Discovery Of Toxic Factors By Learning Chemical Structural Features

Adverse drug reaction(ADR)is one of major clinical problems in modern healthcare.ADRs significantly increase morbidity and mortality risk,which not only impose a huge burden on the health care system,but also bring many challenges to new drug development.Up to date,many efforts have been made for drug toxicity prediction by construction of mathematical models from various aspects of drug physiochemical properties,drug-target interactions and so on.However,few approaches have built consensus associations between drug structural features and clinical ADRs.Therefore,we carried out systematic ADR association analyses in this study.We represented drug structures in both forms of scaffold and fingerprint using different tools like Scaffold Network Generator,PaDEL-Descriptor,Open Babel and so on.At the meanwhile,we obtained the drug-ADR relations from the Adverse Drug Reaction Classification System(ADReCS)v2.0,which includes 1,313 marketed drugs and 4,670 ADR Preferred Terms(PT).Of 4,837 scaffolds for 1,219 drugs in the scaffold network,more than one-fifth of the drugs enriched in 22 core scaffolds,especially the benzene rings.Association analysis showed that 191 scaffolds were significantly associated with at least one ADR(q<0.01).ADR similarity comparison revealed that common scaffolds only contributed to only a small portion of common ADRs.In further association analysis by PubChem Fingerprint,we found 12,095 significant fingerprint-ADR pairs(q<0.01).Of these fingerprint-ADR associations,408 fingerprints served as risky factors(OR>1)for most ADRs and 174 were safe fingerprints(OR<1).In addition to association analysis,we used information theory to identify the top 50 toxicophores for each ADR.Furthermore,we constructed the models of agranulocytosis and drug abuse using machine learning algorithms of support vector machine(SVM)and random forest.The random forest algorithm exhibited satisfied performance that the accuracy in both the ten-fold cross validation and the external evaluation exceeded 60%.Finally,we analyzed the mechanism of adverse drug reactions from the perspective of biological activation.In summary,we systematically explored associations between drug structural features and ADRs in this study.Our findings may provide useful clues in mechanistic understanding of ADR from drug itself.It will be of value in building a reliable and robust mathematical model for broad spectrum ADR assessment.