Prediction device and prediction method

Abstract

Understand the effects of manipulations on biological components. [Solution] The prediction device includes a first acquisition unit that acquires a first dataset including a first set of a first sequence of a first gene involved in a trait and its phenotype, a second acquisition unit that acquires a second sequence, a first setting unit that sets an editing cost for a reference sequence based on the change between each phenotype in the first set and the reference phenotype in the reference sequence, a second setting unit that updates the editing cost, a first generation unit that generates a first list of first sequences included in the second sequence, a second generation unit that generates a second list of first sequences including the second sequence, a first identification unit that identifies a third sequence in the first list that has the smallest effect of genome editing, a second identification unit that identifies a fourth sequence in the second list that has the largest effect of genome editing, and a prediction unit that predicts a new phenotype under the second sequence based on the proportional relationship between the editing distance between different sequences and the change in each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost.

Description

[Technical Field]

[0001] The present invention relates to a prediction device and a prediction method. [Background technology]

[0002] There are known techniques for determining the effects of manipulations on biological components. For example, Non-Patent Document 1 discloses a technique for determining the classification to which a protein belongs based on the amino acid sequence contained in the protein using a trained model. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Leslie et al. 2004, Fast String Kernels using Inexact Matching for Protein Sequences, Journal of Machine Learning Research, vol.5, 1435-1455 Summary of the Invention [Problem to be solved by the invention]

[0004] In the technology described in Non-Patent Document 1 (hereinafter also referred to as the prior art), in order to grasp the above-mentioned classification with high accuracy, it is necessary to prepare a large amount of training data for learning a model. Here, for example, when grasping the effects of genome editing on an organism using the prior art, it is necessary to generate training data after growing each organism whose genome has been edited. Therefore, in such cases, there is a problem in that it is difficult to efficiently grasp the effects on the organism.

[0005] An object of one aspect of the present invention is to more efficiently understand the effects of manipulations applied to biological components. [Means for solving the problem]

[0006] In order to solve the above problem, a prediction device according to one aspect of the present invention includes a first acquisition unit that acquires a first dataset including a plurality of first pairs, each of which is a pair of a first sequence in an expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence; a second acquisition unit that acquires a second sequence in the expression regulatory region; a first setting unit that sets an editing cost for a reference sequence that is a predetermined first sequence, based on a change between each phenotype included in the first pair and a reference phenotype under the reference sequence; a second setting unit that updates the editing cost based on a first threshold for the editing cost; and a list of the first sequences included in the second sequence. The system includes a first generation unit that generates a first list, a second generation unit that generates a second list that is a list of the first sequences including the second sequence, a first identification unit that identifies a third sequence among the sequences included in the first list that will have the smallest effect of genome editing, a second identification unit that identifies a fourth sequence among the sequences included in the second list that will have the largest effect of genome editing, and a prediction unit that predicts a new phenotype under the second sequence based on the proportional relationship between the edit distance between two different sequences and the change between two phenotypes under each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost.

[0007] In order to solve the above problem, a prediction method according to one aspect of the present invention includes a first acquisition process for acquiring a first dataset including a plurality of first pairs, each pair being a pair of a first sequence in an expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence; a second acquisition process for acquiring a second sequence in the expression regulatory region; a first setting process for setting an editing cost for a reference sequence that is a predetermined first sequence, based on a change between each phenotype included in the first pair and a reference phenotype under the reference sequence; a second setting process for updating the editing cost based on a first threshold for the editing cost; and a list of the first sequences included in the second sequence. The method includes a first generation process that generates a first list; a second generation process that generates a second list, which is a list of the first sequences including the second sequence; a first identification process that identifies a third sequence among the sequences included in the first list that has the smallest effect of genome editing; a second identification process that identifies a fourth sequence among the sequences included in the second list that has the largest effect of genome editing; and a prediction process that predicts a new phenotype under the second sequence based on the proportional relationship between the edit distance between two different sequences and the change between two phenotypes under each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost. [Effects of the Invention]

[0008] According to one aspect of the present invention, the influence of an operation applied to a biological component can be grasped more efficiently. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is a block diagram showing an example of the configuration of a prediction device according to a first embodiment of the present invention. [Figure 2] FIG. 1 is a schematic diagram showing an example of an edited site by genome editing. [Figure 3] This is a precondition for an example of an algorithm for rewriting the initial value of the editing cost. [Figure 4] 10 is a process flow of an example of an algorithm for rewriting the initial value of the editing cost. [Figure 5] 10 is a flowchart illustrating an example of a flow of processing executed by the prediction device. DETAILED DESCRIPTION OF THE INVENTION

[0010] [Embodiment 1] An embodiment of the present invention will be described below.

[0011] (Outline of Prediction Device 1) As an example, prediction device 1 is a device that predicts how much the phenotype under a mutant sequence whose phenotype is to be predicted will change from the phenotype under a reference sequence, depending on how far the mutant sequence is from a reference sequence (a sequence that has an inclusive relationship at the site where the mutation is introduced by genome editing or a wild-type sequence) in terms of editing distance.

[0012] Both the reference sequence and the mutant sequence are sequences in the expression regulatory region of a gene involved in a trait of interest. Here, the expression regulatory region according to this embodiment may be a silencer or an enhancer.

