KR20210110241A - Prediction system and method of cancer immunotherapy drug Sensitivity using multiclass classification A.I based on HLA Haplotype - Google Patents

Abstract

The present invention relates to a system and method for predicting cancer immunotherapy drug indication and sensitivity with high precision by converting the reactivity of a cancer immunotherapy drug into a dielectric property of a target dielectric substance and a human leucocyte antigen haplotype (hereinafter called the HLA haplotype). According to the present invention, the method comprises the following steps of: receiving analysis target information including the target dielectric substance and drug data; calculating the target HLA haplotype information of the dielectric substance included in the analysis information, and calculating the target configuration information of the drug included in the analysis information; and calculating a reactivity prediction result of the drug with respect to the dielectric substance included in the analysis target information in the reaction correlation between the analysis target dielectric substance information and the analysis target configuration information by the reactivity prediction algorithm and outputting the calculated results. The target configuration information and the HLA haplotype information respectively correspond to the configuration information and the HLA haplotype information applied to each machine learning process. As such, according to the present invention, the present invention is able to use the HLA haplotype information, use a large number of variations in the base sequence collected from the target dielectric substance for the drug reactivity prediction using the HLA haplotype information, predict the HLA haplotypes, calculate the results of the LASSO regression model between the HLA haplotype and the drug reactivity information, calculate the correlation with the gene expression volume by using the GBM regression, and predict drug reactivity for the in silico clinical trial which filters cancer drugs with high sensitivity among drugs in an effective manner.

Description

Human leukocyte antigen haplotype-based multiclass classification A.I based on HLA Haplotype

The present invention relates to a system and method for predicting high-precision immunotherapy indications and sensitivity by converting the reactivity of an immuno-oncology agent into a genetic characteristic of a target genome and a human leukocyte antigen haplotype (hereinafter, HLA haplotype).

Recently, innovations in next generation sequencing (NGS) technology have made great strides in understanding complex and diverse cancers. In addition, an international consortium effort has developed and published catalogs of somatic mutations in these carcinomas as well as a comprehensive database of driver mutations. Due to these international consortium research achievements, expectations for cancer-customized treatment for specific genomic fingerprints of individual tumors have rapidly increased. However, there are still not enough new, personalized cancer therapies that are currently approved and used clinically by cancer patients and all stakeholders in the medical community, including the pharmaceutical industry. Therefore, there is a need for an efficient and systematic approach to predict the personalized association between genomic information and anticancer drug response.

Several collaborative consortia have been established to integrate molecular profiling data of cancer cell lines and drug toxicity data (www.lincsproject.org). The most important goal of this consortium is to discover genomic biomarkers that can predict anticancer drug toxicity and personalized drugs.

Among the cell line genome databases for drug toxicity in cancer, GDSC (Genomics of Drug Sensitivity in Cancer) is a publicly available database. GDSC is a public database that experimentally measured drug toxicity information of 1,070 human cancer cells for 265 anticancer compounds. The cell line project of GDSC used here was published at the following site (CCLP: COSMIC Cell Lines Project, http:/ /cancer.sanger.ac.uk/cell_lines). These shared resources can be utilized to realize genome-based precision cancer therapeutics.

However, despite the potential value of these databases, the high level of data and complexity present many technical challenges for integrated analysis. Accordingly, many computational methods have been developed to systematically identify molecular biomarkers in anticancer drug toxicity, but despite these efforts, drug toxicity is limited to some specific cell lines and a given set of genetic mutations. This is because the genetic information of all people is different, and since the common mutations are only a small part of the whole, it is impossible to determine the correlation.

With recent developments in information technology, methods increasingly used to solve the aforementioned complex problems are called deep learning models, or deep learning models. Deep learning learning method is a field of technology that performs deep machine learning from a large amount of high-dimensional raw data. Until recently, there were many limitations directly due to the limitation of the amount of computation for learning. However, methodological improvements and the use of powerful machines by parallel computing have made it possible to train deep learning learning models in different layers containing thousands of hidden units. Accordingly, various types of structural information such as pharmacological, genomic, transcriptomic and epigenomic data and their drug reactivity data can be learned, making it suitable for predicting drug-target interactions from diversity information.

The pharmaceutical industry is starting to show great anticipation for learning deep learning that leverages this type of data to develop new drugs. Recently, several promising results have been demonstrated using artificial intelligence in drug development. Drug repositioning with superior predictive accuracy compared to drug-target profiles and other traditional machine learning models is also possible. However, most of the approaches have been limited to proof-of-concept, and production-ready solutions for drug discovery through deep learning learning are currently lacking.

Currently, PubChem (pubchem.ncbi.nlm.nih.gov) is operated by the National Center for Technology Information (NCBI) in the United States, and provides information on about 100 million compounds, 200 million substances, and bioassays. (en.wikipedia.org/wiki/PubChem).

In addition, there are many ways to express this compound as a pharmacophore descriptor. Among them, the PaDELL method can be expressed in drugs with 1,875 (1,444 1D and 2D, and 431 3D features), and 12 fingerprints (total about 16,092 bits). In addition, mutations in the genome can extract various features.

In this prior art, QSAR (Quantitative structureactivity relationship), drug development using drug cytotoxicity data, expression control of whole genome sequencing based on Deep Learning, structural variation, etc. are independently applied. Since it is only used, there was a problem that an accurate correlation could not be predicted.

On the other hand, recently, drug reactivity studies using expression information and biomarker information of drug and target gene drug given in a systematic statistical manner and target gene expression and target gene expression were recently published. Although there was no study on biomarkers in this paper, the inhibition of several major drugs against major targets using known gene biomarkers and expression information was evaluated by CCLE (Cancer Cell Line Encyclopedia), CTRP (Cancer Therapeutics). Response Portal, Broad) and some GDSCs were used to predict (CARE) and verify the results.

In addition, many major studies have been published on the disease or phenotypic association between expression and quantitative trait loci (eQTL), that is, expression and genomic variation. In addition, many major studies have been published on the disease or phenotypic association between protein and Quantitative Trait Loci (pQTL), that is, protein production and genomic mutations.

However, in such previous studies, studies on the drug response or drug phenotype association between cQTL (Copy Number Variation and Quantitative Trait Loci), that is, CNV (copy number variation), a mutation that increases or decreases drug reactivity and gene copy number, and genomic mutations were not It has not been announced yet.

Therefore, in the GBLscan (genetic biomarker label scan) method according to the present invention, (a) eQTL (expression and Quantitative Trait Loci), that is, disease or phenotype association between expression and genomic variation, (b) cQTL (Copy Number Variation and Quantitative Trait) Loci), i.e., gene copy number mutation (copy amplification) and copy deletion (copy deletion) (b) as there is no reference yet, so the term quantitative locus (QTL) is used in a similar manner to (a) in this patent. It is characterized by the progress of the work of

Forbes, S. A., et al. COSMIC: somatic cancer genetics at high-resolution. Nucleic Acids Research. 45, 777-783 (2017).

The present invention has been devised to solve the problems described above, and the present invention relates to known cell line genome gene expression information and in vivo drug response to drugs-cell lines-toxicity (IC50). ) to provide a system that can reliably predict through information-based linear regression modeling and various machine learning methods.

