Targeted gene-based prediction model for success rate in drug clinical trials

Abstract

To reflect effects of drugs on cell populations and genes in humans to predict the success or failure of clinical trials, which can help to solve problems caused by excessive clinical trials, such as overdose of drugs.SOLUTION: The first aspect of the present application is a device for predicting a success rate of clinical trials of new drugs using machine learning, comprising: an information acquisition unit that acquires information about a target gene for a candidate new drug; and a prediction unit that inputs the information about the target gene into a previously trained clinical trial success rate prediction model, and predicts the success rate of the clinical trial of the candidate new drug.SELECTED DRAWING: Figure 17

Description

[Technical field]

[0001] The present application relates to a target gene-based drug clinical trial success rate prediction model. [Background technology]

[0002] Many new drugs pass preclinical trials but fail clinical trials due to safety concerns, which greatly increases the cost of drug development, and clinical trial failures can reduce patients' quality of life and cause additional health problems.

[0003] Selecting safe drug targets in the early stages of new drug development can increase the success rate of clinical trials. Previously, genetic perturbation effects such as CRISPR-Cas9-based genome editing have been used to evaluate the safety of drug targets under the assumption that targeting genes sensitive to genetic perturbation (GPA) is likely to cause serious side effects. This assumption is based on the concept that genes capable of causing genetic perturbation effects in some cases are likely to play a wide range of biological roles and be involved in important protein-protein interactions.

[0004] However, in reality, the perturbation effects between cell groups and populations are not consistent, so clinical trials based on the above assumptions may cause clinical failure due to safety issues of new drug candidates. Even if a particular drug has a small perturbation effect on a particular cell group, the drug may cause a large perturbation effect on the population, which may cause serious safety issues and lead to the failure of the clinical trial.

[0005] For example, the appetite suppressant sibutramine did not significantly affect cell viability. However, the drug was withdrawn from the market in 2010 due to serious side effects in the human population, including neurotoxicity and cardiotoxicity. Indeed, sibutramine inhibits the dopamine transporter (SLC6A3), and perturbations to SLC6A3 may not affect cell viability. However, SLC6A3 has been implicated in serious human diseases such as neuropsychiatric and cardiac disorders.

[0006] Meanwhile, Korean Patent No. 10-2274363 discloses a method for developing a pharmaceutical process based on artificial intelligence (machine learning), and U.S. Patent No. 1123041 discloses a method for developing a new drug or treatment method customized for a patient based on a machine learning algorithm, but most of the prior literature deals with inventions related to the pre- and post-clinical trial processes.

[0007] As a result, there is an emerging need for precise prediction technology that can identify candidate new drugs and prevent serious safety issues from arising during the clinical trial process, thereby saving industry enormous costs and time and preventing health risks for patients. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] Korean Patent No. 10-2274363 [Patent Document 2] US Patent No. 1123041 Summary of the Invention [Problem to be solved by the invention]

[0009] The present application aims to predict the success rate of clinical trials by considering the difference in perturbation effects of drugs between the non-clinical stage targeting a cell group and the clinical stage targeting a population. In this regard, it should be construed that all problems within the scope of understanding of a person skilled in the art are included. [Means for solving the problem]

[0010] In order to solve the above-mentioned problems, a first aspect of the present application provides an apparatus for predicting the success rate of clinical trials of new drugs using machine learning, the apparatus including: an information acquisition unit that acquires information about a target gene for a candidate new drug; and a prediction unit that inputs information about the target gene into a pre-trained clinical trial success rate prediction model to predict the success rate of the clinical trial of the candidate new drug.

[0011] A second aspect of the present application provides a method for predicting a success rate of a clinical trial of a new drug using machine learning, the method comprising: acquiring information about a target gene for a candidate new drug; and inputting the information about the target gene into a pre-trained clinical trial success rate prediction model to predict the success rate of the clinical trial of the candidate new drug.

[0012] The above-mentioned means are merely examples, and should be construed as including all solutions within the scope of what a person skilled in the art can understand. Effect of the Invention

[0013] This application can predict the success or failure of clinical trials by reflecting the effects of drugs on cell groups and genes in humans, thereby solving problems caused by excessive clinical trials, such as overuse of drugs.

[0014] The above-mentioned effects are merely examples, and other effects that are understandable to a person skilled in the art may exist, and all of these may be included. [Brief description of the drawings]

