Study On The Prediction Model For Drug Induced-liver Toxicity Based On The Whole Cell Phenotype

Drug induced liver injury（DILI）is caused by prescription drugs,over-the-counter preparations and herbal medications。It is clinically manifested as a variety of acute and chronic liver diseases,severe cases can directly lead to acute liver failure or even death,is a common adverse drug reaction.DILI not only seriously endangers the health of drug users,but also the main reason for the failure of drug development,the use restriction and withdrawal of the marketed drugs.Therefore,DILI has always been an important issue closely followed by pharmaceutical companies,regulatory agencies,and clinicians.In the past ten years,with the rapid development of modern life sciences,computers,bioinformatics and other technologies,various in vitro hepatotoxicity model had been established by scientists,including biological test models based on kinds of human-derived liver cells and in silico models that integrate various information of drugs,and life omics technologies are also applied.However,the lack of understanding of the DILI mechanism and the technical limitations of in vitro physiology-like models have made existing prediction models unable to accurately predict DILI,especially IDILI drugs,and there is no one suitable for the screening of candidate drugs for hepatotoxicity in the early drug development,and widely accepted and licensed by the industry and regulatory agencies.HCA technology based on 2D cell can obtain multi-dimensional information of the toxic mechanism and effect of tested drug in a single experiment,and is one of the important technologies that are worth choosing to develop a new method for hepatotoxicity prediction in vitro.Up to now,the hepatotoxicity prediction method based on HCA is the only one that has been verified by batch drugs and is suitable for the screening of candidate drugs for hepatotoxicity in early drug development.However,the existing in vitro hepatotoxicity prediction methods based on HCA have some problems,such as different selection and classification criteria for drugs,the combination of test parameters has not been systematically screened,and only a few simple toxicity-related parameters have been tested,which seriously affected the credibility of the model.In view of the fact that HCA-based cellular phenotype profiling of drugs can obtain a large number of measurement values of relevant indicators,and can effectively enrich the connotation of hepatotoxicity in vitro testing.It can be used to identify the toxicity of the test compound through mechanism/pattern recognition,and is not subject to the lack of understanding of DILI toxicology mechanism and the limitation of the in vitro model.Therefore,this study intends to build a new prediction system based on 2D cells using HCA,which is suitable for the early stages of drug discovery,efficient DILI drug identification and prediction,and provides support technology for reducing the cost and improving the efficiency of new drug development.In this study,223 market drugs with various indications and different degrees of clinical DILI were selected as the test drugs followed the public literatures and the hepatotoxicity database at home and abroad.Construction of drugs information database by collecting the SMILES expression of the chemical structure,the fat solubility coefficient of fat-soluble log P value,daily dose,the C_max value,the classification in DILIrank database,the"possibility score"in livertox database,and the types of liver injury of drugs.And then,according to the DILIrank and livertox database classification criteria,the test drugs are divided into four categories:serious DILI（s DILI）drugs,moderate DILI（m DILI）drugs,ambiguous DILI（a DILI）drugs and non-DILI（n DILI）drugs.Besides,in order to meet the requirements of modeling and validating,the test drugs were randomly divided into test sets（120 drugs,including 50 s DILI,50 m DILI and 20 n DILI drugs）and validation sets（103 drugs,31s DILI,31 m DILI,30 a DILI and 11 n DILI drugs）.Then,HCA technology was used to determine the effects of drugs on the cytotoxic phenotype profile,including 1)Cell adaptive stress response parameters,including cellular autophagy-related protein LC3B,and Nrf2,ATF6,Hif1αand NF-κB（the transcription factors of oxidation stress pathway,endoplasmic reticulum stress pathway,hypoxia stress pathway and inflammatory stress pathway,respectively）;2)Common cytotoxic mechanism and toxicity endpoint effect