JP7551189B1 - Method for predicting promoter activity and method for modifying promoter based on the results of the prediction - Google Patents

Abstract

[Problem] The present disclosure aims to provide a method for obtaining a promoter sequence modified to have a desired transcription activity. [Solution] The method for obtaining a promoter sequence modified to have a desired transcription activity includes preparing an original promoter sequence to be modified, generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence, predicting the transcription activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model, and selecting the modified promoter sequence predicted to have the desired transcription activity. [Selected Figure] Figure 2

Description

This disclosure relates to predicting promoter activity and modifying promoters based on the results of that prediction.

Currently, the main method for conferring traits to plants through genome editing is the loss-of-function method, in which a mutation is introduced into the coding region (CDS) to cause a frameshift, resulting in the loss of gene function.

On the other hand, methods can be considered to enhance function by increasing the expression level of the target gene, or to suppress the side effects of knockout by leaving only a small amount of expression (rather than reducing expression to zero). Since it will be possible to apply research on enhancing function by introducing transgenes and knowledge gained from functional analysis using RNAi, such methods are important as they are expected to greatly expand the variety of traits that can be conferred by genome editing.

Zhang, Yi, et al. "Applications and potential of genome editing in crop improvement." Genome biology 19 (2018): 1-11.

One such method is to edit the promoter region instead of the CDS. Since the strength of gene expression is determined by regions such as the promoter and enhancer, it is thought that by modifying the base sequence of this region, it is possible to increase or decrease the expression level while maintaining the function of the gene.

In fact, the dynamic range of gene expression at the mRNA level measured by the RNA-seq method is very large. Some mRNAs contain only a few copies in a cell, while others contain about 10 ⁵ copies. Such differences in gene expression are thought to be mainly caused by promoters and enhancers, and are thought to show the potential of methods that modify the base sequence of promoter regions. Therefore, one of the objectives of the present disclosure is to provide a technology for precisely controlling gene expression levels by genome editing.

Compared to CDS, the sequence-function relationship of "this part has this function" is vaguer in regions such as promoters and enhancers. Several Cis regulatory elements with highly conserved consensus sequences, such as the TATA box, have been discovered and are now in a database. For example, software such as PromoterCAD from the RIKEN Institute and NEW PLACE from the National Agriculture and Food Research Organization are systems that search for Cis regulatory elements based on this knowledge and predict whether transcription factors will bind to promoters, but the accuracy is insufficient for the purpose of predicting and designing expression levels.

Therefore, the present inventors developed a method of learning the relationship between base sequences and expression levels by machine learning, rather than a database method in which individual elements are registered. As a result, a high value of ^R2 = 0.75 was obtained as the coefficient of determination indicating the correlation between the actual value and the predicted value. Then, a sequence was designed using the prediction system, and an experiment using plants was performed to demonstrate the accuracy of the prediction. If an arbitrary base sequence can be given to a computer and the transcription activity can be predicted, it becomes possible to "design" the expression level. As shown in the present disclosure, the present inventors have demonstrated that the function (expression level) of a gene can be increased or decreased by editing the promoter of the gene based on computer prediction.

The present disclosure builds on these findings and includes the following aspects:
[Aspect 1]
A method for obtaining a promoter sequence modified to have a desired activity, comprising the steps of:
Providing an original promoter sequence to be modified;
Generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence;
predicting the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model;
selecting a modified promoter sequence predicted to have a desired activity.
[Aspect 2]
The method according to aspect 1, wherein the desired activity is a higher gene expression induction activity than the original promoter sequence, or a lower gene expression induction activity than the original promoter sequence.
[Aspect 3]
2. The method of embodiment 1, wherein the promoter sequence is a plant cell promoter sequence.
[Aspect 4]
2. The method of embodiment 1, wherein the original promoter sequence comprises a core promoter sequence and its upstream sequence.
[Aspect 5]
The method according to aspect 1, wherein the genome editing technique is a genome editing technique using a CRISPR/Cas system.
[Aspect 6]
6. The method of embodiment 5, wherein the set of multiple modified promoter sequences is generated by sequence deletion caused by cleavage induced by a combination of guide RNA sequences designed based on two PAM recognition sequences.
[Aspect 7]
2. The method of embodiment 1, wherein the set of modified promoter sequences comprises at least 1000 different sequences.
[Aspect 8]
The method according to aspect 1, wherein the machine learning model is a regression model trained to predict gene expression induction activity from a promoter sequence using gene expression induction activity data of a plurality of promoter sequences in plant cells as training data.
[Aspect 9]
The method of embodiment 1, further comprising the step of visualizing on a computer display the activity of each modified promoter sequence contained in the generated set of modified promoter sequences.
[Aspect 10]
An information processing device for predicting a promoter sequence modified to have a desired activity, comprising:
a sequence input section for receiving input of an original promoter sequence to be modified;
A modified sequence generating unit that generates a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence;
an activity prediction unit that predicts the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model;
An information processing device comprising a sequence selection unit that selects a modified promoter sequence predicted to have a desired activity.
[Aspect 11]
A non-transitory computer-readable medium having instructions stored thereon, the instructions, when executed by a processor, performing the following steps:
receiving an input of an original promoter sequence to be modified;
A step of generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence;
predicting the activity of each of the modified promoter sequences included in the generated set of modified promoter sequences using a machine learning model;
A computer readable medium capable of carrying out the step of selecting modified promoter sequences predicted to have a desired activity.
[Aspect 12]
On the computer,
A function to accept input of the original promoter sequence to be modified;
A function of generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence;
A function of predicting the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model;
A program that performs the function of selecting a modified promoter sequence predicted to have a desired activity.
[Aspect 13]
A genome editing method for a cell for regulating the expression level of a desired gene, comprising:
Obtaining a modified promoter sequence having a desired activity by the method according to embodiment 1;
Preparing cells to be subjected to genome editing;
editing the genome of said cell to create said modified promoter sequence.
[Aspect 14]
A method for producing a genome-edited plant in which the expression level of a desired gene is regulated, comprising:
Editing the genome of a desired plant cell by the method of embodiment 13;
A method comprising obtaining a plant individual derived from a genome-edited cell.
[Aspect 15]
A machine learning model that predicts gene expression induction activity in plant cells from a promoter sequence, the machine learning model being a regression model trained to predict gene expression induction activity from a promoter sequence using gene expression induction activity data of multiple promoter sequences in plant cells as teacher data.
[Aspect 16]
A method for generating a machine learning model that predicts gene expression induction activity in plant cells from a promoter sequence, the method comprising training the model using gene expression induction activity data of a plurality of promoter sequences in plant cells as training data.
[Aspect 17]
17. The method of claim 16, wherein a transformer-based pre-trained base model is used as the model to be trained.

