4 in a sample.

[0129] Alternatively, the detection method according to the embodiment may measure a fluorescence intensity at a specific wavelength. Also, the detection method according to the embodiment may measure a fluorescence spectrum in a specific wavelength range. The specific wavelength or specific wavelength range to be measured may be appropriately determined depending on the compound or a salt thereof used as the ligand according to the embodiment.

[0130] The G4-binding ligand of the embodiment has a fluorescence wavelength of 510 nm when excited at 406 nm, and therefore the Stokes shift exceeds 100 nm, making it easier to detect the presence of a base sequence that has formed G4.

[0131] For example, when it is intended to detect the presence of a base sequence forming G4 in a sample using the compound 6OTD-Py, it is preferable, but not limited to, to use an excitation wavelength of 406 nm and measure the fluorescence intensity at wavelengths in the range of 500 to 600 nm.

[0132] The concentration of the ligand used for detection can be adjusted appropriately depending on the concentration of the base sequence, such as the DNA sequence, that may be contained in the sample to be measured. The optimal ligand concentration for fluorescence measurement is not particularly limited, but may be about 1 / 70 to about 50 times the ratio of the base sequence. In addition, the concentration of the ligand in the fluorescence measurement sample is not particularly limited, but can be suitably measured if it is within the range of 0.1 μM to 0.5 μM, preferably 0.1 μM to 0.2 μM.

[0133] The measurement of fluorescence intensity in the detection method according to the embodiment is not particularly limited, but can be carried out at room temperature under pH conditions of, for example, pH 7.0 to 8.0, preferably pH 7.3 to 7.7. The salt concentration in the sample is not particularly limited as long as it is under conditions that allow the formation of G4, and is preferably adjusted to within a range of 10 to 2000 mM, preferably 20 to 150 mM.

[0134] The G4-binding ligand according to the embodiment does not show significant binding to base sequences such as DNA sequences that do not form G4, and therefore can specifically detect the presence of a base sequence that has formed G4. In addition, the fluorescence obtained by the G4-binding ligand according to the embodiment increases depending on the concentration of the base sequence that has formed G4, so it can also be used to quantify the base sequence that has formed G4 in a sample by comparing it with the fluorescence in the absence of the sample (base sequence). In this case, the fluorescence intensity depending on the concentration of the base sequence that has formed G4 can be measured in advance for each compound used as a ligand, and a calibration curve can be created.

[0135] As described above, in the detection method according to the embodiment, a "light-up ligand" is used as a G4-binding ligand, and the occurrence of fluorescence directly indicates the presence of a base sequence that has formed G4 in the sample, so that the concentration of a base sequence that has formed G4 can be quantified more quickly and accurately. Since the fluorescence obtained by the light-up ligand is generated only by binding to a base sequence that has formed G4 and increases depending on the concentration, it can be used to quantify the base sequence that has formed G4 in a sample without comparing it with the fluorescence in the absence of a sample (base sequence) (step (n)). EXAMPLES

[0136] Hereinafter, the embodiment of the present invention will be further described with reference to examples.

[0137] [Example 1: Construction of a prediction device] A G4 formation sequence prediction device according to the second embodiment was constructed. In this case, a prediction model stored in a storage unit of the G4 formation sequence prediction device was constructed by using the machine learning method according to the second embodiment.

[0138] When constructing the prediction model by using the machine learning method, the G4-forming sequences and some of the non-G4-forming sequences (SEQ ID NO: 1 to SEQ ID NO: 20) from which the training data used in the machine learning were obtained are as follows. The non-G4-forming sequences from which the training data used in the machine learning were obtained are the sequences verified not to form G4, and the 5'-N y0 G x1 N y1 G x2 N y2 G x3 N y3 G x4 N y4 -3'(N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G with two or more bases.) were selected.