[0013] The edit distance according to this embodiment may be any distance between two sequences of equal length. Examples of the edit distance according to this embodiment include, but are not limited to, the Levenshtein distance, the Damerau-Levenshtein distance, and the extended Hamming distance. In the following description, the edit distance will be described as the Levenshtein distance.

[0014] (Configuration of prediction device 1) The configuration of the prediction device 1 will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example configuration of the prediction device 1. The prediction device 1 includes a control unit 10 and a storage unit 11. In one example, the control unit 10 may be a CPU (Central Processing Unit). The control unit 10 reads a control program, which is software stored in the storage unit 11, and expands it into a memory such as a RAM (Random Access Memory) to execute various functions.

[0015] The storage unit 11 can be configured by a non-volatile storage device such as a hard disk or a flash memory.

[0016] The control unit 10 includes a first acquisition unit 101, a second acquisition unit 102, a first setting unit 103, a second setting unit 104, a first generation unit 105, a second generation unit 106, a first identification unit 107, a second identification unit 108, a prediction unit 109, a first calculation unit 110, a second calculation unit 111, a third calculation unit 112, a fourth calculation unit 113, a fifth calculation unit 114, and an update unit 115.

[0017] (Basic principle) As an example, the prediction device 1 operates based on the following basic principles P1 and P2: In the following, as an example, a case where the expression regulatory region is a silencer will be considered.

number

number

[0018] However, P1 takes precedence over P2.

[0019] If it is an enhancer, the inequality signs on the right-hand side of P1 and P2 are reversed. To determine whether it is an enhancer or silencer, at least one genome edit must be performed on the expression regulatory region and the transcription level (or phenotype) must be measured. We will proceed assuming this information is known.

[0020] Furthermore, in the first to fourth heuristics described below, inference rules that are invoked only when predetermined conditions are met are added.

[0021] The first acquisition unit 101 acquires a first dataset including a plurality of first pairs, each pair being a first sequence in an expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence.

[0022] The first dataset includes a plurality of first pairs, each pair consisting of a first sequence in an expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence. Here, the first sequence is a sequence for which the corresponding phenotype is known. The first sequence may be the sequence of the expression regulatory region before editing. The reference sequence described above may also be a predetermined first sequence. For example, the first dataset may be a collection of pairs of several sequences (such as a wild-type sequence and several mutant sequences obtained by genome editing) of the expression regulatory region of a gene involved in a trait of interest and the phenotypes under those sequences.

[0023] Specific examples of the first dataset include wild-type and mutant sequences of the expression regulatory regions of the rice TAWAWA1 (hereinafter referred to as TAW1) gene and the OsGA20ox2 / SEMIDWARF1 (hereinafter referred to as SD1) gene, and phenotypic data associated with these sequences. TAW1 and SD1 are genes that control the number of grains and plant height in rice, respectively. Specific examples of the first dataset include those based on the rice TAW1 gene and SD1 gene, but the first dataset may also be based on other genes, including those from varieties other than rice. Specific examples of phenotypes include the number of secondary rachis-branches per number of primary rachis-branches and plant height.

[0024] Table 1 shows examples of phenotypic measurements for wild-type and mutant sequences of TAW1. [Table 1] Table 1 shows, as an example, the measurement results for the wild-type sequence and 26 mutant sequences of TAW1. The measurements listed in Table 1 are the number of secondary rachis branches per number of primary rachis branches corresponding to each sequence ID included in the table.

[0025] The sequence IDs shown in Table 1 above and Tables 3 and 4 below are all IDs assigned to each sequence for the purpose of explanation in this specification. The correspondence between each sequence ID described in this specification and the actual sequence is as shown in the attached sequence listing.

[0026] Figure 2 is a schematic diagram showing an example of edited sites by genome editing. Figure 2 shows the edited sites in the wild-type sequence and 26 mutant sequences of the TAW1 expression regulatory region corresponding to Table 1.

[0027] Each rectangle in Figure 2 indicates the position where a deletion or insertion was introduced, and each rectangle schematically shows the position of introduction in the genome.

[0028] In Figure 2, the strings of characters shown near the edited regions correspond to the IDs of the mutant sequences obtained by the edits, and the IDs of each sequence shown in Figure 2 correspond to the sequence IDs shown in the first column of Table 1.

[0029] In Figure 2, the character strings representing the IDs of each sequence have been omitted from the "TAW1 / " in the sequence IDs in Table 1. Also, the "Δ" at the beginning of the character strings representing the sequence IDs in Figure 2 corresponds to the "d" in the sequence IDs in Table 1.

[0030] In Figure 2, the entire TAW1 silencer represents the wild-type sequence, and the values ​​shown near the edited region represent the phenotypes under the mutant sequence obtained by editing, corresponding to the phenotypic measurements shown in the second column of Table 1.

[0031] For example, "+TT" in Figure 2 indicates that the residue TT has been inserted at the corresponding position in the TAW1 silencer by genome editing. Also, for example, "ΔAGTA" in Figure 2 indicates that the residue AGTA has been deleted at the corresponding position in the TAW1 silencer by genome editing.

[0032] In addition, the inverted triangle in Figure 2 indicates the insertion site when one residue is inserted into the wild-type sequence of the TAW1 silencer. For example, "+A" in Figure 2 indicates that residue A has been inserted at that site. For example, "+T" in Figure 2 indicates that residue T has been inserted at that site.