Figure 1 is a schematic diagram showing the structure of the artificial synthetic gene. Seven promoters were selected from the high expression group, eight promoters from the medium expression group, and four promoters from the low expression group, and artificial synthetic genes were created by connecting them to the luciferase (LUC) gene. The promoters have 19 different sequence patterns. The sequences of the downstream LUC gene and the upstream cauliflower mosaic virus (CaMV) 35S enhancer are common to each artificial gene. Figure 2 is a scatter plot showing the relationship between the measured expression level of the LUC gene and the predicted promoter strength. The vertical axis shows the predicted promoter strength. The higher the value, the greater the expected expression level of downstream genes. The horizontal axis shows the measured LUC expression level. Note that this was normalized with the transcription activity of the CaMV promoter, which is a positive control. FIG. 3 shows the results of calculating and comparing the LUC expression levels corresponding to the predicted transcriptional activity. The expression levels of LUC are plotted on the vertical axis. However, the values are shown as relative values with the luminescence intensity of the positive control taken as 1. As a positive control, the cauliflower mosaic virus (CaMV) 35S promoter, which is known to exhibit high transcriptional activity, was used. The horizontal axis shows the numbers of 19 types of promoters. Promoter numbers 3 to 10 were classified into the "high expression" group, which was predicted to have high expression intensity, promoter numbers 13 to 21 into the "moderate" group, and promoter numbers 22 to 25 into the "low expression" group. The black bars show the actual values of the LUC assay. The white bars show the predicted transcriptional activity values by the prediction system of the present disclosure. Figure 4 shows the results of searching for editing patterns that can reduce transcriptional activity. We searched for editing patterns that can reduce transcriptional activity for promoter numbers 3, 4, 5, 6, 9, 10, and 21. The plot of black triangles shows the predicted LUC assay score (expression level) calculated based on the predicted new transcriptional activity. As a result, it was predicted that the transcriptional activity would be about 14% to 1% compared to the original promoter. Figure 5 shows the actual measured expression levels of the new promoter LUC in shaded bars overlaid on Figure 4. The vertical axis is expanded because the expression levels were very low. Promoter 5 had missing values. For the six promoters for which measured values were available, a significant decrease in expression level was observed, which was consistent with the predicted values. Figure 6 shows the results of searching for promoters whose predicted transcription activity increases after editing. Promoter numbers 13, 17, 22, 23, 24, and 25 were selected for their increased transcription activity, and the predicted expression levels are indicated by white circles. Compared to the original promoter, the predicted LUC expression levels were 7.4 to 125 times higher. Figure 7 shows the results of artificially synthesizing genes with theoretical promoter sequences, transfecting them into protoplasts in the same manner, and measuring the expression levels using a plate reader. Of the six promoters, the expression levels of five were increased compared to the original promoter, but only the 13th promoter showed an increase in expression level beyond the predicted value. Figure 8 shows an example of a visualization method for analyzing the entire image of an input base sequence X. The vertical axis corresponds to the predicted transcription activity, and the horizontal axis corresponds to the position of the base sequence. Figure 9 shows an example of visualization by another method. The vertical axis corresponds to the position of the guide RNA (proximal guide) designed near the target gene, and the horizontal axis corresponds to the position of the guide RNA (distal guide) designed far from the target gene. The color of the plot of each position is changed based on the predicted value of transcription activity. For example, in a color chart, light gray is predicted to be high transcription activity, and black is predicted to be low transcription activity. Figure 10 shows the results of an investigation into increasing the transcription activity of a specific soybean gene. Based on the promoter of a soybean gene, two types of promoter sequences after genome editing were created to increase gene expression (edit1, edit2). The transcription activity of edit1 was predicted to be 2.950119. In addition, the LUC expression level was predicted to be 13.2% compared to the P control by linear regression. The actual LUC assay value of this promoter was 19.0% compared to the P control. The transcription activity of edit2 was predicted to be 2.432977. In addition, the LUC expression level was predicted to be 7.27% compared to the P control. The actual LUC assay value of this promoter was 41.6% compared to the P control. FIG. 11 shows an exemplary configuration of an information processing device that predicts a promoter sequence modified to have a desired activity. FIG. 12 shows an exemplary flowchart of a program that enables a computer to perform the functions of accepting input of an original promoter sequence to be modified, generating a set of multiple modified promoter sequences that can be created using genome editing technology based on the promoter sequence, predicting the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model, and selecting a modified promoter sequence predicted to have a desired activity. FIG. 13 is a schematic diagram of a computer that can be used to implement aspects of the present disclosure.

In the process of breeding crops, breeders hope to enhance or suppress the expression of specific genes. As gene functions are analyzed, the relationship between genes and traits is becoming clearer one after another. For example, in the case of semi-dwarf rice, the trait of semi-dwarfness was set as the breeding goal for rice, and in 1966 the semi-dwarf variety "Reimei" was developed. Semi-dwarf rice is less likely to grow taller even when fertilizer is applied, and the yield does not decrease. When fertilizer is applied, rice usually grows vertically, but this means that it is easily lodged by typhoons or strong winds, making it impossible to harvest or reducing the quality. Therefore, applying fertilizer beyond a certain level will cause lodging and will not lead to an increase in yield. Semi-dwarf rice is less likely to grow taller, so it is resistant to lodging even when a large amount of fertilizer is applied, and the yield increases. It is now known that the mutation that gave the semi-dwarf trait to Leucanthemum moniliforme is present in a gene encoding the G20 oxidase enzyme involved in gibberellin biosynthesis. Semi-dwarf crops were a major invention that brought about enormous advances in the expansion of food production.

The relationship between genes and traits is being clarified one after another, and knowledge is being accumulated. As a result, target genes are often set to obtain desired traits. For example, if one wishes to introduce semi-dwarfism into rice, a mutant of the above gene (the sd1 mutant) can be selected. It is possible to search for naturally occurring mutants, but it is also possible to randomly introduce mutations by chemical or radiation exposure. The loss-of-function allele of sd1 is thought to be recessive, but it is possible to breed individuals in which both alleles are loss-of-function sd1 by planned breeding.

With the advent of recombinant gene technology, it became possible to artificially design genotypes. In other words, by introducing any gene sequence into the target crop, it became possible to give it functions and traits. A typical example is corn that has been given resistance to herbicides (Roundup Ready (registered trademark)). Roundup (registered trademark) (glyphosate) is a herbicide that is in fact an amino acid analog. When some plants and bacteria take in glyphosate, amino acid synthesis is inhibited and they die. By introducing a gene that neutralizes glyphosate from bacteria, plants were given resistance to glyphosate, which was a major invention that significantly reduced the cost of weeding. On the other hand, the production and use of genetically modified plants that introduce foreign genes was limited, and all crops on the planet were not replaced by genetically modified plants. There was particularly strong opposition in Europe, and even today, the cultivation of genetically modified crops is greatly restricted in Europe, with only one variety being cultivated (corn in Spain).

Since the invention of genome editing, the situation has changed even further. It has become possible to directly introduce mutations into target genes without genetic modification. This can be considered a historic invention. For example, if you want to introduce semi-dwarfness into Koshihikari, you can induce a frameshift and cause a loss of function by introducing a small indel of about 1 to 10 bases into the SD1 locus. This method allows for faster development than chemical or radiation-based mutation introduction (followed by 5 to 7 backcrosses), and it has a major advantage in terms of cost, as it can avoid the expression of unintended traits caused by linkage drugs. An example of a food product produced using this method and sold in Japan is red sea bream with a deletion of the myostatin gene, sold by Regional Fish Co.

In addition to small indels of 1 to 10 bases, genome editing can also induce medium-sized deletions of up to several thousand bases. In this case, two guide RNAs are simultaneously delivered into the cell, and if the two cutting sites that cut the target sequences of the two DNA are linked together or physically close to each other, the gap between the two cutting sites may be removed and repaired. A specific example is Corteva Agriscience's waxy corn, which has a deletion of approximately 4 kb, including the entire coding sequence of the Wx locus.

The above mentioned mutations in the varieties developed through genome editing are the kind that can occur in nature, and the resulting crops are essentially the same as those produced by conventional breeding. In contrast, genome editing can also be used to cause homologous recombination, which allows the replacement of a target position in the genome with any DNA sequence. This can be used to introduce various mutations such as point mutations, deletions, and insertions, and can also cause genetic recombination by replacing with a foreign gene. In this way, even though genome editing is simply a term, it can create mutants of essentially different categories, such as mutations that can occur in nature and genetic recombinations that cannot occur at all in nature. To avoid confusion, varieties created through genome editing are classified into three categories, SDN1 to 3. The above examples from Regional Fish and Corteva are both classified as SDN1. In genome-edited crops, procedures such as notification to government agencies and prior consultation are required in many countries, including Japan, when creating new varieties. In SDN1, since the newly introduced mutations are essentially equivalent to naturally occurring mutations, and since there is a history of precedent, it is expected that the process will proceed quickly in many countries, which will be beneficial for industry.

By the way, as mentioned above, genome editing can induce frameshifts, which makes it easy to create a mutant with a specific gene's function. However, it is not only mutants with a gene's function that are useful to breeders. Mutations with increased gene expression or mutants with decreased gene expression have been selected intentionally or unintentionally in countless numbers. In addition, a huge number of gene silencing (RNAi, KD) experiments have been reported in the history of plant science. Phenotypes may differ between knockout (KO) and knockdown (KD) mutants. There are also cases where a small amount of remaining expression avoids a harmful phenotype. In this way, a significant decrease in gene expression, which mimics a KD mutant, is also often useful. The method disclosed herein has the advantage of making it possible to design such increases and decreases in gene expression using genome editing of SDN1, which can be commercialized in Japan simply by submitting a notification to the relevant government agency.

One method to enhance function is to introduce a mutation into a gene. For example, it is thought that activity can be increased by destroying the autoinhibitory domain of an enzyme. For example, in the case of a crop sold in Japan, the high GABA accumulation tomato from Sanatec Seed Co., Ltd. has its GABA synthesis activity enhanced by destroying the autoinhibitory domain at the C-terminus of the GABA synthesis enzyme GAD by frameshifting. This method is very effective, but it can only be used when the target protein (gene) has an autoinhibitory domain. Since there are relatively few proteins with an autoinhibitory domain, the range of applications is narrow. Other methods include destroying the protein's localization signal to make it always active, replacing the amino acid that is phosphorylated with glutamic acid, and introducing a mutation into the active site of the enzyme to improve reactivity. However, these methods are not generally applicable to any gene.