[0015] [Figure 1a] The database contains information on 1,064 unapproved drugs and 1,400 approved drugs for 7,695 targets, categorized by ATC code. [Figure 1b] 1 is a histogram analyzing the relationship between drugs and drug targets. [Figure 2a] FIG. 1 is a diagram illustrating the hypotheses used in this application. [Figure 2b] According to one embodiment of the present application, we show the mapping of CGE and OGE to approved and unapproved drugs. Significance was measured by a two-tailed study. [Figure 2c] 1 illustrates a general prediction process according to one embodiment of the present application. [Figure 2d] According to one embodiment of the present application, the AUPRCs of clinical trial success rates predicted using OGE+CGE, OGE, and CGE for unapproved drugs (drugs with maximum clinical stage reached in phase 1, phase 2, and phase 3, respectively) and approved drugs are shown. The statistical significance of the difference between AUPRCs was measured by Mann-Whitney U test. [Figure 2e] 1 shows the position and number of target genes in the correlation between CGE and OGE according to one embodiment of the present application. In the figure, the y-axis is expressed in log2. [Diagram 3] 1 shows the results of predicting the probability of clinical trial success for a specific new drug candidate according to one embodiment of the present application. The results are categorized according to the clinical stage of each candidate drug. The statistical significance between the probabilities of clinical trial success was measured using the Mann-Whitney U test. [Figure 4a] According to one embodiment of the present application, the precision and recall of predicting the probability of clinical trial success of a specific drug candidate is shown (Monte Carlo test, 1,000 times). The red dots connected by the red line represent the prediction performance (precision and recall). The statistical significance of each precision and recall was measured by paired sample t-test. [Figure 4b] According to one embodiment of the present application, the precision and recall of predicting the probability of clinical trial success of a specific drug candidate is shown (Monte Carlo test, 1,000 times). The red dots connected by the red line represent the prediction performance (precision and recall). The statistical significance of each precision and recall was measured by paired sample t-test. [Figure 5a] The AUPRCs for unapproved drugs (Phase 1) and approved drugs when OGE+CGE, OGE, and CGE were used for prediction are shown (Monte Carlo test, 1,000 times). Statistical significance between each AUPRC was measured by the Mann-Whitney U test. [Figure 5b] According to one embodiment of the present application, the precision and recall for unapproved drugs (clinical phase 1) and approved drugs are shown (Monte Carlo test, 1,000 times). The red dots connected by the red line represent the prediction performance (precision and recall). The statistical significance of each precision and recall was measured by paired sample t-test. [Figure 6a] 1 illustrates the process of classifying risk and safety targets according to one embodiment of the present application, where red represents targets that show biased effects against unapproved drugs, blue represents targets that show biased effects against approved drugs, and grey dots represent targets that show no biased effects against either unapproved or approved drugs. [Figure 6b] According to one embodiment of the present application, each classified target (drug) was compared through hypergeometric testing with drugs that failed due to toxicity in actual clinical trials and drugs with “Death” side effects (MoleculeNet-ClinTox). [Figure 6c] According to one embodiment of the present application, each classified target (drug) was compared through hypergeometric testing to drugs that failed due to toxicity in actual clinical trials and drugs with "Death" side effects (KM Gayvert et al.). [Figure 6d] According to one embodiment of the present application, each classified target (drug) was compared through hypergeometric testing with drugs that failed due to toxicity in actual clinical trials and drugs with "Death" side effects (ADReCS). [Figure 6e] According to one embodiment of the present application, each classified target (drug) was compared through hypergeometric testing with drugs that failed due to toxicity in actual clinical trials and with drugs that have the side effect of "Death" (DrugCentral). [Figure 7a]In accordance with one embodiment of the present application, when the criteria for statistical significance of drug approval bias were relaxed and the number of dangerous and safe targets was increased, it was confirmed whether the safe drugs classified by the method of the present application significantly overlapped with actual toxic drugs. [Figure 7b] According to one embodiment of the present application, each classified target (drug) was compared with drugs that failed due to toxicity in actual clinical trials and drugs with "Death" side effects (MoleculeNet-ClinTox) through a hypergeometric test that relaxed the statistical significance criteria for drug approval bias. [Figure 7c] According to one embodiment of the present application, each classified target (drug) was compared with drugs that failed due to toxicity in actual clinical trials and with "Death" side effects (KM Gayvert et al.) through a hypergeometric test that relaxed the statistical significance criteria for drug approval bias. [Figure 7d] According to one embodiment of the present application, each classified target (drug) was compared with drugs that failed due to toxicity in actual clinical trials and drugs with "Death" side effects (ADReCS) through a hypergeometric test with relaxed criteria for statistical significance of drug approval bias. [Figure 7e] According to one embodiment of the present application, each classified target (drug) was compared with drugs that failed due to toxicity in actual clinical trials and with "Death" side effects (DrugCentral) through a hypergeometric test that relaxed the statistical significance criteria for drug approval bias. [Figure 8a] According to one embodiment of the present application, the number of related drugs per drug subject and the degree of drug approval bias were investigated (Spearman rho=-0.38). The correlation coefficient was measured using Spearman's correlation coefficient (Spearman rank correlation), the x-axis was quantified using log10, and the statistical significance of the difference between the x-axis values ​​was measured using the Mann-Whitney U test. [Figure 8b]According to one embodiment of the present application, the number of related drugs per drug subject and the degree of drug approval bias were investigated (Spearman rho=-0.43). The correlation coefficient was measured using Spearman's correlation coefficient (Spearman rank correlation), the x-axis was quantified using log10, and the statistical significance of the difference between the x-axis values ​​was measured using the Mann-Whitney U test. [Figure 9a] 1 shows a classification of genes based on the differential effects of genetic perturbations between cell populations and humans, according to one embodiment of the present application. [Figure 9b] In Figure 9a, the classified genes are divided into N2E, N2N, E2E, and E2N, respectively, and the enrichment degree for each risk or safety target is shown. [Figure 10a] According to one embodiment of the present application, the distribution of CGE for dangerous and safe targets is shown. The statistical significance of the differences was measured by the Mann-Whitney U test. [Figure 10b] 1 shows the distribution of OGE for dangerous and safe targets according to one embodiment of the present application. The statistical significance of the differences was measured by the Mann-Whitney U test. [Figure 11a] According to one embodiment of the present application, Behan, Fiona M. et al., demonstrates the degree of enrichment for hazardous and safe targets using OGE. [Figure 11b] According to one embodiment of the present application, DepMap,OGE is used to show the degree of enrichment for dangerous and safe targets. [Figure 11c] According to one embodiment of the present application, hPSCs, OGE are used to demonstrate the degree of enrichment for risky and safe targets. [Figure 11d] According to one embodiment of the present application, CGE, pLI(ExAC) is used to show the degree of enrichment for risky and safe targets. [Figure 11e] According to one embodiment of the present application, Behan, Fiona M. et al., demonstrate the degree of enrichment for risk and safety targets using pLI (ExAC). [Figure 11f]According to one embodiment of the present application, DepMap, pLI (ExAC) is used to show the degree of enrichment for risky and safe targets. [Figure 11g] According to one embodiment of the present application, hPSC, pLI(ExAC) is used to demonstrate the degree of enrichment for risky and safe targets. [Figure 12a] According to one embodiment of the present application, a check was performed to see whether results identical or similar to those in Figures 11a to 11g were observed when the statistical significance threshold for drug approval bias was relaxed. [Figure 12b] According to one embodiment of the present application, a check was performed to see whether results identical or similar to those in Figures 11a to 11g were observed when the statistical significance threshold for drug approval bias was relaxed. [Figure 12c] According to one embodiment of the present application, a check was performed to see whether results identical or similar to those in Figures 11a to 11g were observed when the statistical significance threshold for drug approval bias was relaxed. [Figure 12d] According to one embodiment of the present application, a check was performed to see whether results identical or similar to those in Figures 11a to 11g were observed when the statistical significance threshold for drug approval bias was relaxed. [Figure 13a] According to one embodiment of the present application, the effectiveness of the present invention was confirmed by comparing CGE and OGE to determine the extent to which N2E, N2N, E2E, and E2N genes are targeted among clinically failed drugs. [Figure 13b] Figure 13a shows the target gene classification of clinically failed drugs according to their side effects (cardiotoxicity, vascular toxicity, psychiatric toxicity, and neurotoxicity). [Figure 14a] According to one embodiment of the present application, the N2E gene is shown to be associated with certain genetic pathways (or how it is involved in the body). [Figure 14b]According to one embodiment of the present application, this shows which genetic pathways (or how it is involved in the body) the N2N gene is associated with. [Figure 14c] According to one embodiment of the present application, it is shown which genetic pathways (or how they are involved in the body) the E2E gene is associated with. [Figure 14d] According to one embodiment of the present application, it is shown which genetic pathways (or how they are involved in certain bodily functions) the E2N gene is associated with. [Figure 15a] 1 illustrates components of drug targets and chemical information used to predict the success rate of clinical trials, according to one embodiment of the present application. [Figure 15b] According to one embodiment of the present application, the performance of clinical trial success rate prediction models using some or all of OGE, CGE, network, and expression information is compared. [Figure 15c] According to one embodiment of the present application, the results of predicting the success rate of drug clinical trials using drug target information (OGE+CGE+Network+Expression), chemical information (Chemical), or all of them are compared. [Figure 15d] 1 shows the results of a search by ATC code for drugs whose clinical success rates were predicted using only drug target information and chemical information according to one embodiment of the present application. [Figure 16] According to one embodiment of the present application, the effect of the present invention is improved by orthogonality between drug target and chemical information, which is information used to predict the success rate. The vectors shown in the figure represent the direction and degree of drug approval predictability. [Figure 17] FIG. 1 is a diagram of an apparatus according to an embodiment of the present application. [Figure 18] 1 is a flowchart according to an embodiment of the present application. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Hereinafter, the present application will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present application. However, the present application may be embodied in various different forms and is not limited to the embodiments described herein. In the drawings, in order to clearly explain the present application, parts that are not related to the description are omitted, and similar parts are designated by similar reference numerals throughout the specification.