[0033] For example, the first data set acquired by the first acquiring unit 101 may include the measurement result data shown in Table 1 as phenotype data, or may include measurement result data for other mutant sequences as phenotype data. Below, as an example, a case will be described in which the first data set acquired by the first acquiring unit 101 includes the measurement result data shown in Table 1 as phenotype data.

[0034] In this embodiment, the first data set is also referred to as a training data set.

[0035] The second acquisition unit 102 acquires a second sequence s in the expression regulatory region. The second sequence s is a sequence for which a phenotype is to be predicted. The second sequence s may be a sequence after genome editing.

[0036] The first setting unit 103 sets an editing cost for a reference sequence s0, which is a predetermined first sequence, based on the amount of change between each phenotype included in the first set and a reference phenotype π(s0) under the reference sequence s0. The second setting unit 104 sets a first threshold c for the editing cost. E The editing cost is updated based on the first heuristic. The first setting unit 103 and the second setting unit 104 will be described later in the section on first heuristics.

[0037] When a second sequence s of an expression regulatory region of a gene involved in a trait of interest is input under P1 and P2, the prediction device 1 predicts π(s) in the following procedure.

[0038] Step 1. Create a list L1 of the sequences included in s. In other words, the first generation unit 105 generates a first list L1, which is a list of the first sequences included in the second sequence. Note that L1 may be empty.

[0039] Step 2: Create a list L2 of sequences that includes s. In other words, the second generation unit 106 generates a second list L2 that is a list of first sequences that includes the second sequence.

[0040] Step 3. Sequence with minimal effect of genome editing in L1

number

[0041] Step 4. Sequence with the greatest effect of genome editing in L2

number

[0042] The prediction unit 109 predicts a new phenotype π(s) under the second sequence s based on the proportional relationship between the edit distance between two different sequences and the change between the two phenotypes under each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost.

[0043] According to the above configuration, the influence of a manipulation, such as genome editing, on a biological component on a phenotype can be grasped by using the proportional relationship between the edit distance between two different sequences and the change between two phenotypes under each of the two sequences. In other words, the influence of a manipulation on a biological component can be grasped more efficiently.

[0044] Also, as mentioned above, the edit distance may be, for example, the Levenshtein distance, the Damerau-Levenshtein distance, or the extended Hamming distance.

[0045] According to the above configuration, the type of edit distance to be used can be selected depending on the sequence to be predicted.

[0046] Editing costs will be discussed later in the first heuristic.

[0047] In the following steps, a description will be given of each step in which the prediction unit 109 predicts a new phenotype based on the proportional relationship described above. The editing cost will be described later in the section on heuristics 1.

[0048] Step 5. s1 (min) If exists,

number

[0049] Step 6. s1 (min) does not exist, and s2 (max) s2 where ≠s0 (max) If exists,

number

[0050] Step 7. s1 (min) does not exist, and s * =argmin t d L For (s,t) s * If ≠s0, then

number

[0051] Step 8. s1 (min) does not exist, and s * =argmin t d L For (s,t) s * If =s0,

number

[0052] According to the above configuration, when predicting the effect of genome editing on a phenotype, a prediction process can be selected depending on the sequence to be used.

[0053] (First heuristic: Edit-effective base conservation) As shown in Figure 2, even when mutations of the same length are introduced by genome editing, the amount of change in phenotype will differ depending on which part of the expression regulatory region is edited. In the Levenshtein distance, an amount called the editing cost is defined for each character in the two strings to be edited. The prediction device 1 uses the editing cost in the Levenshtein distance to address the above problem as follows. Since the edit distance in this prediction problem is unknown and unobservable, an initial value for the editing cost is initially assigned appropriately, or the base conservation calculated by defining a range of related species is used as the initial value. A specific example of an appropriate initial value for the editing cost is 0.5 for all bases, but other values ​​may also be used.

[0054] Table 2 shows an example of the degree of base conservation of the 113-base expression regulatory region of TAW1, calculated using monocotyledons as related species. [Table 2] Next, given a training dataset, an initial value of the editing cost is set according to a predetermined algorithm.

[0055] Figures 3 and 4 are examples of a predetermined algorithm for rewriting the initial value of the editing cost. Figure 3 shows the prerequisites for the predetermined algorithm. Figure 4 shows the processing flow for the predetermined algorithm. The following notation is used in the algorithms shown in Figures 3 and 4. Vectors are written in bold, and their elements are written in thin type with the same characters, and the elements are given subscripts indicating their numbers. (Note that in the main text of the specification, vectors are written in thin type for convenience of description.) The length of vector x is represented by len(x). The length of string u is represented by |u|. The jth character of the ith element (string) of string type vector u is u. i,j To save on lines, we use the symbol × for the Cartesian product, treating vectors as sets.

[0056] In other words, the first setting unit 103 sets the editing cost for a predetermined first sequence, that is, a reference sequence s0, based on the change between each phenotype included in the first set and the reference phenotype π(s0) under the reference sequence s0.

[0057] Furthermore, the constant c E ≧0 and define c E Rewrite the conservation degree below the value of to 0. Find E=(E1,E2,E3,E4) and c that minimize the prediction error in the training dataset. E Find the value of .

[0058] 3 and 4 show, as an example, an algorithm in which there are four types of editing cost values: E1, E2, E3, and E4. However, this is not limited to this case, and the number of types of editing cost values ​​is arbitrary as long as it is plural.

