KR102336311B1 - Model for Predicting Cancer Prognosis using Deep learning - Google Patents

Abstract

The present invention relates to a prognosis prediction system capable of predicting the prognosis of cancer from omics data of a cancer patient. It is possible to determine the importance of a single gene in predicting the prognosis of cancer.

Description

Cancer prognosis prediction model using deep learning {Model for Predicting Cancer Prognosis using Deep learning}

The present invention relates to a system for effectively predicting the prognosis of cancer by identifying factors affecting the prognosis of cancer based on deep learning using multi-omics data of cancer patients.

Predicting the prognosis of cancer, that is, predicting the survival period of cancer patients after surgery is a very important factor for cancer patients. Past prognosis prediction models used only clinical information such as the patient's age, cancer stage, and sex to predict the prognosis in a linear relationship only. However, this method has a limitation in predicting performance because it does not reflect the fact that the characteristics of each patient are different even for the same cancer.

In order to solve this problem, research on the diversity of cancer is being conducted recently. The Cancer Genome Atlas (TCGA) project, which was completed in 2018, produced various omics data that can confirm the cancer characteristics of individual patients. Recently, studies have been conducted to overcome the limitations of existing prognostic prediction models by applying such omics data to cancer prognosis prediction.

However, the prognosis prediction model of cancer studied so far has some limitations. First, the most widely used prognostic prediction method is a linear method. Although this method has the advantage of being able to intuitively check the effect on the prognosis, it has a limitation in that it is difficult to explain the biological complexity with a simple linear model, and thus the prognostic performance is slightly lowered.

Second, most cancer prognosis predictions use only a single omics. Until recently, many omics methods for identifying cancer characteristics have been developed and used for analysis, but only the omics method for confirming gene expression, such as RNA sequencing, is mostly used for prognosis prediction. However, cancer is caused by complex causes such as gene mutation as well as gene expression and affects the prognosis. Therefore, there is a problem that it is difficult to explain the diversity of cancer with only a single omics.

Third, in the prior art, it is difficult to select a final marker gene by confirming the importance of a single gene on prognosis. In the prior art, multiple multi-omics characteristics per one gene are independently assigned. However, this method has a disadvantage in that it is difficult to select which gene can be used as a marker for prognosis prediction because important genes are different for each omics characteristic.

In order to solve the above limitations, the present inventors have developed a system for predicting the prognosis of cancer that can learn biological diversity with a non-linear model, comprehensively consider molecular biological diversity of cancer using multi-omics, and determine the importance per single gene. developed and completed the present invention.

1. SCIENTIFIC REPORTS | 7: 16954 2. Clin Cancer Res; 24(6) March 15, 2018

It is an object of the present invention to provide a system for predicting the prognosis of cancer that can effectively predict the prognosis of cancer by integrally considering molecular and biological diversity of cancer based on multi-omics data.

In order to achieve the above object, an aspect of the present invention provides a system for predicting the prognosis of cancer using deep learning comprising the steps of:

(1) a learning module for learning the correlation between omics data collected from cancer patients and cancer prognosis;

(2) a storage module for storing the cancer prognosis prediction algorithm learned in the learning module; and

(3) A prediction module for receiving analysis information and calculating a prognosis of a cancer patient whose prognosis is unknown by using the cancer prognosis prediction algorithm stored in the storage module.

In one embodiment of the present invention, the system for predicting the prognosis of cancer may be used to predict the survival period after treatment or the survival period after surgical operation of a cancer patient.

As used herein, the term "omics" refers to various data generated at various molecular levels such as genome, transcriptome, proteomic, metabolite, etc. Multi-omics data is being generated according to the rapid development of information processing ability according to the development of industry.

In one embodiment of the present invention, the omics data is a genome, a transcriptome, a proteome, a metabolome, an epigenome, a lipodome, and a methyl body. (methylome) may be selected from the group consisting of, but is not limited thereto. In the present invention, a plurality of omics data may be used, and preferably, at least three omics data may be used, for example, genome, transcript, and methyl data may be used.

In one embodiment of the present invention, the learning module may include the following steps.

(a) a learning data generator for generating learning data for deep learning from omics data of cancer patients who know the prognosis;

(b) a deep learning performing unit learning the correlation between the omics data and the cancer prognosis from the generated learning data; and

(c) an algorithm generating unit for generating an algorithm for predicting the prognosis of cancer from the results of the deep learning performing unit.

As used herein, the term "deep learning" is also called deep learning, and is one of the technologies that are emerging in the field of machine learning recently. A plurality of hidden layers and a plurality of units included in them ( It is a neural network composed of hidden units.

