KR20220074201A - Machine Learning System based on Convolutional Neural Network that Mimic Human Visual Inspection for Analysis of Epigenetic Data and its Operation Method - Google Patents

Abstract

A CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis and an operating method thereof are presented. The operation method of a CNN-based machine learning system that mimics a human visual inspection for epigenetic data analysis performed through a computer device according to an embodiment is a method in which a CNN (Convolutional Neural Network) model learns cancer genome data analysis result data. step; and calling the peak to search for a functional region that can be a cancer development mechanism for new cancer genome data that the learned CNN model is not trained on.

Description

CNN-based machine learning system based on Convolutional Neural Network that Mimic Human Visual Inspection for Analysis of Epigenetic Data and its Operation Method

The following embodiments relate to epigenetic data analysis, and more particularly, to a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis and an operating method thereof.

In the post-genomic era, understanding epigenetic regulation is one of the most important challenges in biology and medicine. In order to elucidate the pathological mechanism and accurately identify the target gene for developing therapeutics, it is important to evaluate the malformation of the genome-wide interaction of proteins and genomic elements that cause gene dysregulation. In this regard, chromatin immunoprecipitation followed by sequencing (ChIP-seq) is a widely used technique to identify genomic binding sites of epigenetic regulators including histones, transcription factors, and DNA/RNA binding proteins. to be. ChIP-seq makes it possible to discover the interactions between protein complexes and DNA regulatory elements and their gene regulatory networks. Analysis of ChIP-seq data revealed how histone modifications and nucleic acid interactions with proteins regulate important factors in gene regulation, cell lineage determination and maintenance.

One of the major computational difficulties in analyzing ChIP-seq data is identifying peaks in genomic regions where aligned reads are enriched when mapping sequence reads to a given reference genome. This task is challenging for several reasons. Sequence errors and local biases due to structural changes complicate the peak-calling problem. Diverse data patterns caused by biological variability, different regions, experimental environments, sequence coverage, etc. make the problem more difficult. A sensitive and reliable computational method to determine peaks in a noisy background is especially important for medical research where patient samples may be limited, so data quantity and quality are sub-optimal.

Several software tools for recalling peaks in ChIP-seq data have been developed based on various probabilistic and unsupervised learning methods such as MACS2, HOMER, SICER, and SPP. Some of these tools show high sensitivity to call peaks, but suffer from high false positive error rates. Other tools may require additional information as input data (eg, displayability scores of portions that overlap with one or more uniquely displayable portions read in the genome).

In the neural network model, it has been proposed to denoise ChIP-seq data and increase peak call performance. An ensemble approach to improve the accuracy of calling ChIP-seq peaks has also been proposed. The ensemble approach uses the output of existing multiple peak calling software tools to eliminate outliers among multiple peak calling decisions. Although this approach has increased the sensitivity to calling true peaks, it still suffers from high false positive rates. In particular, human cancer cell lines are too complex to understand human malignancies using ChIP-seq because the data patterns vary greatly depending on the patient's primary cancer tissue. The false positive rate calling ChIP-seq peaks in human cancer cell lines is worse than in other datasets.

To address these issues, experts use visualization tools such as the UCSC Genome Browser and Integrated Genome Viewer (IGV) to label actual peaks. False positive peaks can also be corrected by expert researchers. However, it would be very inefficient for scientists to find every peak in the entire genome for a large amount of ChIP-seq data. Hocking proposed a supervised learning approach based on grid search to optimize the parameters (e.g., cutoff values) that the user should set when running a peak calling tool such as MACS2. They use data labeled by experts, learn parameter values, and apply optimized parameters to the rest of the data set. However, this requires complete labeling for each individual ChIP-seq data set and each peak calling tool.

Landt, S. G. et al. ChIP-seq guidelines and practices of the ENCODE and modENCODE consortia. Genome Research 22, 1813-1831 (2012).

The embodiments describe a CNN-based machine learning system that mimics human visual inspection for epigenomic data analysis and an operating method thereof, and more specifically, the ChIP-seq peak detection algorithm, which is one of the core processes of cancer epigenetic data analysis. It provides a convolutional neural network (CNN)-based deep learning model that mimics the visual inspection of existing human researchers.

In the embodiments, the deep learning model learns the cancer genome data analysis result data of the human researcher, and the trained deep learning model searches the functional area that can be the cancer development mechanism for the new untrained cancer genome data, thereby making the human researcher An object of the present invention is to provide a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis and an operation method thereof, which can perform fast analysis while having the accuracy of

The operation method of a CNN-based machine learning system that mimics a human visual inspection for epigenetic data analysis performed through a computer device according to an embodiment is a method in which a CNN (Convolutional Neural Network) model learns cancer genome data analysis result data. step; and calling the peak to search for a functional region that can be a cancer development mechanism for new cancer genome data that the learned CNN model is not trained on.