[0059] In other words, the second setting unit 104 sets the first threshold c E Update the editing costs based on the

[0060] As a concrete example, the Levenshtein distance d between two sequences s1 and s2 is LIn the calculation of (s1, s2), the term +0.2×1+0.4×1+0.6×1+··· appears, and c E When =0.6, these terms become +0×1+0×1+0×1+..., and d L It no longer contributes to (s1, s2).

[0061] According to the above configuration, the effect of genome editing on the phenotype can be understood depending on the site of the expression regulatory region where a mutation has been introduced by genome editing.

[0062] (Second heuristic: editing effect saturation) First, we explain our motivation using the data in Table 1. Phenotype data is rounded to one decimal point. We want to predict π(TAW1 / dTA (SEQ ID NO: 7)) (π(TAW1 / dTA (SEQ ID NO: 7)) = 1.5). There are six sequences contained in TAW1 / dTA (SEQ ID NO: 7): TAW1 / dAGTA (SEQ ID NO: 10), TAW1 / d10 (SEQ ID NO: 19), TAW1 / d22 (SEQ ID NO: 21), TAW1 / d28 (SEQ ID NO: 23), TAW1 / d77 (SEQ ID NO: 25), and TAW1 / d1k (SEQ ID NO: 27). Of these, TAW1 / d22 (SEQ ID NO: 21) has the smallest effect of genome editing, with π(TAW1 / d22 (SEQ ID NO: 21)) = 1.4. Furthermore, the only sequence containing TAW1 / dTA (SEQ ID NO: 7) is the wild-type sequence, with π(TAW1 / WT (SEQ ID NO: 4)) = 0.6. Therefore, from P1, the predicted value of π(TAW1 / dTA (sequence number 7)) is defined as a lower bound of π(TAW1 / WT (sequence number 4)) = 0.6 and an upper bound of π(TAW1 / d22 (sequence number 21)) = 1.4.

[0063] In this case, if we use equation (1) to predict, d L (TAW1 / WT (SEQ ID NO: 4), TAW1 / dTA (SEQ ID NO: 7)) <<d L(TAW1 / dTA (SEQ ID NO: 7), TAW1 / d22 (SEQ ID NO: 21)), π(TAW1 / dTA (SEQ ID NO: 7)) is predicted to be much closer to 0.6 than 1.4 (i.e., it is significantly underestimated). From Figure 1, it can be seen that in the deleted region of TAW1 / d28 (SEQ ID NO: 23), the phenotype increases from 0.6 to 1.5 due to the deletion of 28 bases, but since π(TAW1 / dTA (SEQ ID NO: 7)) = 1.5 and π(TAW1 / dGT (SEQ ID NO: 5)) = 1.4, the effect of increasing the phenotype by deleting two bases stops increasing.

[0064] To address this phenomenon, the concept of saturation of the editing effect is introduced. Specifically, the prediction device 1 is made to perform the following process.

[0065] Step 1. When a second array s is input, first create a list L1 of arrays that s contains and a list L2 of arrays that contain s, just as in the basic principle.

[0066] Step 2. For each array t∈L1, a list of arrays that includes t and has a length equal to or greater than s is created, and the resulting lists are merged. In other words, the second generation unit 106 creates a list of arrays that includes the array t and has a length equal to or greater than the length of the second array s for each array t∈L1 included in the first list, and then merges the lists for each created array to create a second list L2.

[0067] Step 3. Search for the sequence with the greatest effect of genome editing in this list, and if so, re-select it as a sequence. * It is expressed by

[0068] Step 4. In step 3, * exists, then π(s * ) is set as the lower bound of the predicted value of π(s), and the formula (1) can be modified to obtain

number

[0069] Using the concept of saturation of this editing effect, the predicted value of π(TAW1 / dTA (sequence number 7)) is defined as having a lower bound of π(TAW1 / dGT (sequence number 5)) = 1.4 and an upper bound of π(TAW1 / d22 (sequence number 21)) = 1.4, and the predicted value of π(s) is 1.4.

[0070] According to the above configuration, even if the change in the phenotype does not correspond to the change in the edit distance, it is possible to predict the phenotype.

[0071] (Third heuristic: Enhancing prediction accuracy using data from other genes) Suppose we perform genome editing on the expression regulatory regions of several genes, create mutant sequences, measure the phenotypes under those sequences, and retain the data sets of sequences and phenotypes. Then, we use the retained data to predict the phenotype under a second sequence in the expression regulatory region of another gene in the following way.

[0072] Below, we will describe the procedure when the number of genes, i, is 2, but there is no essential change in the procedure when the number of genes is m. For each i=1, 2, s i,0 is the reference sequence (e.g., wild-type sequence) of the expression regulatory region of gene i, i,l (l=1,…,n i ) are other sequences. The phenotypes under these sequences are assumed to be known, and learning is performed using these as a training dataset using the procedure described below.

[0073] In other words, the first acquisition unit 101 further acquires a second dataset including a plurality of second pairs, each pair being a sixth sequence in the expression regulatory region of a second gene that is involved in the trait of interest and different from the first gene, and a phenotype under the sixth sequence.