[0017] Throughout this specification, when a member is said to be "on" another member, this includes not only when the member is in contact with the other member, but also when there is another member between the two members.

[0018] Throughout the specification of this application, when a part is said to "comprise" a certain element, this means that it may further include other elements, but not to the exclusion of other elements, unless specifically stated to the contrary.

[0019] The terms of degree, such as "about," "substantially," and the like, used throughout the specification of this application are used to mean a numerical value or close to a numerical value when manufacturing and material tolerances inherent in the recited meaning are given, and are used to prevent an unconscionable infringer from taking unfair advantage of the disclosure in which precise or absolute numerical values ​​are recited to aid in the understanding of this application. The terms of degree, such as "a step of" or "a step of" used throughout the specification of this application, do not mean "a step for".

[0020] Throughout the specification of this application, the term "combination(s) of" contained in a Markush form expression means a mixture or combination of one or more selected from the group of elements set forth in the Markush form expression, and is meant to include one or more selected from the group of elements set forth above.

[0021] Throughout the specification of this application, the statement "A and / or B" means "A or B, or A and B."

[0022] Throughout the specification of this application, the term "individual" includes humans to whom drugs are administered for safety evaluation in the clinical stage after the preclinical stage of a candidate new drug. It should be interpreted to include not only humans, but also all applicable organisms such as mice, rabbits, dogs, cows, horses, sheep, etc. that can be used for drug experiments.

[0023] Throughout this specification, the term "clinical test failure" refers to a particular drug passing preclinical trials but failing to pass clinical trials in a population, including, but not limited to, all of the above and similar situations.

[0024] Throughout the specification of this application, the term "perturbation" or "perturbation effect" refers to a change at the gene level caused by a substance or stimulus. This can occur mainly due to environmental stimuli, the action of a drug, gene knockdown, etc. For example, it includes the loss-of-function of a target gene as a result of the action of a drug.

[0025] Throughout this specification, "tolerant" or "loss-of-function tolerant" means that the survival of a population, such as a cell population or a human, is not significantly affected after a perturbation occurs. Conversely, "intolerant" or "loss-of-function intolerant" means that the effect of the perturbation cannot be tolerated by the population, such as a cell population or a human, and survival is significantly affected.

[0026] Throughout the specification of this application, the term "drug-likeness rule" refers to a criterion used to preliminarily select and predict substances similar to drugs from various compounds, and to develop compounds suitable for new drugs from such prediction results. Its components include molecular weight, the number of hydrogen bond donors or acceptors, polar surface area, etc. The drug-likeness rule proposed by Lipinski (Rule of 5; Ro5), Veber, and Ghose has been considered a useful guideline for preferred drugs with low toxicity.

[0027] Throughout this specification, "machine learning" refers to an artificial intelligence application in which a computer program uses an algorithm to look for patterns in data provided to it. It primarily refers to the field of training computers to learn from data and improve through experience. The machine learning algorithm used in this application is only one example, and should be construed to include all machine learning methods or types that can be used for the purposes of this invention. For example, machine learning methods may include (1) supervised learning, (2) unsupervised learning, (3) reinforcement learning, (4) semi-supervised learning, and more specifically, may include all of Naive Bayes Classification, Logistic Regression, decision tree, Random forest, boosting (XGBoost / ensemble boosting / AdaBoost / Gradient Boost / LightGBM / CatBoost, etc.), Perceptron, Support Vector Machine, Quadratic classifiers, clustering (K-means clustering, Bayesian network clustering, etc.), but are not limited to these.

[0028] Throughout the specification of this application, the term "non-clinical trials" refers to trials to confirm side effects, toxicity, and efficacy of a newly developed new drug candidate substance when it is used on animals or cell groups before being used on a population. This includes, but is not necessarily limited to, pharmacokinetic studies that study how a drug is absorbed, distributed, and excreted in the body, and pharmacological studies, and should be interpreted as including all processes required for testing.