The step of the CNN model learning the cancer genome data analysis result data may include converting the input data of the CNN model into a vector using a preprocessing module.

The step of converting the input data of the CNN model into a vector using the pre-processing module includes: ChIP-seq read mapping data, data labeled at peak positions for at least some regions of the ChIP-seq read mapping data, and RefSeq annotation information The method may include converting at least any one or more of the input data into a vector through a convolutional layer.

The step of converting the input data of the CNN model into a vector using the pre-processing module is to reduce the dimension of the vector transformed through a max-pooling layer in order to reduce the noise of the input data. may include steps.

The step of the CNN model learning the cancer genome data analysis result data includes learning the CNN model through a plurality of inception modules including at least one filter of convolution, max pooling, and average pooling. can do.

The step of learning the CNN model through the plurality of inception modules includes a peak pattern and a strong peak signal in the CNN model through at least one first inception module including a convolution, max pooling and average pooling filter. providing information about the scale; preventing the problem of increasing vector sizes of max pooling and average pooling layers through at least one or more second inception modules including convolution and max pooling filters; and providing at least one third inception module including a convolutional filter having a wider size and a longer stride than the convolutional filters of the first inception module and the second inception module. .

In the step of the CNN model learning the cancer genome data analysis result data, a residual connection structure may be used between the plurality of inception modules to prevent a vanishing gradient problem.

In the step of the CNN model learning the cancer genome data analysis result data, an average pooling layer following a fully connected layer may be added to the end of the plurality of inception modules to reduce dimensions.

In the CNN model learning the cancer genome data analysis result data, the optimal threshold value of the genomic region may be learned based on ChIP-seq read mapping data of a window selected through the CNN model.

In the step of the CNN model learning the cancer genome data analysis result data, since the size of the output vector of the CNN model becomes smaller than the input vector, the size of the output vector is expanded to be the same as the input vector, and at each individual location. The presence or absence of a peak can be predicted.

In the step of the CNN model learning the cancer genome data analysis result data, the importance of the peak may be measured by assigning a score to each peak using the sigmoid activation of the output data of the CNN model.

The step of calling the peak to search for a functional region that can be a cancer development mechanism for the new cancer genome data that the learned CNN model is not trained includes inputting new cancer genome data into the trained CNN model to select the peak Peaks can be predicted by calling and scoring each peak.

A CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis according to another embodiment includes a CNN model learning unit in which a CNN (Convolutional Neural Network) model learns cancer genome data analysis result data; and a CNN model peak prediction unit that calls a peak to search for a functional region that can be a cancer-generating mechanism for new cancer genome data that has not been trained on the CNN model.

The CNN model learning unit, at least one of ChIP-seq read mapping data, data labeled at peak positions for at least some regions of the ChIP-seq read mapping data, and RefSeq annotation information, input data of the CNN model as a vector It may include a preprocessing module that converts.

The CNN model learning unit may include a plurality of inception modules including filters of at least any one of convolution, max pooling, and average pooling.

The plurality of inception modules include at least one first inception module configured to include a convolution, max pooling, and average pooling filter in order to provide information on a peak pattern and a magnitude of a strong peak signal to the CNN model. ; at least one second inception module, configured to include a convolution and a max pooling filter, in order to avoid the problem of increasing the vector size of the max pooling and average pooling layer; and at least one or more third inception modules configured to include a convolutional filter having a wider size and a longer stride than the convolutional filters of the first inception module and the second inception module.

A vanishing gradient problem may be prevented by using a residual connection structure between the plurality of inception modules.

In the CNN model learning unit, an average pooling layer following a fully connected layer may be added to the end of the plurality of inception modules to reduce a dimension.

The CNN model learning unit may measure the importance of the peak by assigning a score to each peak using sigmoid activation for the output data of the CNN model.

The CNN model peak prediction unit may predict a peak by inputting new cancer genome data to the learned CNN model and calling a peak and assigning a score to each peak.

According to embodiments, the deep learning model learns the cancer genome data analysis result data of a human researcher, and the learned deep learning model searches for a functional area that can become a cancer development mechanism for new untrained cancer genome data, It is possible to provide a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis, which can perform fast analysis while having the accuracy of a human researcher, and an operating method thereof.

1 is a diagram for explaining visualization of a labeling peak using ChIP-seq read mapping data according to an embodiment.
2A to 2C are diagrams illustrating an example of a filter of a CNN structure according to an embodiment.
3 is a diagram illustrating a CNN structure according to an embodiment.
4 is a diagram illustrating a peak calling process of a learned CNN model according to an embodiment.
5 is a diagram for explaining a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis according to an embodiment.
6 is a flowchart illustrating an operation method of a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis according to an embodiment.