Another method of enhancing function that can be generally used for many genes is to introduce mutations into the region that regulates the expression level of a gene. In other words, the idea is to increase the expression level by editing the promoter, and enhance the function, but this idea has a drawback. That is, even in the case of promoters, the correspondence between sequence and function is ambiguous. In genes, the correspondence between codons and amino acid sequences is strict, and the functional domains created by specific amino acid sequences are also highly conserved, so the function can be inferred from the sequence. For this reason, it is possible to destroy the target domain with a high degree of accuracy. On the other hand, with promoters (non-ORFs), it is difficult to make a hypothesis that "if you do this part like this, this will happen."
Of course, there is also a relationship between sequence and function. Cis regulatory elements, such as the TATA box, are understood as binding sites for basic transcription factors and transcription factors, and there have been successful cases of predicting transcription factors based on sequences and estimating binding sites from transcription factors. For example, web applications such as PromoterCAD from the RIKEN Institute and NEW PLACE from the National Agriculture and Food Research Organization are tools for searching such elements. While such tools have been proposed, there seem to be only a limited number of successful breeding cases using them. Song et al.,2022 reported the improvement of gene expression by genome editing of promoters (Song, X., Meng, X., Guo, H. et al. Targeting a gene regulatory element enhances rice grain yield by decoupling panicle number and size. Nat Biotechnol 40, 1403-1411 (2022). https://doi.org/10.1038/s41587-022-01281-7). This report reports that an increase in yield can be achieved by editing the promoter or 5'UTR of the rice IPA1 gene. The procedure involves creating lines with various deletion patterns using genome editing, evaluating expression patterns and traits, and identifying important control elements. This shows how difficult it is to design functions from control elements. One reason for this difficulty is the fluctuation of control elements (which typically tolerate a mismatch of several bases from the consensus sequence). Recently, there have been reports of applying deep learning to estimate Cis control elements, but there have been no examples of this being linked to actual crop traits.

The inventors adopted a strategy to directly infer expression levels from DNA sequences, rather than the conventional method of inferring function from Cis regulatory elements. To achieve this, they limited the target to promoters. Although the relationship between sequence and expression level is a black box, methods using deep learning often perform well even when the principles and theories are unknown. For example, in the field of machine translation, until the 1990s, methods were explored to convert language by collecting vocabulary and building a dictionary linked to function (semantic information) (e.g., converting English to German), but high performance was not obtained. In contrast, around 2011, when WATSON, which applied neural networks (NN), demonstrated high performance in natural language processing, it became possible to translate with high accuracy without inputting the meaning of individual sentences by having the computer learn huge amounts of text data, rather than having humans create dictionaries and have the computer understand the meaning. Today's Google Translate and DeepL are typical examples of machine translation using NNs, and both provide very high usefulness. In this way, machine learning using NNs can sometimes demonstrate high performance without humans understanding or organizing the principles and theories.

DNABERT is a model that has been pre-trained to handle DNA sequences, based on the natural language model BERT, which was created for the purpose of natural language processing. DNABERT was developed with the aim of analyzing non-coding regions based on DNA sequences, and has shown high performance in tasks such as discovering promoters and splice sites. For this reason, we decided to use it for the task of this invention, which is to predict the transcriptional activity of core promoters. However, there were two problems with applying DNABERT to this invention.

First, the human genome was used as pre-training data, and it was unclear whether DNABERT could handle plant genomes. In plants, the core elements that make up the core promoter may have plant-specific elements such as Y patch, CA, and GA, in addition to common elements with animals such as the TATA box, Initiator, and Kozak. In addition, in the promoters of animal and plant genes, transcriptional activity may be controlled by DNA methylation, but in animals, cytosines in sequences called CG (CpG) are often methylated. In contrast, in plants, cytosines in non-CG sequences are also methylated. Thus, although the basic mechanism of transcription is common between animals and plants, the components such as core elements and methylation sites are different. Thus, it was unclear whether DNABERT would perform well in plants with promoters whose structures are significantly different from those of humans.

Secondly, DNABERT has been shown to perform well in tasks such as finding transcription start sites and splice sites, but these tasks were binary classification problems, in which the task was to determine whether an element of interest was "present" or "absent" in a given sequence. In contrast, the inventors' task was to predict gene expression intensity downstream of a core promoter sequence, and it was required to output continuous values. It was completely unclear whether DNABERT would be effective or perform well in such regression problems. When dealing with continuous values related to DNA sequences, it was common to use CNN. Based on reports that the natural language processing model BERT can also be used to handle regression problems, the inventors modified the DNABERT code. Specifically, they added a class and modified it so that it could handle continuous values.

Despite these issues, the inventors were able to fine-tune the DNABERT pre-training model using a procedure described separately, and successfully predict promoter strength with high accuracy. Furthermore, the inventors have also successfully developed a program that incorporates this learning model and changes promoter function through genome editing.

In one aspect, the present disclosure relates to a method for obtaining a promoter sequence modified to have a desired activity. In some embodiments, the method of the present disclosure may include preparing an original promoter sequence to be modified, generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence, predicting the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model, and selecting the modified promoter sequence predicted to have the desired activity.

A promoter is a DNA sequence located (usually) upstream of a gene; it is not itself transcribed, but has the function of determining the amount of transcription (expression level) of the gene, and in multicellular organisms, the site and tissue specificity of expression. In this specification, the region that interacts with basic transcription factors, particularly in the vicinity of the gene, is called a "core promoter." A core promoter can be defined as a position between -200 and +50, when the base position of the transcription start point is represented as +1. More preferably, it can be defined as a position between -180 and +25. Even more preferably, it can be defined as a position between -170 and +17. Most preferably, a position between -165 and +5 can be defined as a core promoter. It is believed that a core promoter has the function of determining the basic amount of transcription.

In some embodiments, the promoter sequence is a promoter sequence of a plant cell. It is known that the types of core elements contained in the core promoters of animals and plants are different. In this specification, the term "plant" is not particularly limited. For example, a wide range of plants can be mentioned, including mosses, ferns, gymnosperms, magnolias of angiosperms, monocots, and eudicots (Rosa I, Rosa II, Chrysanthemum I, Chrysanthemum II, and their outgroups). More specific examples of plants include eggplants such as tomatoes, bell peppers, chili peppers, eggplants, tobacco, and torvum; melons such as cucumbers, pumpkins, melons, and watermelons; vegetables such as cabbage, broccoli, Chinese cabbage, and kale; raw and spicy vegetables such as shiso, celery, parsley, and lettuce; onions such as green onions, onions, and garlic; other fruit vegetables such as strawberries and melons; taproots such as radishes, turnips, carrots, and burdock; tubers such as taro, cassava, potato, sweet potato, and Chinese yam; grains such as rice, corn, wheat, sorghum, barley, rye, wheatgrass, and buckwheat; soybeans, adzuki beans, mung beans, cowpeas, and yams. Examples of such plants include beans such as columbine, peanut, pea, and broad bean; soft vegetables such as asparagus, spinach, and mitsuba; flowers such as lisianthus, rose, stock, carnation, and chrysanthemum; turf grasses such as bentgrass and zoysiagrass; oil crops such as rapeseed, camelina, rapeseed, jatropha, sesame, and perilla; fiber crops such as cotton, rush, and hemp; forage crops such as clover, dent corn, and alfalfa; deciduous fruit trees such as apple, pear, grape, and peach; citrus fruits such as satsuma mandarin, orange, lemon, and grapefruit; and woody plants such as azalea, azalea, cedar, poplar, and rubber tree. The promoter may be site-specific or non-site-specific. The site-specific promoter may be, for example, one that controls specific expression in leaves or roots.
In some embodiments, the desired activity of the modified promoter may be a higher gene expression induction activity than the original promoter sequence, or a lower gene expression induction activity than the original promoter sequence. The gene expression induction activity higher than the original promoter sequence may be, for example, 1.1 to 1000 times the original activity, for example, 1.1 times, 1.2 times, 1.5 times, 2 times, 3 times, 4 times, 5 times, 10 times, 20 times, 50 times, 100 times, 200 times, 300 times, 400 times, 500 times, 600 times, 700 times, 800 times, 900 times, or 1000 times. The gene expression induction activity lower than that of the original promoter sequence can be, for example, anywhere from 90% to 0.01% of the original activity, such as 90%, 80%, 50%, 10%, 5%, 1%, 0.5%, 0.1%, 0.05%, 0.04%, 0.03%, 0.02%, or 0.01%.

In some embodiments, the desired activity of the modified promoter may be determined by comparison with the activity of another reference promoter sequence, rather than the original promoter sequence.

