Biomedical Discovery through Data Mining and Data Science November - PowerPoint PPT Presentation

Implicit correction of age differences in exposed vs non-exposed zanamivir memantine atomoxetine rivastigmine actinomycin D galantamine ethosuximide donepezil 6-thioguanine bicalutamide retinoic acid flutamide methylphenidate verteporfin thiotepa acenocoumarol PGE2 Original darifenacin Corrected N-butyldeoxynojirimycin amiodarone -40 -20 0 20 40 (Average Age of Cases) - (Average Age of Controls)

Bias, corrected. Missing data? If there are no observations then no associations can be found.

Diseases can be identi fj ed by the side e ff ects they elicit 21

Diseases can be identi fj ed by the side e ff ects they elicit Diabetes 21

Diseases can be identi fj ed by the side e ff ects they elicit level of detection Diabetes 21

Diseases can be identi fj ed by the side e ff ects they elicit level of detection Diabetes unmeasured severe effect 21

Diseases can be identi fj ed by the side e ff ects they elicit measured level of minor effects detection Diabetes unmeasured severe effect 21

Diseases can be identi fj ed by the side e ff ects they elicit • physicians use observable side e ff ects to form hypothesis about the underlying disease measured level of minor effects detection Diabetes unmeasured severe effect 21

Diseases can be identi fj ed by the side e ff ects they elicit • physicians use observable side e ff ects to form hypothesis about the underlying disease • e.g. you can’t see diabetes, but you can measure blood glucose measured level of minor effects detection Diabetes unmeasured severe effect 21

Severe ADE’s can be identi fj ed by the presence of more minor (and more common) side e ff ects measured minor effects level of detection Adverse unmeasured Event severe effect 22

Severe ADE’s can be identi fj ed by the presence of more minor (and more common) side e ff ects • First, identify the common side e ff ects that are harbingers for the underlying severe AE measured minor effects level of detection Adverse unmeasured Event severe effect 22

Severe ADE’s can be identi fj ed by the presence of more minor (and more common) side e ff ects • First, identify the common side e ff ects that are harbingers for the underlying severe AE • Then, combine these side e ff ects together to form an “e ff ect pro fj le” for an adverse event measured minor effects level of detection Adverse unmeasured Event severe effect 22

Severe ADEs can be identified by the presence of more minor (and more common) side effects Pain Increased Numbness Blood measured level of Glucose minor effects detection unmeasured T2DM severe effect

DDI prediction validation Table S3 Novel drug-drug interaction predictions for diabetes related adverse events. Minimum Randomization Known DDI Rank Drug A Drug B Score Rank exists 38 PAROXETINE HCL PRAVASTATIN SODIUM 11.351896014962 72 DIOVAN HCT HYDROCHLOROTHIAZIDE 7.1786599539 89 94 CRESTOR PREVACID 4.7923771645 148 107 DESFERAL EXJADE 3.97220625 129 159 COUMADIN VESICARE 0.8928376683 169 160 DEXAMETHASONETHALIDOMIDE 0.8928376683 168 CRITICAL 170 FOSAMAX VOLTAREN 0.5033125 1138 175 ALIMTA DEXAMETHASONE 0.2442375 197 • Focus on top hit from diabetes classifier • paroxetine = depression drug, pravastatin = cholesterol drug • Popular drugs, est. ~1,000,000 patients on this combination!

Analyzed blood glucose values for patients on either or both of these drugs To the electronic health records…

200 11 Pravastatin (N = 2,063) 10 180 Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mg/dl) 9 160 8 140 7 120 6 100 5 80 Baseline After Treatment Tatonetti, et al. Clinical Pharmacology & Therapeutics (2011)

200 200 11 11 Pravastatin (N = 2,063) Pravastatin (N = 2,063) Paroxetine (N = 1,603) 10 10 180 180 Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mg/dl) Blood Glucose Concentration (mg/dl) 9 9 160 160 8 8 140 140 7 7 120 120 6 6 100 100 5 5 80 80 Baseline Baseline After Treatment After Treatment Tatonetti, et al. Clinical Pharmacology & Therapeutics (2011)