Hereinafter, embodiments will be described with reference to the accompanying drawings. However, the described embodiments may be modified in various other forms, and the scope of the present invention is not limited by the embodiments described below. In addition, various embodiments are provided in order to more completely explain the present invention to those of ordinary skill in the art. The shapes and sizes of elements in the drawings may be exaggerated for clearer description.

The examples below show a Convolutional Neural Network (CNN) deep that mimics the visual inspection of existing human researchers using the ChIP-seq peak detection algorithm, which is one of the core processes of cancer epigenetic data analysis. It provides a learning model. Here, the epigenome refers to a set of substances involved in the function of regulating the expression of genes. According to embodiments, it is possible to achieve the level of accuracy of a manual test of a professional researcher analyzing ChIP-seq epigenetic data.

Embodiments may perform development, design, and verification of a deep learning model that mimics the visual inspection of a human researcher. In particular, the embodiments provide a loss function design different from the existing deep learning learning method in order to accurately search a relatively very small functional region in the cancer genome, and the latest deep learning methods such as batch normalization and residual structure. technique can be used.

Existing epigenetic analysis tools have a problem in that, in the case of classical techniques such as the Poisson assumption, in the case of cancer genomes, there are many false positive errors due to the high level of mutation within the genome, making it difficult to accurately analyze them. Here, the Poisson distribution assumption refers to the discrete probability distribution used in the assumption necessary for predicting the occurrence pattern of regulatory factors in the genome.

Embodiments may use a deep learning model that learned epigenetic data manually inspected by a human researcher to develop an analysis tool capable of performing rapid analysis of classical analysis tools while maintaining the accuracy of a human researcher. Below, a deep learning model learns data from human researcher's cancer genome data analysis result data, and the learned deep learning model is an epigenome that can explore functional areas that can become cancer mechanisms for new untrained cancer genome data. A CNN-based machine learning system that mimics human visual inspection for data analysis and its operation method will be described in more detail.

ChIP-seq is one of the key experimental resources available to understand genome-wide epigenetic interactions and to identify disease-related functional elements. Although the analysis of ChIP-seq data is important, it is computationally difficult due to irregular noise and various levels of bias. Although many peak-calling methods have been developed, current computational tools still often require human manual inspection using data visualization. However, due to the sheer volume of ChIP-seq data, it is nearly impossible for researchers to find all peaks manually. A recently developed convolutional neural network (CNN) can achieve human-like classification accuracy and can be applied to this problem. In this embodiment, we design a new supervised learning method for ChIP-seq peak identification using CNN, and integrate it into a software pipeline called CNN-Peaks. Here, we annotated the presence or absence of peaks in some genomic regions using the data indicated by the researchers. The trained model can then be applied to predict previously unseen peaks in genomic regions in multiple ChIP-seq datasets, including benchmark datasets commonly used for validation of peak calling methods. According to the embodiments, it is possible to confirm better performance than the conventional method.

Recently, supervised machine learning methods based on deep neural networks such as convolutional neural networks (CNN) have been successfully applied to epigenetics, regulatory genomics, and systems biology. The present embodiments provide a new peak call software pipeline based on CNN named CNN-Peaks that uses data that researchers have partially labeled. The embodiments are used to reduce false positive peak calls by identifying local peak call thresholds that may vary from genomic region to genomic region, thus resolving local bias due to conformational changes in human cancer cell lines. Note that some peaks in genomic regions with frequent copy number variations are expected to have a higher mapping depth than others. As training data for CNN-Peaks, we use data labeled by experts as well as read count information from the preprocessing step of the original ChIP-seq read mapping. The software tool trains a model to determine the appropriate peak detection cutoff value in a specific genomic region by considering the read mapping pattern of the ChIP-seq neighborhood.

CNNs have been successfully used to solve the problems of image processing and natural language translation using data patterns with regionally dependent features. We also use the unified annotation information (RefSeq) available at NCBI ( https://www.ncbi.nlm.nih.gov/ ) as training input data for constructing predictive CNNs. RefSeq data include the genetic location, transcription, and protein encoding of genes that can help determine the presence or absence of peaks for human visual inspection.

The CNN-Peaks software package according to the embodiments consists of three main modules. One is to feed the original input data (labeled data, genomic annotations, read count information) to the CNN structure, the other is to train the model using training data, and the last one is to This is for peak prediction. Here, we evaluate the performance of CNN-Peaks using labeled data reserved as test data. In addition, we test CNN-Peaks against the ChIP-seq benchmark dataset commonly used for evaluation, and compare its performance with other major peak calling tools. In addition, we use the CNN-Peaks tool to analyze various real-world data sets for histone modifications such as H3K27ac and binding of transcription factors such as GATAD2.