When low level features are input to the deep learning model, these basic features are transformed into high level features that can better explain the problem to be predicted while passing through a plurality of hidden layers. In this process, since prior knowledge or intuition of an expert is not required, the subjective factor in feature extraction can be removed, and a model with higher generalization ability can be developed.

In addition, in the case of deep learning, since feature extraction and model construction are composed of one set, it has the advantage of forming a final model through a simpler process compared to the existing machine learning theory.

In the present invention, the deep learning may be performed based on a deep neural network (DNN), for example, may be performed with Google open source TensorFlow.

In the present invention, the learning data generating unit may align the omics data of the cancer patient with known prognosis in a matrix form for individual genes. By arranging in a matrix form, the prognosis of cancer is not predicted only with information on individual genes and single omics, but the prognosis of cancer can be predicted in a non-linear way by learning common characteristics of omics data for individual genes. The genes used are not selected separately, but all genes included in the omics data can be used.

In one embodiment of the present invention, the deep learning performing unit may include the following steps:

(b-1) learning common characteristics of omics data for individual genes from the aligned omics data;

(b-2) removing insignificant information from common features of omics data for individual learned genes;

(b-3) selecting the highest value per individual gene as a representative value for predicting the prognosis of cancer from the result of (b-2);

(b-4) confirming the accuracy of predicting cancer prognosis by combining the representative values for each individual gene selected in (b-3); and

(b-5) selecting a representative value combination with the highest accuracy from the result of (b-4).

In one embodiment of the present invention, step (b-1) may be performed by performing convolution with the aligned omics data, and step (b-2) is an error for each learned individual gene. This may be performed by applying a rectified linear unit (ReLU) function to a common feature of the mix data.

In one embodiment of the present invention, in the step (b-4), the negative log partial induction of the representative value per individual gene and the prognosis of cancer patients using a Cox proportional hazards model (negative log partial) likelihood) can be calculated.

Another aspect of the present invention provides a method for predicting the prognosis of cancer using the system for predicting the prognosis of cancer. Using the cancer prognosis prediction system, individual genes that have the greatest impact on cancer prognosis can be selected from multi-omics data collected from cancer patients with unknown prognosis, so the prognosis of cancer patients with unclear prognosis can be effectively predicted. have. In addition, by predicting the prognosis of cancer, if the prognosis is expected to be poor, it is possible to provide personalized medical care in such a way as to perform aggressive treatment.

According to the system for predicting cancer prognosis of the present invention, by using a plurality of omics data, the prognosis of a cancer patient can be predicted by considering the molecular and biological diversity of cancer in an integrated manner, and the importance of a single gene in predicting the prognosis of cancer is determined can do.

1 is a block diagram illustrating the configuration of a cancer prognosis prediction system according to the present invention.
2 shows an example of a learning data generator of a learning module in the cancer prognosis prediction system according to the present invention.
3 shows an example of the deep learning performing unit of the learning module in the cancer prognosis prediction system according to the present invention.
4 shows a result of comparing the prognosis prediction performance of the cancer prognosis prediction system and the existing cancer prognosis prediction model according to an embodiment of the present invention.
5 shows a result of comparing the prognosis prediction system for cancer according to an embodiment of the present invention and the prognosis prediction performance of a gene known as a cancer prognosis predictor marker.

Hereinafter, the present invention will be described with reference to the accompanying drawings. However, the present invention may be embodied in several different forms, and thus is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

When a part "includes" a certain component, it means that other components may be additionally provided without excluding other components unless otherwise stated.

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

1 exemplarily shows the configuration of the cancer prognosis prediction system of the present invention.

The cancer prognosis prediction system

(a) a learning module 100 for learning the correlation between the omics data collected from cancer patients and the prognosis of cancer;

(b) a storage module 200 for storing the cancer prognosis prediction algorithm learned in the learning module; and

and (c) a prediction module 300 that receives the analysis information and calculates a prognosis of a cancer patient who does not know the prognosis by using the cancer prognosis prediction algorithm stored in the storage module.

The learning module 100 of (a) is a part that learns by deep learning the correlation between the cancer prognosis, that is, the omics data of the cancer patient who already knows the survival period after treatment, and the cancer prognosis, and the learning data It includes a generating unit 110 , a deep learning performing unit 120 , and an algorithm generating unit 130 .

The omics data may be collected from a public database such as the Cancer Genome Atlas (TCGA), and may also be collected from the National Center for Biotechnology Information (NCBI). A plurality of the omics data may be used to integrate the molecular and biological diversity of cancer, and preferably three or more omics data may be used.