Alternatively, the desired activity of the modified promoter may be determined by stratification, such as high expression, medium expression, and low expression. In some embodiments, any number or range of stratifications may be performed.

In some embodiments, the modification of the promoter sequence may be by deletion, substitution, or insertion of a portion of the sequence. In some embodiments, the modification of the promoter sequence may be by deletion of a portion of the sequence by making two truncations in the promoter sequence.

In some embodiments, the original promoter sequence to be modified may be prepared by obtaining it from a database such as GenBank. Examples of promoter sequences include any DNA region including the transcription start point within a DNA region of -9900 to +100, preferably -4950 to +50 or -2975 to +25, more preferably -1995 to +5 around the transcription start point of the target gene. Alternatively, any DNA region between the transcription start point of the target gene and the transcription end point of the adjacent gene may be used. In other words, a region including the core promoter and its upstream sequence may be used as the promoter sequence. The length of the sequence upstream of the core promoter may be, for example, at least 200bp, 400bp, 600bp, 800bp, 1000bp, 1200bp, 1400bp, 1600bp, 1800bp, 2000bp, 3000bp, 4000bp, 5000bp, 6000bp, 7000bp, 8000bp, or 9000bp. If there are multiple transcription start points, the multiple transcription start points can be selected. In addition, any range of 100 to 1800 bases included in -1995 to +5 around the transcription start point may be used. The length of the promoter (number of base pairs: bp) can be, for example, at least 200 bp, 400 bp, 600 bp, 800 bp, 1000 bp, 1200 bp, 1500 bp, 2000 bp, 3000 bp, 4000 bp, 5000 bp, 6000 bp, 7000 bp, 8000 bp, or 9000 bp, and the range can be 100 to 10000 bp, preferably 200 to 5000 bp, and more preferably 500 to 3000 bp.

In some embodiments, a set of multiple modified promoter sequences is generated that can be created using genome editing techniques based on the promoter sequence to select a modified promoter sequence with a desired activity.

Genome editing techniques that can be used include ZFNs, TALEN, and CRISPR-Cas systems. Examples of CRISPR-Cas systems include CRISPR-Cas3 (Class 1 Type I), Cas9 (Class 2 Type II), Cas12a (Class 2 Type V), Cas12f (Cas14a) (Class 2 Type V), and Cas13a (Class 2 Type VI), regardless of their functional mechanism or classification. Cas proteins derived from various bacteria can also be used. For example, Cas9 includes SpCas9 derived from Streptococcus pyogenes, SaCas9 derived from Staphylococcus aureus, FnCas9 derived from Francisella novicida, CjCas9 derived from Campylobacter jejuni, and Cas12a includes AsCas12a (AsCpf1) derived from Acidaminococcus sp., LbCas12a (LbCpf1) derived from Lachnospiraceae bacterium, and ErCas12a derived from Eubacterium rectale, and the origin is not limited. In addition, modified nucleotide sequences encoding Cas proteins or amino acid sequences of Cas proteins, fused with other proteins, functional domains, peptides, or amino acid sequences, or modified with compounds may also be used.

In the CRISPR-Cas system, Cas nuclease and guide RNA (gRNA) are used to cleave the target sequence. Guide RNA consists of two elements, crRNA and tracrRNA, which may exist as individual RNAs or be linked together to exist as a single-stranded RNA. Single-stranded guide RNA is also called sgRNA. A base sequence of 1 to 10 bases, 10 to 50 bases, 50 to 100 bases, or 100 to 500 bases may be added to the 5' and 3' ends of the guide RNA. Guide RNA binds complementarily to a specific DNA sequence and guides Cas nuclease to a specific position in the genome. This allows Cas nuclease to cleave a specific DNA sequence and enable genome editing. For example, when using CRISPR, a list of cleavable positions is created by searching for PAM sequences in the promoter sequence. The PAM sequence depends on the Cas protein used; for example, NGG for SpCas9, NGRRT or NGNRRN for SaCas9, NNNNGATT for NmeCas9, NNNNRYAC for CjCas9, TTTV for LbCas12a(Cpf1), TTTV for AsCas12a(Cpf1), TTN for AacCas12a(Cpf1), ATTN, TTTN, or GTTN for BhCas12b v4. However, these PAM sequences may contain mismatches of 1, 2, 3, 4, 5, 6, 7, or 8 bases if the Cas protein can recognize the sequence. The presence of the PAM sequence is important for increasing the accuracy and specificity of genome editing in the CRISPR system. Without the PAM sequence, Cas nuclease cannot bind to the DNA sequence specified by the guide RNA. This characteristic makes it possible to precisely target a specific gene region for genome editing. Also, as described above, by using different types of Cas nucleases, regions with different PAM sequences can be targeted. For example, typically, the cleavage position is 3 to 4 bases from the PAM sequence for Cas9, and 18 to 23 bases from the PAM sequence for Cas12a. A person skilled in the art can design an appropriate guide RNA based on the PAM sequence in the target genome. Typically, about 50 cleavage positions are found in 2000 bases. For example, there are nC ₂ combinations that specify two of the n cleavage positions. In other words, there are 1250 possible combinations for the 50 cleavage positions. Thus, in some embodiments, in order to select a modified promoter sequence having a desired activity, a set of base sequences that is truncated (deleted) between the two cleavage positions for nC ₂ combinations of PAM sequences in the promoter sequence to be modified can be generated. Thus, in some embodiments, a set of multiple modified promoter sequences is generated by sequence deletion caused by cleavage induced by a combination of guide RNA sequences designed based on two PAM recognition sequences.

Other promoter editing methods that can be used include deleting the space between two nicks created by Cas9 nickase, using base editing to replace a targeted base in the genome with another base, and using a prime editor that can delete, insert, or replace any base sequence consisting of 1 to 20 bases, 20 to 50 bases, or 50 to 100 bases in the genome.

The number of different sequences included in the set of modified promoter sequences may be at least 5, 10, 20, 50, 100, 150, 200, 300, 500, 700, 1000, 1200, 1500, 2000, 3000, 4000, or 5000. To fully explore sequences that affect activity, it is preferable that the number of different sequences included in the set of modified promoter sequences is 1000 or more. In addition, to ensure a sufficient number of different sequences included in the set of modified promoter sequences, it is preferable that the length of the original promoter sequence is, for example, 1800 bp or more.

After a set of multiple modified promoter sequences is generated, in some embodiments, the activity of each modified promoter sequence included in the set of modified promoter sequences generated is predicted by a machine learning model. For example, for each generated base sequence, 170 bases on the 3' end are obtained and input to the machine learning model. In some embodiments, 100 to 250 bases on the 3' end may be obtained and input to the machine learning model. In some embodiments, a portion of the sequence included in the 170 bases on the 3' end, for example, 100 to 169 bases, may be input. A properly trained machine learning model can output the strength (transcriptional activity) of a given base sequence as a core promoter.

In some embodiments, the activity of each modified promoter sequence in the generated set of modified promoter sequences is visualized on a computer display. Visualization can be performed, for example, as shown in FIG. 8 or FIG. 9.

Then, by selecting a modified promoter sequence predicted to have the desired activity from the prediction results of the machine learning model, a promoter sequence modified to have the desired activity can be obtained.

In some embodiments, the machine learning model may be a regression model trained to predict gene expression induction activity from a promoter sequence using gene expression induction activity data of multiple promoter sequences in plant cells as teacher data. In some embodiments, data from Jores et al. (2021) may be used for learning (Jores, T., Tonnies, J., Wrightsman, T. et al. Synthetic promoter designs enabled by a comprehensive analysis of plant core promoters. Nat. Plants 7, 842-855 (2021). https://doi.org/10.1038/s41477-021-00932-y). Thus, one aspect of the present disclosure relates to a method for generating a machine learning model that predicts gene expression induction activity in plant cells from a promoter sequence, comprising training the model using gene expression induction activity data of multiple promoter sequences in plant cells as teacher data. In some embodiments, a transformer-based pre-trained basic model may be used as the model to be trained.

In some embodiments, deep learning models such as Transformers, particularly models based on BERT, may be used to build the machine learning model. BERT (Bidirectional Encoder Representations from Transformers) is one of the machine learning models widely used in natural language processing (NLP). It was developed by Google in 2018 and has shown remarkable performance in understanding and generating text. The main features of BERT include 1) bidirectional context understanding, 2) use of Transformer architecture, 3) pre-training and fine-tuning, and 4) diverse application possibilities. First, BERT is a "bidirectional" model that understands words in a given text in both left and right context. While conventional models only consider context in one direction (from left to right or vice versa), BERT can comprehensively understand the entire text. In addition, BERT is built on a neural network architecture called Transformer. Transformer uses an "attention mechanism" to capture the relationship between each word in the input text. This enables more complex and sophisticated text understanding. In addition, BERT is "pre-trained" on a large amount of text data to acquire a general understanding of language. It can then be "fine-tuned" for a specific task (e.g., sentiment analysis or question answering) to optimize it for a specific use. BERT is applicable to a wide range of NLP tasks, and is used in a wide range of fields, including text classification, question answering, sentiment analysis, and machine translation.