Users can install a CNN-Peaks pipeline using Docker images and run the package using their own data along with a trained model to predict human narrow histone modifications and transcription factor binding sites. Even experienced users can use CNN-Peaks to build new predictive models trained on self-labeled data. The software package according to the embodiments includes a desktop application that supports experts to easily create labeled data by labeling peaks of randomly selected genomic regions from given ChIP-seq read mapping data. The pipeline can also be applied to other types of high-penetration sequence data, such as DNase-seq and ATAC-seq. The package is available through the Github repository http://github.com/odb9402/CNNPeaks .

Data description

1 is a diagram for explaining visualization of a labeling peak using ChIP-seq read mapping data according to an embodiment.

According to the embodiments, several ChIP-seq read mapping data sets were obtained from ENCODE data portal in BAM format, labeled data were generated, and CNN-Peeks were validated. These data include several ChIP-seq datasets to examine histone modifications such as histones H3K36me3, H3K4me3, H3K27me3, H2AFZ and H3K9ac, as well as the binding of transcription factors such as GATAD2, POLR2A, SMARCE1, and transcription factors. In addition, according to the examples, ChIP-seq data were downloaded from the leukemia cell line K562 and labeled background binding of genomic regions to train a CNN model.

Some genomic regions were randomly selected from the ChIP-seq read mapping data in BAM format, and as shown in Figure 1, the positions of the peaks can be labeled using a visualization tool with BAM alignment data. .

In addition to labeled data and ChIP-seq read mapping BAM files, curated, non-overlapping collections of genomic, transcriptional, and protein sequence records to human reference sequences (RefSeq) from NCBI can be used as additional input vectors for CNN models. RefSeq data may include protein code positions and similar products. When researchers manually examine ChIP-seq data to determine peaks, genomic annotation information such as transcripts and corresponding protein records (RefSeq annotation information) will usually be displayed along with the given ChIP-seq read mapping data. can This information often allows us to estimate where the peaks are. We can take this into account to mimic human inspection, and build our model to use these other annotation types as input. Here, this annotation information can be added as a binary vector indicating the presence or absence of transcription and protein at each genomic location.

Data preprocessing

The CNN-Peaks pipeline according to an embodiment may include a preprocessing module. These pre-processing modules can transform any input data, including ChIP-seq read mapping BAM files (data), labeled data, RefSeq annotation information, into well-shaped vectors to be fed into the CNN structure. Bins of a fixed size (12,000 bins by default) can be used to normalize the different window sizes of the labeled area to the same size. If the window is smaller than the target window size, CNN-Peaks users must use a visual inspection tool to label additional regions to add to this window, adjust the parameters of the target window size, or simply remove the window.

The preprocessing module also includes the ability to smooth read count patterns to reduce noise in read-aligned data. This smoothing step also helps make the CNN model close to human visual inspection. Without the smoothing and correction steps, the (raw) depth pattern is very noisy. For smoothing, max-pooling and Gaussian filter convolution operations can be used. Gaussian filters are commonly used for denoising and smoothing in various domains such as image processing. The equations of the convolution and max pooling operations can be expressed as the following equations, respectively.

[Equation 1]

[Equation 2]

는 각 연산의 필터 크기이고, v, u는 입력 벡터이며, v_i는 v의 i 번째 요소이다. X는 ChIP-seq 읽기 매핑 데이터의 벡터가 되고, 가우스 필터를 필터링하며, X_smoothing은 X의 평활화 벡터이다. 이는 다음 식과 같이 표현될 수 있다.Here, * denotes a convolution operation. The stride of convolution and max pooling is 1. Over

is the filter size of each operation, v and u are the input vectors, and v _i is the i-th element of v. X becomes a vector of ChIP-seq read mapping data, filtering Gaussian filter, and X _smoothing is a smoothing vector of X. This can be expressed as the following equation.

[Equation 3]

[Equation 4]

CNN architecture

2A to 2C are diagrams illustrating an example of a filter of a CNN structure according to an embodiment. More specifically, FIG. 2a shows an A-module, FIG. 2b shows a B-module, and FIG. 2c shows a C-module.

2A to 2C, an example of an inception module-based filter of a CNN-peaks model according to an embodiment is shown. The CNN-peaks model can be built based on the inception module used in GoogLeNet, which extracts various features from data through filters of different sizes. Here, the convolutions 201, 211, and 221 each represent a convolution operation with a filter size of 1*N, and concatenate filters 204, 213, 222 are several convolutions 201, 211, 221. and pulling (202, 203, 212) outputs.

A-module can use three types of filters: convolution 201 , max pooling 202 and average pooling 203 . The B-module may use two types of filters, convolutional 211 and max pooling 212 without average pooling, and may be composed of three types of convolutional filters 211 . In addition, the C-module may be composed of four types of convolutional filters 221 having a wider size and a longer stride than the filters used in the A-module and B-module.

3 is a diagram illustrating a CNN structure according to an embodiment.