And the present invention is to solve the problems of the prior art as described above, A) pre-calculating gene RNA expression information that is difficult to extract clinically from cancer patients as quantitative trait location (QTL), B) cell line DB (GDSC) After obtaining gene RNA expression information and drug reactivity correlation experiment information using public data such as it is for

In addition, the present invention is to predict drug reactivity information of cancer patients by calculating the HLA haplotype including functional information based on genomic information that can be easily obtained from cancer patients.

According to a feature of the present invention for achieving the above object, the present invention comprises the steps of: receiving analysis target information including genome and drug data to be analyzed; calculating the analysis target HLA haplotype information of the genome included in the analysis information, and calculating the analysis target composition information of the drug included in the analysis information; Calculating the reactivity prediction result of the drug with respect to the genome included in the analysis target information in the reaction correlation between the analysis target genome information and the analysis target configuration information by a reactivity prediction algorithm, and outputting the calculated result Performed: The analysis target configuration information and HLA haplotype information are information corresponding to configuration information and HLA haplotype information applied to each machine learning.

Meanwhile, in the present invention, CDRscan (Cancer drug response scanning), an AI deep learning method that integrates heterogeneous characteristic information (genomic information, QSAR information and expression information) into drug-cell lines-toxicity (IC50) data The prediction accuracy is further improved compared to previous computer modeling approaches. In particular, the present invention can provide an interaction model of virtual drugs vs. virtual cells (cell lines) or target proteins. Here, two different heterogeneous characteristic virtual information is explained by the PaDELL method in the case of the first drug.

And in the present invention, it can be described as a literature method for the genomic fingerprint, or a set of mutation features of the whole genome (or target protein), and an accurate drug response prediction model and drug reuse / repositioning (drug repositioning), It can be used as a clinical decision supporting system for screening chemical substances, discovering new anticancer drug candidates, and selecting patient-specific anticancer drugs.

On the other hand, the present invention, in the immuno-oncology indication and response prediction method for receiving the analysis information and calculating the reactivity prediction result of the drug with respect to the genome included in the analysis information, the analysis including the genome and drug data to be analyzed receiving target information; calculating the analysis target HLA haplotype information of the genome included in the analysis information, and calculating the analysis target composition information of the drug included in the analysis information; Calculating the reactivity prediction result of the drug with respect to the genome included in the analysis target information in the reaction correlation between the analysis target genome information and the analysis target configuration information by a reactivity prediction algorithm, and outputting the calculated result Performed: The analysis target configuration information and HLA haplotype information may be information corresponding to configuration information and HLA haplotype information applied to each machine learning.

In this case, the analysis target composition information and HLA haplotype information may be functional group information constituting the drug and HLA haplotype information included in the genome.

In addition, the analysis target composition information and HLA haplotype information may be functional group information constituting a drug and characteristic information on mutations included in the genome.

Here, the reactivity prediction algorithm may be an algorithm learned by multi-classification machine learning of the reactivity correlation of constituent information constituting the drug with respect to the HLA haplotype information included in the genome from the collected learning information.

And the machine learning is a LASSO regression process that classifies and aligns the learning target data according to a plurality of cell lines and a plurality of drugs, and learns all genomic variables based on the reactivity (IC50) based on the convolutional product. class; GBM regression process for classifying and arranging learning target data according to multiple cell lines and multiple drugs, and learning based on convolution based on expression level information of all genes; And LASSO classification that learns based on convolution with the HLA haplotype information calculated from the LASSO regression process and the GBM regression process and the parameters for each drug reactivity and gene expression level merged ( classification) process.

In addition, the machine learning may be performed by LASSO machine learning that performs machine learning using HLA haplotype information and drug responsiveness information (IC50), which are final analysis targets, as learning elements.

And the LASSO machine learning comprises the steps of: (A) collecting, by a learning data generator, learning information indicating a degree of response to each drug for each cell line genome; (B) generating genetic information for HLA haplotypes included in the learning information; (C) generating configuration information constituting the drug included in the learning information; (D) generating pre-processed learning data indicating a degree of response to a group of constituent information constituting a drug to the HLA haplotype information group included in the learning information; (E) deriving a response correlation of individual configuration information to individual HLA haplotype information through machine learning on the pre-processed learning data; and (F) generating a drug reactivity prediction algorithm composed of constituent information about the genome including HLA haplotype information through the response correlation of the constituent information to the learned HLA haplotype information. It may be performed by including ;.

Meanwhile, the machine learning may be performed by GBM machine learning that performs machine learning using HLA haplotype information and the expression level of a specific gene as learning factors.

And the GBM machine learning comprises the steps of: (a) collecting, by the learning data generator, learning information indicating the degree of response to each drug for each cell line genome; (b) generating genetic information for HLA haplotypes included in the learning information; (c) generating HLA haplotype information pre-processing data indicating a drug's reactivity to the HLA haplotype information group included in each genome; (d) deriving a drug response correlation to each genetic information through multi-classification machine learning for the genetic information learning information; (e) generating configuration information constituting the drug included in the learning information; (f) generates compositional information preprocessing learning data indicating the reactivity of the compositional information group constituting the drug for each genome, and the response correlation of each compositional information for each genome through the GBM algorithm for the compositional information learning data deriving; (g) Deriving the response correlation of individual gene expression level information to individual haplotype information through the correlation between the drug response to each gene expression level information and each HLA haplotype information for each genome It may be carried out including;

Here, the learning information may be test result data on the reactivity of various drugs to various cell lines.

In addition, the HLA haplotype information may be mutation information or characteristic information on mutations.

In addition, the composition information may be functional group information constituting the drug.

And the criterion of drug reactivity in the step (E) may be determined based on the acceptance inhibition index IC50.

On the other hand, the present invention is an immuno-oncology indication and response prediction system that receives analysis information and calculates a result of predicting drug reactivity with respect to the genome included in the analysis information, analysis including the target genome and drug data an input unit for receiving target information; a comparison data generation unit for generating comparison data comprising gene expression data and drug data included in the analysis target information in a form corresponding to HLA haplotype information and configuration information used for machine learning, respectively; And, by the reactivity prediction algorithm derived by the reactivity prediction algorithm component part, the human leukocyte antigen haplotype-based multi-classification comprising a prediction result generating unit that calculates the response prediction result of the drug to the genome included in the analysis target information It includes an immuno-oncology indication and response prediction system using an artificial intelligence model.

In this case, the HLA haplotype information may be calculated by a program that allows a plurality of polymorphic markers and SNPs using likelihood-based inference.

The following effects can be expected from the system and method for predicting immunotherapy indications and responses using the human leukocyte antigen haplotype-based multi-classification artificial intelligence model according to the present invention as discussed above.

That is, in the present invention, in order to calculate a drug reactivity prediction model by using HLA haplotype information, a step of predicting HLA haplotypes using a plurality of mutations in a nucleotide sequence collected from a genome to be tested and HLA haplotype The step of calculating the solution of the LASSO regression model between the drug and drug reactivity information and the correlation with the gene expression level using this solution, using GBM regression to effectively filter out highly sensitive cancer drugs among drugs (In Silico Clinical Trial) It has the effect of predicting drug reactivity for

1 is an exemplary view illustrating a system for implementing a method for multiple classification of drug reactivity and gene expression level based on HLA haplotype according to the present invention.
Figure 2 is an exemplary view showing the result of calculating the correlation between the cell line drug response value (IC50) and HLA haplotype according to the present invention by LASSO regression.
Figure 3 is an exemplary view showing the correlation between the reactivity information (IC50) and gene expression level of the drug according to the present invention.
4 is an exemplary diagram illustrating the correlation of HLA haplotypes predicted and tested through the multi-classification machine learning model according to the present invention.
5 to 9 are diagrams showing the results of testing the sensitivity between the HLA haplotype predicted through the multi-class machine learning model according to the present invention and five cancer drugs.