200 200 200 11 11 11 Pravastatin (N = 2,063) Pravastatin (N = 2,063) Pravastatin (N = 2,063) Paroxetine (N = 1,603) Paroxetine (N = 1,603) Combination (N = 135) 10 10 10 180 180 180 Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mg/dl) Blood Glucose Concentration (mg/dl) Blood Glucose Concentration (mg/dl) 9 9 9 160 160 160 +18 mg/dl incr. 8 8 8 140 140 140 p < 0.001 7 7 7 120 120 120 6 6 6 100 100 100 5 5 5 80 80 80 Baseline Baseline Baseline After Treatment After Treatment After Treatment Tatonetti, et al. Clinical Pharmacology & Therapeutics (2011)

no diabetics 200 11 Pravastatin (N = 2,063) Paroxetine (N = 1,603) Combination (N = 135) 10 180 Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mg/dl) 9 160 8 140 7 120 6 100 5 80 Baseline After Treatment Tatonetti, et al. Clinical Pharmacology & Therapeutics (2011)

no diabetics including diabetics 200 200 11 Pravastatin (N = 2,063) Pravastatin Paroxetine (N = 1,603) Paroxetine Combination (N = 135) Combination (N=177) 180 10 180 Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mg/dl) Blood Glucose Concentration (mg/dl) 9 160 160 8 140 140 7 120 120 6 100 100 5 80 80 Baseline After Treatment Baseline After Treatment Tatonetti, et al. Clinical Pharmacology & Therapeutics (2011)

no diabetics including diabetics 200 200 11 Pravastatin (N = 2,063) Pravastatin Paroxetine (N = 1,603) Paroxetine Combination (N = 135) Combination (N=177) 180 10 180 Blood Glucose Concentration (mmol/L) Blood Glucose Concentration (mg/dl) Blood Glucose Concentration (mg/dl) 9 160 160 8 140 ~60mg/dl 140 increase 7 120 120 6 100 100 5 80 80 Baseline After Treatment Baseline After Treatment Tatonetti, et al. Clinical Pharmacology & Therapeutics (2011)

Informatics methods have taken us far, skeptics remain

Informatics methods have taken us far, skeptics remain • Insulin Resistant Mouse Model

Informatics methods have taken us far, skeptics remain • Insulin Resistant Mouse Model • 10 control mice on normal diet (Ctl Ctl)

Informatics methods have taken us far, skeptics remain • Insulin Resistant Mouse Model • 10 control mice on normal diet (Ctl Ctl) • 10 control mice on high fat diet (HFD)

Informatics methods have taken us far, skeptics remain • Insulin Resistant Mouse Model • 10 control mice on normal diet (Ctl Ctl) • 10 control mice on high fat diet (HFD)

Informatics methods have taken us far, skeptics remain • Insulin Resistant Mouse Model • 10 control mice on normal diet (Ctl Ctl) • 10 control mice on high fat diet (HFD) Simulating Pre-Diabetics

Informatics methods have taken us far, skeptics remain • Insulin Resistant Mouse Model • 10 control mice on normal diet (Ctl Ctl) • 10 control mice on high fat diet (HFD) Simulating Pre-Diabetics

Informatics methods have taken us far, skeptics remain • Insulin Resistant Mouse Model • 10 control mice on normal diet (Ctl Ctl) • 10 control mice on high fat diet (HFD) Simulating Pre-Diabetics

Informatics methods have taken us far, skeptics remain • Insulin Resistant Mouse Model • 10 control mice on normal diet (Ctl Ctl) • 10 control mice on high fat diet (HFD) • 10 mice on pravastatin + HFD • 10 mice on paroxetine + HFD • 10 mice on combination + HFD

Summary of fasting glucose levels Average ITT Fasting Glucose (mg/dl) 100 120 140 160 180 60 80 C t l C t l P r a v a s t a t i n P a r o x e t i n e C o n t r o l C o m b i n a t i o n ~60mg/dl increase

Replication is vital to science • In biology we would never trust a result that hasn’t been replicated • Why should algorithms be any different?

Drug-drug interactions and acquired Long QT Syndrome (LQTS) • Long QT syndrome (LQTS): congenital or drug-induced change in electrical activity of the heart that can lead to potentially fatal arrhythmia: torsades de pointes (TdP) • 13 genes associated with congenital LQTS • Drug-induced LQTS usually caused by blocking the hERG channel ( KCNH2 ) AL George, J. Clin. Invest. (2013) From Berger et al., Science Signaling (2010)

Identify acquired LQTS drug-drug interactions using Latent Signal Detection AFib measured tachycardia bradycardia minor effects level of detection unmeasured LQTS severe effect Lorberbaum, et al. Drug Safety (2016)

Latent Signal Detection of acquired LQTS Top Prediction: Ceftriaxone + Lansoprazole • Ceftriaxone — common in-patient cephalosporin antibiotic • Lansoprazole — proton-pump inhibitor used to treat GERD, one of the most commonly taken drugs in the world • In the EHR: Patients on the combination have QT intervals 11ms longer, on average and are 1.5X as likely to have a QT interval > 500ms Lorberbaum, et al. Drug Safety (2016) 
 Lorberbaum, et al. JACC (In press)