Referring to FIG. 3 , the structure of the CNN-peaks model 310 may provide three types of inception modules having a slimmer structure than the existing inception modules. Here, the CNN-peaks model 310 may be included in the CNN model. CNN-peaks model 310 includes inception modules 313, 314, 315 as shown in FIGS. 2a to 2c, as well as several hidden layers, such as pooling 312, 316, convolutional layer 311, etc. It may be composed of a hidden layer.

The CNN-peaks model 310 learns an optimal threshold value for calling a peak in the local genomic region of the ChIP-seq data 301, and subtracts the threshold value from each input signal value 340 and sig A mode operation 350 may be used to determine the presence or absence of a peak. That is, the peak existence probability 360 may be obtained through the sigmoid operation 350 or the peak may be predicted 390 from the CNN-peaks model. In addition, a cross-entropy 380 may be used as a loss function for learning the CNN-Peaks model 310 . In FIG. 3 , the residual connection between the inception modules 313 , 314 , and 315 is expressed using arrows, and the expansion of the output vector 317 is expressed through “Expand”.

On the other hand, when the inception modules 313 , 314 , and 315 are disposed at the beginning of the CNN-Peaks model 310 , it is known that it causes inefficiency in terms of prediction accuracy. To avoid this potential problem, a convolution filter 311 and a max pooling layer 312 may be provided at the beginning of the CNN-Peaks model 310 . Details of the CNN-Peaks model 310 according to an embodiment are as follows.

First, the ChIP-seq read mapping pattern 301 and the input vector for annotation information (RefSeq) 302 may be converted into vectors through the convolutional layer 311 . The max pooling layer 312 may then reduce the dimension of this vector. And there are seven inception modules (313, 314, 315). Each inception module 313 , 314 , 315 has a layer that connects the output list of the filter. Two of these inception modules 313 , 314 , 315 are called A-modules 313 and can use three types of filters: convolution, max pooling, and average pooling. The A-module 313 may provide information on the peak pattern and magnitude of the strong peak signal to the CNN-Peaks model 310 . However, since max pooling and average pooling have the same number of features as the previous layer, this increases the vector size exponentially as a function of the depth of the pooling layer. To solve this problem, another three inception modules called B-module 314 can be used. The B-module 314 is composed of three types of convolutional filters without averaging pooling. Finally, two inception modules called C-module 315 can be used. The C-module 315 has four types of convolutional filters having a wider size and longer stride than the filters used in the A-module 313 and B-module 314 .

An average pooling layer 316 following a fully connected layer may be added to the end of the inception module 315 to reduce dimensions. Also, a residual structure between the inception modules 313 , 314 , and 315 may be used. This can avoid the vanishing gradient problem. Also, batch normalization can be used to avoid overfitting during training.

Output layer of CNN structure

In order to determine the presence or absence of a peak at each genomic location, the output layer 317 of the CNN-Peaks model 310 requires a number of neurons equal to the number of genomic bases. A large number of neurons in the output layer 317 usually causes a significant degradation in learning performance. Rather than calculating the probability of the presence of a p-value or peak signal at each individual genomic location to reduce the number of neurons, the optimal The CNN-Peaks model 310 can be designed to learn the threshold of . This can significantly reduce the number of neurons required for the output layer 317 of the CNN-Peaks model 310 and prevent performance degradation.

Since the size of the output vector 317 is smaller than that of the input vector, the size of the output vector 317 is expanded to be the same as that of the input vector, so that the presence or absence of a peak at each individual position can be predicted. Such an extension vector 320 is implemented using the "broadcasting vector" standard in Tensorflow and Numpy, and operations between vectors of different sizes are possible. The peak calling process of the CNN-Peaks model 310 will be described with reference to FIG. 4 .

Loss function

Determining the presence or absence of a peak signal is a binomial classification problem. Here, a cross-entropy 380 may be used as a loss function for learning the CNN-Peaks model 310 . Most methods for classification problems require a balancing act between sensitivity and specificity of performance. Likewise, you need to be careful not to focus on just one of them. In the peak call problem of ChIP-seq data, peaks are relatively sparse compared to the overall genome size. If a particular method tends to call it 'no-peak' (ie, it has a low false-positive error rate), it will miss important peaks but will have high accuracy. Therefore, the weighted cross-entropy can be used as a loss in the following equation.

[Equation 5]

는 주석 벡터,

는 모델의 파라미터 집합, y_i는 라벨링된 데이터에 대한 입력 i 번째 요소,

는 평가에서 거짓 양성 호출을 기준으로 한 거짓 음성(false-negative) 호출을 중요시하는 가중치,

은 주어진 입력 X, 모델 파라미터

에 대해 예측된 출력에서 i 번째 요소이다. 가중치

는 주어진 데이터에 대해 음의 영역(no-peak)과 양의 영역(peak) 사이의 비율로 결정될 수 있다.where X is the input read count vector,

is the annotation vector,

is the parameter set of the model, y _i is the input i-th element for the labeled data,

is the weight giving importance to false-negative calls based on false positive calls in the evaluation;

is the given input X, the model parameter

is the i-th element in the predicted output for . weight

may be determined as a ratio between a negative region (no-peak) and a positive region (peak) for a given data.