The BERT model is trained on two main tasks. The MLM task is a task that randomly "masks" words from the input text and asks BERT to predict the masked words. The main goal of this task is to have BERT use context to understand the meaning of words. To take into account bidirectional context, the model uses information from the entire sentence to predict the masked words. The NSP task asks BERT to determine whether two sentences are consecutive. The model is given a sentence (A) and another sentence (B) and is asked to predict whether B comes immediately after A. Half the time, B is the actual sentence that follows A, and the other half is a randomly chosen unrelated sentence. The goal of the NSP task is to teach BERT to understand the relationships between sentences. This ability is especially useful in tasks where understanding the connections and semantic flow of sentences is important, such as question answering and text summarization. These tasks train BERT to understand not only at the word level, but also the relationships between whole sentences and multiple sentences. As a result, BERT is able to perform well in a variety of natural language processing tasks.

In some embodiments, the machine learning may utilize DNABERT, a transformer-based pre-trained base model (Yanrong Ji, Zhihan Zhou, Han Liu, Ramana V Davuluri, DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome, Bioinformatics, Volume 37, Issue 15, August 2021, Pages 2112-2120). DNABERT is a pre-trained bidirectional encoder representation that can capture a global understanding of genomic DNA sequences based on the context of upstream and downstream base sequences. DNABERT does not perform the NSP task used in BERT, but only trains on the MLM task. A certain percentage of the DNA sequence is masked, and the k-mer token of the masked site is predicted. The training data used for pre-training DNABERT is DNA sequences sampled from the human genome. The inventors have demonstrated that by fine-tuning a pre-trained basic model such as DNABERT, it is possible to generate a machine learning model that predicts gene expression induction activity in plant cells from a promoter sequence. The machine learning model disclosed herein can handle not only binary classification problems but also regression problems, and as a result, can predict the level of gene expression induction activity.

One aspect of the present disclosure relates to a method for genome editing of a cell to regulate the expression level of a desired gene. In some embodiments, the method may include obtaining a modified promoter sequence having a desired activity by a method for obtaining a promoter sequence modified to have a desired activity according to the present disclosure, preparing a cell to be subjected to genome editing, and editing the genome of the cell to produce the modified promoter sequence.

In some embodiments, the cells are plant cells, including cells from a wide range of plants including bryophytes, ferns, gymnosperms, and angiosperms such as magnolias, monocots, and eudicots (Rosa I, Rosa II, Chrysanthemum I, Chrysanthemum II, and outgroups thereof). More specific examples of plants include eggplants such as tomatoes, bell peppers, chili peppers, eggplants, tobacco, and torvum; melons such as cucumbers, pumpkins, melons, and watermelons; vegetables such as cabbage, broccoli, Chinese cabbage, and kale; raw and spicy vegetables such as shiso, celery, parsley, and lettuce; onions, onions, and garlic; other fruit vegetables such as strawberries and melons; taproots such as radishes, turnips, carrots, and burdock; tubers such as taro, cassava, potato, sweet potato, and Chinese yam; grains such as rice, corn, wheat, sorghum, barley, rye, wheatgrass, and buckwheat; soybeans, adzuki beans, mung beans, cowpeas, and kidney beans. These include legumes such as beans, peanuts, peas, and broad beans; soft vegetables such as asparagus, spinach, and mitsuba; flowers such as lisianthus, roses, stocks, carnations, and chrysanthemums; turfgrass such as bentgrass and zoysiagrass; oil crops such as rapeseed, camelina, rapeseed, jatropha, sesame, and perilla; fiber crops such as cotton, rush, and hemp; forage crops such as clover, dent corn, and alfalfa; deciduous fruit trees such as apples, pears, grapes, and peaches; citrus fruits such as satsuma mandarins, oranges, lemons, and grapefruit; and woody plants such as azalea, azalea, cedar, poplar, and rubber tree.

One aspect of the present disclosure relates to a polynucleotide having promoter activity, the polynucleotide having a sequence obtained by the method of the present disclosure. Such a polynucleotide may have, for example, the sequence of SEQ ID NO: 2 or 4.

One aspect of the present disclosure relates to a method for producing a genome-edited plant in which the expression level of a desired gene is regulated. In some embodiments, the method may include editing the genome of a desired plant cell by a cellular genome editing method for regulating the expression level of a desired gene according to the present disclosure, and obtaining an individual plant derived from the genome-edited cell. Furthermore, one aspect of the present disclosure relates to a genome-edited plant produced by the production method according to the present disclosure.

When obtaining genome-edited cells, the genome editing enzyme can be introduced into plant tissues or cells in the form of DNA, RNP, or protein. Examples of sites where the genome editing enzyme can be introduced in plants include flowers (egg cells, pollen, petals, etc.), stems (cambium, pith, cortex, etc.), leaves (including leaf primordia), roots, shoot tips, lateral buds, flower buds, root tips, protoplasts, etc.

The introduction method is not particularly limited and can be selected appropriately depending on the plant species and the target cells/tissues. Examples of introduction methods include the Agrobacterium method, the particle gun method, the whisker method, the nanopipette method, and virus-mediated nucleic acid delivery.

In some embodiments, obtaining an individual plant derived from a genome-edited cell may include, for example, steps such as (a) culturing the genome-edited cell or tissue containing the edited cell under appropriate culture conditions, (b) proliferating the cells to form a callus (an undifferentiated cell mass), (c1) inducing a shoot (a sprout) from the callus, (c2) directly forming a shoot from the tissue containing the edited cell, (d) inducing roots in the shoot to regenerate a plant, (e1) obtaining next-generation seeds from the regenerated plant, or (e2) propagating a part of the regenerated plant by cutting.

One aspect of the present disclosure relates to an information processing device that predicts promoter sequences that have been modified to have a desired activity (see FIG. 11).

In some embodiments, the present device 010 may be an information processing device including a sequence input unit 011 that accepts input of an original promoter sequence to be modified, a modified sequence generation unit 012 that generates a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence, an activity prediction unit 013 that predicts the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model, and a sequence selection unit 014 that selects a modified promoter sequence predicted to have a desired activity. The sequence input unit 011 may be connected to an input device such as a communication device with an external network or a keyboard. The sequence selection unit 014 may be connected to an output device such as a display or a printer, a communication device with an external network, or the like. A person skilled in the art would be able to understand the configuration required for the present device to carry out the sequence acquisition method according to the present disclosure in light of the disclosure of this specification.

One aspect of the present disclosure relates to a non-transitory computer-readable medium having instructions or programs stored thereon.

In some embodiments, the computer-readable medium may be a computer-readable medium having stored thereon instructions or a program that, when executed by a processor, can execute the following steps: step S100 of accepting an input of an original promoter sequence to be modified; step S110 of generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence; step S120 of predicting the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model; and step S130 of selecting the modified promoter sequence predicted to have the desired activity. FIG. 12 shows an exemplary flowchart of such a program.

Step S110 of generating a set of multiple modified promoter sequences that can be created by genome editing technology may further include step S112 of comprehensively searching for cleavable sites (PAM sequences) by Cas9 from the input sequences, step S114 of combinatorially selecting all two of the searched cleavable sites (PAM sequences), and step S116 of deleting the sequence between the two cleavage sites selected for all combinations to generate a set of modified promoter sequences. In addition, step S120 of predicting the activity of each modified promoter sequence using a machine learning model may be performed using only the core promoter portion (150 to 200 bp, for example about 170 bp, from the 3' end) of each modified promoter sequence.

One aspect of the present disclosure relates to a computer program that predicts promoter sequences that have been modified to have a desired activity.

In some embodiments, the program may be a program that causes a computer to perform the functions of accepting input of an original promoter sequence to be modified, generating a set of multiple modified promoter sequences that can be created using genome editing technology based on the promoter sequence, predicting the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model, and selecting a modified promoter sequence predicted to have a desired activity.

FIG. 13 is a schematic diagram showing an exemplary embodiment of the device of the present disclosure. In FIG. 13, 100 is a computer, which includes a control unit 101, a memory unit 102, a peripheral device I/F unit 103, an input unit 104, a display unit 105, and a communication unit 106, which are connected by a bus 110. Note that this configuration is merely an example, and various configurations can be adopted as appropriate.