[0029] Throughout the specification of this application, the term "clinical trial" means, but is not necessarily limited to, a test or study conducted on human subjects to verify the pharmacokinetic / pharmacodynamic / pharmacodynamic / clinical effects of an investigational drug and to investigate any abnormal reactions in order to prove the safety and efficacy of the investigational drug, and should be interpreted as including all steps necessary for the trial.

[0030] Throughout the specification of this application, "success rate" refers to the probability of success in a clinical trial, calculated as the area under the AUROC of the model's performance prediction results. The success rate set in this application is set to 0.5, but users may set a different success rate within a range that ensures accuracy according to their needs.

[0031] In the entire specification of this application, "cellular gene essentiality (CGE)" and "organismal gene essentiality (OGE)" refer to the calculation of the degree of perturbation effect on a gene in a cell group or a population, respectively. This is the core component of the technical idea of ​​this application, and this application aims to predict the success rate of clinical trials based on the difference in the effect of genetic perturbation on a target by a drug.

[0032] Throughout this specification, "protein-protein interaction networks (PPIN)" refers to the mathematical representation of physical contacts between proteins in cells. Such contacts are highly defined, occur in restricted binding regions between proteins, and have specific biological meanings (specific functions).

[0033] Hereinafter, embodiments and examples of the present application will be described in detail with reference to the accompanying drawings, but the present application is not limited to the embodiments and examples and drawings.

[0034] A first aspect of the present application provides an apparatus for predicting the success rate of clinical trials of new drugs using machine learning, the apparatus including: an information acquisition unit that acquires information about a target gene for a candidate new drug; and a prediction unit that inputs information about the target gene into a pre-trained clinical trial success rate prediction model to predict the success rate of the clinical trial of the candidate new drug.

[0035] Here, the information on the target gene may be CGE (cellular gene essentiality) or OGE (organismal gene essentiality).

[0036] A second aspect of the present application provides a method for predicting a success rate of a clinical trial of a new drug using machine learning, the method comprising: acquiring information about a target gene for a candidate new drug; and inputting the information about the target gene into a pre-trained clinical trial success rate prediction model to predict the success rate of the clinical trial of the candidate new drug.

[0037] The commonalities between the different aspects all apply together.

[0038] Example 1. Experimental information on predicting success rate using target gene information (1) Obtaining information on perturbation effects on humans and cell populations The usefulness of target gene information in predicting the success rate of clinical trials was confirmed.

[0039] CGE information was inferred from the fitness score in the ProjectSCORE database (Behan, FM et al. Prioritization of cancer therapeutic targets using CRISPR-Cas9 screens. Nature 568, 511-516 (2019).). Fitness scores were measured by whole-genome CRISPR-Cas9 screening for 324 cell lines. Genes with an average CGE of 0 or more were considered essential genes, and the opposite was considered non-essential genes.

[0040] To confirm various metrics of CGE, we used the group of fitness genes by Behan, FM et al., the probability of dependency by DepMap (20Q4 v2), and the fitness score of hPSCs (human pluripotent stem cells). The pan-cancer core fitness genes and core fitness genes defined by Behan, FM et al. were considered as essential genes, and the rest were considered as non-essential genes.

[0041] DepMap also screened 789 cell lines through CRISPR-Cas9 and compared with ProjectSCORE to investigate the dependency probability of genes on cell lines as an independent indicator. For each gene, the average probability was calculated across 789 cell lines. Genes with an average probability of 0.5 or higher were considered essential genes, and genes with an average probability of less than 0.5 were considered non-essential genes.

[0042] To perform a general survey for cell populations, not just cancer cell lines, fitness scores were obtained for human pluripotent stem cells (hPSCs) in three different media conditions (mouse embryonic fibroblasts feeder cells, laminin, and vitronectin) from Mair, B. et al. The fitness scores of genes for each hPSC cell were expressed as false discovery rates (FDRs). The three fitness scores for each gene were combined through Fisher's combined probability test using python scipy.stats.combine_p values. Genes with combined FDRs < 0.01 were considered essential genes, and genes with combined FDRs > 0.01 were considered non-essential genes.

[0043] OGE was derived from the loss-of-function observed / expected upper bound fraction (LOEUF) of the gnomAD database (Karczewski, KJ et al. The mutational constraint spectrum quantified from variation in 141,456 humans. Nature 581, 434-443 (2020).). LOEUF was estimated by modeling the mutation rate from up to 140,000 population base sequence data, and was calculated by comparing the expected number of mutations with the amount of observed mutations of a gene. For example, a low LOEUF, indicating a significant reduction in the number of observed mutations compared to the expected mutations of a gene, is considered to be an intolerant gene perturbation effect due to negative selection.

[0044] OGE was directionally aligned with CGE by subtracting the LOEUF of each gene from 1.996 (the maximum LOEUF for all genes). According to the criteria for intolerant gene perturbation effects defined by Karczewski, KJ et al. (LOEUF<0.35), genes with OGE ≥ 1.646 were designated as essential genes in human populations, whereas genes with OGE < 1.646 were considered as non-essential genes in human populations.

[0045] As another index of OGE, the probability of being loss-of-function intolerance (pLI) from the ExAC database (Lek, M. et al. Analysis of protein-coding genetic variation in 60,706 humans. Nature 536, 285-291 (2016).) was used. Genes with a pLI of 0.9 or more were considered essential genes, and genes with a pLI of less than 0.9 were considered non-essential genes.

[0046] (2) Mapping of CGE and OGE with drugs From the ChEMBL (v30; Mendez, D. et al. ChEMBL: towards direct deposition of bioassay data. Nucleic Acids Res. 47, D930-D940 (2019).) database, information on 2,464 drugs whose clinical trials had been completed at any one of the phases 1 to 4 was used. Drugs in the phase 4 stage were considered as approved drugs, and drugs in phases 1, 2, and 3 were considered as unapproved drugs. Drugs for cancer treatment were excluded due to their inherent cytotoxicity. The molecular types were restricted to small molecules. These exclusion criteria were defined by Nguyen et al. (Nguyen, PA, Born, DA, Deaton, AM, Nioi, P. & Ward, LD Phenotypes associated with genes encoding drug targets are predictive of clinical trial side effects. Nat. Commun. 10, 1579 (2019).) and Duffy, Aine (Duffy, A. et al. Tissue-specific genetic features inform prediction of drug side effects in clinical trials. Sci. Adv. 6, eabb6242 (2020).). Drug-target interaction information was obtained from the STITCH5 (Duffy, A. et al. Tissue-specific genetic features inform prediction of drug side effects in clinical trials. Sci. Adv. 6, eabb6242 (2020).) database (binding interaction score >= 700). Finally, we collected information on 1,064 unapproved drugs and 1,400 approved drugs for 7,695 targets (see Figures 1a and 1b). Drugs were classified by Anatomical Therapeutic Chemical Codes (ATC codes).