[0139] (G4 forming sequence) SEQ ID NO: 1: GTGAGGCTAGGGGGTGGTGGTGGTGGGGGTGAGGAAGGGAAAGAT SEQ ID NO: 2: ATAGCGCCAGTGGGTGGGCTTAGATCTGGGAAGGGCGGGACAGAG SEQ ID NO: 3: GTCCAGGAAGGCTAGAGGTGGGGGGGCGGGTACCGGTGAAGGGAG SEQ ID NO: 4: TAAAGAGGAGGGTGGCGGTAGTGGAGGGGGGGGTTGGAGTTGGTT SEQ ID NO: 5: CTTGTGTTCGGGAAGGGGGGGGGGGGAGCGTTGGAAACGCATGC SEQ ID NO: 6: ACTTGGGTAGGCGGGCTTGGGAGGCTCCGTTGGACGTGGGGTCTA SEQ ID NO: 7: CACTCTTTTTGGGGTTGGGACTGAGGTGGCGGTCACGCGAGGATC SEQ ID NO: 8: AGTAGGGAGCTTTGGAATGAGGGATTAACACTTTGGGGGACTTAGTC SEQ ID NO: 9: AGGGGTAGGTTGGGTGGTAAGAGGTGGTAAGCGGAGCGGCTGCTG SEQ ID NO: 10: TGGGGAGGCTGGGTGGAGGACAAAGTAGAAATAAAAATACGGAAGTAGAT

[0140] (G4 non-forming sequence) SEQ ID NO: 11: CAAATACATACCGACTCTCCGGGGACCAAGGGGAA SEQ ID NO: 12: TTCCGCGCCTAGCCAAGGGGAGGAACGGGGCAGGA SEQ ID NO: 13: CACCCACAAGCTCTGGGATGGGGCTGGGAGTGGTC SEQ ID NO: 14: CTACAGCAAATGATGGAAATTAACAAGGCAGGAGG SEQ ID NO: 15: TTTCCACCTCAGGAGAAAGGGCGGGAGGGAAGGAG SEQ ID NO: 16: CTAGACTGCTGGAAGGGGGAGACCGGCGTCGACAG SEQ ID NO: 17: ACGCAAACGGGGAGGGCCGAGACGCGGATTCTCGA SEQ ID NO: 18: ACCTGAGAGGCACCCGGGGAGGAAGGAAAATGTGT SEQ ID NO: 19: AGTTCCGGAGGTAGGCGCGCACTTGAAGGAGAAG SEQ ID NO: 20: CCAGAGTCAGGCTGCGGGCCCACGGGGGTGCAGGC

[0141] Furthermore, when constructing the prediction model by using the machine learning method, the optimal value of the RBF kernel hyperparameter γ and the optimal value of the cost parameter C were examined using a grid search method, and the optimal value of γ was γ=0.01 and the optimal value of C was C=1000, with the accuracy rate at γ=0.01 and C=1000 being 96.45% (the highest value).

[0142] [Evaluation of prediction accuracy of the prediction device of Example 1] The prediction accuracy of the G4-forming sequence prediction device (hereinafter sometimes abbreviated as "prediction device") constructed in Example 1 was evaluated. FIG. 4 is a diagram that shows an outline of a process flow for predicting the presence or absence of G4 formation in a DNA sequence (whether or not the DNA sequence is a G4-forming sequence) when evaluating the accuracy of the prediction device constructed in Example 1. FIG. 5 is a table showing the results of the prediction results of the prediction device constructed in Example 1 determined by a verification experiment. A specific explanation will be given below with reference to these figures.

[0143] [Prediction of G4-forming and non-G4-forming sequences] In the evaluation of the prediction accuracy of the prediction device constructed in Example 1, first, 5'-N sequences that are expected to form G4 were selected from the genome of SARS CoV2. y0 G x1 N y1 G x2 N y2 Gx3 N y3 G x4 N y4 -3' DNA sequence (in the sequence, N y0 ~N y4 represents any sequence fragment with 0 or more bases, and G x1 ~G x4 represents a sequence fragment containing only G with two or more bases.) A search was conducted for DNA sequences of 15 to 40 bases as prediction targets. As a result of the search, it was found that the genome of SARS CoV2 contains a "G x1 N y1 G x2 N y2 G x3 N y3 G x4 " was present in 132 locations. In contrast, there were 5,726 types of predicted DNA sequences.

[0144] Next, the prediction device of Example 1 was used to predict the presence or absence of G4 formation (whether it is a G4-forming sequence or not) for 5726 types of DNA sequences to be predicted.

[0145] In this case, when predicting the presence or absence of G4 formation in each DNA sequence (whether each DNA sequence is a G4 forming sequence or not), first, sequence data of the DNA sequence was input via the input unit of the prediction device. Next, a feature vector x indicating the characteristics of the DNA sequence was determined from the sequence data by a determination unit of a processing unit of the prediction device. Next, the G4 formation prediction unit of the processing unit of the prediction device predicted the presence or absence of G4 formation in the DNA sequence (whether the DNA sequence is a G4 forming sequence or not) from the feature vector x by using a prediction model stored in the storage unit. Specifically, a value calculated by inputting the feature vector x into the discrimination function f(x) expressed by the above formulas (7-1) to (7-3) was obtained as a predicted value of the presence or absence of G4 formation.

