KR20200105069A - Method for identifying condition-specific micro rna targets with big data - Google Patents

Abstract

The present invention relates to a method of investigating a micro-RNA target by conditions by using big data, capable of proposing a novel algorithm with improved prediction accuracy. According to the present invention, the method of investigating the miRNA target by conditions by using the big data includes: a first step of generating an MIR profile; a second step of extracting a seed bicluster from the MIR profile; a third step of generating a temporary extended bicluster by adding a vector having a zero rate (Z_(cut,s)) among vectors which are able to be added except for the seed bicluster in the MIR profile; a fourth step of generating an extended bicluster by removing row and column vectors, which are less dense, from the temporary extended bicluster based on a threshold value; and a fifth step of creating and updating a merged bicluster by adding the extended bicluster to the seed bicluster.

Description

Micro RNA target investigation method by condition using big data {METHOD FOR IDENTIFYING CONDITION-SPECIFIC MICRO RNA TARGETS WITH BIG DATA}

The present invention relates to a method of investigating microRNA (miRNA) targets for each condition using big data, and more particularly, it is possible to find micro RNA (miRNA) targets for each condition by analyzing large-capacity biodata, which is big data, through bike clustering. It is a description of an algorithm that exists.

Micro RNA (hereinafter referred to as miRNA) is a small non-coding RNA molecule that regulates gene expression by binding to a miRNA-reactive enzyme in mRNA after transcription. Since their discovery, extensive research has been shown to play a key role in regulating cell cycle and differentiation, chronic disease, cancer progression and other processes. Since the function of miRNAs is characterized by target genes, efforts have been made to systematically identify these target genes based on the binding sequence. While these methods provided hundreds to thousands of potential targets, they did not yield multiple false positives and suggest specific targets related to the cellular conditions being tested.

Paired expression profiles of miRNA and mRNA (denoted as miRNA-mRNA profiles) have been used to select more specific mRNA targets for each miRNA. In other words, it is a method using anticorrelation between miRNA and target mRNA. In addition to the simple Pearson and Spearman correlation methods, many computational methods have been developed that incorporate both binding sequences and miRNA-mRNA profiles to detect miRNA-mRNA relationships, including penalized regression and Bayesian methods. Many of these methods used multivariate linear models in which multiple miRNAs regulate common target genes. This inverse correlation-based method has improved target prediction, but requires a very costly miRNA-mRNA profile, and currently, only a limited number of data sets have been published.

Another approach to improving miRNA target prediction is to estimate the miRNA regulatory module. Based on the binding sequence information, a graph consisting of two parts between miRNA and mRNA was constructed, and the maximum biliques were identified.

The bike was further refined for specific cellular conditions by incorporating the miRNA-mRNA profile as it represents a miRNA regulatory module in which several miRNAs can match a common target. Due to the modular nature of cellular processes, this module is considered to exhibit a stable miRNA regulation pattern. This approach suggests targets of a group of miRNAs for a fixed cell condition. However, this prior art provided an unstable miRNA target group.

The present invention proposes a novel approach to identify miRNA targets for various cellular conditions by cycling large amounts of mRNA profiles for sequence-specific targets. That is, in the conventional case, under a given cellular condition, miRNA and mRNA targets are bike-clustered to a miRNA regulatory module, whereas in the present invention, mRNA targets and various cellular conditions are bike-clustered for a specific miRNA. To this end, 5,518 human microarray expression data sets were collected from the Gene Expression Omnibus (GEO) database, and mRNA fold change data between the mRNA target and the cellular condition for the miRNA of interest were analyzed by cycling. In the conventional case, as shown in FIG. 3, miRNA and mRNA targets (g1, g2, …, gn) for miRNA regulatory modules (R1, R2, …, Rn) under given cellular conditions (C1, C2, …, Cn) ) To bike clustering.

However, the present invention provides a more stable group of miRNA targets that are tightly regulated across various cellular conditions without the use of miRNA-mRNA profiles. As shown in Figure 4, the present invention bikes the mRNA targets (g1, g2, …, gn) and cellular conditions (C1, C2, …, Cn) for a given miRNA. The inventive bike cluster was evaluated using experimentally validated targets and showed significantly improved accuracy compared to pure sequence-based methods. Moreover, the accuracy was further improved by selecting targets that functionally interact with other target genes. In particular, these gains (enhancements) were obtained using publicly available gene expression and protein function interaction data, but showed higher accuracy than those obtained from expensive inverse correlation-based methods. In addition, through the present invention, a bike cluster database called BiMIR, which can search for 459 human miRNA regulatory targets and predicted results for cellular conditions, was constructed.

US registered patent US8712935 “EVOLUTIONARY CLUSTERING ALGORITHM”

An object of the present invention is to provide a novel approach capable of predicting the mRNA target of miRNA and related cellular conditions by biclustering the mRNA change data under various cellular conditions for a sequence-specific target.

In addition, an object of the present invention is to provide an approach through biclustering that predicts targets of sequence recognition regulators such as miRNA, transcription factor, and RNA binding protein by combining gene expression data and target recognition sequence information.

In addition, it is an object of the present invention to propose a new algorithm with improved prediction accuracy that extends a bicluster to identify a maximum block through an appropriate permutation.

In addition, an object of the present invention is to propose an algorithm capable of preventing the bicluster from being elongated in one direction when expanding.

The technical problems to be solved by the invention are not limited to the technical problems mentioned above, and other technical problems that are not mentioned will be clearly understood by those of ordinary skill in the technical field to which the present invention belongs from the following description. I will be able to

Micro RNA target irradiation method for each condition using big data according to the present invention comprises: a first step of generating an MIR profile;

A second step of extracting a seed bike cluster from the MIR profile;

A third step of generating a temporary extended bike cluster by adding a vector having a zero rate (Z _cut,s ) from among vectors that can be added excluding the seed bike cluster from the MIR profile;

A fourth step of generating an extended bicycle cluster by removing less dense row and column vectors from the temporary extended bicycle cluster based on a threshold value;

A fifth step of creating and updating a merge bike cluster by adding the extended bike cluster to the seed bike cluster;

The seed bike cluster is characterized in that it is gradually expanded by repeating the third to fifth steps R times.