[0074] The second dataset includes a plurality of second pairs, each pair consisting of a sixth sequence in an expression regulatory region of a second gene that is involved in a trait of interest and different from the first gene, and a phenotype under the sixth sequence. Here, the sixth sequence is a sequence for which the corresponding phenotype is known. The sixth sequence may be the sequence of the expression regulatory region before editing. For example, like the first dataset, the second dataset may be a collection of pairs of several sequences (such as a wild-type sequence and several mutant sequences obtained by genome editing) of the expression regulatory region of a gene involved in a trait of interest and the phenotypes under those sequences.

[0075] Also, here,

number

[0076] Second sequence s of the expression regulatory region of gene i i,l (l=n i +1), for constants ε1, ε2>0,

number

number

[0077] The above can be restated as follows:

[0078] (Calculation of the first ratio and the first group of ratios) The first calculation unit 110 calculates the edit distance d between the first sequence and the reference sequence. L (s i,k ,s i,0 ) is calculated as the first maximum value.

[0079] The second calculation unit 111 calculates the second array s i,l (l=n i +1) and the reference sequence. L (s i,l )(l=n i +1).

[0080] The third calculation unit 112 calculates the ratio of the edit distance between the first sequence and the reference sequence to the first maximum value for each first sequence, and calculates each calculated ratio D L (s i,j ) to calculate the first group of ratios.

[0081] (Calculation of the second ratio group) The fourth calculation unit 113 calculates a second maximum value, which is the maximum value of the edit distance between the sixth sequence and the reference sequence.

[0082] The fifth calculation unit 114 calculates the ratio of the edit distance between the sixth sequence and the reference sequence to the second maximum value for each sixth sequence that does not have the maximum edit distance from the reference sequence, and calculates each calculated ratio D L (s i’,k ) to calculate a second group of ratios.

[0083] (Condition determination using equations (2) and (3)) The prediction unit 109 predicts the values ​​of each ratio included in the first ratio group within a first range (D L (s i,l )-ε1,D L (s i,l )+ε1)(l=n i +1) (the above-mentioned formula (2)), and the second range (D L (s i,l )-ε2,D L (s i,l )+ε2)(l=n i +1) (Equation (3) above) is used to predict novel phenotypes.

[0084] According to the above configuration, it is possible to determine whether the acquired training data set is suitable for use in predicting a phenotype.

[0085] (The third heuristic learning stage) In the learning stage, the training dataset {(s i,j ,π(s i,j )):i=1,2,j=0,…,n i} The values ​​of the parameters ε1 and ε2 are calculated so as to minimize the prediction error under the above-mentioned prediction procedure in

[0044] In other words, the update unit 115 updates at least one of the first range and the second range based on the first data set and the second data set.

[0086] (Updated part 115) Moreover, the update unit 115 according to this embodiment updates at least one of the first threshold, the first range, and the second range based on the first data set and the second data set.

[0087] According to the above configuration, each parameter used in predicting a phenotype can be updated so as to reduce the prediction error.

[0088] (Fourth heuristic: Assessing the reliability of the upper and lower bounds of the forecast and correcting the forecast) The predictor given by the basic principle equation (1) (A) sets upper and lower bounds on the phenotype predictor, (B) (i) calculates the Levenshtein distance between the mutant sequence that gives the upper bound (or lower bound) and the mutant sequence whose phenotype is to be predicted, and (ii) calculates the Levenshtein distance between the mutant sequences that give the upper and lower bounds, and (iii) multiplies the difference between the upper and lower bounds by the ratio of these to estimate the increase in the phenotype under the mutant sequence to be predicted from the lower bound phenotype. Therefore, the accuracy of the predictor depends on the reliability of the upper and lower bounds. When reliable upper and lower bounds cannot be set, it may be better to use a predictor from another known technology. Therefore, the prediction device 1 evaluates the reliability of the upper and lower bounds and uses them to learn the conditions for using another known technology instead of the prediction device 1, as follows.

[0089] The condition C for using other known techniques to predict the phenotype π(s) of a mutant sequence s for constants U1, U2, V1, and V2≧0 is formulated as follows:

number

[0090] In the prediction stage, if the above-mentioned condition C is met, π(s) is predicted by a known technique, and if not, π(s) is predicted by the prediction device 1.

[0091] In other words, for example, the prediction unit 109 determines whether the edit distance between the second sequence and the third sequence exceeds the second threshold U1 and the edit distance between the second sequence and the fourth sequence exceeds the third threshold V1 under the condition [d L (s,s1 (min) )>U1∧d L (s,s2 (max) )>V1], and the condition that the edit distance between the second and third sequences exceeds the fourth threshold U2 [d L (s,s1 (min) )>U2], and the condition that the edit distance between the second sequence and the fourth sequence exceeds the fifth threshold V2 [dL (s,s2 (max) )>V2], novel phenotypes may be predicted.

[0092] According to the above configuration, it is possible to select the prediction process to be used when predicting a phenotype.

[0093] Furthermore, for example, the update section 115 may further update at least one of the second threshold, the third threshold, the fourth threshold, and the fifth threshold based on the first data set.

[0094] According to the above configuration, it is possible to determine the suitability of the prediction process used to predict a phenotype so as to reduce the prediction error.

[0095] (Prediction method executed by prediction device 1) The processing flow of the prediction method executed by the prediction device 1 will be described with reference to Fig. 5. Fig. 5 is a flowchart showing an example of the processing flow of the prediction method executed by the prediction device 1.

[0096] (First acquisition process S1) First, the first acquisition unit 101 acquires a first dataset including a plurality of first pairs, each pair being a pair of a first sequence in the expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence.