[0146] In this manner, the presence or absence of G4 formation (whether it is a G4-forming sequence or not) was predicted for 5726 types of DNA sequences to be predicted, and a predicted value for the presence or absence of G4 formation was obtained for each DNA sequence. As described above, according to the predicted value for the presence or absence of G4 formation, with 0.5 as the boundary, the closer the predicted value is to 1, the more likely the DNA sequence is to be a G4-forming sequence, and the closer the predicted value is to 0, the more likely the DNA sequence is to be a non-G4-forming sequence.

[0147] Next, the predicted DNA sequence was divided into two parts, the Tether sequences at both ends of which were removed, called "G x1 N y1 G x2 N y2 G x3 N y3 G x4 In order to be able to verify both the sequence of the portion predicted to be a G4-forming sequence from only the portion represented by ", and the DNA sequence predicted to be a G4-forming sequence from the DNA sequence to be predicted, G4-forming sequences and non-G4-forming sequences were predicted from 5,726 types of DNA sequences to be predicted, and candidates for G4-forming sequences and candidates for non-G4-forming sequences were extracted.

[0148] Specifically, the "G" in the 5726 types of DNA sequences to be predicted x1 N y1 G x2 N y2 G x3 N y3 G x4 From each of the 132 groups that share a common portion represented by ", we selected the DNA sequence that predicted the presence or absence of G4 formation to be closest to 1. We then predicted that the 132 selected DNA sequences were G4-forming sequences and extracted 132 types of candidates for G4-forming sequences.

[0149] Furthermore, the G x1 N y1 G x2 N y2 G x3 N y3 G x4From the 132 groups that share a common portion represented by "", at least one group was selected in which the predicted value of the presence or absence of G4 formation for all of the DNA sequences in the group was less than 0.5. Then, all of the DNA sequences in at least one of the selected groups were predicted to be non-G4 forming sequences and extracted as candidates for non-G4 forming sequences.

[0150] [Verification experiment] Thirteen types of candidates for verification of G4-forming sequences (verification candidates 1 to 13 of G4-forming sequences) were selected from the above 132 types of candidates for G4-forming sequences, and whether the selected candidates for verification were G4-forming sequences was verified by experiments. In addition, nine types of candidates for verification of non-G4-forming sequences (verification candidates 1 to 9 of non-G4-forming sequences) were selected from the above non-G4-forming sequence candidates, and whether the selected candidates for verification were G4-forming sequences was verified by experiments.

[0151] Specifically, first, candidates for verification of 13 types of G4-forming sequences (verification candidates 1 to 13 for G4-forming sequences) were prepared as 13 types of oligonucleotides, and candidates for verification of 9 types of non-G4-forming sequences (verification candidates 1 to 9 for non-G4-forming sequences) were prepared as 9 types of oligonucleotides.

[0152] Next, the fluorescence intensity of samples with various concentrations was measured for each of the 22 types of oligonucleotides, including 13 candidates for validation of G4-forming sequences and 9 candidates for validation of non-G4-forming sequences. In this case, 6OTD-Py (0.2 μM) was dissolved in lithium cacodylate (10 mM, pH 7.4, 100 mM KCl, 0.01% Tween-20), and then oligonucleotides were added at various concentrations (0 to 5 μM) and incubated for 24 hours. The fluorescence intensity was then measured for the samples with various concentrations obtained in this manner. At this time, the fluorescence intensity was measured with a microplate reader (TECAN, Spark) using a 96 or 384 well plate (Perkin Elmer, 6002290; Greiner, 781900). At this time, 40 μL of samples with various concentrations were placed in each well of a 96 or 384 well plate, and the fluorescence intensity was measured at an excitation wavelength of 406 nm and an emission wavelength of 510 nm.

[0153] Next, for each of the 22 types of oligonucleotides, including 13 types of candidates for verification of G4-forming sequences (candidates 1 to 13 for verification of G4-forming sequences) and 9 types of candidates for verification of non-G4-forming sequences (candidates 1 to 9 for verification of non-G4-forming sequences), the fluorescence intensity was measured in samples of various concentrations, and the logarithm of the dissociation constant of 6OTD-Py, log(K d [M]) was calculated. Then, the logarithm of the dissociation constant, log(K d If [M]) < -7, the corresponding candidate is determined to be a G4-forming sequence, and the logarithm of the dissociation constant log(K d Candidates corresponding to the case where [M]) ≧−7 were determined to be G4 non-forming sequences.