Repeating the third to fifth steps R times,

Wherein (Z _{cut, s)} updated merged zero rate of by-cluster in step 5 is characterized in that the extension is repeated until it reaches a zero rate (Z _{cut, s)} of the temporary extension by cluster.

The zero rate (Z _cut,s ) is performed by [Equation 1].

[Equation 1]

In addition, when the direction of any one of the row or column vectors in the merged bike cluster generated in the fifth step is longer, the correction zero rate (Z _cut,m ) of [Equation 3] is applied to the longer vector. Step 6; may further include.

[Equation 3]

In addition, the microRNA target irradiation apparatus for each condition using big data according to the present invention comprises: an extraction unit for generating a seed bike cluster by competitively adding rows and columns by BICLIQUE;

An extension unit 10 for generating a temporary extended cluster by attaching a vector having a zero rate (Z _cut,s ) among vectors that can be added in the MIR profile;

A trim unit 20 for generating an extended bike cluster by removing less dense rows and columns from the temporary extended cluster based on a threshold value;

And an update unit 30 for gradually expanding the bike cluster by updating the extended bike cluster in the seed bike cluster.

In addition, the micro RNA target irradiation apparatus for each condition using big data according to the present invention may further include an expansion prevention unit 40 preventing lengthening in one direction in the generated merged bike cluster.

In the micro RNA target irradiation apparatus for each condition using the big data, the zero rate (Z _cut,s ) is characterized in that [Equation 1].

[Equation 1]

The expansion preventing unit 40 is characterized in that when the direction of any one of the row or column vectors in the merge bike cluster is longer, a correction zero rate (Z _cut,m ) is applied to the longer vector.

[Equation 3]

By means of solving the above problems, the present invention can provide a novel approach for predicting human mRNA targets and related cellular conditions by biclustering large amounts of mRNA change data for sequence-specific targets.

In addition, the present invention can provide a bike cluster that predicts a target of a regulator such as miRNA, transcription factor, or RNA binding protein by combining gene expression data and target recognition sequence information.

In addition, the present invention can propose a new algorithm with improved prediction accuracy that extends a bicluster to identify a maximum block through appropriate permutation.

In addition, the present invention can propose an algorithm capable of preventing the bicluster from being elongated in one direction when expanding.

In addition, the algorithm proposed by the present invention has high accuracy and sensitivity compared to the prior art.

1 is a schematic diagram illustrating a miRNA target irradiation method for each condition using big data according to the present invention.
Figure 2 is a flow chart showing a micro RNA target irradiation method for each condition using big data of the present invention.
3 is a schematic diagram illustrating an approach for discovery of a conventional miRNA regulatory module.
4 is a schematic diagram illustrating an approach for discovery of a miRNA regulatory module in the present invention.
5 is a schematic diagram illustrating prediction of miRNA targets based on biclustering.
6 shows a simulation test for a biclustering algorithm, which is an embodiment of a simulation profile.
7 is an experimental example showing the precision of the biclustering method performed in FIG. 6 compared with the prior art and the present invention.
8 is an experimental example showing the sensitivity of the bi-clustering method implemented in FIG. 6 compared with the prior art and the present invention.
9 is a graph showing the miRNA target performance predicted by using a binding sequence, biclustering and functional network, and showing the biclustering results with blue nodes.
FIG. 10 is a graph showing the average sensitivity and specificity of the target network for different node orders together with FIG. 9.
11 is a graph showing the accuracy improvement of methods using combined sequence information, bike clustering, and network information together with FIG. 9.
12 is a graph comparing performance between biclustering and decorrelation method.
13 is a diagram showing the PI3K/Akt pathway (breast cancer) predicted by the biclustering method, miRNAs predicted to regulate it, and target genes.
14 is a graph showing the survival analysis rate without recurrence for 210 breast cancer patients with high and low levels of miR-29a, miR-29b and miR-29c in the PI3K/Akt pathway of FIG. 13 (breast cancer).
FIG. 15 is a graph analyzing the transcription levels of miR-29 target gene candidates by qRT-PCR in the PI3K/Akt pathway (breast cancer) of FIG. 13.
Figure 16 is a graph analyzing the levels of pAKT, AKT, pFAK and FAK in the tumor microenvironment of total cell lysates extracted from miR-29b-3p transfected cells in the PI3K/Akt pathway (breast cancer) of Figure 13 to be.
FIG. 17 is a graph analyzing the levels of pAKT, AKT, pFAK, and FAK in the tumor microenvironment of total cell lysates extracted from miR-29c-3p transfected cells in the PI3K/Akt pathway (breast cancer) of FIG. 13 to be.
18 is a graph schematically showing a comparison of the Progressive Bicluster Extension (PBE) algorithm (a) and the conventional algorithm (b to g) applied to the method of researching micro RNA targets for each condition using big data according to the present invention.
19 is a graph showing the number distribution of conditions, genes, and density in bike clusters having three fold-change zero rates (Z _cut,s ).
20 is a diagram showing the BiMIR database.

The terms used in the present specification will be briefly described, and the present invention will be described in detail.

The terms used in the present invention have been selected from general terms that are currently widely used while considering functions in the present invention, but this may vary depending on the intention or precedent of a technician working in the field, the emergence of new technologies, and the like. Therefore, the terms used in the present invention should be defined based on the meaning of the term and the overall contents of the present invention, not a simple name of the term.

When a part of the specification is said to "include" a certain element, it means that other elements may be further included rather than excluding other elements unless specifically stated to the contrary.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art may easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein.

Specific matters, including the problems to be solved, means for solving the problems, and effects of the present invention, are included in the following examples and drawings. Advantages and features of the present invention, and a method of achieving them will become apparent with reference to the embodiments described below in detail together with the accompanying drawings.

Hereinafter, the present invention will be described in more detail with reference to the accompanying drawings.

The present invention aims to propose an approach to identify new modules, which are mRNA targets and related cellular conditions, of human microRNAs (hereinafter, miRNAs) by cycling many mRNA change data for sequence-specific targets.