In addition, the Top-K method can be applied to the loss function. In the Top-K method, sensitivity is considered more important than specificity for high values of K, whereas specificity is more important for low values of K. To strike a balance between sensitivity and specificity, we set L = K/2, where L is the magnitude of the output vector. The final loss function can be expressed as the following equation.

[Equation 6]

은 모든 l_i 중에서 n 번째 가장 큰 개별 가중 교차 엔트로피 손실이다. 손실 함수

는 역전파(backpropagation)을 사용하여 파라미터

를 조절하는 아담 최적화(Adam optimizer)를 사용하여 최적화되었다.here,

is the nth largest individually weighted cross-entropy loss among all l _i . loss function

is a parameter using backpropagation

was optimized using an Adam optimizer that adjusts

Scoring peaks using sigmoid activation

An important task of the peak calling algorithm is to give each peak a significant score. Calculate the p-value for each genomic position in the Poisson distribution (counts for each position in the genomic range tag data of the ChIP experiment are known to follow the Poisson distribution), and compute these p-values in the CNN-Peaks model (310) The importance of the peak can be measured by combining it with the sigmoid value from the output layer 317 of . The sigmoid function 350 in the output layer 317 of the CNN-Peaks model 310 may generate a value that can be interpreted as a peak existence probability. The score value of each peak called by the CNN-Peaks model 310 may be determined by the product of the sigmoid activation value of the corresponding peak and the p-value of the Poisson distribution. This score can help users rate the importance of a particular peak.

4 is a diagram illustrating a peak calling process of a learned CNN model according to an embodiment.

Referring to Figure 4, it is the peak calling process using the trained CNN-Peaks model 410, where the signal is the read mapping depth of the ChIP-seq input data, and the hatched (gray box) part under the signal is in the RefSeq annotation information. indicates that the gene is present.

401 is a window with both ChIP-seq read mapping signals and RefSeq annotation information in the genomic region. The peak (hatched portion of 402) of the window may be predicted using the trained CNN-Peaks model 410, and may be generated in BED format at this time.

5 is a diagram for explaining a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis according to an embodiment.

Referring to Figure 5, which shows the pipeline of a CNN-based machine learning system that mimics human visual inspection for epigenomic data analysis, where each parallelogram represents data, rectangular boxes represent individual modules, and is indicated by dashed lines. The rounded rectangle represents the trained CNN-Peaks model 540, and the arrow represents the data flow. Here, the CNN-Peaks model may be included in the CNN model, and may be expressed as a CNN model.

A CNN-based machine learning system that imitates human visual inspection for epigenetic data analysis according to an embodiment may include a CNN model learning unit 570 and a CNN model peak prediction unit 580 .

The CNN model learning unit 570 learns the CNN model 530 from the cancer genome data analysis result data 501, 502, and 503, and the CNN model peak prediction unit 580 is the trained CNN model 540 is not trained. A peak may be called 505 to search for a functional region that may be a mechanism of cancer development for new cancer genomic data 504 that has not been identified.

As input data 501, 502, and 503 of training of the CNN model learning unit 570, ChIP-seq data 501 in BAM format, labeled data 502 in text format, and RefSeq annotation information 503 are can be used The input data 501 , 502 , and 503 may be converted into a vector having a shape suitable for the CNN model 530 . The CNN model 530 may be trained based on the data of these vectors.

The trained CNN model 540 can then be applied to unlabeled data 504 in the prediction step. The CNN model peak prediction unit 580 uses the trained CNN model 540 to call peaks for narrow histone modifications and transcription factor binding sites in other human ChIP-seq data 504 without additional labeling and training. can be used In addition, the peaks of ATAC-seq can also be detected using CNN-Peaks.

Unlike most other peak calling methods, CNN-Peaks do not require additional control ChIP-seq data and can generally be used as a background signal to reduce false-positive errors. That is, the user can determine the peak of the target ChIP-seq data 504 without control data.

6 is a flowchart illustrating an operation method of a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis according to an embodiment.

Referring to FIG. 6 , in the operation method of a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis performed through a computer device according to an embodiment, the CNN model learns cancer genome data analysis result data It can be made including a step (S110), and a step (S120) of calling a peak to search for a functional region that can be a cancer development mechanism for new cancer genome data that has not been trained by the learned CNN model.

The operation method of the CNN-based machine learning system may be described in more detail by taking the CNN-based machine learning system described in FIG. 5 as an example.