The learning data generating unit 110 is a part that processes omics data of cancer patients into learning data for deep learning, and as shown in FIG. 2 , each omics data is arranged in a matrix form for individual genes. Sorting in a matrix form is to construct a prognosis prediction system that comprehensively considers the diversity of cancer, such as gene expression and gene mutation, unlike the existing method that predicted the prognosis of cancer using only gene expression data.

The deep learning performing unit 120 is a part that learns the correlation between the omics data aligned for individual genes and the cancer prognosis by using different deep learning functions from the learning data, and as shown in FIG. 3 , It may include the following steps:

(b-1) learning common characteristics of omics data for individual genes from the aligned omics data;

(b-2) removing insignificant information from common features of omics data for individual learned genes;

(b-3) selecting the highest value per individual gene as a representative value for predicting the prognosis of cancer from the result of (b-2);

(b-4) confirming the accuracy of predicting cancer prognosis by combining the representative values for each individual gene selected in (b-3); and

(b-5) selecting a combination of representative values with the highest accuracy from the results of (b-4) as a marker for predicting cancer prognosis.

Step (b-1) may be performed by performing convolution with the omics data arranged in a matrix form. Specifically, according to Equation 1 below, convolution is performed with k kernels for n genes. For example, if k=3, convolution is performed for each of k=1, 2, and 3. Through this process, common characteristics of omics data for individual genes are learned for each kernel. After the deep learning is completed, in Equation 1 below, the size of w indicates the importance in predicting the prognosis of the omics data, and k is a value indicating high performance in the hyperparameter tuning (optimization) process. Hyperparameter means a value that must be set in advance to perform deep learning.

[Equation 1]

는 커널 k로 컨볼루션을 진행한 유전자 n에 대한 값을 의미함)(m=omics, n=genes, Xmn =matrix consisting of genes and omics features, w=weight multiplied by multi-omics data of individual genes, k=kernel performing convolution , and

denotes the value for gene n convolved with kernel k)

)만 남게 된다.The step (b-2) is a step of removing unimportant information by applying Equation 2 below to the common characteristics of the learned omics data for individual genes, and in this process, the resultant value (

) will remain.

[Equation 2]

The rectified linear unit (ReLU) function used in Equation 2 is an activation function used in a deep learning process. As described in Equation 2, when x is a value less than 0, 0 is used, and vice versa, 0 If it is a larger value, it is a function that uses the value as it is. In addition, b denotes a bias, and is automatically determined in the deep learning process.

)으로 선별하는 단계이다.In the step (b-3), max pooling according to Equation 3 is performed on the result value for the gene having high importance from (b-2), and the highest value per individual gene is obtained for the prognosis of cancer. Representative values for prediction (

) is the stage of selection.

[Equation 3]

The step (b-4) is a step of confirming the accuracy of predicting the prognosis of cancer by combining the representative values for each individual gene selected in (b-3). Specifically, the prognosis of cancer predicted by the combination of the representative values per individual gene and the Check the accuracy by comparing the known prognosis. At this time, using the Cox proportional hazards model according to Equation 4 below, the negative log partial likelihood between the representative value per individual gene and the prognosis of cancer patients was calculated and learning was carried out.

[Equation 4]

=매트릭스,

=콕스 비례 위험 모델의 파라미터,

=검열이 안된 샘플들(uncensored samples) 집단,

=검열된 샘플들(at-risk samples) 및

=생존 추적(follow up) 기간으로

보다 더 길 때를 의미함)(

= matrix,

= parameters of the Cox proportional hazards model,

= population of uncensored samples,

=at-risk samples and

= to follow up period

means longer than)

Step (b-5) is a step of selecting the combination of representative values with the highest accuracy from the result of (b-4) as a marker for predicting cancer prognosis. or a combination of genes is selected.

The storage module 200 is a part for storing the cancer prognosis prediction algorithm learned by the learning module, and is configured to include a cancer prognosis prediction algorithm database (database, DB) 220, and is configured for each learned individual gene. The omics data-prognosis related DB 210 for storing the correlation between the omics data and the prognosis of cancer may be further included.

The prediction module 300 includes an analysis information receiving unit 310 that receives information to be analyzed, and a prognosis result output unit 320 that outputs a prognosis calculated from the analysis information using a prognosis prediction algorithm stored in the storage module.

The analysis information may be omics data of a cancer patient with an uncertain prognosis, and the prognostic result may be a survival period after cancer treatment.

Hereinafter, the present invention will be described in more detail by way of Examples.

Example: Prognosis prediction system validation of cancer

The cancer prognosis prediction system (hereinafter, referred to as the prognosis prediction system) of the present invention uses genetic agent, transcriptome and methyl data among TCGA liver cancer patient data for learning. Specifically, 80% of the TCGA liver cancer patient data was used for learning, validation, and hyperparameter optimization of the prognostic prediction system, and the remaining 20% was used for testing the prognostic prediction system.