The control unit 101 is composed of a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The CPU loads and executes programs stored in the memory unit 102, the ROM, a recording medium, etc. into a work memory area on the RAM, drives and controls each device connected via the bus 110, and realizes the processing performed by the computer. The ROM is a non-volatile memory that holds the boot program of the computer 100, programs such as the BIOS, and data, etc. The RAM is a volatile memory that temporarily holds programs, data, etc. loaded from the memory unit 102, the ROM, a recording medium, etc., and has a work area that the control unit 101 uses when performing various processing. The memory unit 102 is, for example, a hard disk drive (HDD), and stores the programs executed by the control unit 101 and various other data.

The peripheral device I/F (interface) unit 103 is a port for connecting the computer 100 to a peripheral device. The peripheral device I/F unit 103 is configured with USB, IEEE1394, RS-232C, etc. The connection with the peripheral device may be wired or wireless. The input unit 104 has input devices such as a keyboard, a pointing device such as a mouse, and a numeric keypad, and issues operational instructions, operation instructions, data input, etc. to the computer 100. The display unit 105 is a logic circuit or device driver for displaying videos, images, etc. on a display device such as a liquid crystal panel. The input unit 104 and the display unit 105 can also be configured as an integrated touch display.

The communication unit 106 has a communication control device, a communication port, etc., and is a wired or wireless communication interface that mediates communication with the network 120. The bus 110 is a communication path that mediates the transmission and reception of control signals, data signals, etc. between each device. The network 120 may further be connected to an external server 130 and net storage 140.

For example, a program recorded on a computer-readable medium according to the present disclosure can be loaded into the device of FIG. 13, and the computer can function as an information processing device including a sequence input unit that accepts input of an original promoter sequence to be modified, a modified sequence generation unit that generates a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence, an activity prediction unit that predicts the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model, and a sequence selection unit that selects a modified promoter sequence predicted to have a desired activity.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Although any methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present invention, some possible, preferred methods and materials are now described. All publications mentioned herein are incorporated herein by reference to disclose and describe the methods and/or materials in connection with which the publications are cited. It is understood that the present disclosure supersedes the disclosure of the incorporated publications in the case of any conflict.

When a range of values is described, unless the context clearly indicates otherwise, it is understood that each intervening value, to one-tenth of the unit of the lower limit, between the upper and lower limits of that range is also specifically disclosed. Each smaller range between any stated or intervening value in a stated range and any other stated or intervening value in that stated range is also encompassed by the disclosure. The upper and lower limits of these smaller ranges may be independently included or excluded, and each range in which either, either or both limits are included in the smaller range is also encompassed by the invention, but with the reservation of any specifically excluded limit in the stated range. When a stated range includes one or both limits, ranges excluding either or both of the included limits are also included by the invention. The term "about" with respect to numerical values means within 5%.

The embodiments described herein are intended to be merely exemplary, and those skilled in the art will be able to make numerous variations and modifications without departing from the spirit of the present invention. Also, certain variations and modifications may produce less than optimal results, but still be satisfactory. All such variations and modifications are intended to be within the scope of the present invention as defined by the appended claims. Also, any combination of the components disclosed herein, and any conversion of the expressions of the present disclosure between methods, devices, systems, computer programs, data structures, recording media, etc., are also valid aspects of the present disclosure. Thus, details described with respect to the methods of the present disclosure may be applied to systems, computer programs, data structures, recording media, etc.

Example 1: Construction of a machine learning model A machine learning model that predicts gene expression induction activity in plant cells from a promoter sequence was constructed as follows. First, the data shown in the paper by Jores et al. was obtained as the learning source data (Jores, T., Tonnies, J., Wrightsman, T. et al. Synthetic promoter designs enabled by a comprehensive analysis of plant core promoters. Nat. Plants 7, 842-855 (2021). https://doi.org/10.1038/s41477-021-00932-y). The overview of the dataset was a dataset of about 70,000 items, consisting of a 170-base DNA sequence and an expression intensity pair.

For machine learning, we used DNABERT, a transformer-based pre-trained basic model. DNABERT has been reported, for example, in the paper by Ji et al. mentioned above. DNABERT is primarily focused on dealing with binary classification problems, and has been shown to be able to determine, for example, whether a given base sequence has a splicing site or not. A binary classification problem is a problem of classifying into two categories, such as 1 if it has a splicing site and 0 if it does not. In contrast, a problem such as "predicting downstream gene expression levels for a given base sequence" is called a regression problem in contrast to a classification problem, and requires the output of continuous values. Therefore, in order to handle continuous values in DNABERT, a class was added to the transformer and the model was trained.

Example 2: Prediction and classification of promoter strength by trained model Promoter sequences are often located upstream of genes and control the transcription amount (expression amount) of genes located directly downstream. The expression amount of a gene is considered to be controlled by the promoter sequence. The expression amount of a downstream gene is defined as the promoter strength. Each promoter sequence is considered to show a unique strength.

As explained in Example 1, a machine learning model was constructed to predict transcription activity based on the base sequence. Using this model, a list was created that evaluated the transcription activity of all Arabidopsis genes.

All promoters were classified into three groups based on their predicted transcriptional activity:
1. A "high expression group" predicted to show high promoter strength;
2. a "low expression group" predicted to have promoter strengths that are very low or almost non-existent; and 3. a "moderate expression group" showing promoter strengths intermediate between those.

Example 3: Comparison of model predictions and actual gene expression Seven promoters were selected from the high expression group, eight promoters from the medium expression group, and four promoters from the low expression group, and artificial synthetic genes were created by connecting them to the luciferase (LUC) gene.

The gene structures are shown in Figure 1. The promoters have 19 different sequence patterns. The downstream LUC gene and the upstream cauliflower mosaic virus (CaMV) 35S enhancer sequences are common to all artificial genes.

The plasmid vector carrying this artificial gene was transfected into protoplasts taken from Arabidopsis leaves using the polyethylene glycol method. The protoplasts expressed the LUC gene at various levels. To evaluate the LUC expression level, luminescence intensity was measured using a plate reader. Figure 2 shows the relationship between the measurement results and the predicted promoter strength. A scatter plot shows the relationship between the actual measured expression level of the LUC gene and the predicted promoter strength. The vertical axis shows the predicted transcription activity. The higher the value, the greater the expected expression level of the downstream gene. The horizontal axis shows the actual measured LUC expression intensity, which was logarithmically transformed. However, this was standardized using the expression level of the CaMV promoter, which was the positive control (P control).

The results shown in Figure 2 show that the actual measured values are divided into three groups, consistent with the groups based on the predicted values. In this way, a clear correlation was confirmed between the predicted values by the constructed machine learning model and the actual measured values. This figure shows the high level of predictability. In this figure, linear convergence means that the predicted values and actual measured values match, and the model has high predictive performance. When this predictive performance was evaluated using the correlation coefficient R, it was found to be R = 0.9296673.

Next, based on the correlation shown in the scatter plot above, we calculated the LUC expression level corresponding to the predicted transcription activity by linear regression prediction. A comparison of these is shown in Figure 3 below.

The luminescence intensity of LUC is plotted on the vertical axis. The luminescence intensity of the P control is set to 1, and is shown as a relative value. The cauliflower mosaic virus (CaMV) promoter, known for its high transcriptional activity, was used as the P control. The horizontal axis shows the numbers of 19 types of promoters. Promoter numbers 3 to 10 were classified into the "high expression" group, which was predicted to have high expression intensity, promoter numbers 13 to 21 into the "moderate" group, and promoter numbers 22 to 25 into the "low expression" group. The black bars show the actual values (expression levels) of the LUC assay. The white bars show the expression intensity predicted by the machine learning model. In Figure 3, the black bars show a relatively high expression level in the "high expression" group and a low expression level in the "low expression" group. The "moderate" group showed intermediate values.

Next, we checked how accurately the expression levels were predicted for individual promoters. Comparing the black and white bars for each promoter, we can see that promoters 3, 6, 10, and 18 were predicted with a high degree of agreement. On the other hand, for promoter 4, the actual value was greater than the predicted value. For promoter 5, the opposite was true, with the predicted value greater than the actual value.

In summary, the constructed model was able to predict general trends such as high and low expression. Low-expressing promoters were generally predicted with high accuracy. In the high-expressing group, expression could be predicted within the range of 50% to 200% of the actual measured value.

In this way, the machine learning model disclosed herein can predict gene expression levels based on promoter base sequences.

Example 4: Prediction and demonstration of promoters with reduced activity Next, we devised a method to reduce the transcriptional activity of promoters. Each promoter has many Cas9 PAM recognition sequences (NGG). We searched these sequences and listed approximately several thousand possible editing patterns.