Hereinafter, with reference to the accompanying drawings, a system and method for predicting immunotherapy indications and responses using a human leukocyte antigen haplotype-based multi-classification artificial intelligence model according to a specific embodiment of the present invention will be described.

Hereinafter, the main analysis contents of the immuno-oncology indication and response prediction system according to the present invention will be reviewed.

1. Evaluation of markers and haplotypes

Populations of individuals with genetic diversity do not have the same genome. Rather, the genome exhibits sequence variability between individuals at multiple locations within the genome. That is, there are multiple polymorphic sites within the population. In some cases, in the absence of a reference allele, different alleles of one polymorphic site may appear. Alternatively, a reference sequence may be referenced for a particular polymorphic site.

By “marker” is meant a characteristic genomic sequence of a particular variant allele. Markers may include alleles of any variant type found in the genome, including SNPs, microsatellites, insertions, deletions, duplications and translocations.

"Haplotype" means a segment of a genomic DNA strand characterized by a particular combination of genetic markers (alleles) arranged along the segment. For example, a haplotype may comprise one or more alleles, two or more alleles, three or more alleles, four or more alleles, or five or more alleles. As used herein, the term “reactivity” encompasses both increased sensitivity and decreased sensitivity. Thus, certain markers and haplotypes of the invention may be characterized by increased sensitivity. Markers and haplotypes that confer increased sensitivity are also considered to confer drug responsiveness.

2. Haplotype analysis and prediction

An approach to haplotype analysis involves using likelihood-based inference applied to a nested model. The method is implemented in HLAscan, a program that accepts multiple polymorphic markers and SNPs. The method is specifically designed for case-control studies whose purpose is to identify a group of haplotypes that confer different characteristics. It is also a tool for studying LD structures.

In this case, the maximum likelihood estimate, the likelihood ratio, and the parameter p-value are calculated directly, and the data treats it as a deleted-data problem.

And to produce a valid p-value, because of the uncertainty of the deleted genotype, one can rely on a likelihood ratio test based on the likelihood calculated directly on the observed data in which the information loss occurred, i.e. how much information is incomplete because the information is incomplete. It is important to know if is lost.

3. Machine Learning Methods for Association Analysis

In the field of machine learning, it is largely divided into supervised learning with correct answers and unsupervised learning. Supervised learning is learning that proceeds by giving the correct answer. Therefore, the correct answer should always be provided along with the data when learning. In supervised learning, the words correct answer, actual value, label, target, class, and y value are often used interchangeably, but they all mean the same thing. It is used when you need to predict the label of new data using the given data and labels. It has the advantage of being able to easily evaluate the performance of the model because it is labeled with the data. Classification and regression are typical examples of supervised learning. The biggest difference between classification and regression is that when data is input, classification predicts discrete values and regression predicts continuous values. The data of the present invention are divided into reactive and non-reactive.

In the present invention, analysis is performed using LASSO and Gradient Boosting Model (GBM), which are machine learning methods specialized for classification.

LASSO (least absolute shrinkage and selection operator) machine learning is an analysis method using variable selection and regularization to increase the accuracy of estimation.

This has the effect of partially improving the problem of overfitting due to too many estimation coefficients by adding a constraint to the basic OLS estimation equation and increasing the accuracy of the estimation. In LASSO, since the constraint region is square, the estimate value reaches the edge and the coefficient value is likely to be 0. That is, automatic variable selection is made by estimating the coefficients of input variables without explanatory power to 0. In addition, LASSO has the advantage of continuously selecting variables through reduction of specific input variables.

GBM is a compound word of gradient descent and boosting, and is a powerful algorithm that combines boosting with gradient descent. Gradient descent is an optimization method that reduces errors by moving the error function in the opposite direction to the differential value, and boosting refers to a method of creating more accurate and powerful learners by combining simple and weak learners. Even if the accuracy is low, it is a method to create a model first, and then the model compensates for the calculated error. Unlike random forest, GBM builds trees sequentially in a way that compensates for errors in the previous tree. Usually one to five shallow trees are used, so it uses less memory and predicts faster. It is an algorithm that connects multiple simple models of such a shallow tree structure.

4. Composition of Analysis Objects and Algorithms

As shown in FIG. 1, the immuno-oncology indication and response prediction system according to the present invention is based on the human leukocyte antigen haplotype and the corresponding HLA haplotype and the drug (anticancer agent) reactivity (IC50) and specific genes of the disease. It is a machine learning system that predicts the response of a drug by analyzing the relationship with the expression level.

The method discovers a corresponding anticancer drug and a specific HLA haplotype predicted by different independently designed machine learning functions (3), and finally calculates whether the drug is reactive or not.

At this time, the different machine learning functions are broadly classified: 1) a machine learning function that performs machine learning using HLA haplotype information and drug responsiveness information (IC50), which are the final analysis target, as learning elements, and 2) HLA under A machine learning function that performs machine learning using flow type information and expression levels of specific genes as learning factors, and 3) machine learning using genome and haplotype information as learning factors to calculate the first-learned relationship. Then, it is divided into a machine learning function that performs secondary learning on these information.

Hereinafter, the configuration and method of the present invention for implementing and performing this will be described with reference to the drawings.

The immuno-oncology indication and response prediction system according to the present invention comprises a prediction module, wherein the prediction module receives analysis information and calculates a result of predicting drug reactivity with respect to the genome included in the analysis information, To this end, the prediction module is configured to include an input unit, a comparison data generation unit, and a prediction result generation unit.

In this case, the input unit receives analysis target information, and the input target information refers to information including genome and drug data to be analyzed.

And the comparison data generation unit is a part that generates comparison data composed of gene expression data and drug data included in the analysis target information in a form corresponding to HLA haplotype information and configuration information used for machine learning, respectively. That is, when the machine learning is performed with HLA haplotype information and functional group information, the comparison data generator calculates HLA haplotype data of the genome included in the analysis target information, and the functional group included in the analysis target information yield information.

Of course, when the machine learning is performed with gene expression level information and functional group information, the comparison data generation unit calculates characteristic information about the gene expression level of the genome included in the analysis target information, and includes Functional group information is calculated.

And the prediction result generation unit is a part that calculates the prediction result of the drug response to the genome included in the analysis target information by the reactivity prediction algorithm derived by the reactivity prediction algorithm component.

On the other hand, the method for predicting the indication and response of an immuno-oncology drug according to the present invention starts with the input unit receiving the analysis target information including the genome and drug data to be analyzed.

Then, the comparison data generation unit calculates the analysis target HLA haplotype information of the genome included in the analysis information, and calculates the analysis target composition information of the drug included in the analysis information.

At this time, as described above, the analysis target configuration information and HLA haplotype information correspond to the configuration information and HLA haplotype information applied to each machine learning, and are included in the functional group information constituting the drug and the genome, respectively. It may be HLA haplotype information or characteristic information on mutations included in the genome.

And the reactivity prediction algorithm calculates the reactivity prediction result of the drug with respect to the genome included in the analysis target information in the reaction correlation between the analysis target genome information and the analysis target composition information, and outputs the calculated result.