[0154]

number

[0155] F: Fluorescence intensity m0: [OLG] (oligonucleotide concentration [μM]) m1: [6OTD-Py] (6OTD-Py concentration [μM]) m2:K d (Dissociation constant [μM]) A: Maximum fluorescence intensity of the complex of 6OTD-Py and G4

[0156] As for the G4-forming sequences, the logarithm of the dissociation constant log(K) was calculated for the following known sequences: telo24 (SEQ ID NO: 21), bcl2 (SEQ ID NO: 22), and k-ras (SEQ ID NO: 23) in the same manner as above. d [M]) is required, and both are log(K d The logarithm of the dissociation constant log(K d [M]) is required, and both are log(K d [M] was ≥-7.

[0157] telo24 (SEQ ID NO: 21): TTAGGGTTAGGGTTAGGGTTAGGG bcl2 (SEQ ID NO: 22): GGGCGCGGGGAGGAATTGGGCGGG k-ras (SEQ ID NO: 23): AGGGCGGTGTGGGAAGAGGGAAGAGGGGGAGG

[0158] Triplex-2 (SEQ ID NO: 24): CTGGGAGGGAGGGA Triplex-3 (SEQ ID NO: 25): GGTTGGTGTGG dsDNA (SEQ ID NO: 26): TATAGCTATATTTTTTTATAGCTATA ssDNA (SEQ ID NO: 27): TTAGAGTTAGAGTTAGAGTTAGAG

[0159] [Evaluation of prediction accuracy] As a result of the judgment in the validation experiment, 11 of the 13 validation candidates for G4-forming sequences (validation candidates 1 to 13 for G4-forming sequences) were judged to be G4-forming sequences as expected, and 8 of the 9 validation candidates for non-G4-forming sequences (validation candidates 1 to 9 for non-G4-forming sequences) were judged to be non-G4 sequences as expected, as shown in Figure 5. It was confirmed that the prediction of these 22 candidates in total was possible with an AUC (Aria under ROC curve) of 0.90 and an accuracy rate of 86.36%.

[0160] [Example 2: Construction of a prediction device] A G4 formation sequence prediction device according to the first embodiment was constructed. In this case, a prediction model stored in a storage unit of the G4 formation sequence prediction device was constructed by using the machine learning method according to the first embodiment.

[0161] When performing machine learning using training data to construct the prediction model by using the machine learning method, the dissociation constant K of the compound 6OTD-Py of the G4-forming sequence in the training data for various G4-forming sequences is d The value of was set to the value [M] actually obtained by a general method for the oligonucleotide of the G4-forming sequence. In addition, the dissociation constant K of the compound 6OTD-Py of the G4-non-forming sequence in the training data for various G4-non-forming sequences was set to 0.01. d The value of the dissociation constant K d Considering the possibility that the value of may not be accurately obtained, the dissociation constant K of the compound 6OTD-Py of the G4 non-forming sequence in the training data for all types of G4 non-forming sequences for the training data was d The value of 1×10 -5.2 A constant value greater than M (1×10 -4 M) was set.

[0162] Furthermore, when constructing the prediction model by using the machine learning method, the optimal value of the error dead zone parameter ε, the optimal value of the RBF kernel hyperparameter γ, and the optimal value of the cost parameter C were examined using a grid search method, and the optimal value of ε was ε = 0.1, the optimal value of γ was γ = 0.01, and the optimal value of C was C = 100. The dissociation constant K of the compound 6OTD-Py of the DNA sequence calculated when the values ​​of ε, γ, and C are set to the optimal values ​​of ε, γ, and C, respectively, was d The predicted value [M] and the actual value (dissociation constant K of the compound 6OTD-Py of the DNA sequence in the training data) d The correlation coefficient calculated using the value of [M]) was 0.7826 (the highest value). d Predicted value x 10 6 Logarithm of [μM] (log(K d Predicted value x 10 6 [μM]) and K d The logarithm of the actual measured value (log(K d The correlation coefficient between the ε, γ, and C was calculated. In addition, when the values ​​of ε, γ, and C were set to the optimal values ​​of ε, γ, and C, respectively, the correlation coefficient between the ε, γ, and C was calculated. d The logarithm of the predicted value of (log(K d The predicted value of [M]) and K d The logarithm of the actual measured value (log(K d The RMSE calculated from [M]) was 2.5073 (the highest value).