More specifically, the steps are as follows:
1. Comprehensively search for sites (PAM sequences) that can be cleaved by Cas9 within a given base sequence;
2. Choose two locations from the list;
3. Cut at the two cut sites and simulate the new core promoter sequence that appears after repair;
4. Estimate the core promoter strength for the simulated sequences;
5. The above steps are carried out for all possible cleavage site pairs;
6. Among all the simulated core promoters, choose the one showing the lowest score (or the highest score if you are selecting promoters with increased transcriptional activity).

From the above, it is possible to determine which cleavage site should be selected when enhancing/suppressing the expression level. In this way, by predicting the promoter transcription activity for the base sequence of each editing pattern, editing patterns that can reduce the promoter transcription activity were searched for. The results are shown in Figure 4. Editing patterns that can reduce promoter strength were searched for for promoter numbers 3, 4, 5, 6, 9, 10, and 21. The black triangular plot shows the predicted LUC assay score calculated based on the predicted new promoter strength. As a result, it was predicted that the LUC expression level would be approximately 14% to 1% of the transcription activity compared to the original promoter.

Next, genes with these theoretical promoter sequences were artificially synthesized, and protoplasts were transfected in the same manner, and the expression levels were measured using a plate reader. The results are shown in Figure 5. The actual measured values of the expression levels of LUC for promoters with new sequences are shown in shaded bars superimposed on Figure 4. Because the expression levels were very low, the vertical axis has been enlarged. Promoter number 5 was a missing value. For all six promoters for which measured values were obtained, a significant decrease in expression level was observed, which matched the predicted values. For example, the sequence of promoter number 3 was
CGGAAACTTGTCACTTCCTTTACATTTGAGTTTCCAACACCTAATCACGACAACAATCATATAGCTCTCGCATACAAACAAACATATGCATGTATTCTTACACGTGAACTCCATGCAAGTCTCTTTTCTCACCTATAAATACCAACCACACCTTCACCACATTCTTCACT (SEQ ID NO: 1)
The predicted transcription activity was 5.43, and the LUC expression level predicted by linear regression from this value was 111.5% compared to the P control. The actual LUC expression level was 68.6%. In contrast, the edited sequence of promoter 3 was
GAAACTGATTAGCTCCTATCAGTTCAGCAAACCACAAGCTGAAGAATCCAAGACTTGAGAAACAAATTTACAAAAGCCCATGTTCCAATCAAAACTGTTACCAAACATCTGAAATAGATCTAAATGAGCGTTGGTATAATTGAAACTTACCGAAGGCCCACATTCTTCAC (SEQ ID NO: 2)
This occurs when the 845 bases between -852 and -7 are truncated. The predicted transcriptional activity was 0.393, and the LUC expression level predicted by linear regression from this value was 7.98% of the P control. The actual LUC expression level was 3.41% of the P control.

In this way, expression can be reduced by editing the promoter sequence based on the predictions of the machine learning model disclosed herein.

Example 5: Prediction and Demonstration of Promoters with Increased Activity Finally, a method for increasing promoter activity was devised. As in Example 4, among the promoter base sequences of possible editing patterns, those in which the predicted value of promoter strength increases after editing were searched for. The results are shown in Figure 6. For promoter numbers 13, 17, 22, 23, 24, and 25, those with increased promoter strength were selected, and the predicted expression levels are shown with white circles. It was predicted that the LUC expression level would be 7.4 to 125 times higher than the original promoter.

Next, genes with these theoretical promoter sequences were artificially synthesized, and protoplasts were transfected in the same manner, and the expression levels were measured using a plate reader. The results are shown in Figure 7. Of the six promoters, five showed increased expression levels compared to the original promoter, but only promoter number 13 showed an increase in expression level beyond the predicted value. For example, the sequence of promoter number 13 is
TCAAGCAATCATTATCGACTACGGTCGTTCGTTAAAGATCATGCATGTGCTTAGTGGCAATACCCTACGCATCTTGATTCGTTACTGCGGCACGTGTCATGACCATGCACATGAATGATGATTAATGTTTAGTACATATAATGTTCACGCAAACGCATAGTGTTAGGAAA (SEQ ID NO: 3)
The predicted transcription activity was 2.00, and the LUC expression level predicted by linear regression from this value was 6.94% compared to the P control. The actual LUC expression level was 6.60% compared to the P control. In contrast, the edited sequence of promoter 13 was
GAAACTTGAAAATCAAATCAGTGAGTCGCAAGTAAGACTTTGTGGTTGTTGTATCAGATTTCGCCGTGCGCATCTTGATTCGTTACTGCGGCACGTGTCATGACCATGCACATGAATGATGATTAATGTTTAGTACATATAATGTTCACGCAAACGCATAGTGTTAGGAA (SEQ ID NO: 4)
The predicted transcriptional activity was 4.18, and the LUC expression level predicted by linear regression from this value was 36.7% compared to the P control, while the actual LUC expression level was 69.5% compared to the P control.

In this way, expression can be increased by editing the promoter sequence based on the predictions of the machine learning model disclosed herein.

Example 6: Program for changing promoter function by genome editing The present inventors developed a program used for changing promoter function by genome editing in the following procedure.

1. Input the base sequence X. For example, the 2000 bases from -1995 to +5 around the transcription start site of the target gene.
2. Search for the PAM sequence in the base sequence X. (Cas9: NGG or Cas12a: TTTV sequence)
3. Based on the PAM sequence, a list of cleavage positions is stored in the array. Here, for Cas9, the cleavage positions are 4 bases downstream of the PAM sequence, and for Cas12a, the cleavage positions are 18 bases downstream of the PAM sequence. For example, about 50 cleavage positions are found in a 2000-base sequence.
4. Of the n cleavage sites, there are nC ₂ combinations for specifying two sites. For example, for 50 cleavage sites, there are 1250 possible combinations. For each of the nC ₂ combinations, a base sequence is generated that is truncated (deleted) between the two cleavage sites.
5. For each generated base sequence, obtain 170 bases from the 3' end and input them into the learning model.
6. The learning model outputs the strength (transcriptional activity) of the given base sequence as a core promoter.
7. In summary, we obtain three types of information:
A. Transcription activity of base sequence X B. A set of cleavage sites where the promoter activity of the sequence obtained by truncating the region between the two cleavage sites is greater than that of A. C. A set of cleavage sites where the promoter activity of the sequence obtained by truncating the region between the two cleavage sites is less than that of A. 8. For example, to increase the expression of a target gene, two types of guide RNAs targeting the cleavage sites obtained in 7.B. are introduced into cells at the same time, and genome editing is performed using CRISPR-Cas9 or -Cas12a, which is expected to result in an individual with a deletion between the two cleavage sites.

In particular, it should be noted that the simultaneous use of two different guide RNAs is intended to induce small to medium-sized deletions within the category SDN1.

Example 7: Visualization of predicted promoter activity Finally, the program described in Example 6 gives us a set of predicted promoter activity values for the positions of the two guide RNAs and the deleted sequence, but this is not intuitively understandable. Therefore, the present inventors have devised a method to visualize these values.

As one of the visualization methods, we devised a method to analyze the whole picture of the input base sequence X. The vertical axis shows the predicted promoter strength, and the horizontal axis shows the position of the base sequence. Each plot shows the value when the DNA sequence at a certain position is cut out with a window of 170 bases and the transcription activity is predicted. For example, the point at position 0 on the horizontal axis is the transcription activity predicted by taking 170 bases from the 5' end of the input base sequence X. As a result, the transcription activity was 1.24 (this 170 base sequence is not adjacent to a gene, so it is unlikely to function as a core promoter, but if a gene is connected directly below this 170 base sequence, it is predicted that it will transcribe with a strength of 1.24). Therefore, it was plotted at the position of the point (0, 1.24). Next, the window was slid by one base, and the base sequence from the 2nd base to the 171st base of the base sequence X was taken and predicted in the same way. As a result, the transcription activity was estimated to be 1.37. Therefore, it was plotted at the point (1, 1.37). This process was repeated to plot a total of 1,830 points (Figure 8).

In Figure 8, the point on the far right (1829, 1.47) is an evaluation of the 170 bases immediately upstream of the gene. If the sequence between the point with a value higher than this value and the 3' end can be removed by genome editing, it is expected that gene expression will increase. Similarly, it is thought that gene expression can also be decreased. Using this diagram allows us to get an overview of the entire picture of sequence X, and makes clear which part the guide RNA should hybridize to in order to obtain the desired expression level.

The visualization method described above may be effective in getting a rough overall picture of the target sequence. However, it does not show the strength of the promoter and the relative positions of the two guide RNAs that should be designed. To overcome this issue, we devised a different visualization method. Figure 9 below shows a visualization using a different method.