[0097] (Second acquisition process S2) Next, the second acquisition unit 102 acquires the second sequence s in the expression regulatory region.

[0098] (First setting process S3) Next, the first setting unit 103 sets an editing cost for a reference sequence s0, which is a predetermined first sequence, based on the change between each phenotype included in the first set and the reference phenotype π(s0) under the reference sequence s0.

[0099] (Second setting process S4) Next, the second setting unit 104 sets a first threshold c EUpdate the editing costs based on the

[0100] (First generation process S5) Next, the first generation unit 105 generates a first list L1, which is a list of first sequences included in the second sequence.

[0101] (Second generation process S6) Next, the second generation unit 106 generates a second list L2, which is a list of first sequences including the second sequence.

[0102] (First identification process S7) Next, the first identifying unit 107 identifies a third sequence s1 that has the smallest effect of genome editing among the sequences included in the first list L1. (min) Identify the following.

[0103] (Second identification process S8) Next, the second identification unit 108 identifies the fourth sequence s2, which has the greatest effect of genome editing among the sequences included in the second list L2. (max) Identify the following.

[0104] (Prediction process S9) Next, the prediction unit 109 predicts a new phenotype π(s) under the second sequence s based on the proportional relationship between the edit distance between the two different sequences and the change between the two phenotypes under each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost.

[0105] [Software implementation example] The functions of the prediction device 1 (hereinafter referred to as the "device") can be realized by a program that causes a computer to function as the device, and a program that causes a computer to function as each control block of the device (particularly each part included in the control unit 10).

[0106] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.

[0107] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0108] Furthermore, some or all of the functions of the control blocks can be realized by logic circuits. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of the control blocks can also be realized by, for example, a quantum computer.

[0109] Furthermore, each process described in each of the above embodiments may be executed by AI (Artificial Intelligence). In this case, the AI ​​may run on the control device or on another device (for example, an edge computer or a cloud server).

[0110] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

[0111] 〔summary〕 A prediction device according to a first aspect of the present invention includes a first acquisition unit that acquires a first dataset including a plurality of first pairs, each pair being a pair of a first sequence in an expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence; a second acquisition unit that acquires a second sequence in the expression regulatory region; a first setting unit that sets an editing cost for a reference sequence that is a predetermined first sequence, based on a change between each phenotype included in the first pair and a reference phenotype under the reference sequence; a second setting unit that updates the editing cost based on a first threshold for the editing cost; and a first list that is a list of the first sequences included in the second sequence. a first generation unit that generates a list of the first sequences including the second sequence; a second generation unit that generates a second list that is a list of the first sequences including the second sequence; a first identification unit that identifies a third sequence among the sequences included in the first list that will have the smallest effect of genome editing; a second identification unit that identifies a fourth sequence among the sequences included in the second list that will have the largest effect of genome editing; and a prediction unit that predicts a new phenotype under the second sequence based on the proportional relationship between the edit distance between two different sequences and the amount of change between two phenotypes under each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost.

[0112] In a prediction device according to a second aspect of the present invention, in addition to the configuration of the prediction device according to the first aspect, when the third sequence is present, the prediction unit predicts the new phenotype based on a second phenotype under the fourth sequence, a change between the first phenotype and the second phenotype under the third sequence, an edit distance between the third sequence and the fourth sequence, and an edit distance between the second sequence and the fourth sequence; when the third sequence is not present and the fourth sequence different from the reference sequence is present, the prediction unit predicts the new phenotype based on the reference phenotype, a change between the second phenotype and the reference phenotype, and an edit distance between the fourth sequence and the reference sequence. The novel phenotype is predicted based on the difference between the third sequence and the reference sequence and the edit distance between the second sequence and the reference sequence. If the third sequence does not exist and the fifth sequence having the smallest edit distance from the second sequence is different from the reference sequence, the novel phenotype is predicted based on the reference phenotype, the change between the third phenotype and the reference phenotype under the fifth sequence, the edit distance between the fifth sequence and the reference sequence, and the edit distance between the second sequence and the reference sequence. If the third sequence does not exist and the sequence having the smallest edit distance from the second sequence is the reference sequence, the reference phenotype is predicted as the novel phenotype.

[0113] In a prediction device according to aspect 3 of the present invention, in addition to the configuration of the prediction device according to aspect 1 or 2, the second generation unit generates, for each sequence included in the first list, a list of sequences that includes the sequence and has a length equal to or greater than the length of the second sequence, and then merges the lists generated for each sequence to generate the second list.

[0114] In a prediction device according to a fourth aspect of the present invention, in addition to the configuration of the prediction device according to any one of the first to third aspects, the first acquisition unit further acquires a second dataset including a plurality of second pairs, each pair being a pair of a sixth sequence in an expression regulatory region of a second gene that is involved in the trait of interest and different from the first gene, and a phenotype under the sixth sequence; and further includes a first calculation unit that calculates a first maximum value that is the maximum value of the edit distance between the first sequence and the reference sequence; a second calculation unit that calculates a first ratio that is the ratio of the edit distance between the second sequence and the reference sequence to the first maximum value for each of the first sequences, and calculates each calculated ratio as a function of the edit distance between the first sequence and the reference sequence to the first maximum value for each of the first sequences. the sixth sequence and the reference sequence; and a fifth calculation unit that calculates, for each of the sixth sequences whose edit distance with the reference sequence is not the greatest, a ratio of the edit distance between the sixth sequence and the reference sequence to the second maximum value, thereby calculating a second ratio group including each of the calculated ratios. The prediction unit predicts the novel phenotype based on the condition that, for each ratio value included in the first ratio group, there is no value that falls within a first range centered on the first ratio, and the condition that, for each ratio value included in the second ratio group, there is a value that falls within a second range centered on the first ratio.