[0163] [Evaluation of prediction accuracy of the prediction device of Example 2] The prediction accuracy of the G4 formation sequence prediction device (hereinafter sometimes abbreviated as "prediction device") constructed in Example 2 was evaluated.

[0164] [Prediction of G4-forming and non-G4-forming sequences] In evaluating the prediction accuracy of the prediction device constructed in Example 2, the prediction device of Example 2 was used to predict the stability of G4 in each of 22 types of DNA sequences (base sequences) consisting of validation candidates 1 to 13 for G4-forming sequences and validation candidates 1 to 9 for G4 non-forming sequences, which were verified in the [Verification experiment] in [Evaluation of prediction accuracy of the prediction device of Example 1] above.

[0165] In this case, first, sequence data of the DNA sequence was input via the input unit of the prediction device. Next, a feature vector x indicating the characteristics of the DNA sequence was determined from the sequence data by a determination unit of the processing unit of the prediction device. Next, the G4 stability prediction unit of the processing unit of the prediction device predicted the stability of G4 in the DNA sequence from the feature vector x by using a prediction model. Specifically, the value calculated by inputting the feature vector x into the regression equation f(x) of the prediction model represented by the above formulas (3-1) and (3-2) was used as the dissociation constant K of the compound 6OTD-Py of the DNA sequence. d The predicted value [M] was used. Next, the output section of the prediction device outputs the dissociation constant K d The predicted value was output.

[0166] In this way, the stability of G4 in each of the 22 types of DNA sequences to be predicted was predicted, and the dissociation constant K d The predicted value [M] of the dissociation constant K d According to the predicted value of, the lower the predicted value, the higher the possibility of G4 formation in the DNA sequence and the higher the stability of the G4 formed in the DNA sequence is predicted. Also, the higher the predicted value, the lower the possibility of G4 formation in the DNA sequence and the lower the stability of the G4 formed in the DNA sequence is predicted.

[0167] [Evaluation of prediction accuracy] For the 22 types of DNA sequences to be predicted, the dissociation constant K d Predicted value x 10 6 Logarithm of [μM] log(K dPredicted value x 10 6 [μM]) was calculated. FIG. 6 shows the dissociation constant K d 7(a) is a table showing the dissociation constant K of teacher data obtained when predicting the stability of G4 in a DNA sequence using the prediction device of Example 2, together with the predicted value of the presence or absence of G4 formation obtained when predicting the stability of G4 in a DNA sequence using the prediction device of Example 2. d The logarithm of the value of log(K d FIG. 7(b) is a graph showing the predicted stability of G4 against [M], and FIG. 7(b) is a graph showing the dissociation constant K of the training data when the presence or absence of G4 formation in a DNA sequence is predicted using the prediction device of Example 1. d The logarithm of the value of log(K d 7(a) and 7(b) are graphs showing predicted values ​​of the presence or absence of G4 formation for each log(K d [M]) values ​​greater than -6 were indicated as -6.

[0168] As shown in FIG. 6 and FIG. 7(a), when the stability of G4 in each DNA sequence was predicted using the prediction device of Example 2, log(K d Predicted value x 10 6 The predicted stability value of G4, which is the dissociation constant K d The logarithm of the value of log(K d [M]) increases, and the predicted stability value of G4 and log(K d [M]) and, in particular, log(K d In the range of [M] < -7, the predicted stability of G4 (log(K d Predicted value x 10 6 [μM]) and log(K d The correlation coefficient between log(K d In the range of [M]<-7, K d The logarithm of the predicted value of (log(K d The predicted value of [M]) and Kd The logarithm of the measured value (log(K d The RMSE calculated from the log(K [M]) was 2.5073. Therefore, it was confirmed that the prediction device of Example 2 can predict the stability of G4 in DNA sequences with high accuracy. On the other hand, as shown in Fig. 6 and Fig. 7(b), when the prediction device of Example 1 is used to predict the presence or absence of G4 formation in each DNA sequence, the predicted value of the presence or absence of G4 formation and log(K d There was no correlation between [M].