A guide RNA designed close to the target gene is called a proximal guide, and a guide RNA designed farther away is called a distal guide. Visualization was done based on the positions of designable guide RNAs (position of the PAM sequence). However, only the range of 1000 bases immediately upstream of the gene is shown.

The horizontal axis shows the position of the base hybridized by the distal guide in sequence X. In other words, it shows the distance between the transcription start point of the gene and the distal guide. Similarly, the vertical axis shows the distance between the proximal guide and the transcription start point of the gene. Here, the proximal guide is never designed to be farther away than the distal guide, so half of the plot in Figure 9 is hidden. At points on the line y = x, the proximal guide and the distal guide are both located at almost the same distance from the gene, indicating that they are close to each other. In this case, the bases deleted are small, so editing is minimal and desirable. 170 bases were taken from the 3' end of the new sequence created when the sequence between the proximal guide and the distal guide was deleted, and the transcription activity was predicted. The color of the plot is changed based on the predicted value. In the color chart, light gray indicates that the transcription activity is predicted to be high, and black indicates that the transcription activity is predicted to be low.

For example, let us use the plot of point (849, 17) as an example. The distal guide position of this point is 849 bases away from the gene. The proximal guide position is 17 bases away from the gene. From base sequence X, the 832 bases between these two points are deleted to create base sequence X'. 170 bases were taken from the 3' end of base sequence X', and the predicted transcription activity was 4.02. A light gray plot was made according to the color chart. The above operations were performed for all combinations of distal and proximal guides to create the plot.

If you want to increase the expression level of this gene, you can select a point with a transcription activity greater than the original 2.34, and read the position of the guide RNA to be designed from the positions of the distal and proximal guides. For example, if you select point (849, 17) or point (841, 29), you can expect high transcription activity. Also, if you select point (432, 110) or point (271, 77), you can expect low transcription activity. Visualization in this way may make it easier to design guide RNAs that achieve the desired promoter structure.

After obtaining the position of the target sequence of the guide RNA in this way, the base sequence of that position was obtained, and 20 or 23 bases were obtained to design the guide RNA. The sequence of this guide RNA was evaluated and verified using tools such as CRISPOR to evaluate the level of specificity and off-targets. If a suitable guide RNA was successfully designed, a vector carrying the guide RNA was created using artificial DNA synthesis and PCR. From this point on, genome editing can be performed according to standard methods.

Example 8: Example of soybean gene promoter The training data for the machine learning model was promoter sequences from Arabidopsis thaliana, sorghum, and corn. Therefore, it was unclear whether the prediction system according to the present invention would function as expected for promoter sequences from organisms other than these three species. Therefore, we investigated how to enhance the transcription activity of a specific soybean gene. The results are shown in FIG. 10.

The transcription activity of a 170-base sequence in the core promoter of a soybean gene was predicted to be 0.927494. Based on this value, the predicted LUC expression level was 1.26% of the positive control. This promoter sequence was artificially synthesized and introduced into protoplasts isolated from Arabidopsis leaves, and a LUC assay was performed. The result was 4.15%.

Based on this promoter, two types of promoter sequences were created after genome editing to increase gene expression. They are called edit1 and edit2. The transcription activity of edit1 was predicted to be 2.950119. The LUC expression level was predicted to be 13.2%. The actual measured value of the LUC assay for this promoter was 19.0%.

The transcriptional activity of edit2 was predicted to be 2.432977. The LUC expression level was predicted to be 7.27%. The actual LUC assay value for this promoter was 41.6%.

010: Information processing device 011: Sequence input section 012: Modified sequence generation section 013: Activity prediction section 014: Sequence selection section 100: Computer 101: Control section 102: Memory section 103: Peripheral device I/F section 104: Input section 105: Display section 106: Communication section 110: Bus 120: Network 130: External server 140: Database

Claims (12)

A method for obtaining a promoter sequence modified to have a desired activity, comprising the steps of:
Providing an original promoter sequence to be modified;
Generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence;
predicting the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model;
selecting a modified promoter sequence predicted to have a desired activity;
the promoter sequence is a plant cell promoter sequence,
The original promoter sequence includes a core promoter and a sequence upstream thereof, wherein the core promoter is a sequence included in positions −200 to +50 when the base position of the transcription start site is represented as +1;
The genome editing technology is a genome editing technology using the CRISPR/Cas system,
A set of multiple modified promoter sequences is generated by sequence deletion caused by cleavage induced by a combination of guide RNA sequences designed based on two PAM recognition sequences, where at least one of the two PAM recognition sequences is located within a core promoter;
A method , wherein the original promoter is at least 2000 bp in length . The method of claim 1, wherein the desired activity is a gene expression induction activity higher than that of the original promoter sequence, or a gene expression induction activity lower than that of the original promoter sequence. The method of claim 1, wherein the original promoter sequence includes a core promoter sequence and an upstream sequence thereof. The method of claim 1, wherein the set of modified promoter sequences includes at least 1000 different sequences. The method according to claim 1, wherein the machine learning model is a regression model trained to predict gene expression induction activity from a promoter sequence using gene expression induction activity data of multiple promoter sequences in plant cells as training data. The method of claim 1, further comprising a step of visualizing the activity of each of the modified promoter sequences included in the generated set of modified promoter sequences on a computer display. An information processing device for predicting a promoter sequence modified to have a desired activity, comprising:
a sequence input section for receiving input of an original promoter sequence to be modified;
A modified sequence generating unit that generates a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence;
an activity prediction unit that predicts the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model;
a sequence selection section for selecting a modified promoter sequence predicted to have a desired activity;
the promoter sequence is a plant cell promoter sequence,
The original promoter sequence includes a core promoter and an upstream sequence thereof, wherein the core promoter is a sequence included in positions −200 to +50 when the base position of the transcription start site is represented as +1;
The genome editing technology is a genome editing technology using the CRISPR/Cas system,
A set of multiple modified promoter sequences is generated by sequence deletion caused by cleavage induced by a combination of guide RNA sequences designed based on two PAM recognition sequences, where at least one of the two PAM recognition sequences is located within a core promoter;
The original promoter is at least 2000 bp in length. A non-transitory computer-readable medium having instructions stored thereon, the instructions, when executed by a processor, performing the following steps:
receiving an input of an original promoter sequence to be modified;
A step of generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence;
predicting the activity of each of the modified promoter sequences included in the generated set of modified promoter sequences using a machine learning model;
A step of selecting modified promoter sequences predicted to have a desired activity can be carried out,
the promoter sequence is a plant cell promoter sequence,
The original promoter sequence includes a core promoter and a sequence upstream thereof, wherein the core promoter is a sequence included in positions −200 to +50 when the base position of the transcription start site is represented as +1;
The genome editing technology is a genome editing technology using the CRISPR/Cas system,
A set of multiple modified promoter sequences is generated by sequence deletion caused by cleavage induced by a combination of guide RNA sequences designed based on two PAM recognition sequences, where at least one of the two PAM recognition sequences is located within a core promoter;
A computer readable medium, wherein the original promoter is at least 2000 bp in length. On the computer,
A function to accept input of the original promoter sequence to be modified;
A function of generating a set of multiple modified promoter sequences that can be created by genome editing technology based on the promoter sequence;
A function of predicting the activity of each modified promoter sequence included in the generated set of modified promoter sequences using a machine learning model;
A function for selecting modified promoter sequences predicted to have a desired activity is realized,
the promoter sequence is a plant cell promoter sequence,
The original promoter sequence includes a core promoter and a sequence upstream thereof, wherein the core promoter is a sequence included in positions −200 to +50 when the base position of the transcription start site is represented as +1;
The genome editing technology is a genome editing technology using the CRISPR/Cas system,
A set of multiple modified promoter sequences is generated by sequence deletion caused by cleavage induced by a combination of guide RNA sequences designed based on two PAM recognition sequences, where at least one of the two PAM recognition sequences is located within a core promoter;
The original promoter length is at least 2000 bp, program. A method for genome editing of a plant cell for regulating the expression level of a desired gene, comprising:
Obtaining a modified promoter sequence having a desired activity by the method of claim 1;
Preparing a plant cell to be subjected to genome editing;
editing the genome of said plant cell to generate said modified promoter sequence. A method for producing a genome-edited plant in which the expression level of a desired gene is regulated, comprising:
Editing the genome of a desired plant cell by the method of claim 10 ;
A method comprising obtaining a plant individual derived from a genome-edited cell. The method of claim 1, wherein the prediction of the activity of each modified promoter sequence by a machine learning model is performed using only the core promoter portion of each modified promoter sequence.