Bicluster targets were evaluated using experimentally identified mRNA targets and showed an increase in accuracy (sensitivity + specificity) of 17.0% (median 19.4%) on average. When using a functional network, the overall accuracy improved up to 32.0% (median 33.4%). In the present invention, by analyzing cancer-related biclusters, it was found that the PI3K/Akt signaling pathway is a statistically significant concentration of several miRNA targets of breast cancer and diffuse B-cell lymphoma. In particular, five independent prognostic miRNAs were identified, and inhibition of bicluster target and pathway activity by miR-29 was experimentally verified. For a total of 459 human miRNAs, 29,898 bike clusters were collected from the BiMIR database, and bike clusters can search for miRNAs, tissues, diseases, keywords and target genes.

More specifically, the method of irradiating micro RNA targets according to conditions using big data according to the present invention may be carried out by the following method as shown in FIGS. 1, 2 and 5.

First, the first step (S10) generates an MIR profile.

In general, biclustering algorithms seek to elucidate a complete association (i.e., biclique) between two subsets of nodes (e.g., a subset of target genes and a subset of cellular conditions). In consideration of the noise of the microarray data, the present invention allows a small portion of the irrelevant connection between two subsets of nodes, but a progressive bicluster extension (PBE) algorithm that creates a much larger sized bike cluster. Was developed.

In the first step (S10), the PBE algorithm identifies a bike using the bimax algorithm. The bike is a matrix block full of 1s in data composed of 0s and 1s. By using the bike as a seed, the column and row with the lowest density of 0 values are competitively attached to the seed, while the'density' row and column having a small ratio of the zero value are competitively added and expanded. .

More specifically, as shown in (a) to (c) of FIG. 5, first, 5,158 mRNA microarray data sets with two sample groups (experimental group/control) were collected from the GEO database, and the corresponding logFC (mRNA change value) The logarithmic conversion value of, the base value is 2.) Data sets 20,639 human genes as column vectors and 5,158 fold-change cell conditions as row vectors. These logFC data are quantized into genes (denoted 1, 0, and -1, respectively) that are up, neutral, and down-regulated using ± log1.3 (hereinafter referred to as 1.3 FC) thresholds. Excluding the noisy data, the 1.3 FC was regarded as an appropriate threshold representing the change in target expression due to miRNA regulation covering many finely tuned mRNA targets.

5(a) shows an overview of gene expression fold-change, and sequence specific targets for each of the miRNAs were obtained from 7 miRNA target databases (FIG. 5(b)). For each miRNA, sequence-specific targets predicted from at least 3 of the 7 miRNA target databases were selected and displayed as sequence-based miRNA targets corresponding to the blue ellipses in FIG. 5(b). .

Then, the logFC profile for each cell condition was accumulated in the background set based on the enrichment of the regulated genes-a 1.3-fold increase in the background set (initial distribution test, FDR <5%). The logFC sub-matrix obtained here is converted to a binary matrix by replacing -1 with 0 and is called the MIR profile for a given miRNA.

Next, the second step (S20) extracts the seed bike cluster from the MIR profile.

More specifically, as shown in FIG. 5(d), a small bike cluster completely filled with 1 (hereinafter, seed biclusters) was obtained by applying the bimax biclustering algorithm to the MIR profile. The seed bike cluster is gradually expanded using the PBE algorithm, which will be described below.

Next, in the third step (S30), a vector having a zero rate (Z _cut,s ) is added from among the vectors that can be added, excluding the seed bike cluster from the MIR profile to generate a temporary extended bike cluster. The zero rate (Z _cut,s ) has the same meaning as a threshold value or a threshold value for the expansion of the bike cluster.

The zero rate (Z _cut,s ) is characterized in that [Equation 1].

[Equation 1]

Next, the fourth step (S40) repeats the gradual expansion R times in the temporary expansion bike cluster, with a zero rate (Z _{cut, s} ) for each step. After completing the expansion by pasting the columns and rows with fewer zeros, the columns and rows of the expanded bike cluster are examined again, and columns and rows with a ratio of 0 higher than the zero ratio (Z _cut,s ) are removed to create an expanded bike cluster. do.

That is, in the fourth step (S40), the same threshold value is applied to the temporary expansion bike cluster in the expansion and removal processes. This process is necessary because columns and rows with a ratio of '0' higher than the zero rate (Z _cut,s ) may exist in the temporary expansion bike cluster due to an accidental arrangement of '0' values in the temporary expansion bike cluster. . The extended bike cluster is created by removing less dense row and column vectors based on the threshold value.

Next, in the fifth step (S50), as shown in FIG. 5(e), a merge bike cluster is created and updated by adding the extended bike cluster to the seed bike cluster.

The bike cluster obtained by expanding and removing the seed bike cluster filled with 1 extracted in the second step (S20) is referred to as the expanded bike cluster. For a given miRNA, many of these expanded bike clusters appear, and some of these expanded bike clusters are substantially overlapped, and the process as shown in FIG. 5(e) is performed to merge the overlapping bike clusters. The final bike cluster thus obtained is referred to as the merge bike cluster. Finally, the bike clusters used for prediction are the merged bike clusters. In summary, the seed bike clusters are expanded to obtain the expanded bike clusters, and merged bike clusters are obtained by merging them through the hierarchical clustering.

The seed bike cluster is expanded by repeating the third step (S30) to the fifth step (S50) R times. That is, it is preferable to repeat the extension until it reaches the fifth merged zero rate of by-cluster updated at step _{(S50) (Z cut, s} ) zero rate (Z _{cut, s)} of the temporary extension by cluster Do.

The merged bike cluster was clustered using average connection hierarchical clustering (extended bike cluster) to avoid redundant connections. Among the three methods of measuring the distance between groups in the hierarchical clustering, the average linkage method is the most commonly used method. In the case of two different expansion bike clusters A and B, the distance between A and B is as shown in [Equation 2] below.