[0047] For CGE and OGE, the average of the CGE and OGE for each drug target was calculated and mapped to each drug. All drugs used in one embodiment of this application were filtered to include the CGE and OGE of more than 90% of the drug targets.

[0048] (3) The progress of machine learning A random forest classifier was used as the clinical trial success rate prediction model, which was implemented using Scikit-learn (v0.24.2) in Python. The classifier created 1,000 trees as a base setting. CGE and OGE of drugs for each clinical stage were used to train the clinical trial success rate prediction model. To effectively measure the prediction performance, Monte Carlo cross-validation (1,000 times) was performed and the dataset was randomly divided into a training (90%) and a test (10%) set. The success probability of clinical trials for new drug candidates was calculated as the average probability calculated by the clinical trial success rate prediction model for the test set through 1,000 cross-validations. Drugs with an average approval probability of 0.5 or more may be considered approved drugs, and drugs with an average approval probability of less than 0.5 may be considered unapproved drugs. To analyze the predictive performance of the clinical trial success rate prediction model, we measured the area under the precision-recall curve (AUPRC) of the predicted results of the test set.

[0049] Example 2. Explanation of a success rate prediction experiment using target gene information The applicant of the present application has focused on the fact that the cause of failure of a drug in a clinical trial is the difference in response to each gene between a nonclinical model and a clinical population. It has been hypothesized that drugs that fail in a clinical trial tend to be associated with genes that show a tolerable perturbation effect on a cell group, but actually show an intolerable perturbation effect on a population such as a human (see FIG. 5a). Here, the gene perturbation effect on a cell group can be measured using CRISPR-Cas9 knockout. However, the gene perturbation effect on a population cannot be measured by applying such a knockout. Instead, the gene perturbation effect on a population can be estimated by modeling the loss-of-function mutation rates for genes using exome and genome base sequence analysis data of a large population.

[0050] Genes that exhibit resistant perturbation effects are conserved in populations for a long time, so genes with a low loss-of-function mutation rate may be considered to be resistant to the perturbation effect. Conversely, genes with many loss-of-function mutations may be considered to be resistant to the perturbation effect. This application uses this assumption to evaluate the safety of drugs and treatments for populations.

[0051] (1) CGE / OGE To investigate the relationship between the difference in the perturbation effect of a drug in a cell group and a human population and the approval of a clinical trial, the present application calculated the cellular gene essentiality (CGE) and organic gene essentiality (OGE) for a total of 17,662 genes as an example (see Example 1-(1)). Genes with high CGE and OGE scores show an intolerable perturbation effect on both cell groups and humans, meaning that they are essential genes. A low score means the opposite.

[0052] (2) Cell viability test and comparison The 7,695 targets of 2,464 drugs (1,064 unapproved drugs and 1,400 approved drugs) in the STITCH5 database were sorted by drug approval status through the ChEMBL database, and the CGE and OGE of the drug targets were matched to the drugs (see Example 1-(2)). Because the unapproved and approved drugs had passed cell viability tests in non-clinical trials, the median CGE for the unapproved and approved drugs was significantly lower than the random distribution (see Figure 5b; median CGE for unapproved drugs = -2.08, Two-tailed test P = 5.3 × 10 -58 ;P=6.3×10 -81 However, the median OGE for unapproved drugs was significantly higher than the random distribution (median OGE for unapproved drugs = 1.23, P = 5.3 × 10 -36 ), the median OGE for approved drugs was significantly lower than the random distribution (median OGE for approved drugs = 1.11, P = 3.1 × 10 -33 ), indicating that unapproved drugs that pass cell viability tests generally have intolerable effects in humans.

[0053] Given that clinical trial approval of a drug is associated with discrepancies in the effects of drug target perturbation between cell groups and human populations, the applicants of the present application anticipate that such discrepancies may be an important feature in predicting drug approval or non-approval in clinical trials.

[0054] (3) Establishment of machine learning predictive models To confirm this, we developed a clinical trial success rate prediction model as a machine learning classifier that uses OGE and CGE of drug targets to predict the drug approval probability (see FIG. 5b). In addition, the applicant of the present application developed two stand-alone classifiers that use only OGE or CGE to evaluate the prediction model of the present invention.

[0055] The area under the precision-recall curve (AUPRC) of our clinical trial success rate prediction model using the discrepancy between CGE and OGE of drug targets (OGE+CGE) was 0.70 ± 0.002 (see Fig. 5d; mean ± 95% confidence interval), which was significantly higher than the AUPRC of two stand-alone classifiers using only OGE or CGE (0.62 ± 0.002, 0.60 ± 0.002, Mann-Whitney U test, P = 4.4 × 10 -221 , P = 2.0 × 10 -275 ).

[0056] In addition, the accuracy of the classifier of the present application was confirmed by evaluating the correlation between the approval probability predicted in the present application and the clinical stage of the drug. In general, the approval probability in the early clinical stage is lower than that in the later clinical stage. As a result of the actual prediction, drugs predicted to have a low approval probability belonged to the early clinical stage, while drugs predicted to have a high approval probability belonged to the later clinical stage (see FIG. 3; see Example 1-(3)).

[0057] This means that the predictive performance of the clinical trial success rate prediction model according to the present application is quite accurate and can be used to predict the success rate of actual clinical trials.

[0058] In addition, the use of the clinical trial success rate prediction model according to the present application can reduce the amount of drugs wasted in clinical trials. Even drugs that were determined to be safe in cell viability tests often failed to pass clinical trials due to intolerable perturbation effects on humans. The clinical trial success rate prediction model according to the present application significantly reduced erroneous predictions, such as drugs that were initially predicted to be approved due to genetic perturbation effects on cell groups but were ultimately not approved. The recall (recall in this specification means the proportion of drugs that were correctly predicted to be unapproved among unapproved drugs) when predicting clinical success rates using OGE and CGE was significantly lower than the recall when predicting clinical success rates using only CGE (see FIG. 4a; Paired sample t-test; P=2.4×10 -47).