In step S110 , the CNN model learning unit 570 may learn the CNN model 530 from the cancer genome data analysis result data. The CNN model learning unit 570 may convert the input data 501 , 502 , and 503 of the CNN model 530 into vectors by using the preprocessing module 510 . More specifically, the pre-processing module 510 of the CNN model learning unit 570 is at least one of ChIP-seq read mapping data, data labeled at peak positions for at least some regions of ChIP-seq read mapping data, and RefSeq annotation information. One or more input data 501 , 502 , and 503 may be transformed into vectors through a convolutional layer. In addition, the preprocessing module 510 of the CNN model learning unit 570 reduces the dimension of the transformed vector through a max-pooling layer in order to reduce the noise of the input data 501 , 502 , and 503 . can do it

The CNN model learning unit 570 may include a plurality of inception modules including filters of at least any one of convolution, max pooling, and average pooling. Here, the plurality of inception modules may include a first inception module, a second inception module, and a third inception module, each of the first inception module, the second inception module, and the third inception module. may consist of at least one or more.

The first inception module may be configured to include a convolution, max pooling, and average pooling filter to provide the CNN model 530 with information on the peak pattern and magnitude of the strong peak signal. The second inception module may be configured to include a convolution and max pooling filter in order to avoid the problem of increasing the vector size of the max pooling and average pooling layer. In addition, the third inception module may include a convolutional filter having a wider size and a longer stride than the convolutional filters of the first inception module and the second inception module. For example, the first inception module, the second inception module, and the third inception module may correspond to the A-module, B-module, and C-module described with reference to FIGS. 2 and 3 . A vanishing gradient problem may be prevented by using a residual connection structure between the plurality of inception modules.

In the CNN model learning unit 570, an average pooling layer following a fully connected layer may be added to the end of the plurality of inception modules to reduce dimensions. Since the size of the output vector of the CNN model 530 becomes smaller than the input vector, the CNN model learning unit 570 expands the size of the output vector to be the same as the input vector, so that the presence or absence of a peak at each individual position can be predicted. . In addition, the CNN model learning unit 570 may measure the importance of the peak by giving the output data of the CNN model 530 a score to each peak using sigmoid activation. The CNN model learning unit 570 may learn the optimal threshold value of the genomic region based on the ChIP-seq read mapping data 501 of the window selected through the CNN model 530 .

In step S120, the CNN model peak prediction unit 580 calls the peak so that the learned CNN model 540 searches a functional region that can be a cancer-generating mechanism for the new untrained cancer genome data 504. (505) You can. More specifically, the CNN model peak prediction unit 580 may predict a peak 505 by inputting new cancer genome data 504 to the trained CNN model 540 and calling the peak and scoring each peak. .

As described above, according to embodiments, the deep learning model learns the data as a result of analysis of cancer genome data of a human researcher, and the learned deep learning model is a functional area in which new cancer genome data that has not been trained can become a cancer occurrence mechanism. It is possible to perform fast analysis with the accuracy of a human researcher by exploring

The device described above may be implemented as a hardware component, a software component, and/or a combination of the hardware component and the software component. For example, devices and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It may be implemented using one or more general purpose or special purpose computers, such as a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For convenience of understanding, although one processing device is sometimes described as being used, one of ordinary skill in the art will recognize that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that can include For example, the processing device may include a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as parallel processors.

Software may comprise a computer program, code, instructions, or a combination of one or more thereof, which configures a processing device to operate as desired or is independently or collectively processed You can command the device. The software and/or data may be any kind of machine, component, physical device, virtual equipment, computer storage medium or apparatus, to be interpreted by or to provide instructions or data to the processing device. may be embodied in The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

As described above, although the embodiments have been described with reference to the limited embodiments and drawings, various modifications and variations are possible from the above description by those skilled in the art. For example, the described techniques are performed in an order different from the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (20)