[Equation 2]

In the present invention, a threshold value is applied to the metrological clustering dendrogram (Zcut,s is a threshold value used for bike cluster expansion and a threshold value that determines at which height the tree is cut) 0.3, 0.5, 0.7. It was set to and tested. When a high threshold is applied to the tree diagram, many small clusters are generated, and in the opposite case, a small number of large clusters are obtained. In the fifth step (S50), the seed bike cluster and the extended bike cluster are merged to create a merge bike cluster, and then row or column vectors containing 10% or more of 0 are individually removed to finally create a merge bike cluster. do.

The merged bike cluster represents a cell condition clustered in a row vector and a target gene predicted in a column vector. The down-regulated bike cluster in the merged bike cluster is also created in a symmetric manner. In other words, the up-regulated (or up-regulated) combined bike cluster means that the corresponding miRNA is down-regulated (or up-regulated) under cellular conditions included in the bike cluster.

In the present invention, the analysis results for the 1.3FC threshold were mainly reported, but the merged bike cluster was generated at ±log1.5 and ±log2.0 thresholds (represented by 1.5FC and 2.0FC thresholds, respectively), and is more specific. A strong miRNA regulation was collected. In total, for a list of sequence specific targets for a given miRNA, two MIR profiles (top and bottom) are generated for each threshold (1.3, 1.5 and 2.0). The three up-regulated (and down-regulated) MIR profiles have different condition counts with the same gene count. Accordingly, the resulting seed bike cluster and the merge bike cluster differ according to different threshold values.

Next, in the sixth step (S60), if the direction of any one of the row or column vectors in the merged bike cluster generated in the fifth step (S50) is longer, Apply the rate (Z _cut,m ). For example, if the column of the bike cluster is 3 times longer than the row, Zcut,m is applied instead of Zcut,s. As r=3 is included, it becomes more difficult to add a new column and the probability of adding a row is relatively It becomes higher.

[Equation 3]

Below, the pathway of cancer-related biclusters and the prognosis of related miRNAs were analyzed and the approach of the present invention was verified through confirmation experiments. The algorithm for investigating miRNA targets for each condition using big data of the present invention was named PBE algorithm in the following experiments and drawings.

(1) Materials and methods

1) Expression fold-change data collection

I downloaded the CEL file for the 2,019 GEO series built using the Affymetrix U133 Plus 2.0 chip. A robust multiple array mean (RMA) normalization was applied to each CEL file using the R'affy'package's' justRMA' function. The intensity of the probe for each gene collapsed to an average value. Next, for each experimental series, two sample groups (test/control) were selected and the logarithmic FC (expressed as logFC) of the mean expression for each group was calculated. A total of 5,158 (test/control) logFC profiles for cell conditions were collected for 20,639 human genetic symbols. Information on logFC matrix and cell conditions can be obtained from the bimir R package (https://github.com/unistbig/bimir).

2) sequence specific miRNA target

Sequence-specific miRNA targets were obtained from 7 sequence-based target prediction databases (TargetScan, miRanda, mirSVR, PITA, DIANA-microT, miRDB and TargetRank).

3) Experimental verification of miR-29b/c regulation in breast cancer

Ⅰ) miRNA transfection

miR-29n-3p and miR-29c-3p mimics and miRNA scramble controls were purchased from Genolution. Each miRNA (100nM) was infected with MDA-MB231 using G-fectin Reagent (Genolution). All experiments were performed 48 hours after cell infection.

Ii) Real-time quantitative PCR

Total RAN 1mg from MDA-MB231 cells was reverse transcribed with oligo dT and M-MLV RT reverse transcriptase. Real-time quantitative PCR was performed using GENETBIO SYBR Green Prime Q-master Mix and QuantStudio 5 PCR system (ThermoFisher). All tests were accompanied by an internal control B2M or HPRT gene. Since both reference genes produced very similar results, only the B2M results are shown in FIG. 6. Samples were replicated and run and normalized to B2M or GAPDH using a dd cycle threshold-based algorithm to provide arbitrary units representing relative expression.

Iii) Immunoblotting

The harvested cells were lysed in RIPA buffer and subjected to centrifugation before the supernatant was collected. The protein concentration was measured using the BCA Protein Analysis Kit (Pierce), and the same amount of the protein was analyzed using 10% or 12% sodium dodecyl sulfate (SDS)-polyacrylamide gel electrophoresis (PAGE), and then a nylon membrane (GE). Healthcare, Amersham). Primary such as rabbit anti-human FAK (1:1000 cell signaling), phospho-FAK (1:1000 cell signaling), Akt (1:1000 cell signaling) and mouse anti-human GAPDH (1:1000, cell signaling) The target protein was observed by incubation with the antibody and the infrared fluorescent dye-conjugated secondary antibody. HRP-conjugated secondary antibodies were purchased from Cell Signaling Technology. Immune detection was performed using an Odyssey CLx scanner (Li-COR Biosciences).

(2) result

1) Comparison with other bike clustering algorithms

As shown in Fig. 19, compared with the seed bike cluster of the present invention, the PBE algorithm of the present invention yields a much larger bike cluster by allowing a small amount of noise. 18, ISA, QUBIC, FABIA, BiBit, and HOCCLUS2, which are conventional biclustering algorithms, and the PBE algorithm of the present invention are compared.

First, in [Table 1], the size and signal density of the bike cluster generated from the actual up-regulated MIR profile of hsa-let-7c-5p were compared.