[0169] [Verification experiment of G4 stability] The stability of G4 was experimentally verified for the following six types of G4-forming sequences among the 83 types of G4-forming sequences from which training data was obtained for constructing the prediction model in Example 2. The following six types of G4-forming sequences have the same dissociation constant K d Low K value d Three types of value groups (NCOR2(log(K d [M])=-9.7), telo24(log(K d [M])=-7.1), and BIN1(log(K d [M]) = -7.5)), and the dissociation constant K of compound 6OTD-Py d High value of high K d Three types of value groups (OTP(log(K d [M])=-6.0), WT1(log(K d [M])=-6.1), ZN318(log(K d [M])=-5.9)).

[0170] NCOR2 (SEQ ID NO: 28): ACTGGCATGGTGGGGTGCAAGGTTCTTTGCTGGGGTGCGGGCGAT telo24 (SEQ ID NO: 29): TTAGGGTTAGGGTTAGGGTTAGGG BIN1 (SEQ ID NO: 30): AGGATGGCAGAGATAGGGGAGCAAGGGGGTGACGGCGGGGAAGATC OTP (SEQ ID NO: 31): AAGGAGGGAAGGGGAGGATTAGAGCACGGAGTAGGTATTGTGCTC WT1 (SEQ ID NO: 32): AGTAGGGAGCTTTGGAATGAGGGATTAACACTTTGGGGGACTTAGTC ZN318 (SEQ ID NO: 33): AGGAGAGGGGTGGTCTCTCTGCATCACGTGGTCTCCGGGGGAGTA

[0171] When the stability of G4 in the above six types of G4-forming sequences was verified by experiments, specifically, the UV absorption spectrum was measured for each of the oligonucleotides of the above six types of G4-forming sequences. In this case, each oligonucleotide (10 μM) was added to a solution in which KCl (100 mM) was added to a lithium cacodylate buffer solution (10 mM, pH 7.4). Next, the UV absorption spectrum (280 nm to 320 nm) of the sample solution obtained was measured at 25°C and 95°C using a UV-visible spectrophotometer (V-630) manufactured by JASCO Corporation, and the UV absorption difference spectrum under these two temperature conditions was obtained. Furthermore, the high K d For the sample solutions to which the three types of the value group (OTP, WT1, and ZN318) were added, the compound L2H2-6OTD (20 μM), which is a known G4 binding ligand, was further added, and the sample solutions after addition were subjected to similar measurements to obtain UV absorption difference spectra. Figure 8 is a graph showing the UV absorption difference spectra of six types of G4-forming sequences (NCOR2, telo24, BIN1, OTP, WT1, and ZN318) before and after the addition of the compound L2H2-6OTD, respectively, as solid lines. d In the graphs for the three types of value groups (OTP, WT1, and ZN318), the UV absorption difference spectra after addition of compound L2H2-6OTD are represented by dashed lines.

[0172] As shown in Figure 8, the low K dThe UV absorption difference spectra (solid lines) of three types of the value group (NCOR2, telo24, and BIN1) before and after the addition of L2H2-6OTD showed UV absorption at 295 nm, indicating the formation of G4. d The UV absorption difference spectra (solid lines) of the three types of the high-K group (OTP, WT1, and ZN318) before the addition of L2H2-6OTD showed no UV absorption at 295 nm. d In the UV absorption difference spectra (dashed lines) of the three types of low-K groups (OTP, WT1, and ZN318) after addition of L2H2-6OTD, UV absorption at 295 nm was observed, indicating the formation of G4. d The three oligonucleotides in the high K group spontaneously formed G4s, whereas the high K d It was found that the three oligonucleotides in the value group did not spontaneously form G4s, but only formed G4s when a G4-binding ligand was added. Therefore, in the verification experiment, the dissociation constant K d It was confirmed that the lower the value, the higher the stability of G4.

[0173] [Example 3: Synthesis of 6OTD-Py] A new compound, 6OTD-Py, was synthesized according to the following method.

[0174] 1. Synthesis of Compound 16 Compound 16 (macrocycle 16) described in JP 2020-176108 A was synthesized according to the method described in the section "Example 1: Synthesis of Compound 1" in JP 2020-176108 A.