[0115] A prediction device according to aspect 5 of the present invention, in addition to the configuration of the prediction device according to aspect 4 above, further comprises an update unit that updates at least one of the first threshold, the first range, and the second range based on the first data set and the second data set.

[0116] In a prediction device according to aspect 6 of the present invention, in addition to the configuration of the prediction device according to any one of aspects 1 to 5 above, the prediction unit predicts the new phenotype based on the following conditions: the edit distance between the second sequence and the third sequence exceeds a second threshold and the edit distance between the second sequence and the fourth sequence exceeds a third threshold; the edit distance between the second sequence and the third sequence exceeds a fourth threshold; and the edit distance between the second sequence and the fourth sequence exceeds a fifth threshold.

[0117] In a prediction device according to aspect 7 of the present invention, in addition to the configuration of the prediction device according to aspect 6, the update unit further updates at least one of the second threshold, the third threshold, the fourth threshold, and the fifth threshold based on the first data set.

[0118] A prediction device according to an eighth aspect of the present invention has the same configuration as the prediction device according to any one of the first to seventh aspects, and the edit distance is a Levenshtein distance, a Damerau-Levenshtein distance, or an extended Hamming distance.

[0119] A prediction method according to a ninth aspect of the present invention includes a first acquisition process for acquiring a first dataset including a plurality of first pairs, each pair being a pair of a first sequence in an expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence; a second acquisition process for acquiring a second sequence in the expression regulatory region; a first setting process for setting an editing cost for a reference sequence that is a predetermined first sequence, based on the amount of change between each phenotype included in the first pair and a reference phenotype under the reference sequence; a second setting process for updating the editing cost based on a first threshold for the editing cost; and a first list that is a list of the first sequences included in the second sequence. a first generation process for generating a list of the first sequences including the second sequence; a second generation process for generating a second list that is a list of the first sequences including the second sequence; a first identification process for identifying a third sequence among the sequences included in the first list that will have the smallest effect of genome editing; a second identification process for identifying a fourth sequence among the sequences included in the second list that will have the largest effect of genome editing; and a prediction process for predicting a new phenotype under the second sequence based on the proportional relationship between the edit distance between two different sequences and the amount of change between two phenotypes under each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost. [Example]

[0120] An embodiment of the present invention will now be described.

[0121] In this example, wild-type and mutant sequences of the expression regulatory regions of the rice TAW1 and SD1 genes, as well as phenotype data under these sequences, were used to verify the learning and prediction accuracy of the prediction device 1. As described above, TAW1 and SD1 are genes that control the number of grains and plant height in rice, respectively, and here, we focused on the number of secondary rachis-branches per number of primary rachis-branches and plant height as phenotypes.

[0122] The data used will be explained further using the TAW1 gene. TAW1 is located on rice chromosome 10 at positions 17,888,297 to 17,889,724. In this study, we used the wild-type sequence of the 113-base expression regulatory region downstream of TAW1, as well as 26 mutant sequences obtained by genome editing, along with phenotypic data for these regions. This region is known to act as a silencer, suppressing TAW1 expression.

[0123] The prediction accuracy of Predictor 1 was evaluated using the TAW1 and SD1 datasets with the leave-one-out method.

[0124] Table 3 shows the prediction results for TAW1. [Table 3] The measured values ​​listed in Table 3 are the measured values ​​of the number of secondary rachis-branches per number of primary rachis-branches corresponding to each of the sequence IDs included in the table. The predicted values ​​listed in Table 3 are the predicted values ​​of the number of secondary rachis-branches per number of primary rachis-branches corresponding to each of the sequence IDs included in the table.

[0125] Table 4 shows the prediction results for SD1. [Table 4] The measured values ​​listed in Table 4 are the measured plant height (cm) corresponding to each of the sequence IDs included in the table. The predicted values ​​listed in Table 4 are the predicted plant height (cm) corresponding to each of the sequence IDs included in the table.

[0126] The average relative errors in predicting TAW1 and SD1 using predictor 1 were 0.1287 and 0.0461, respectively (see the figures at the bottom right of Tables 3 and 4).