Hereinafter, the basic configuration of machine learning for the implementation of the system and method for predicting immuno-oncology indications and responses according to the present invention will be described.

(1) Genotyping and haplotype analysis

First, as shown in FIG. 1 , the drug indication and response prediction according to the present invention is performed including the HLA haplotype prediction step, the machine learning step, and the statistical analysis step.

In this case, the HLA haplotype prediction step predicts through HLAscan, a program that allows multiple polymorphic markers and SNPs using likelihood-based inference.

The technical configuration of such HLA haplotype prediction has also been partially disclosed in Patent Registration No. 10-1815529 (Human Haplotyping System and Method), which is an earlier application of the present applicant.

(2) Machine learning method for association analysis

The machine learning step is a part that learns the reactivity correlation of the constituent information constituting the drug to the HLA haplotype information contained in the genome from the collected learning information by multi-classification machine learning.

Information on drug responsiveness to cell lines, expression levels of genes before and after drug treatment for cell lines, and drug sensitivity to cancer cells are collected from the genomics database (GDSC).

In order to perform such a function, the machine learning model is configured to include LASSO regression, GBM regression, and LASSO classification.

Here, the LASSO regression is a part that generates learning data for multi-classification machine learning from the collected learning information, and the GBM regression performs multi-classification machine learning through a plurality of learning data generated from the learning data generator. The LASSO classification is a part that generates a reactivity prediction algorithm for predicting the reactivity of a drug to HLA haplotype information from the results learned from the deep learning machine learning unit.

In this case, the LASSO method and the GBM method may be set in various ways according to data information received as input in the corresponding step. That is, the parameters of the LASSO method and the GBM method may be set to sub-unit information constituting the genome and drug, respectively, or various information included therein.

That is, in the present invention, the accuracy of the reactivity prediction result is improved as there are more common elements between the target for which machine learning is performed and the target of information input for analysis. If the unit is set, the accuracy of the prediction result of the reactivity to the unknown drug may be improved.

In a specific embodiment of the present invention, when mutation information and characteristic information are set as the HLA haplotype information, when mutation information is set as the HLA haplotype information and the corresponding machine learning is performed , the more common elements between the mutations in the cell line included in the learning information, which is the subject of learning, and the mutations in the genome input as the analysis target, the higher the accuracy of the analysis.

On the other hand, when the characteristic information is set as the genetic information and multi-classification machine learning is performed, there is at least a common mutation between the mutation of the cell line included in the learning information that is the learning target and the mutation of the genome input as the analysis target, Reactivity can be accurately predicted according to the similarity of each mutation.

Accordingly, the HLA haplotype information may be mutation information included in the genome or characteristic information on mutations included in the genome.

In this case, when the genetic information is mutation information, the learning data is a plurality of pieces of information indicating a degree of response to a group of functional group information constituting a drug with respect to a group of mutation information included in the cell line.

(3) LASSO regression regularization and machine learning method through association matrix

On the other hand, the machine learning model learns the response correlation of each constituent information constituting the drug with respect to each HLA haplotype information included in the cell line through specialized machine learning for the learning data.

Here, the machine learning model may be performed by LASSO machine learning and GBM machine learning.

First, LASSO machine learning is a method of performing machine learning using HLA haplotype information and drug responsiveness information (IC50), which are final analysis targets, as learning elements.

The LASSO machine learning starts with the learning data generating unit collecting learning information indicating the response to each drug for each cell line genome.

In this case, the learning information refers to test result data on the reactivity of various drugs to various cell lines. Thereafter, the learning data generator generates genetic information for HLA haplotypes included in the learning information. Here, the HLA haplotype information may be mutation information or characteristic information on mutations.

And the learning data generation unit generates the configuration information constituting the drug included in the learning information. In this case, the composition information may be functional group information constituting the drug.

Next, the learning data generation unit generates pre-processed learning data showing the response to the configuration information group constituting the drug with respect to the HLA haplotype information group included in the learning information. Here, the data combined in a form for applying the learning information to the LASSO model is theoretically generated as much as the number of cell lines x the number of drugs included in the learning information.

Thereafter, the machine learning unit derives a response correlation of individual configuration information to individual HLA haplotype information through LASSO machine learning for the preprocessed learning data. Here, the criterion for predicting and reactivity of the drug may be determined based on the receptivity inhibitory index IC50. The IC50 refers to the concentration of a drug required to kill 50% of cells in a cell line, and the lower the IC50 value, the higher the reactivity of the drug.

Next, the reactivity prediction algorithm configuration unit, through the response correlation of the configuration information to the HLA haplotype information learned by the LASSO machine learning unit, configuration information on the genome including HLA haplotype information We create an algorithm for predicting drug reactivity consisting of In this case, the LASSO machine learning unit may be configured to calculate a final predicted value from the average of each predicted value after performing the deep learning machine learning in a plurality of methods.

Next, the GBM machine learning unit, which is the second method of multi-classification machine learning according to the present invention, also starts with the learning data generating unit collecting learning information indicating the response to each drug for each cell line genome.

And the learning data generation unit generates genetic information for the genomes included in the learning information. Also in this case, the HLA haplotype information may be mutation information or characteristic information on mutations.

Thereafter, the learning data generator generates HLA haplotype information pre-processing data representing the drug's reactivity to the HLA haplotype information group included in each genome.

In addition, the multi-classification machine learning unit derives a drug response correlation to each genetic information through multi-classification machine learning for the genetic information learning information.

Next, the learning data generation unit generates the configuration information constituting the drug included in the learning information.

Thereafter, the composition information pre-processing learning data indicating the reactivity of the composition information group constituting the drug for each genome is generated, and the response correlation of each composition information for each genome through GBM machine learning for the composition information learning data to derive

And the GBM machine learning unit response correlation of the individual gene expression level information to the individual haplotype information through the correlation of the drug response to each gene expression level information and the correlation of each HLA haplotype information for each genome draw a relationship

The second method of such GBM machine learning is to distribute the machine learning process when the number of HLA haplotype information and drug composition information included in the genome is large, so that the processing process can be dualized as well as correlation It can improve the accuracy of the relationship.

To summarize the multi-classification machine learning method as described above,

Multi-classification machine learning according to the present invention is performed including a LASSO regression process, a GBM regression process, and a LASSO classification process.

At this time, the LASSO regression process refers to a process of classifying and aligning the learning target data according to a plurality of cell lines and a plurality of drugs, and learning all genomic mutations based on the reactivity (IC50) based on convolution.

And the GBM regression process refers to a process of classifying and arranging learning target data according to multiple cell lines and multiple drugs, and learning based on convolution based on expression level information of all genes.

Finally, the LASSO classification process refers to a process of convolution-based learning while merging HLA haplotype information calculated from the LASSO regression and GBM regression processes and parameters for drug reactivity and gene expression levels.

Specifically, the first step is 1) calculating the expression correlation between mutations (4 million in total) and genes (20,000), and 2) gene expression of drugs (265) and cell line genomes (1,000) ( Each of the 20,000) correlation calculation step, 3) 1) and 2) collecting the same gene expression pattern associated with multiple or single mutations, and 4) determining the HLA haplotype of the genome with drug reactivity It is carried out including the step of extracting together.