PBE Row Column Density N 23.3 38.9 98.1 17 QUBIC Consistency Row Column Density N 1.0 17.2 41.9 1.0 46 0.98 16.9 42.0 0.9997 44 0.95 14.2 43.4 0.9989 36 0.92 15.2 61.3 0.9796 24 BiBit (minimum row and column size = 10) Row Column Density N 13.5 12.4 1.0 1227 FABIA Continuous input Binary input Sparseness loading Row Column Density N Row Column Density N 0.01 80.1 219.4 0.318 55.5 226.3 0.358 30 22 0.05 64.1 195.5 0.330 25.8 291.1 0.548 26 27 0.1 25.9 189.6 0.355 15 285.2 0.653 26 25 0.15 28.5 198.3 0.359 10.9 265.6 0.714 28 23 0.2 25.8 198.7 0.368 25 12.2 268.1 0.723 15 0.25 19.1 201.2 0.388 14.4 272 0.701 27 11 0.3 28.2 195.8 0.361 11.9 266.8 0.719 23 10 ISA Continuous input Binary input TG TC Row Column Density N Row Column Density N 1.0 1.0 174.0 119.3 0.500 4 192.7 116.2 0.464 6 1.5 176.7 60.7 0.526 7 191.8 70.2 0.498 9 2.0 196.5 27.7 0.493 15 200.8 39.5 0.534 13 2.5 202.8 11.4 0.546 22 216.9 18.9 0.532 22 3.0 189.5 5.6 0.672 28 221.0 10.2 0.582 18 1.5 1.0 106.4 118.6 0.486 7 95.0 118.0 0.486 9 1.5 106.7 60.1 0.530 12 94.8 66.8 0.511 9 2.0 100.1 28.3 0.582 15 113.2 39.5 0.560 11 2.5 101.6 12.4 0.623 22 127.7 21.2 0.591 24 3.0 105.2 6.2 0.707 23 133.6 11.1 0.647 16 2.0 1.0 58.1 112.0 0.482 11 59.2 113.8 0.509 12 1.5 52.5 58.4 0.554 17 58.3 69.5 0.533 13 2.0 52.3 27.0 0.621 21 72.4 43.2 0.605 11 2.5 54.3 12.3 0.641 28 66.6 20.9 0.569 29 3.0 59.2 8.1 0.744 18 74.1 13.0 0.676 18 2.5 1.0 25.8 110.8 0.529 30 28.6 109.0 0.529 24 1.5 25.8 58.3 0.632 28 32.7 71.3 0.581 23 2.0 32.0 28.3 0.703 22 33.0 42.1 0.599 29 2.5 34.9 14.3 0.719 21 38.6 22.0 0.635 30 3.0 37.2 8.7 0.770 13 45.7 10.1 0.694 44 3.0 1.0 14.7 105.2 0.638 37 15.0 120.6 0.619 41 1.5 17.1 52.7 0.701 33 16.1 72.5 0.645 45 2.0 18.3 27.5 0.744 32 18.4 42.7 0.658 42 2.5 23.0 14.9 0.804 19 21.2 25.2 0.654 46 3.0 26.1 9.2 0.806 9 26.9 12.3 0.669 43 HOCCLUS2 Level Beta Row Column Density N One 0.4 13 10 1.0 60 0.5 13 10 1.0 60 0.6 13 10 1.0 60 0.7 13 10 1.0 60 0.8 13 10 1.0 60 0.9 13 10 1.0 60 2 0.4 23.5 19 0.844 30 0.5 23.5 19 0.844 30 0.6 23 19 0.855 31 0.7 23 18.5 0.863 32 0.8 19 16.5 0.986 40 0.9 12 11 1.0 53 3 0.4 45 38 0.687 15 0.5 45 38 0.687 15 0.6 41.5 33 0.742 18 0.7 24 19 0.805 25 0.8 18 17 1.0 35 0.9 12 10.5 1.0 52 4 0.4 80 63 0.578 8 0.5 71 58 0.575 9 0.6 45 40 0.682 13 0.7 23 19 0.797 22 0.8 18 17 1.0 34 0.9 12 10 1.0 51

As shown in Fig. 18, the PBE of the present invention yields a large bike cluster with a high density, while the conventional algorithm yields a bike cluster with a smaller size or a lower density. The PBE algorithm better captured stem cell specific bike clusters than conventional algorithms.

Next, as shown in FIG. 19, the sensitivity and specificity of the biclustering algorithm was tested using the simulated binary profile reflecting the average size and density of the actual MIR profiles (700 rows, 300 columns, and 20% density). The simulated profile contains 7 bike clusters randomly selected between 20 and 80 row and column sizes, and each bike cluster contains 1 to 3 zeros (noise). Some of the bike clusters overlap each other with less than 20% of the bike cluster size. The simulation was repeated 50 times. Here,'true element' indicates an element included in the seven bike clusters, and'false element' indicates an element outside the bicycle cluster. Therefore, after executing the conventional bike-clustering algorithm, the sensitivity was measured by dividing the predicted number of “real elements” in the bike cluster by the number of all “true elements”. Precision was measured as the percentage of actual elements in the predicted bike cluster. The PBE showed perfect precision (median: 100%) and high sensitivity (median: 95.6%).

The performance of the ISA depends on the threshold values of the row (TG) and column (TC). That is, when TG = TC = 1, high sensitivity was observed (median value: 97.2%), but precision (median value = 87.7%) was relatively low. Increasing both TG and TC to 2 increased precision (median = 96.8%) but decreased sensitivity (median = 86.1%).

The QUBIC result was influenced by the consistency parameter c. As this value increases, the precision increases and the sensitivity decreases. Best performance was observed when using the basic parameters (c = 0.95, medium precision = 80.8%, medium sensitivity = 100%).

The BIMAX and BiBit did not allow 0 in the bike cluster and exhibited very low sensitivity (median BIMAX sensitivity = 10.2%, median BiBit median = 14.5%). However, when applying 30 iterations to the BIMAX, the sensitivity increased to 86.7%.

The FABIA yielded very noisy bike clusters for the tested sparsity parameters due to its low precision (median ≤46.6%) and sensitivity (≤66.0%). Results for a = 0.01 and 0.05 are shown in FIGS. 7 and 8. When a ≥ 0.1, no bike cluster was created.

The HOCCLUS2 also did not create a bike cluster in this simulation setup, so it was tested in FIGS. 7 and 8 but was excluded. The HOCCLUS2 detected bike clusters in sparse data (12% or less). This result indicates that the PBE, which is the present invention, is an efficient method for identifying a bike cluster from binary data with a lot of noise.