[0127] Furthermore, compared to SVR combined with a spectral kernel, which is an example of a known technique, the prediction device 1 reduced the prediction error by approximately 0.3826 times in the TAW1 dataset and by approximately 0.2394 times in the SD1 dataset. [Explanation of symbols]

[0128] 1 Prediction device 10 Control Unit 11 Storage section 101 First acquisition part 102 Second acquisition part 103 First Setting Section 104 Second Setting Section 105 1st generation part 106 Second generation part 107 1st Specific Part 108 Second Specific Part 109 Prediction Department 110 First Calculation Unit 111 Second calculation unit 112 Third Calculation Unit 113 Fourth Calculation Section 114 5th Calculation Section 115 Update Department

Claims

1. a first acquisition unit that acquires a first dataset including a plurality of first pairs, each pair being a first sequence in an expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence; a second acquisition unit that acquires a second sequence in the expression regulatory region; a first setting unit that sets an editing cost for a reference sequence that is a predetermined first sequence, based on a change between each phenotype included in the first set and a reference phenotype under the reference sequence; a second setting unit that updates the editing cost based on a first threshold value for the editing cost; a first generator that generates a first list that is a list of the first arrays included in the second array; a second generator that generates a second list that is a list of the first arrays including the second array; A first identification unit that identifies a third sequence having the smallest effect of genome editing among the sequences included in the first list; A second identification unit that identifies a fourth sequence having the greatest effect of genome editing among the sequences included in the second list; a prediction unit that predicts a new phenotype under the second sequence based on a proportional relationship between an edit distance between two different sequences and a change between two phenotypes under each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost; A prediction device comprising:

2. The prediction unit When the third sequence is present, predicting the new phenotype based on a second phenotype under the fourth sequence, a change between the first phenotype and the second phenotype under the third sequence, an edit distance between the third sequence and the fourth sequence, and an edit distance between the second sequence and the fourth sequence; When the third sequence does not exist and the fourth sequence different from the reference sequence exists, predicting the new phenotype based on the reference phenotype, the changes between the second phenotype and the reference phenotype, the edit distance between the fourth sequence and the reference sequence, and the edit distance between the second sequence and the reference sequence; When the third sequence does not exist and a fifth sequence having a minimum edit distance with the second sequence is different from the reference sequence, predicting the new phenotype based on the reference phenotype, the change between the third phenotype and the reference phenotype under the fifth sequence, the edit distance between the fifth sequence and the reference sequence, and the edit distance between the second sequence and the reference sequence; If the third sequence does not exist and the sequence having the smallest edit distance from the second sequence is the reference sequence, the reference phenotype is predicted as the novel phenotype. The prediction device according to claim 1 .

3. The second generation unit For each sequence included in the first list, generating a list of sequences that includes the sequence and has a length equal to or greater than the length of the second sequence, and then merging the lists for each of the generated sequences to generate the second list; The prediction device according to claim 2 .

4. the first acquisition unit further acquires a second dataset including a plurality of second pairs, each pair being a pair of a sixth sequence in an expression regulatory region of a second gene that is involved in the trait of interest and different from the first gene, and a phenotype under the sixth sequence; a first calculation unit that calculates a first maximum value that is a maximum value of the edit distance between the first sequence and the reference sequence; a second calculation unit that calculates a first ratio, which is a ratio of an edit distance between the second sequence and the reference sequence to the first maximum value; a third calculation unit that calculates, for each of the first sequences, a ratio of the edit distance between the first sequence and the reference sequence to the first maximum value, thereby calculating a first ratio group including each calculated ratio; a fourth calculation unit that calculates a second maximum value that is a maximum value of the edit distance between the sixth sequence and the reference sequence; a fifth calculation unit that calculates a ratio of the edit distance between the sixth sequence and the reference sequence to the second maximum value for each of the sixth sequences that does not have the maximum edit distance from the reference sequence, thereby calculating a second ratio group including each calculated ratio; the prediction unit predicts the novel phenotype based on a condition that, for each ratio value included in the first ratio group, there is no value included in a first range having the first ratio as a center value, and a condition that, for each ratio value included in the second ratio group, there is a value included in a second range having the first ratio as a center value. The prediction device according to claim 2 .

5. an update unit that updates at least one of the first threshold, the first range, and the second range based on the first data set and the second data set. The prediction device according to claim 4 .

6. the prediction unit predicts the novel phenotype based on a condition that an edit distance between the second sequence and the third sequence exceeds a second threshold and an edit distance between the second sequence and the fourth sequence exceeds a third threshold, a condition that an edit distance between the second sequence and the third sequence exceeds a fourth threshold, and a condition that an edit distance between the second sequence and the fourth sequence exceeds a fifth threshold. The prediction device according to claim 5 .

7. the update unit further updates at least one of the second threshold, the third threshold, the fourth threshold, and the fifth threshold based on the first data set. The prediction device according to claim 6 .

8. The edit distance is the Levenshtein distance, the Damerau-Levenshtein distance, or the extended Hamming distance. The prediction device according to any one of claims 1 to 7.

9. a first acquisition process for acquiring a first dataset including a plurality of first pairs, each pair being a first sequence in an expression regulatory region of a first gene involved in a trait of interest and a phenotype under the first sequence; a second acquisition process for acquiring a second sequence in the expression regulatory region; a first setting process for setting an editing cost for a reference sequence that is a predetermined first sequence, based on the amount of change between each phenotype included in the first set and a reference phenotype under the reference sequence; a second setting process for updating the editing cost based on a first threshold value for the editing cost; a first generation process for generating a first list that is a list of the first arrays included in the second array; a second generation process for generating a second list that is a list of the first arrays including the second array; A first identification process for identifying a third sequence having the smallest effect of genome editing among the sequences included in the first list; A second identification process for identifying a fourth sequence having the greatest effect of genome editing among the sequences included in the second list; a prediction process for predicting a new phenotype under the second sequence based on a proportional relationship between the edit distance between two different sequences and the change between two phenotypes under each of the two sequences, the first dataset, the third sequence, the fourth sequence, and the editing cost; A prediction method, including:

Images