Through such a learning process, when learning in step 1 and new genome genetic information and drug property information of a drug are input in step 2, the degree of reactivity (IC50) of drug property data input from the genome mutation information is calculated in step 3 can be predicted as

Hereinafter, embodiments of the drug indication and response prediction system and method according to the present invention will be described in detail.

(1) Reproducible embodiment of selected haplotype relevance to drug

As described above, the reproduction of the selected haplotype relevance to the drug is an embodiment consisting of two consecutive steps in the multi-classification machine learning of the present invention, in the first step, the genome sequence data of the tumor and the anticancer drug From the chemical properties, 30 HLA haplotype information and 1,072 drug reactivity information features are extracted, respectively.

Next, each combination of haplotype information and drug reactivity information sets independently generates a specific haplotype set for each specific drug reactivity using the LASSO regression model.

On the other hand, two different input sources are used in the multi-classification machine learning model, each representing genomic HLA haplotype information of individual cancer cell lines and chemical properties of anticancer drugs.

(2) Implementation example of machine learning data configuration

In the multi-classification machine learning model according to the present invention, machine learning is performed by merging heterogeneous information, wherein the heterogeneous information is genomic mutation and HLA haplotype information and functional group information or phenotype information of the drug, respectively, which are distinguished by cell line and drug. can

That is, these different pieces of information are arranged according to the criteria used by cell lines and drugs, respectively, and these merged data are learned by multi-class machine learning.

On the other hand, in the present invention, a functional group of a drug is used in the multi-classification machine learning and prediction process, and the use of the drug functional group improves the efficiency of learning and analysis, compared to applying a polymer compound for the drug as it is.

Meanwhile, as described above, the training data for multi-classification learning of the present invention is extracted from the GDSC database. It provides comprehensive public information on the genomic profile of human cancer cell lines and drug susceptibility assays. The GDSC includes drug sensitivity analysis results of more than 1,000 cancer cell lines and 265 anticancer drugs. The full data set of these databases includes 686,312 mutation sites and 265 drugs in 1001 cell lines.

Meanwhile, in the present disease, these data are filtered and used according to the following criteria.

First, a mutation belonging to a gene included in Cancer Gene Census is used, and the mutation is determined from a list of 567 genes associated with cancer.

Second, for use in the present invention, only cancer types represented by at least 10 different cell lines are included.

Of the 31 cancer types comprising 1001 cell lines, 25 cancer types with a total of 787 cell lines are included in the data set.

On the other hand, specific cancer types may be excluded, for example, when a specific cancer type is expressed in a relatively small number of cell lines, these cancer types may be excluded from judgment.

Here, an embodiment of the present invention selected sequence variation information at 28,328 positions from 567 genes in the COSMIC cancer gene census.

The GDSC provides IC50 values from drug sensitivity assays for over 200,000 drug-cancer cell line pairs.

In this case, the IC50 is a criterion for determining the activity for drug reactivity, and is usually used as a criterion of 50%, but data set according to another criterion may be applied.

In addition, GDSC used the same set of 1001 genetically characterized cell lines, and 265 anticancer therapies were included in the test, including those under investigation in FDA approval.

Although these individual pairs of individuals were 9 pairs, since the IC50 values were different in all pairs, the 9 pairs could be regarded as 18 drugs to perform learning.

That is, in the example of the present invention, there are 244 drugs representing 235 small chemical substances as the final data set, and the final matrix of cell lines and drugs used in deep learning machine learning consists of a total of 152,594 instances.

(3) Implementation example of machine learning

Among the examples of multi-classification machine learning according to the present invention, the example of predicting the activity of 25 specific cancers, 1000 cancer cell lines, and 250 pharmaceuticals can be performed through the following process.

1) By analyzing/extracting all available data from the GDSC database derived from COSMIC data, a total of 200,000 cancer cells vs. We secure cell activity (potential as cancer treatment) data of about 250 pharmaceuticals.

2) Next, on a total of 200,000 clinical/experimental observation data, machine learning is performed using the multi-classification machine learning model as described above using LASSO machine learning and GBM machine learning.

3) And to verify the performance, the performance is evaluated by the 5-fold-cross-validation evaluation method for all the data in the first step.

At this time, in the embodiment of the present invention, accuracy of 0.9 or more of the Pearson correlation index was confirmed in all 25 cancer cell types.

Through this, in the present invention, the reliability of machine learning can be objectively confirmed.

Hereinafter, the accompanying drawings will be described in detail.

1 shows a system for implementing a multiple classification method of drug reactivity and gene expression level based on HLA haplotype. That is, the process of extracting HLA haplotypes by extracting multiple mutations from the integrated DB of Cell Lines (c1 to cN) composed of genotypes and the relationship between single or multiple mutations and gene expression and drug response data by the QTL method is shown. . In addition, the process of the system is explained in detail in terms of HLA haplotype and gene expression level.

2 is a correlation matrix between the cell line drug response value (IC50) and the HLA haplotype. Finally, the correlation between cell lines, drug reactivity information, and HLA haplotype is calculated by LASSO regression. Among 265 drugs, 72 drugs highly correlated with HLA haplotype were classified and selected, and 30 HLA haplotypes were already related to drug response based on quantitative trait location (QTL) information in each cell line genome. show that you have

3 shows that the drug reactivity information (IC50) and the gene expression level have a correlation with each other, and also shows that in many cases, the gene expression level has a tendency to correlate with a large number of genes. This is a drug reactivity calculation schema using the HLA haplotype, where most known active drugs are divided into a drug-sensitive group and a drug-resistant group, where sensitivity and resistance are closely related to the HLA haplotype. has been

Meanwhile, Table 1 is an example of a highly correlated HLA haplotype, and the correlation between gene expression level and HLA haplotype is collected by setting p-value < 10^3 and p-adj < 10^2 or less.