[0059] The proportion of drugs that were ultimately approved but were incorrectly predicted to be unapproved also decreased. The proportion of drugs that were predicted to be unapproved among unapproved drugs was also predicted with improved accuracy when OGE and CGE were used to predict clinical success rates compared to when CGE alone was used to predict clinical success rates (see Figures 4b and 4c; P = 2.5 × 10 -284 ).

[0060] Predicting whether a drug will be approved in clinical phase I is necessary to confirm the predictive ability of the clinical trial success rate prediction model of the present application (using OGE and CGE) in evaluating drug safety. This is because the largest proportion of drugs fail due to safety issues in phase I. In addition, the first safety study for healthy clinical trial volunteers is conducted in phase I. Even when the classifier of the present application was trained only on drugs that were not approved in clinical phase I or were finally approved, the prediction accuracy was improved when OGE and CGE were used compared to when only CGE was used, and the number of drugs incorrectly predicted to be approved by CGE was significantly reduced. This suggests that the classifier of the present application makes more sophisticated and superior predictions of drug safety by reflecting OGE, which is a population intolerance perturbation effect (see Figures 5a and 5b).

[0061] By using the mismatch of drug target perturbation effect between cell groups and populations, the predictive performance of the clinical trial success rate prediction model of the present application can be greatly improved. Through the present invention, many genes that have a tolerable perturbation effect in cell groups but an intolerable perturbation effect in humans have been discovered. By using such genes, it is possible to more accurately predict whether a drug will be clinically approved based on the association between the mismatch between cell groups / humans and drug approval. There are 2,512 genes, accounting for about 14% of the 17,662 genes (see Figure 2e). In addition, OGE showed a weak correlation with CGE (Spearman's rho = 0.08), suggesting that using both OGE and CGE for prediction is key information for predicting whether a drug will be approved in clinical trials.

[0062] Example 3. Verification using various indicators (1) N2E, N2N, E2E, E2N To understand the association between gene sets (hazardous / safe targets; N2E, N2N, E2E, and E2N) and drugs with safety issues, we investigated drugs that failed due to toxicity, side effects, and safety issues in clinical trials.

[0063] We obtained information on drugs that failed in clinical trials due to toxicity from two datasets. The ClinTox dataset from MolecularNet (Wu, Z. et al. MoleculeNet: a benchmark for molecular machine learning. Chem. Sci. 9, 513-530 (2018).) contains drugs that failed in clinical trials due to toxicity. Another toxic drug dataset was obtained from Gayvert et al. (Gayvert, KM, Madhukar, NS & Elemento, OA Data-Driven Approach to Predicting Successes and Failures of Clinical Trials. Cell Chem. Biol. 23, 1294-1301 (2016).). We obtained information on failed or approved drugs from two datasets.

[0064] Information on drugs with adverse drug reactions was obtained from the ADReCS (v3.1; Cai, M.-C. et al. ADReCS: an ontology database for aiding standardization and hierarchical classification of adverse drug reaction terms. Nucleic Acids Res. 43, D907-D913 (2015).) and Drug Central (v2021; Avram, S. et al. DrugCentral 2021 supports drug discovery and repositioning. Nucleic Acids Res. 49, D1160-D1169 (2021).) databases. In the Drug Central database, drug adverse drug reactions were provided with probabilities and probability thresholds for important drug-adverse drug reactions relationships. Important drug-adverse drug reactions relationships were filtered at twice the probability threshold. Gender specificity of adverse drug reactions was not considered. Drugs with the side effect of death were defined as the Medical Dictionary For Regulatory Activities (MedDRA) term "Death" (MedDRRA term: "Death", MedDRA ID: 10011906).

[0065] Drug data that failed due to safety issues were obtained from drug safety data curation in the ChEMBL database (Hunter, FMI et al. Drug Safety Data Curation and Modeling in ChEMBL: Boxed Warnings and Withdrawn Drugs. Chem. Res. Toxicol. 34, 385-395 (2021)) and Onakpoya, IJ et al. (Onakpoya, IJ, Heneghan, CJ & Aronson, JK Post-marketing withdrawal of 462 medicinal products because of adverse drug reactions: a systematic review of the world literature. BMC Med. 14, 10 (2016)). The reasons for failure were classified according to the type of organ in which the side effects occurred, such as cardiotoxicity and neurotoxicity. ChEMBL provided a failure grade for side effects. The reasons for failure in Onakpoya., IJ et al. were classified manually.

[0066] All drugs with safety issues were mapped to STITCH5 identifiers using a mapping file for STITCH5 and external database identifiers in PubChemPy (v1.0.4, https: / / pypi.org / project / PubChemPy).

[0067] (2) Risky and safe targets Dangerous and safe targets were classified to investigate the relationship between the risk of drug targets and the mismatch between cell groups and people. Each target was classified by the method according to the present application. Specifically, drug targets associated with unapproved drugs were considered as dangerous targets, and drug targets associated with approved drugs were considered as safe targets. Drug approval bias was measured by the OR (odds ratios) and p-values ​​of Fisher's exact test for each drug target. Only when the p-value was less than 0.05, targets were classified as dangerous targets when the OR was greater than 1, and as safe targets when the OR was less than 1, respectively. As a result of the measurement, information on 141 dangerous targets and 313 safe targets was obtained (see FIG. 6a).

[0068] The accuracy of the method of the present application was determined by comparing each target classified by the method of the present application with drugs that failed due to toxicity in actual clinical trials (toxic drugs) and drugs with "Death" side effects. Dangerous drugs targeting dangerous targets were shown to significantly overlap with toxic drugs (see Figures 6b and 6c; hypergeometric test; P = 3.6 × 10 -10 ;P=1.2×10 -17 ), whereas safe drugs did not significantly overlap with toxic drugs (P = 1.0; P = 0.2). "Death" refers to serious side effects that result in death or serious injury. Dangerous drugs classified by our method significantly overlap with actual toxic drugs (see Fig. 6d, Fig. 6e; P = 1.7 × 10 -8 ;P=8.1×10 -6 ), and the safe drugs classified by the method of the present application did not significantly overlap with actual toxic drugs (P = 0.5; P = 0.1). Even when the statistical significance criteria for drug approval bias were relaxed and the number of risky and safe targets was increased, these results were consistently maintained (see Figures 7a to 7e). This indicates that the method of the present application can effectively classify whether a drug is risky or not and can effectively predict the success rate of clinical trials.