[0175] 2. Synthesis of compound New-1 [ka]

[0176] Under an argon atmosphere, 1-ethenylpyrene (12.8 mg, 55.6 mol) and Grubbs second generation catalyst (1.50 mg, 1.70 mol) were added to a solution of compound 16 (5.00 mg, 5.56 mol) in dry THF (tetrahydrofuran) (500 L) at room temperature, and the reaction mixture was heated to 90°C. The reaction mixture was then stirred for 29 hours and concentrated under reduced pressure. The residue was purified by silica gel column chromatography (chloroform / methanol = 200:1) to obtain compound New-1 (4.20 mg, 69%).

[0177] TIFF0007650525000021.tif43170

[0178] 3. Synthesis of 6OTD-Py [ka]

[0179] To a solution of compound New-1 (4.20 mg, 3.82 mol) in dichloromethane (2 mL), trifluoroacetic acid (1 mL) was added at room temperature and stirred for 10 minutes. The reaction mixture was concentrated in vacuum to give 6OTD-Py (3.00 mg, 87%).

[0180] TIFF0007650525000023.tif43170

[0181] Although the embodiments of the present invention have been described in detail above, the present invention is not limited to the above-described embodiments, and various design modifications can be made without departing from the spirit of the present invention as described in the claims. [Industrial Applicability]

[0182] The present invention can be used in drug discovery and development research for antiviral drugs, antibacterial drugs, antifungal drugs, etc. against emerging and re-emerging infectious diseases, as well as in drug discovery and development research for human genetic diseases (cancer, neurodegenerative diseases, etc.). [Explanation of symbols]

[0183] 10 G4 formation sequence prediction device 2 Input section 4 Storage section 6 Processing section 6a Decision section 6b G4 stability prediction section 8 Output section 20 G4 formation sequence prediction device 12 Input section 14 Storage section 16 Processing section 16a Decision section 16b G4 formation prediction part 18 Output section 100 Computers 110 Processing device (computer) 120 Storage device 130 Input Device 140 Output Device 150 Input / Output Interface (I / F)

Claims

1. 5'-N ｙ０ G ｘ１ N ｙ１ G ｘ２ N ｙ２ G ｘ３ N ｙ３ G ｘ４ N ｙ４ -3' (in the base sequence, N ｙ０ ~N ｙ４ represents any sequence fragment having 0 or more bases, and G ｘ１ ~G ｘ４ represents a sequence fragment containing only G having two or more bases; and a prediction unit that predicts a G-quadruplex formation state in the base sequence from the feature vector by using a prediction model; The present invention relates to a G-quadruplex formation sequence prediction device, and is characterized in that the prediction model is a prediction model constructed by machine learning using training data that is a pair of a feature vector indicating the characteristics of the base sequence and G4 formation state information indicating the formation state of the G-quadruplex in the base sequence.

2. the prediction unit is a G4 stability prediction unit that predicts the stability of the G-quadruplex in the base sequence as the formation state of the G-quadruplex in the base sequence; The device for predicting G-quadruplex forming sequences according to claim 1 , wherein the G4 formation state information is a value of a dissociation constant of a G-quadruplex binding ligand of the base sequence.

3. the prediction unit is a G4 formation prediction unit that predicts whether or not the G-quadruplex is formed in the base sequence as the formation state of the G-quadruplex in the base sequence; The G4 formation state information is a classification label for classifying the presence or absence of the formation of the G-quadruplex in the base sequence.

4. 5'-N ｙ０ G ｘ１ N ｙ１ G ｘ２ N ｙ２ G ｘ３ N ｙ３ G ｘ４ N ｙ４ -3' (in the base sequence, N ｙ０ ~N ｙ４ represents any sequence fragment having 0 or more bases, and G ｘ１ ~G ｘ４ represents a sequence fragment containing only G with two or more bases; and a training data acquisition step of acquiring a set of a feature vector showing the characteristics of the base sequence and G4 formation state information showing the formation state of a G-quadruplex in the base sequence as training data; and a machine learning process for constructing a predictive model for predicting the formation state of the G-quadruplex in the base sequence by performing machine learning using the training data.

5. the G4 formation state information is a value of a dissociation constant of a G-quadruplex binding ligand of the base sequence, The machine learning method according to claim 4, characterized in that the predictive model is a predictive model for predicting the stability of the G-quadruplex in the base sequence as the formation state of the G-quadruplex in the base sequence.

6. the G4 formation state information is a classification label for classifying the presence or absence of the formation of the G-quadruplex in the base sequence, The machine learning method according to claim 4, characterized in that the predictive model is a predictive model for predicting whether or not the G-quadruplex is formed in the base sequence as the formation state of the G-quadruplex in the base sequence.

Images