Allele A_case B_case A_ctrl B_ctrl F_case F_ctrl Freq P_Logit OR L95 U95 P_adj A*01 42 342 147 1013 0.1094 0.1267 0.1224 0.4384 0.8826 0.6436 1.2104 0.6576 A*02 99 285 292 868 0.2578 0.2517 0.2532 0.8316 1.026 0.81 1.2995 0.8756 A*03 59 325 121 1039 0.1536 0.1043 0.1166 0.0217 1.4158 1.0522 1.9049 0.065 A*11 32 352 72 1088 0.0833 0.0621 0.0674 0.1973 1.2909 0.8756 1.903 0.3947 A*24 46 338 210 950 0.1198 0.181 0.1658 0.015 0.6841 0.5038 0.929 0.065 A*26 20 364 63 1097 0.0521 0.0543 0.0538 0.8756 0.9621 0.5935 1.5597 0.8756 B*07 44 340 137 1023 0.1146 0.1181 0.1172 0.8689 0.9733 0.7061 1.3417 0.9937 B*08 21 363 97 1063 0.0547 0.0836 0.0764 0.1059 0.6971 0.4502 1.0795 0.7412 B*15 27 357 66 1094 0.0703 0.0569 0.0602 0.323 1.2744 0.7879 2.0613 0.7538 B*35 28 356 108 1052 0.0729 0.0931 0.0881 0.2761 0.8033 0.5416 1.1914 0.7538 B*40 28 356 80 1080 0.0729 0.069 0.0699 0.812 1.0502 0.7016 1.5719 0.9937 B*44 40 344 121 1039 0.1042 0.1043 0.1043 0.9937 0.9985 0.689 1.4471 0.9937 B*51 27 357 73 1087 0.0703 0.0629 0.0648 0.631 1.1114 0.7224 1.7098 0.9937 C*01 32 352 102 1058 0.0833 0.0879 0.0868 0.8027 0.9534 0.6556 1.3865 0.9771 C*03 60 324 146 1014 0.1562 0.1259 0.1334 0.1776 1.22 0.9137 1.6289 0.5826 C*04 37 347 116 1044 0.0964 0.1 0.0991 0.8505 0.9665 0.678 1.3777 0.9771 C*05 27 357 81 1079 0.0703 0.0698 0.0699 0.9771 1.0059 0.6731 1.5033 0.9771 C*06 29 355 95 1065 0.0755 0.0819 0.0803 0.7113 0.9267 0.6193 1.3867 0.9771 C*07 92 292 279 881 0.2396 0.2405 0.2403 0.972 0.9954 0.7708 1.2856 0.9771 C*12 25 359 103 1057 0.0651 0.0888 0.0829 0.1911 0.7593 0.5024 1.1474 0.5826 C*14 25 359 56 1104 0.0651 0.0483 0.0525 0.2185 1.342 0.84 2.144 0.5826 DPB1*01 19 365 66 1094 0.0495 0.0569 0.0551 0.6484 0.9038 0.5851 1.3961 0.8442 DPB1*02 57 327 178 982 0.1484 0.1534 0.1522 0.8442 0.9734 0.7441 1.2734 0.8442 DPB1*04 176 208 496 664 0.4583 0.4276 0.4352 0.4012 1.0824 0.8997 1.3023 0.8442 DPB1*05 39 345 138 1022 0.1016 0.119 0.1146 0.4544 0.8902 0.6563 1.2074 0.8442 DQB1*02 49 335 167 993 0.1276 0.144 0.1399 0.479 0.8962 0.6618 1.2138 0.8039 DQB1*03 133 251 388 772 0.3464 0.3345 0.3374 0.7022 1.0434 0.8392 1.2973 0.8778 DQB1*04 22 362 80 1080 0.0573 0.069 0.0661 0.4823 0.8565 0.5561 1.3193 0.8039 DQB1*05 69 315 187 973 0.1797 0.1612 0.1658 0.4421 1.1145 0.8453 1.4695 0.8039 DQB1*06 111 273 338 822 0.2891 0.2914 0.2908 0.9382 0.991 0.7893 1.2444 0.9382 DRB1*01 36 348 113 1047 0.0938 0.0974 0.0965 0.8478 0.9654 0.6739 1.383 0.9033 DRB1*03 26 358 105 1055 0.0677 0.0905 0.0848 0.2232 0.7814 0.5255 1.1621 0.4533 DRB1*04 69 315 205 955 0.1797 0.1767 0.1775 0.9033 1.0174 0.7707 1.343 0.9033 DRB1*07 37 347 94 1066 0.0964 0.081 0.0848 0.4149 1.1557 0.8161 1.6365 0.6635 DRB1*08 17 367 84 1076 0.0443 0.0724 0.0654 0.1035 0.6789 0.4259 1.0821 0.4533 DRB1*11 29 355 102 1058 0.0755 0.0879 0.0848 0.4976 0.8745 0.5935 1.2885 0.6635 DRB1*13 49 335 113 1047 0.1276 0.0974 0.1049 0.1239 1.2948 0.9317 1.7993 0.4533 DRB1*15 44 340 159 1001 0.1146 0.1371 0.1315 0.2266 0.7917 0.5422 1.1561 0.4533 DRB1*15 16 80 80 606 0.1667 0.1166 0.1228 0.1395 1.6172 0.8548 3.0595 0.558

And Table 2 shows a tendency similar to the result of FIG. 4 as an example of the HLA haplotype and the expression level of other genes with high correlation.

Allele A_case B_case A_ ctrl B_ ctrl F_case F_ ctrl Freq P_ Logit OR L95 U95 P_ adj A*01 147 1013 51 447 0.1267 0.1024 0.1194 0.2289 1.1969 0.8931 1.604 0.4578 A*02 292 868 132 366 0.2517 0.2651 0.2557 0.6093 0.9457 0.7634 1.1715 0.6533 A*03 121 1039 73 425 0.1043 0.1466 0.117 0.0309 0.7381 0.5602 0.9725 0.0927 A*11 72 1088 37 461 0.0621 0.0743 0.0657 0.4098 0.8567 0.5931 1.2375 0.6147 A*24 210 950 56 442 0.181 0.1124 0.1604 0.0024 1.5497 1.1673 2.0573 0.0147 A*26 63 1097 30 468 0.0543 0.0602 0.0561 0.6533 0.9086 0.5982 1.3802 0.6533 B*07 137 1023 51 447 0.1181 0.1024 0.1134 0.4103 1.1366 0.838 1.5415 0.718 B*08 97 1063 29 469 0.0836 0.0582 0.076 0.1163 1.3608 0.9264 1.9988 0.4071 B*15 66 1094 41 457 0.0569 0.0823 0.0645 0.0513 0.6635 0.4392 1.0024 0.3593 B*35 108 1052 39 459 0.0931 0.0783 0.0887 0.3793 1.1691 0.8253 1.6563 0.718 B*40 80 1080 38 460 0.069 0.0763 0.0712 0.6264 0.9131 0.6331 1.3169 0.7309 B*44 121 1039 57 441 0.1043 0.1145 0.1074 0.5445 0.9027 0.6483 1.2569 0.7309 B*51 73 1087 33 465 0.0629 0.0663 0.0639 0.8112 0.9525 0.639 1.4198 0.8112 C*01 102 1058 36 462 0.0879 0.0723 0.0832 0.3407 1.1908 0.8314 1.7056 0.6814 C*03 146 1014 82 416 0.1259 0.1647 0.1375 0.0629 0.7811 0.602 1.0134 0.5035 C*04 116 1044 53 445 0.1 0.1064 0.1019 0.7163 0.9434 0.6888 1.292 0.7211 C*05 81 1079 38 460 0.0698 0.0763 0.0718 0.6774 0.9274 0.65 1.323 0.7211 C*06 95 1065 38 460 0.0819 0.0763 0.0802 0.7211 1.0687 0.7421 1.5388 0.7211 C*07 279 881 115 383 0.2405 0.2309 0.2376 0.6894 1.0493 0.8288 1.3284 0.7211 C*12 103 1057 33 465 0.0888 0.0663 0.082 0.172 1.2925 0.8944 1.8676 0.6814 C*14 56 1104 30 468 0.0483 0.0602 0.0519 0.331 0.8034 0.5166 1.2492 0.6814 DPB1*01 66 1094 27 471 0.0569 0.0542 0.0561 0.8575 1.0355 0.7074 1.5158 0.8575 DPB1*02 178 982 68 430 0.1534 0.1365 0.1484 0.4623 1.0985 0.855 1.4114 0.6164 DPB1*04 496 664 228 270 0.4276 0.4578 0.4367 0.3643 0.925 0.7817 1.0947 0.6164 DPB1*05 138 1022 51 447 0.119 0.1024 0.114 0.4338 1.116 0.8478 1.469 0.6164 DQB1*02 167 993 74 424 0.144 0.1486 0.1454 0.8269 0.971 0.7455 1.2645 0.9187 DQB1*03 388 772 168 330 0.3345 0.3373 0.3353 0.9187 0.9896 0.81 1.209 0.9187 DQB1*04 80 1080 30 468 0.069 0.0602 0.0663 0.5603 1.1221 0.7615 1.6533 0.9187 DQB1*05 187 973 89 409 0.1612 0.1787 0.1665 0.4235 0.9018 0.7002 1.1615 0.9187 DQB1*06 338 822 137 361 0.2914 0.2751 0.2865 0.5465 1.0668 0.8646 1.3163 0.9187 DRB1*01 113 1047 43 455 0.0974 0.0863 0.0941 0.5199 1.1166 0.798 1.5625 0.6103 DRB1*03 105 1055 39 459 0.0905 0.0783 0.0869 0.4707 1.1349 0.8047 1.6006 0.6103 DRB1*04 205 955 89 409 0.1767 0.1787 0.1773 0.9279 0.9883 0.7657 1.2755 0.9279 DRB1*07 94 1066 50 448 0.081 0.1004 0.0869 0.2587 0.833 0.6066 1.1439 0.6103 DRB1*08 84 1076 22 476 0.0724 0.0442 0.0639 0.0677 1.4828 0.9718 2.2625 0.5413 DRB1*11 102 1058 36 462 0.0879 0.0723 0.0832 0.3468 1.1857 0.8315 1.6906 0.6103 DRB1*13 113 1047 57 441 0.0974 0.1145 0.1025 0.3338 0.8579 0.6287 1.1706 0.6103 DRB1*15 159 1001 63 435 0.1371 0.1265 0.1339 0.534 1.1122 0.7955 1.555 0.6103 DRB1*15 16 80 80 606 0.1667 0.1166 0.1228 0.1395 1.6172 0.8548 3.0595 0.558