parameters,including DNA damage related proteins p H₂AX and oxidative damage related proteins Mn SOD,nuclear membrane permeability（NMP）,GSH,mitochondrial membrane potential（MMP）and inhibition rate;3)Subcellular organelles and cytoskeletal parameters,including nuclear morphology and size,lysosome,cytoskeletal proteins F-actin andα-tubulin.Construct a database of drug cell phenotype profiles by calculating the EC₅₀,LEC and effect value of the test drug on the parameters,and correct it with the maximum blood concentration C_max of the human body.On this basis,two strategies are used to build prediction system:1)Based on the cell phenotype parameter of test set of drugs,the phenotype parameters with high false positives are excluded first,and use machine learning t-class classification and identification software to screen and identify the best test combination of parameters.Select the combination with the highest correlation of prediction accuracy,stability and ROC analysis and form a software system,and use the phenotype data obtained by the verification set of drugs to verify.Finally,a test combination consisting of 7 or 13 phenotypic parameters was obtained,the sensitivity of the method was not less than 80%,and the specificity was 94%.For the7-parameters model,the test can be completed through 3 sets of toxicity phenotype parameter combinations（LC3B,nuclear+Mn SOD+GSH+α-tubulin,nuclear+MMP）,and the 13-parameters model can be completed through 3 sets of toxicity phenotype parameter combinations（LC3B+p H₂AX,nuclear+Mn SOD+GSH+α-tubulin,nuclear+MMP）and 3 groups of the stress pathways（Nrf2,ATF6 and Hif1α）.When this method was combined with RO2,the sensitivity of IDILI drug increased significantly to 84%,but the specificity of the method decreased from 94%to 81%.2)Through clustering comparison of the phenotypic spectrum of tested drugs to predict DILI.Hierarchical clustering is performed based on the maximum effect value of all43 phenotypes of cytotoxic dependence and independence,combined with drug chemical structure,target,indications,damage type and other information to determine the hepatotoxic phenotype characteristic spectrum.The results show that the drug of cell phenotypes profile is closely related to the chemical structure,targets and indications,of which the drug target in most correlated,and then is chemical structure and indications.It suggests that the hepatotoxicity of an unknow drugs can be predicted and evaluated through the location of the similarity of the phenotype spectrum in the case of the exposure concentration in the body is unknow.Finally,the hierarchical clustering of cytotoxicity-independent cell phenotype profiles combined with the analysis of the corresponding cell phenotype profiles associated with toxicity.The results showed that decrease of F-actin＿24h and MMP＿24h,increase of LC3B＿16h,lyso＿16h and p H₂AX＿16h,activation of ATF6＿5h,Nrf2＿6h,NF-κB＿24h and Hif1α＿3h may be an early independent event of hepatocyte damage,and there are early events with the parameters lyso＿16h and LC3B＿16h,lyso＿16h and MMP＿24h increase together,and types without obvious early cell phenotype changes.Furthermore,increase of lyso＿16h,activation of ATF6＿5h,Nrf2＿6h and Hif1α＿3h are often accompanied by the significant changes of late multicellular phenotype,decrease of MMP＿24h,activation of Hif1α＿3h and NF-κB＿24h are strongly correlated with s DILI.The results suggest that there may be some common pathways for DILI,and early toxicity parameters play an important role in the prediction of drug toxicity.In a word,this study used HCA technology to detect the effect of 223 test drugs on the cell phenotype parameters,combined with C_max to determine the characteristics of DILI,and constructed a database of test drug cell phenotype profiles.On this basis,an innovative in vitro hepatotoxicity prediction system based on cell phenotypic pattern analysis was established by using machine learning,biological information analysis technologies and through modeling and verification,phenotype cluster analysis processes.On the one hand,a subject prediction method with 7 or 13phenotypic parameters combined with drug C_max was constructed;on the other hand,hepatotoxicity prediction was carried out in combination with cluster analysis of the constructed drug phenotype profile database.Finally,the comparative analysis based on the unaccompanied and concomitant cellular phenotype profiles provides new clues for a comprehensive understanding of the mechanism of DILI.