In order to confirm that the prognostic prediction system of the present invention does not overfit the data used for learning, but operates on a completely different dataset, the performance was confirmed not only on the TCGA data but also on the Korean liver cancer sample data, which is a completely independent dataset. proceeded

Specifically, gene expression, DNA somatic mutation, DNA germline mutation, copy number variation, and expression regulation (DNA methylation) for 211 TCGA liver cancer patients and 144 Korean liver cancer patients The prognosis of a liver cancer patient was predicted with the prognosis prediction system of the present invention using the omics features of the present invention, and the performance was compared with the prior art Glmnet and RSF (random survival forest) models.

As a result, as shown in FIG. 4 , the Glmnet model using a linear model had the lowest performance (concordance index, C-index) near 0.5 in both datasets. The C-index has a value between 0.0 and 1.0, and the closer to 0.5, the more random prediction is evaluated, and the closer to 1.0, the more accurate prediction is evaluated. The RSF model using the nonlinear model showed high performance in the TCGA liver cancer sample used for training, but the performance was low in the Korean liver cancer sample, confirming the phenomenon of overfitting the dataset used for training. On the other hand, the prognostic prediction system of the present invention showed higher performance than other models in all datasets.

In addition, 3 genes (UPB1, SOCS2 and RTN3), 11 genes (ACSM3, CXCL14, INTS8, LCAT, MARCO, PAMR1, CRHBP, DNASE1L3, FCN2, MT1X and VIPR1) known to predict the prognosis of liver cancer in the paper and Identification of 12 genes (ACSM3, CXCL14, INTS8, LCAT, MARCO, PAMR1, CRHBP, DNASE1L3, FCN2, MT1X and VIPR1) related to cancer associated fibroblasts (CAF) and prediction of the prognosis of liver cancer by these genes Performance was evaluated. As a result, as shown in FIG. 5 , it was confirmed that the prognosis predicting genes of liver cancer presented in the paper are overfitted to the TCGA sample used for learning.

Through these results, it was confirmed that the prognosis prediction system of the present invention showed significantly superior performance in predicting the prognosis of liver cancer compared to other prognostic prediction models due to the integrated learning of the genetic unit using various omics data.

100: learning module
110: training data generation unit
120: deep learning execution unit
130: Algorithm generator
200: storage module
210: omics data-prognosis related DB
220: prognosis prediction algorithm DB
300: prediction module
310: analysis information receiving unit
320: prognostic result output unit

Claims (9)

(1) a learning module for learning a correlation between three or more omics data collected from a cancer patient and a cancer prognosis;
(2) a storage module for storing the cancer prognosis prediction algorithm generated by the learning module; and
(3) a prediction module that receives the analysis information and calculates the prognosis of a cancer patient who does not know the prognosis by using the cancer prognosis prediction algorithm stored in the storage module;
The learning module is
(a) a learning data generator for generating learning data for deep learning by aligning omics data of cancer patients with known prognosis in a matrix form for individual genes;
(b) a deep learning performing unit learning the correlation between the omics data and the cancer prognosis from the generated learning data; and
(c) an algorithm generating unit that generates an algorithm for predicting the prognosis of cancer from the results of the deep learning performing unit; The system for predicting the prognosis of cancer according to claim 1, wherein the prognosis of the cancer is the survival period after cancer treatment. The system according to claim 1, wherein the omics data is selected from the group consisting of a genome, a transcript, a proteomic, a metabolite, an epigene, a lipid, and a methyl body. delete delete delete The method of claim 1, wherein the deep learning performing unit
(b-1) learning common characteristics of omics data for individual genes from the aligned omics data;
(b-2) removing insignificant information from common features of omics data for individual learned genes;
(b-3) selecting the highest value per individual gene as a representative value for predicting the prognosis of cancer from the result of (b-2);
(b-4) confirming the accuracy of predicting the prognosis of cancer by combining the representative values for each individual gene selected in (b-3); and
(b-5) selecting a combination of a representative value with the highest accuracy from the result of (b-4) as a cancer prognosis predicting marker; a system for predicting cancer prognosis. The method according to claim 7, wherein in step (b-4), a negative log partial likelihood of the representative value per individual gene and the prognosis of a cancer patient is obtained using a Cox proportional hazards model. A system for predicting the prognosis of cancer that calculates. A method for predicting the prognosis of cancer using the system for predicting the prognosis of cancer according to any one of claims 1 to 3, 7 and 8.

Images