2) Accuracy of prediction of the bike clustering target

The bicluster target was evaluated using a validated miRNA target. miRTarBase provides hundreds of thousands of experimentally validated miRNA-target relationships using'strong' evidence (reporter analysis or western blot) and'less strong' evidence (pSILAC or microarray experiments). Among the sequence-specific targets (background set) of a given miRNA, targets verified with'strong' evidence were considered gold targets (GP) targets, while strong evidences and weak markers were set as gold targets (GN) targets. For evaluation, in the present invention, miRNAs having 30 or more GPs with a fraction of 5% or more in the background set were selected. As shown in Figure 9, 11 miRNAs meeting this criterion were analyzed.

When a 1.3 FC threshold was used to quantize the logFC data, the average sensitivity and specificity of the 11 miRNAs were 0.704 and 0.466, respectively (sum = 1.170), which was 17.0% (median based on 19.4%) compared to the sequence-based system. Target prediction. 1.3 FC zero rate (Z _cut,s ) (Fig. 9)), while positive gains were obtained for all 11 miRNAs, the relative performance for each miRNA is different FC zero rate (Z _cut,s ) It was quite different depending on. For example, the gain of miR-34a-5p decreased as the FC zero ratio (Z _cut,s ) increased due to a sharp drop in sensitivity (1.3 FC: 20.8%, 1.5 FC: 13.3%, 2.0 FC: 7.2. %). Conversely, the specificity of miR-21-5p was relatively increased and increased as the zero rate (Z _cut,s ) increased (1.3 FC: 16.4%, 1.5 FC: 26.5% and 2.0 FC: 31.3%). These differences indicate different miRNA regulation patterns. The former case corresponds to a'fine tuner' miRNA that properly regulates many genes. Therefore, using a low zero rate (Z _cut,s ) helps to detect subtle changes in the target expression. However, in the latter case, miRNAs appear to more strongly regulate a relatively small number of targets. Of the three thresholds tested, the 1.3 FC showed the best acquisition with the greatest sensitivity.

miRNA targets tend to be functionally related to each other. Therefore, in the present invention, in order to improve the prediction of the miRNA target, the protein function interaction network of the STRING database (edge threshold> 150) was integrated between the bicluster target genes. Among the bicluster targets, samples with k-tables or functional interactions with other targets were selected and the corresponding benefit was measured. Interestingly, the specificity increased rapidly with increasing k (Fig. 10), and when k = 3, the maximum gain reached 32.0% (specificity = 77.8%, Fig. 11). The maximum central gain was much higher (33.4% when k = 4). These results show that the target interaction network can significantly improve miRNA target prediction.

3) Comparison with anticoagulant-based methods in cancer

The miRNA-mRNA pair profile was generally used to predict condition-specific miRNA targets based on the correlation between miRNA and miRNA targets. Therefore, in the present invention, in predicting cancer-specific miRNA targets, the biclustering method and 7 uncorrelated analysis methods (GenMiR++, Pearson correlation, Spearman correlation, Lasso, Elastic Net, IDA, and Tiresias) were compared. . The Pearson/Spearman correlation, Lasso, Elastic Net, and IDA were implemented using the miRLAB R package, and the GenMiR++ and Tiresias were implemented using MATLAB and Perl codes, respectively. For the 11 miRNAs evaluated in the previous section, the accuracy of the predicted target was compared with the anti-correlation based method and the inventive biclustering method. In the case of the antibody-reaction method, the sequence specific targets of each miRNA were classified in the order of the anticorlation score calculated from the TCGA miRNA-mRNA profile by the Pearson/Spearman correlation, Bayesian method, penalized regression, or neural network model. These classification scores were compared to the gold standard positive/negative set that produced the ROC curve.

In the case of the biclustering method, a bike cluster in which 30% or more of the tumors belong to the'tumor versus normal' or'aggressive versus non-aggressive tumor' state was selected. The biclusters represent 33 miRNA-cancer pairs for 5 cancer types (breast, brain, lung, colon or blood cancer). All these cancer types have both miRNA and mRNA data from TCGA (The Cancer Genome Atlas), so that the anticoagulation-based method could be tested in the present invention.

For each miRNA-cancer pair, the corresponding bike cluster targets were pooled in order of proportion of the specific cancer states of each bike cluster. Therefore, at each pooling step, the percentage of true and false positives of the bike cluster targets were plotted instead of a continuous curve (asterisk, Figure 12). After removing the 6 cases in which none of the 7 regions in the ROC curve (AUC) exceeded 0.6 and the maximum biclustering acquisition was less than 1.1, the analysis from 20 cases consistent with the known expression of the corresponding miRNA (quantitative PCR result). Selected clusters. In other words, up-regulated bike clusters were selected when the corresponding miRNA was known to be downregulated in cancer, and vice versa.

Overall, as shown in Fig. 12, the biclustering method was compared advantageously with the miRNA-mRNA profile based method. In 11 of the 20 cases, the biclustering method showed better benefits than the anticorlation-based method. In six different cases, both approaches gave similar results. In the remaining three cases, the biclustering method was inferior to the best correlation analysis-based method because of the lowest sensitivity. As we saw in the previous section, integrating network information tended to increase the specificity and benefit of the biclustering method. Of the seven correlation analysis methods, the Genmir ++ performed best in most cases.

These results indicate that the biclustering approach is generally superior to the reverse correlation-based method when prioritizing miRNA targets for each condition when miRNA expression information is provided. In particular, miRNA expression is relatively easily obtained from literature or quantitative PCR experiments.

4) miRNAs targeting PI3K/Akt signaling in cancer

In the present invention, as shown in Fig. 5, the bicluster target corresponding to 20 cancer-miRNA pairs was further analyzed. Among them, breast cancer and diffuse B-cell lymphoma (DLBCL) produced the largest number of bike clusters. In breast cancer, the bicluster targets of miR-1, miR-29a/b/c, miR-34a and miR-145 were upregulated in aggressive cancer. Targets of miR-29a/b/c, miR-34a and miR-145 were also upregulated in DLBCL. We pooled bicluster targets corresponding to each cancer type and performed pathway enrichment analysis (KEGG annotation) using the DAVID tool to identify 7 and 4 important pathways (FDR<0.05) for breast cancer and DLBCL, respectively. Interestingly, in both cancer types, the bicluster targets were strongly enriched with the PI3K/Akt signaling pathway (FDR = 2.6E-7 for breast cancer and FDR = 5.3E-7 for DLBCL). This pathway is known to be overexpressed in many cancers to promote cell cycle and survival, proliferation, and epithelial-mesenchymal mutations of tumor cells. In addition, extracellular matrix (ECM)-receptor interactions and local adhesion pathways were generally found in both cancer types, but all but two (CAV2, BIRC2) corresponding bicluster targets were also involved in the PI3K/Akt signaling pathway. Included.