As shown in FIG. 6 , the HLA haplotypes predicted and tested for all through the multi-classification machine learning model according to the present invention show a strong correlation. This confirms that the predicted HLA haplotype and the prediction of drug reactivity are accurate for all five drugs.

On the other hand, in FIGS. 5 to 9, the cancer drugs paclitaxel (FIG. 5), BI-2536 (FIG. 6), AT-7519 (FIG. 7), QL-XII-47 (FIG. 8) and GW843682X (FIG. 9) drugs Test results for deriving HLA haplotypes and genetic biomarkers that can predict reactivity are shown, respectively. That is, a more significant result can be obtained when the expression level information and the haplotype information are used as a combination of the two information rather than individually as drug response markers, and it shows that it is possible to distinguish drug response groups through these biomarkers. .

Finally, to summarize the present invention, the present invention predicts HLA haplotypes by utilizing a number of mutations in the nucleotide sequence collected from the genome to be tested in order to calculate a drug reactivity prediction model using HLA haplotype information. The step of calculating the solution of the LASSO regression model between the HLA haplotype and drug reactivity information, and the step of calculating the correlation with the gene expression level using this solution virtual It relates to a method for predicting drug reactivity for in silico clinical trials.

The rights of the present invention are not limited to the embodiments described above, but are defined by the claims, and those skilled in the art can make various modifications and adaptations within the scope of the claims. it is self-evident

The present invention relates to a system and method for predicting high-precision immunotherapy indications and sensitivity by converting the reactivity of the immuno-oncology drug into the genetic characteristics of the target genome and the human leukocyte antigen haplotype (hereinafter referred to as the HLA haplotype). According to the invention, drug reactivity prediction using HLA haplotype information is predicted using multiple mutations in the nucleotide sequence collected from the genome to be tested, and LASSO regression between HLA haplotype and drug reactivity information By calculating the results of the model, there is an effect of predicting drug reactivity for an in silico clinical trial that effectively filters out highly sensitive cancer drugs among drugs using GBM regression to correlate with gene expression level.

Claims (23)