In the operating method of a CNN-based machine learning system that mimics human visual inspection for epigenetic data analysis performed through a computer device,
Learning a convolutional neural network (CNN) model from cancer genome data analysis result data; and
Calling peaks so that the trained CNN model searches functional regions that may be cancer-generating mechanisms for new untrained cancer genome data.
Containing, a method of operation of a CNN-based machine learning system. According to claim 1,
The step of the CNN model learning the cancer genome data analysis result data,
Converting the input data of the CNN model into a vector using a preprocessing module
Containing, a method of operation of a CNN-based machine learning system. 3. The method of claim 2,
The step of converting the input data of the CNN model into a vector using the pre-processing module comprises:
Transforming the input data of at least one of ChIP-seq read mapping data, data labeled at peak positions for at least a partial region of the ChIP-seq read mapping data, and RefSeq annotation information into a vector through a convolutional layer
Containing, a method of operation of a CNN-based machine learning system. 4. The method of claim 3,
The step of converting the input data of the CNN model into a vector using the pre-processing module comprises:
reducing the dimension of the transformed vector through a max-pooling layer to reduce noise of the input data;
Containing, a method of operation of a CNN-based machine learning system. According to claim 1,
The step of the CNN model learning the cancer genome data analysis result data,
Learning the CNN model through a plurality of inception modules including at least one filter of convolution, max pooling, and average pooling
Containing, a method of operation of a CNN-based machine learning system. 6. The method of claim 5,
Learning the CNN model through the plurality of inception modules comprises:
providing information on a peak pattern and a magnitude of a strong peak signal to the CNN model through at least one first inception module including a convolution, max pooling, and average pooling filter;
preventing the problem of increasing vector sizes of max pooling and average pooling layers through at least one or more second inception modules including convolution and max pooling filters; and
providing at least one third inception module comprising a convolutional filter having a wider size and a longer stride than the convolutional filter of the first inception module and the second inception module;
Containing, a method of operation of a CNN-based machine learning system. 6. The method of claim 5,
The step of the CNN model learning the cancer genome data analysis result data,
Using a residual connection structure between the plurality of inception modules to avoid the vanishing gradient problem.
A method of operation of a CNN-based machine learning system, characterized in that. 6. The method of claim 5,
The step of the CNN model learning the cancer genome data analysis result data,
At the end of the plurality of inception modules, an average pooling layer following a fully connected layer is added to reduce the dimension
A method of operation of a CNN-based machine learning system, characterized in that. According to claim 1,
The step of the CNN model learning the cancer genome data analysis result data,
Learning the optimal threshold value of the genomic region based on the ChIP-seq read mapping data of the window selected through the CNN model
A method of operation of a CNN-based machine learning system, characterized in that. The method of claim 1,
The step of the CNN model learning the cancer genome data analysis result data,
Since the size of the output vector of the CNN model becomes smaller than the input vector, the size of the output vector is expanded to be the same as the input vector, and the presence or absence of a peak at each individual position is predicted.
A method of operation of a CNN-based machine learning system, characterized in that. According to claim 1,
The step of the CNN model learning the cancer genome data analysis result data,
Measuring the importance of the peak by giving the output data of the CNN model a score to each peak using sigmoid activation
A method of operation of a CNN-based machine learning system, characterized in that. According to claim 1,
The step of calling the peak to search for a functional region that can be a cancer development mechanism for the new cancer genomic data that the learned CNN model is not trained is,
Predicting peaks through peak calling and scoring of each peak by inputting new cancer genome data to the learned CNN model
A method of operation of a CNN-based machine learning system, characterized in that. In a CNN-based machine learning system that mimics human visual inspection for epigenomic data analysis,
a CNN model learning unit in which a CNN (Convolutional Neural Network) model learns cancer genome data analysis result data; and
CNN model peak prediction unit that calls the peak to search the functional region that can be the mechanism of cancer development for the new untrained cancer genome data of the learned CNN model
Including, the operation system of a CNN-based machine learning system. 14. The method of claim 13,
The CNN model learning unit,
ChIP-seq read mapping data, data labeled at peak positions for at least some regions of the ChIP-seq read mapping data, and RefSeq annotation information A preprocessing module for converting the input data of the CNN model into a vector
Including, the operation system of a CNN-based machine learning system. 14. The method of claim 13,
The CNN model learning unit,
A plurality of inception modules including at least one filter of convolution, max pooling, and average pooling
Including, the operation system of a CNN-based machine learning system. 16. The method of claim 15,
The plurality of inception modules,
at least one first inception module configured to include a convolution, max pooling, and average pooling filter to provide information on a peak pattern and a magnitude of a strong peak signal to the CNN model;
at least one second inception module, configured to include a convolution and a max pooling filter, in order to avoid the problem of increasing the vector size of the max pooling and average pooling layer; and
at least one or more third inception modules including a convolutional filter having a wider size and a longer stride than the convolutional filters of the first inception module and the second inception module
Including, the operation system of a CNN-based machine learning system. 16. The method of claim 15,
Preventing a vanishing gradient problem by using a residual connection structure between the plurality of inception modules
characterized in that, the operation system of a CNN-based machine learning system. 16. The method of claim 15,
The CNN model learning unit,
At the end of the plurality of inception modules, an average pooling layer following a fully connected layer is added to reduce the dimension
characterized in that, the operation system of a CNN-based machine learning system. 14. The method of claim 13,
The CNN model learning unit,
Measuring the importance of the peak by giving the output data of the CNN model a score to each peak using sigmoid activation
characterized in that, the operation system of a CNN-based machine learning system. 14. The method of claim 13,
The CNN model peak prediction unit,
Predicting peaks through peak calling and scoring of each peak by inputting new cancer genome data to the learned CNN model
characterized in that, the operation system of a CNN-based machine learning system.

Images