12 and 13 depict the PI3K/Akt pathway highlighting bicluster targets for breast cancer and DLBCL, respectively. In both cancer types, miRNAs are multiple ligands that contain genes encoding growth factors (e.g. VEGFA and PDGFC targeted by miR-29) and ECMs (e.g. COL1A1, LAMC1 and THBS2 by miR-29). Was targeted. Receptor tyrosine kinase (e.g. MEK and/or PDGFRA by miR-34a), G protein (GNB4 and GNG12 by miR-29), toll-like receptors (TLR4 by miR-34a and miR-145) and integrins (e.g. , ITGB1 by miR-29), NRAS (by miR-29 and miR-145) and CDK6 (by miR-29). In addition, AKT3 was a target of miR-29 in breast cancer, and cytokine receptors (IL2RB and IL6R), a component of the PI3K complex (PIK3R3), were targets of miR-34a and miR-29 in DLBCL. In fact, it has been previously shown that miR-29b upregulation in breast cancer significantly inhibits metastasis by inhibiting related targets.

In the present invention, the well-established metastatic and invasive cancer cell line, the human breast cancer cell line MDA-MB231, was used to experimentally verify the bicluster target of miR-29. The levels of transcription of nine bicluster targets associated with ECM or PI3K were analyzed 2 days after transient transfection with miR-29 or control miRNA. All 9 targets were significantly downregulated compared to the control by miR-29b or -29c transfection (FIG. 15 ). In addition, activation of ECM-related downstream pathways such as focal adhesion kinase (FAK) and AKT was attenuated by miR-29 (Figs. 16 and 17), showing the ability of biclustering assays to capture disease-related pathways.

Finally, the present invention analyzed the prognostic value of these miRNAs using a multivariate Cox Ratio Hazard (mCPH) model for public miRNA expression data sets. Survival without recurrence was tested in 210 breast cancer patients (GEO database, GSE22216). Of the six miRNAs analyzed, three miR-29 family miRNAs showed significant prognosis (mCPH p value of miR-29a = 0.0042, miR-29b = 0.0064, miR-29c = 0.0038, age, tumor size, lymph node ER, grade). Then, the overall survival rate of 116 DLBCL patients (GSE40239) was also analyzed for 5 miRNAs. Two of them showed significant prognosis (mCPH p value of miR-34a = 0.0185 and miR-145 = 0.0041; International Prognostic Index (IPI) and gender). Kaplan-Meier plots contrasting the effects of high and low miRNA levels on survival are also shown in FIGS. 12 and 14.

Overall, cancer biclusters were analyzed to analyze the major pathways (PI3K/Akt signaling, ECM and local attachment) and five related prognostic miRNAs (miR-29a, miR-29b and miR-29c in breast cancer; miR-34a and miR-145. ) Has been shown to inhibit tumor progression (risk ratio 0.593-0.745). In particular, the effect of miR-29b / c on this pathway was experimentally proven (Figs. 15 and 16).

5) BiMIR: Bicluster database for miRNA targets by condition

PBE algorithm of the present invention (13949 for 1.3 FC, 10,999 for 1.5 FC, 4,950 for 2.0 FC threshold) and 459 human miRNAs compiled with the BiMIR database (http://www.btool.org/bimir_dir/) In total, 29,898 bike clusters were created. Here, bikers were able to search for miRNAs, tissues, diseases, keywords, genes of interest, and combinations thereof. The BiMIR can be used to investigate new miRNA functions, targets and related cellular conditions.

Along with the list of searched bike clusters, functional enhancement results for bike cluster targets are provided based on the MSigDB pathway and gene ontology categories. When bike-clusters are searched for a specific organ/tissue or disease, the percentage of that condition for each bike-cluster is also reported. They allow the user to find the bike clusters that are most suitable. Heat maps of each bike cluster are visualized (FIG. 20), and detailed information is provided by hyperlinking the corresponding target genes and cell conditions to Genecards and GEO databases, respectively. For the bicluster target genes, experimental evidence of miRTarBase, network node diagram, and protein network visualization based on the STRING database are provided. All bike clusters can be downloaded from the BiMIR database.

(3) Discussion

In the present invention, we proposed a new framework for prioritizing miRNA targets by biclustering cell-specific targets and cellular conditions, a dimension that has been rarely investigated.

This is based on the idea that miRNA targets, like other cellular molecules, have modular activity and can be repeatedly captured in different cellular conditions. Indeed, bicluster targets exhibited significantly improved accuracy compared to purely sequence-based targets and often enriched well-known pathways characterizing the identified modules. Moreover, functionally linked targets showed much higher accuracy and further confirmed the modular activity of miRNA targets.

In the present invention, we analyzed cancer bike clusters and found that the PI3K/Akt signaling pathway is intensively targeted by several miRNAs of two cancer types. In addition, the prognostic value of miR-29 and the regulatory effect of miR-29 were also verified. These results show that biclustering analysis can reveal a major pathway controlled by miRNAs in disease.

Based on the knowledge of miRNA expression, the prediction of miRNA targets in the present invention was advantageously compared with seven anti-correlation based methods under cancer conditions. This result demonstrates the practical value of the approach of the present invention in that various results can be obtained without generating an expensive miRNA-mRNA profile. The BiMIR database was designed to explore the modular regulatory network of miRNAs by linking miRNAs, cellular conditions (or diseases), mRNA targets, and related pathways. Users can identify candidate miRNAs and target genes for the cellular condition of interest. Knowledge of the level of miRNA expression helps to choose the appropriate orientation for the bike clusters (up or down).