[0069] In addition, to confirm the effect of the present invention, the number of related drugs per drug target and the degree of drug approval bias were investigated. The more significant the bias in drug approval, the greater the number of related drugs (see Figure 8a; Spearman rho = -0.38, P = 1.6 × 10-269 ), and high levels of drug approval bias were observed (see Figure 8b; Spearman rho = -0.43, P < 1.0 × 10 -307 ), which indicates that the risky or safe targets classified in the present method are explicitly divided into unapproved and approved drugs, respectively. Thus, the risk of a drug target can be evaluated by the drug approval bias.

[0070] (3) Linkage of N2E, N2N, E2E and E2N with dangerous targets and safe targets To characterize the perturbed genes into risk and safety targets, we compared the gene perturbation effects between the cell population and human population, and classified 17,662 genes into four groups (see Figure 9a): 2,512 genes were identified as non-essential in the cell population but essential in human (N2E), 13,835 genes were identified as non-essential in both the cell population and human (N2N), 402 genes were identified as essential in both the cell population and human (E2E), and 913 genes were identified as essential in the cell population but non-essential in human (E2N).

[0071] It was confirmed that the risk of drug targets in clinical trials can be explained by the discrepancy in perturbation effects between cell populations and humans. In particular, the risk target was associated with the N2E gene, which showed a tolerable perturbation effect in cell populations but an intolerable perturbation effect in human populations (see Figure 9b; One-tailed test; P = 3.1 × 10 -7 ). The OGE of the dangerous targets was significantly higher than that of the safe targets, and this difference was significantly larger than the CGE (see Fig. 10a and Fig. 10b; Mann-Whitney U test; P = 2.0 × 10 -12 ;P=1.0×10 -1 ). Thus, risk targets can be characterized through their perturbation effects on the population.

[0072] In contrast, the safety target was associated with the N2N gene, which showed a resistant perturbation effect in all cell populations and humans (see Figure 9b; P = 3.0 × 10 -10). E2E and E2N genes showing intolerant perturbation effects on the cell population were not associated with any targets (see FIG. 9b). This suggests that drug targets with severe effects on cell viability were excluded at the preclinical stage.

[0073] By analyzing various indicators of CGE and OGE, we more firmly proved the conclusion that risky targets are associated with N2E genes, whereas safe targets are associated with N2N genes. We extended the analysis for OGE by pLI (probability of being loss-of-function intolerant) of the ExAC database. For CGE, we investigated the types of fitness genes in Behan, FM et al., the probability of dependency in DepMap, and the fitness core of human pluripotent stem cells (hPSCs). Through the same analysis with other gene essentiality datasets, we found that risky and safe targets were significantly enriched in N2E and N2N genes, respectively (see Figures 11a to 11g). Even when we relaxed the statistical significance threshold for drug approval bias to increase the number of risky or safe targets, significant results were consistently observed (see Figures 12a to 12d). As mentioned above, the results obtained in the present application were obtained even when unbiased CGE and OGE were used for the analysis by increasing the number of risk or safety targets, suggesting that the effect of the present invention is emphasized. Therefore, the risk of a drug target in a clinical trial can be evaluated by the discrepancy of the effect of gene perturbation between cell groups and populations, as in the present invention.

[0074] (4) Comparison of N2E with actual clinically failed drugs It was confirmed that the N2E gene is likely to be a dangerous target along with clinically failed drugs with serious side effects. Of the 200 failed drugs, the N2E gene was involved in 139 failed drugs (see Figure 13a; Hypergeometric test; P = 4.0 × 10 -8For example, thalidomide, used to treat morning sickness, was withdrawn due to the risk of birth defects because the N2E gene, the transcription factor SP1, was silenced by thalidomide, which disrupted angiogenesis in the embryo. In another example, sibutramine, used to treat obesity, was withdrawn due to heart attacks and strokes because it also silenced another N2E gene, the dopamine transporter (SLC6A3), which has been linked to a number of diseases, including neuropsychiatric disorders and heart failure.

[0075] It was confirmed that the N2E gene induces side effects related to the central nervous system (CNS) and cardiovascular system, which may lead to the withdrawal of drugs from the market. Such side effects were frequently observed in withdrawn drugs that target N2E. Drugs that were withdrawn due to cardiotoxicity, vascular toxicity, psychiatric, and neurotoxicity had a significant number of N2E genes (Figure 13b; P = 2.1 × 10 -5 ;P=3.2×10 -5 ;P=2.6×10 -3 ;P=2.4×10 -2 In contrast, drugs targeting the N2N, E2E and E2N genes were not among the recalled drugs with such side effects.

[0076] The N2E gene is associated with a number of neuronal and circulatory pathways (see FIG. 14a). For example, N2E plays an important role in axon formation. Perturbation effects on axon formation induce neuropathy. Angiogenesis is also associated with the N2E gene. Cardiotoxicity was induced by perturbation effects on angiogenesis.

[0077] Meanwhile, GPCR signaling pathways and metabolism-related pathways were associated with N2N genes (see Figure 14b). 70% of approved small molecule drugs target GPCR signaling pathways. This suggests that N2N may be a safer target than N2E. E2E and E2N were associated with pathways of cellular processes essential for cell survival, such as RNA processing, DNA replication, and transcription (see Figure 14c-d).

[0078] Example 4. Success rate prediction experiment using additional drug target information and chemical information The present application focuses on the chemical safety of drugs as another cause of failure of candidate new drugs that succeed in non-clinical trials but fail in clinical trials. The chemical safety of drugs refers to the properties of the chemical substances themselves that constitute the drug, including, for example, the fact that the chemical structure of the drug is not stably maintained until it reaches the target gene. The inventors of the present application confirmed that the success rate of clinical trials can be effectively predicted when chemical information is used in addition to the above-mentioned target gene information. In particular, the prediction performance of the method of the present application can be improved by combining additional drug target information and chemical information with the perturbation effect (see FIG. 15a).

