Computational Characterization of DrugInduced Adverse Events with A

Computational Characterization of Drug-Induced Adverse Events with A Focus on Pulmonary Fibrosis Alex Jiang Cincinnati Children’s Hospital Medical Center Cornell University

Disclosure I have no relevant relationships with commercial interests to disclose. AMIA 2018 | amia. org 2

Overview • Spontaneous reporting systems such as FDA Adverse Event Reporting System (FAERS) are a great resource to mine for medication usage data. • FAERS data can be used to find previously unsuspected causal and therapeutic drugs for various conditions. • Our proposition: Combine FAERS data with transcriptional signatures from disease models and drug treatments to generate hypotheses for molecular basis of adverse events (AEs) and find potential treatments. • We test this hypothesis using drug-induced pulmonary fibrosis (DIPF) as a proof-of-concept study. AMIA 2018 | amia. org 3

Focus of Study: Why Pulmonary Fibrosis? • Scarring of lung tissue leading to death in ~5 years • Existing treatments are limited in effect • Many cases are idiopathic (around 5 million globally) • Several are drug-induced – pulmonary adverse events of chemotherapy • Pre-existing models to check against but: • No large scale studies of human patients • Little standardization across mouse studies • Little to no work on the effects of polypharmacy AMIA 2018 | amia. org 4

Methods FAERS Spontaneous reporting • • Mines and organizes data from FAERS Metrics: relative risk, safety signal AERSMine Aim 1 • AERSMine* (Adverse Events Reporting System Mine) DIPF Drug Suspects Daily. Med • Check against Pneumotox DIPF Risk Drug combinations with reduced DIPF risk Genomic analysis Curated gene expression datasets from sources like NCBI GEO, publications • Alternatives: Enrichr, Topp. Gene Aim 2 • Illumina BSCE (Base Space Correlation Engine) • Pneumotox Transcriptome BCSE data Mechanism of Action *Sarangdhar et al. , Nature Biotechnology, 2016 AMIA 2018 | amia. org 5

Bl M eom et ho yc Am tre in io xat G da e em ro ci ne ta b G ine e D fitin oc ib e Pa tax c el O lita xa xe C L yc e lipl l lo flu at ph no in os m La pha ide ns m o id N pra e itr og zol ly e R ce it ri D uxi n ox m o ab Vi rub nb ici n Vi las no tin r e Ta elbi Su mo ne lfa xi sa fen l R azin i Th tuxi e al ma i To dom b c N iliz ide itr u of m D ura ab ro n ne to d in C aro Ac im ne et eti Pr ylcy dine op st ox ein yp e he ne AERSMine Causal Suspect Selection 40 35 30 25 20 15 10 5 0 Relative Risk AMIA 2017 | amia. org Safety Signal 6

AERSMine Causal Results AMIA 2018 | amia. org 7

er id is p ne Relative Risk AMIA 2017 | amia. org C te r. In ne ne 1 a id i lo n ta be fil st at rli iv ud i m La O F M en a Va rd D on Pa e lip er id on D en e os um ab Li na gl ip tin Li ra gl ut C id an e ag lif lo zi n Ef av ire D nz id an Le os vo in e no rg es tre Fa l m pr id in e R as id o Zi pr AERSMine Therapeutic Candidate Selection 2 1 0 -1 -2 -3 -4 -5 -6 Safety Signal 8

AERSMine Therapeutic Results AMIA 2017 | amia. org 9

BSCE Results AMIA 2018 | amia. org 10

Conclusion Identified likely causal drugs with no label warnings: • Rituximab, doxorubicin, tamoxifen, and acetylcysteine • Dozens of other suspects Promising treatments (may lead to discovery of mechanisms): • Antipsychotic drugs: ziprasidone, risperidone, and paliperidone • Anti-diabetes drugs: linagliptin, liraglutide, canagliflozin • Patient group overlap, previous studies have suggested connections EHR (electronic health records) data mining – Computational validation Experimental Validation – both In vitro and In vivo (mouse models) AMIA 2018 | amia. org 11

Acknowledgements • Dr. Anil Jegga (Mentor) • Dr. Mayur Sarangdhar (assisted with AERSMine) • CCHMC SURF Program • Funding: NIH R 21 HL 135368 AMIA 2018 | amia. org 12

Thank you!