In the present invention, the zero rate (Zcut,s) was gradually increased from 0.01 to 0.1 (step size 0.01) during 10 iterations of the bike cluster expansion. This may seem to allow 10% of zero in the end, but due to the trim unit 20 and the fourth step (S40), the final zero ratio is only about 1.5%. Blocking hierarchical clustering of the extended cluster is also a less sensitive parameter. In addition, the bike clusters were created with rather stringent criteria (objects in three or more databases). Therefore, BiMIR can be used to select a small number of targets that are very likely for the cell state of interest.

The biclustering approach presented in the present invention can also be applied to predict condition specific targets of transcription factors or other sequence specific regulatory factors such as RNA binding proteins. In this regard, 5,158 mRNA fold-change profiles for 20,639 genes are provided for general systems biology studies. These mRNA fold-change data are different from GTEx transcript data. This mRNA fold-change data indicates that the GTEx data represents the level of transcription in normal tissue, whereas the fold-change data of the present invention represents the'change' of gene expression for various cellular conditions such as disease, chemotherapy, and tissue. It is different from the corpse data.

While conventional methods of identifying miRNA regulatory modules co-exist with multiple miRNAs and multiple target genes representing neutralization regulatory networks, the present invention provides a high probability of a single miRNA that is commonly detected in multiple cellular conditions. It focuses on prioritizing. The approach of the present invention can be extended to be able to evaluate the miRNA core regulatory network by overlapping bike clusters for other miRNAs. Significant overlap indicates common mRNA targets in several cellular conditions. The approach and data of the present invention will contribute to uncovering the modular structure of a complex regulatory network.

By means of solving the above problems, the present invention can provide a novel approach for predicting human mRNA targets and related cellular conditions by biclustering large amounts of mRNA change data for sequence-specific targets.

In addition, the present invention can provide a bike cluster that predicts a target of a regulator such as miRNA, transcription factor, or RNA binding protein by combining gene expression data and target recognition sequence information.

In addition, the present invention can propose a new algorithm with improved prediction accuracy that extends a bicluster to identify a maximum block through appropriate permutation.

In addition, the present invention can propose an algorithm capable of preventing the bicluster from being elongated in one direction when expanding.

In addition, the algorithm proposed by the present invention has high accuracy and sensitivity compared to the prior art.

As described above, it will be understood that the technical configuration of the present invention described above can be implemented in other specific forms without changing the technical spirit or essential features of the present invention by those skilled in the art.

Therefore, the embodiments described above are to be understood as illustrative and non-limiting in all respects, and the scope of the present invention is indicated by the claims to be described later rather than the detailed description, and the meaning and scope of the claims and the All changes or modifications derived from the equivalent concept should be interpreted as being included in the scope of the present invention.

10. Extension
20. Trim part
30. Update section
40. Expansion prevention part
S10. The first step of creating an MIR profile
S20. The second step of extracting a seed bike cluster from the MIR profile
S30. A third step of generating a temporary extended bike cluster by adding a vector having a zero rate (Zcut,s) among vectors that can be added excluding the seed bike cluster from the MIR profile
S40. The fourth step of generating an extended bicycle cluster by removing less dense row and column vectors from the temporary extended bicycle cluster based on a threshold value.
S50. A fifth step of creating and updating a merge bike cluster by adding the extended bike cluster to the seed bike cluster
S60. The sixth step of applying a correction zero rate (Z _cut,m ) to the longer vector when the direction of any one of the row or column vectors in the merged bike cluster created in the fifth step is longer

Claims (8)

A first step of generating an MIR profile;
A second step of extracting a seed bike cluster from the MIR profile;
A third step of generating a temporary extended bike cluster by adding a vector having a zero rate (Z _cut,s ) from among vectors that can be added excluding the seed bike cluster from the MIR profile;
A fourth step of generating an extended bicycle cluster by removing less dense row and column vectors from the temporary extended bicycle cluster based on a threshold value;
A fifth step of creating and updating a merge bike cluster by adding the extended bike cluster to the seed bike cluster;
The seed bike cluster repeats the third to fifth steps R times and gradually expands . Micro RNA target irradiation method for each condition using big data, characterized in that. The method of claim 1,
Repeating the third to fifth steps R times,
Big data characterized in that the repeatedly extended, until the first merged zero rate of by-cluster (Z _{cut, s)} updated at step 5 is reached in the temporary expansion zero rate of by-cluster (Z _{cut, s)} Micro RNA target investigation method for each condition used. The method of claim 1,
The zero rate (Z _cut,s ) is a micro RNA target irradiation method according to conditions using big data, characterized in that [Equation 1].
[Equation 1]

The method of claim 1,
In the merged bike cluster generated in the fifth step, when the direction of any one of the row or column vectors is longer than a preset range, the correction zero rate (Z _cut,m ) of [Equation 3] is applied to the longer vector. A sixth step of making further expansion difficult by applying; micro RNA targeting method for each condition using big data, characterized in that it further comprises.
[Equation 3]

An extraction unit for generating a seed bike cluster by competitively adding rows and columns by BICLIQUE;
An extension unit 10 for generating a temporary extended cluster by attaching a vector having a zero rate (Z _cut,s ) among vectors that can be added in the MIR profile;
A trim unit 20 for generating an extended bike cluster by removing less dense rows and columns from the temporary extended cluster based on a threshold value;
And an update unit (30) for gradually expanding the bike cluster by updating the expanded bike cluster in the seed bike cluster. A micro RNA target irradiation apparatus for each condition using big data, comprising: a. The method of claim 5,
Micro RNA target irradiation device for each condition using big data, characterized in that it further comprises a; expansion prevention unit (40) for preventing lengthening in one direction in the generated merged bike cluster. The method of claim 5,
The zero rate (Z _cut,s ) is a micro RNA target irradiation method according to conditions using big data, characterized in that [Equation 1].
[Equation 1]

The method of claim 6,
The expansion prevention part 40,
In the merged bike cluster, when the direction of any one of the row or column vectors is longer than a preset range, a correction zero rate (Z _cut,m ) of [Equation 3] is applied to the longer vector. Micro RNA target investigation method by condition using data.
[Equation 3]

Images