[0079] Additional drug target information was obtained using protein interaction networks and tissue expression information. Research has shown that dangerous drugs target genes with high or medium centrality in protein interaction networks, and such genes tend to show high or widespread expression levels across tissues, such as the expression of housekeeping genes. Chemical information was obtained using information derived from drug-likeness rules, such as molecular weight, number of hydrogen bond donors or acceptors, and polar surface area. The applicant of the present application synthesized the above information to create models that use OGE, CGE, networks, expression information, and chemical information in whole or in part, and compared their prediction performance.

[0080] The AUPRC of the clinical trial success rate prediction model, an integrated classifier that uses OGE, CGE, network, and expression information (OGE+CGE+Expression+Network) to predict drug approval, was 0.01 when OGE and CGE were used (OGE+CGE; P = 3.5 × 10 -146 ) and network information (Network;P=9.4×10 -265 ) or expression information (Expression; P < 1.0 × 10 -307 ) results in a much improved prediction performance (see Figure 15b).

[0081] In addition, the AUPRC when chemical information was further included (OGE+CGE+Expression+Network+Chemical) was 1.01% higher than that when OGE, CGE, network, and expression information was used (OGE+CGE+Network+Expression, P = 2.3 × 10 -107 ) AUPRC or chemical information only (Chemical, P = 2.1 × 10 -231 ) was significantly improved over the AUPRC when using

[0082] The orthogonality between drug targets and chemical information is believed to improve drug approval prediction with additional information. In the prediction, chemical information and drug targets predicted different information from each other. The orthogonality between drug targets and chemical information was analyzed using principal component analysis (PCA). The vector of chemical information in PCA had a different direction from the vector of drug target information in PCA (see Figure 16). In addition, when drug targets and chemical information were used in combination, the drug approval or disapproval was more accurately predicted. Classifiers using only either drug target information (OGE+CGE+Network+Expression) or chemical information (Chemical) correctly predicted 862 and 842 approved drugs, respectively. However, when both drug target information and chemical information were used, the total number of correctly predicted approved drugs increased to 1,092 (see Figure 15c). This suggests that the accurate prediction of drug approval can be expanded by integrating the biological information of drug targets and the chemical information of drugs.

[0083] In addition, the prediction performance may vary depending on the organ or system in which the drug acts. Drugs that predicted the success rate of clinical trials using only drug target information or chemical information were investigated using anatomical therapeutic chemical code (ATC code). The cardiovascular system (ATC code: C) and hematology (ATC code: B) related to 250 drugs were predicted through OGE, CGE, network information and expression information, and the dermatology (ATC code: D) and antiparasitic drugs (ATC code: P) related to 230 drugs were predicted through chemical information (see Figure 15d). In other words, since the ATC codes of drug target information and chemical information are different from each other, it is possible to reflect a wider range of information in the prediction, and it is judged that the success or failure of clinical trials can be predicted more accurately. [Explanation of symbols]

[0084] 1710: Acquisition Department 1720: Learning Department 1730: Prediction Department

Claims

1. A device that uses machine learning to predict the success rate of clinical trials for new drugs. An information acquisition unit that acquires information on a target gene for a candidate new drug and information on the chemical safety of the candidate new drug; A prediction unit that inputs information about the target gene and information about the chemical safety of the candidate new drug into a pre-trained clinical trial success rate prediction model to predict the success rate of the clinical trial of the candidate new drug, The apparatus, wherein the information regarding the target gene is cellular gene essentiality (CGE) and organic gene essentiality (OGE).

2. 2. The apparatus according to claim 1, wherein the information acquiring unit further acquires one or more pieces of information selected from the group including protein interaction network information, tissue expression profile information, and drug-likeness rule component information.

3. 3. The device of claim 2, wherein the prediction unit further inputs any one or more pieces of information selected from the group including the protein interaction network information, the tissue expression information, and the similar drug rule component information into the clinical trial success rate prediction model to predict the success rate of the clinical trial of the candidate new drug.

4. a learning unit for learning the clinical trial success rate prediction model using learning data including information on the target gene for each of a plurality of drugs and labeling whether the clinical trial for each of the plurality of drugs is successful; The apparatus of claim 1 further comprising:

5. The apparatus of claim 4 , wherein the learning unit learns the clinical trial success rate prediction model through Monte Carlo cross-validation using the training data.

6. 2. The apparatus of claim 1, wherein the clinical trial success rate prediction model is trained to increase the success rate of a clinical trial when the candidate drug causes a resistance perturbing effect on a cell group or population.

7. A method for predicting a success rate of a clinical trial of a new drug using machine learning, comprising the steps of: (a) obtaining information on a target gene for a candidate drug and information on the chemical safety of the candidate drug; (b) inputting information about the target gene and information about the chemical safety of the candidate new drug into a pre-trained clinical trial success rate prediction model to predict the success rate of the clinical trial of the candidate new drug; The method, wherein the information about the target gene is CGE (cellular gene essentiality) and OGE (organismal gene essentiality).

8. 8. The method of claim 7, wherein the step (a) further comprises obtaining one or more pieces of information selected from the group consisting of protein interaction network information, tissue expression profile information, and drug-likeness rule component information.

9. 9. The method of claim 8, wherein the step (b) further inputs any one or more pieces of information selected from the group including the protein interaction network information, the tissue expression information, and the similar drug rule component information into the clinical trial success rate prediction model to predict the success rate of the clinical trial of the candidate new drug.

10. (c) learning the clinical trial success rate prediction model using learning data including information on the target gene for each of a plurality of drugs and labeled with the success or failure of the clinical trial for each of the plurality of drugs; The method of claim 7 further comprising:

11. The method of claim 10, wherein the step (c) comprises training the clinical trial success rate prediction model through Monte Carlo cross-validation using the training data.

12. 8. The method of claim 7, wherein the clinical trial success rate prediction model is trained to increase the success rate of a clinical trial when the candidate drug causes a resistance perturbing effect on a cell group and a population.

Images