In the immuno-oncology indication and response prediction method for receiving analysis information and calculating a result of predicting the reactivity of a drug with respect to the genome included in the analysis information,
Receiving the analysis target information including the genome and drug data to be analyzed;
calculating the analysis target HLA haplotype information of the genome included in the analysis information, and calculating the analysis target composition information of the drug included in the analysis information;
Calculating the reactivity prediction result of the drug with respect to the genome included in the analysis target information in the reaction correlation between the analysis target genome information and the analysis target configuration information by a reactivity prediction algorithm, and outputting the calculated result Performed:
The human leukocyte antigen haplotype-based multi-classification artificial intelligence model, characterized in that the analysis target configuration information and HLA haplotype information are information corresponding to the configuration information and HLA haplotype information applied to each machine learning. Indications and methods of predicting response.
The method of claim 1,
The analysis target configuration information and HLA haplotype information are,
A method for predicting the indication and response of an immuno-oncology drug using a multi-classification artificial intelligence model based on a human leukocyte antigen haplotype, characterized in that it is information on the functional groups constituting the drug and the HLA haplotype information included in the genome.
The method of claim 1,
The analysis target configuration information and HLA haplotype information are,
A method for predicting the indication and response of an immuno-oncology drug using a multi-classification artificial intelligence model based on a human leukocyte antigen haplotype, characterized in that the functional group information constituting the drug and the characteristic information on the mutations included in the genome.
The method of claim 1,
The reactivity prediction algorithm is
Human leukocyte antigen haplotype-based multi-classification artificial, characterized in that it is an algorithm learned by multi-classification machine learning for the reactivity correlation of constituent information constituting drugs to HLA haplotype information contained in the genome from the collected learning information. A method for predicting the indications and responses of immuno-oncology drugs using an intelligent model.
5. The method of claim 4,
The machine learning is
LASSO regression process for classifying and aligning learning target data according to multiple cell lines and multiple drugs, and learning all genomic rheology based on convolution based on reactivity (IC50);
GBM regression process for classifying and arranging learning target data according to multiple cell lines and multiple drugs, and learning based on convolution based on expression level information of all genes; and
LASSO classification that learns based on convolution with the HLA haplotype information calculated from the LASSO regression process and the GBM regression process and parameters for each drug reactivity and gene expression level merged ), a method for predicting immunotherapy indications and response using a multi-classification artificial intelligence model based on human leukocyte antigen haplotype, characterized in that it is carried out including the process.
5. The method of claim 4,
The machine learning is
Immunity using human leukocyte antigen haplotype-based multi-classification artificial intelligence model, characterized in that it is performed by LASSO machine learning, which performs machine learning using HLA haplotype information and drug reactivity information (IC50), which are the final analysis target, as learning factors. Anticancer drug indication and response prediction method.
7. The method of claim 6,
The LASSO machine learning is
(A) the learning data generating unit collecting learning information indicating the degree of response to each drug for each cell line genome;
(B) generating genetic information for HLA haplotypes included in the learning information;
(C) generating configuration information constituting the drug included in the learning information;
(D) generating pre-processed learning data indicating a degree of response to a group of constituent information constituting a drug to the HLA haplotype information group included in the learning information;
(E) deriving a response correlation of individual configuration information to individual HLA haplotype information through machine learning on the pre-processed learning data; and
(F) generating a drug reactivity prediction algorithm composed of constituent information on a genome including HLA haplotype information through the response correlation of the constituent information to the learned HLA haplotype information; A method for predicting immunotherapy indications and response using a multi-classification artificial intelligence model based on human leukocyte antigen haplotype, characterized in that it is performed, including:
5. The method of claim 4,
The machine learning is
Indications and responses to immuno-oncology drugs using human leukocyte antigen haplotype-based multi-classification artificial intelligence model, characterized in that it is performed by GBM machine learning that performs machine learning using HLA haplotype information and expression levels of specific genes as learning factors Prediction method.
9. The method of claim 8,
The GBM machine learning is,
(a) collecting learning information indicating the degree of response to each drug for each cell line genome by the learning data generator;
(b) generating genetic information for HLA haplotypes included in the learning information;
(c) generating HLA haplotype information pre-processing data indicating a drug's reactivity to the HLA haplotype information group included in each genome;
(d) deriving a drug response correlation to each genetic information through multi-classification machine learning for the genetic information learning information;
(e) generating configuration information constituting the drug included in the learning information;
(f) generates composition information pre-processing learning data indicating the reactivity of the composition information group constituting the drug for each genome, and the response correlation of each composition information for each genome through the GBM algorithm for the composition information learning data deriving;
(g) Deriving the response correlation of individual gene expression level information to individual haplotype information through the correlation between the drug response to each gene expression level information and each HLA haplotype information for each genome A method for predicting immunotherapy indications and response using a multi-classification artificial intelligence model based on human leukocyte antigen haplotype, characterized in that it is performed, including the steps of:
10. The method according to claim 7 or 9,
The learning information is
Indications and response prediction method for immuno-oncology drugs using a multi-classification artificial intelligence model based on human leukocyte antigen haplotype, characterized in that it is test result data for the reactivity of various drugs to various cell lines.
10. The method according to claim 7 or 9,
The HLA haplotype information is,
A method for predicting immunotherapy indications and response using a human leukocyte antigen haplotype-based multi-classification artificial intelligence model, characterized in that it is mutation information or characteristic information on mutations.
10. The method according to claim 7 or 9,
The configuration information is
A method for predicting immunotherapy indications and response using a human leukocyte antigen haplotype-based multi-classification artificial intelligence model, characterized in that it is information on the functional groups constituting the drug.
8. The method of claim 7,
The standard of drug reactivity in step (E) is,
A method for predicting immunotherapy indications and response using a multi-classification artificial intelligence model based on a human leukocyte antigen haplotype, characterized in that it is judged based on the receptive inhibition index IC50.
In the immuno-oncology indication and response prediction system for receiving analysis information and calculating a result of predicting the reactivity of a drug with respect to the genome included in the analysis information,
an input unit for receiving analysis target information including genomic and drug data to be analyzed;
a comparison data generation unit for generating comparison data comprising gene expression data and drug data included in the analysis target information in a form corresponding to HLA haplotype information and configuration information used for machine learning, respectively; and
Human leukocyte antigen haplotype-based, characterized in that it comprises a prediction result generating unit that calculates the response prediction result of the drug to the genome included in the analysis target information by the reactivity prediction algorithm derived by the reactivity prediction algorithm component Immuno-oncology indication and response prediction system using multi-classification artificial intelligence model.
15. The method of claim 14,
The HLA haplotype information is,
Immunotherapy indications and response prediction using a multi-classification artificial intelligence model based on human leukocyte antigen haplotype, characterized in that it is calculated by a program that allows multiple polymorphic markers and SNPs using likelihood-based inference system.
15. The method of claim 14,
The comparison data generation unit,
When the machine learning is performed with HLA haplotype information and functional group information, HLA haplotype data of the genome included in the analysis target information is calculated, and functional group information included in the analysis target information is calculated. Immuno-oncology indication and response prediction system using multi-classification artificial intelligence model based on human leukocyte antigen haplotype.
15. The method of claim 14,
The comparison data generation unit,
When the machine learning is performed with gene expression level information and functional group information, characteristic information about the gene expression level of the genome included in the analysis target information is calculated, and functional group information included in the analysis target information is calculated. Immuno-oncology indication and response prediction system using multi-classification artificial intelligence model based on human leukocyte antigen haplotype.
18. The method according to any one of claims 14 to 17,
The reactivity prediction algorithm is
Human leukocyte antigen haplotype-based multi-classification artificial, characterized in that it is an algorithm learned by multi-classification machine learning for the reactivity correlation of constituent information constituting drugs to HLA haplotype information contained in the genome from the collected learning information. Immuno-oncology indication and response prediction system using intelligent model.
19. The method of claim 18,
The machine learning is
LASSO regression process for classifying and aligning learning target data according to multiple cell lines and multiple drugs, and learning all genomic rheology based on convolution based on reactivity (IC50);
GBM regression process for classifying and arranging learning target data according to multiple cell lines and multiple drugs, and learning based on convolution based on expression level information of all genes; and
LASSO classification that learns based on convolution with the HLA haplotype information calculated from the LASSO regression process and the GBM regression process and parameters for each drug reactivity and gene expression level merged ), an immuno-oncology indication and response prediction system using a multi-classification artificial intelligence model based on human leukocyte antigen haplotype, characterized in that it is carried out including the process.
20. The method of claim 19,
The machine learning is
Immunity using human leukocyte antigen haplotype-based multi-classification artificial intelligence model, characterized in that it is performed by LASSO machine learning, which performs machine learning using HLA haplotype information and drug reactivity information (IC50), the final analysis target, as learning factors Anticancer drug indication and response prediction system.
21. The method of claim 20,
The LASSO machine learning is
(A) the learning data generating unit collecting learning information indicating the degree of response to each drug for each cell line genome;
(B) generating genetic information for HLA haplotypes included in the learning information;
(C) generating configuration information constituting the drug included in the learning information;
(D) generating pre-processed learning data indicating a degree of response to a group of constituent information constituting a drug to the HLA haplotype information group included in the learning information;
(E) deriving a response correlation of individual configuration information to individual HLA haplotype information through machine learning on the preprocessed learning data; and
(F) generating a drug reactivity prediction algorithm composed of constituent information on a genome including HLA haplotype information through the response correlation of the constituent information to the learned HLA haplotype information; Immuno-oncology indication and response prediction system using a multi-classification artificial intelligence model based on human leukocyte antigen haplotype, characterized in that it is performed including
20. The method of claim 19,
The machine learning is
Indications and responses to immuno-oncology drugs using human leukocyte antigen haplotype-based multi-classification artificial intelligence model, characterized in that it is performed by GBM machine learning that performs machine learning using HLA haplotype information and expression levels of specific genes as learning factors prediction system.
23. The method of claim 22,
The GBM machine learning is,
(a) collecting learning information indicating the degree of response to each drug for each cell line genome by the learning data generator;
(b) generating genetic information for HLA haplotypes included in the learning information;
(c) generating HLA haplotype information pre-processing data indicating a drug's reactivity to the HLA haplotype information group included in each genome;
(d) deriving a drug response correlation to each genetic information through multi-classification machine learning for the genetic information learning information;
(e) generating configuration information constituting the drug included in the learning information;
(f) generates compositional information preprocessing learning data indicating the reactivity of the compositional information group constituting the drug for each genome, and the response correlation of each compositional information for each genome through the GBM algorithm for the compositional information learning data deriving;
(g) Deriving the response correlation of individual gene expression level information to individual haplotype information through the correlation between the drug response to each gene expression level information and each HLA haplotype information for each genome Immuno-oncology indication and response prediction system using a human leukocyte antigen haplotype-based multi-classification artificial intelligence model, characterized in that it is carried out including;

Images