• Predicted QT-DDI: ceftriaxone (cephalosporin antibiotic) and lansoprazole (proton pump inhibitor) • Neither drug alone has any evidence of QT prolongation/ hERG block • Negative control: lansoprazole + cefuroxime

• Predicted QT-DDI: ceftriaxone (cephalosporin antibiotic) and lansoprazole (proton pump inhibitor) • Neither drug alone has any evidence of QT prolongation/ hERG block • Negative control: lansoprazole + cefuroxime (another cephalosporin) – no evidence in FAERS of an interaction

• Predicted QT-DDI: ceftriaxone (cephalosporin antibiotic) and lansoprazole (proton pump inhibitor) • Neither drug alone has any evidence of QT prolongation/ hERG block • Negative control: lansoprazole + cefuroxime • Negative control: lansoprazole + cefuroxime (another cephalosporin) – no evidence in FAERS of (another cephalosporin) – no evidence in FAERS of an interaction an interaction

• Predicted QT-DDI: ceftriaxone (cephalosporin antibiotic) and lansoprazole (proton pump inhibitor) • Neither drug alone has any evidence of QT prolongation/ hERG block • Negative control: lansoprazole + cefuroxime • Negative control: lansoprazole + cefuroxime (another cephalosporin) – no evidence in FAERS of (another cephalosporin) – no evidence in FAERS of an interaction an interaction Ceftriaxone Cefuroxime

FAERS Lansoprazole Ceftriaxone+ Side Effect Profile Lorberbaum, et al. In Revision

FAERS Lansoprazole Ceftriaxone+ Side Effect Profile Lorberbaum, et al. In Revision

Cefuroxime+ Ceftriaxone+ Lansoprazole Lansoprazole FAERS Lorberbaum, et al. In Revision

Electronic Health Records Lansoprazole Ceftriaxone+ * * * * * * * * * * * * Cefuroxime+ Lansoprazole * * * * Lorberbaum, et al. In Revision

Electronic Health Records Lansoprazole Ceftriaxone+ * * * * * * * ~10ms longer * * * * * on average Cefuroxime+ Lansoprazole * * * * Lorberbaum, et al. In Revision

What is the mechanism?

MADSS Modular Assembly of Drug Safety Subnetworks Use network analysis to build AE • neighborhoods: a subset of the interactome surrounding AE “seed” proteins Score each protein on • connectivity to seeds using: Mean first passage time • Betweenness centrality • Shared neighbors • Inverse shortest path • Overarching hypothesis: drugs • targeting proteins within an AE neighborhood more likely to be Protein Adverse event (AE) involved in mediating that AE Drug known to cause AE Seed protein Interaction Drug predicted to cause AE • Ran MADSS using 13 LQTS genes as seeds Lorberbaum, et al. Clin. Pharmacol. Ther. (2015)

MADSS Modular Assembly of Drug Safety Subnetworks Use network analysis to build AE • neighborhoods: a subset of the interactome surrounding AE “seed” proteins Score each protein on • connectivity to seeds using: Mean first passage time • Betweenness centrality • Shared neighbors • Inverse shortest path • Overarching hypothesis: drugs • targeting proteins within an AE neighborhood more likely to be Protein Adverse event (AE) involved in mediating that AE Drug known to cause AE Seed protein Interaction Drug predicted to cause AE • Ran MADSS using 13 LQTS • Ran MADSS using 13 LQTS genes as seeds genes as seeds Lorberbaum, et al. Clin. Pharmacol. Ther. (2015)

Putative mechanisms of QT-DDIs Cluster 1 SCN5A LQTS Fosphenytoin ADRB1 CAV3 Metoprolol KCNH2 ATP4A CACNA1C CACNG1 Lansoprazole Ceftriaxone Phenytoin Diltiazem Cluster 7 Cluster 3 Known drug-target binding Predicted drug-hERG binding (DrugBank) (Random Forest classifier)

Automated Patch Clamp • Collaboration with Rocky Kass (CUMC Pharmacology Dept.) • Take HEK293 cells over- expressing the hERG channel • Perform a single-cell patch clamp experiment Nanion Patchliner • control • ceftriaxone alone • lansoprazole alone Voltage protocol: step to +40mV followed by a return to -40mV • combination of ceftriaxone and lansoprazole Lorberbaum, et al. JACC (In press)

Ceftriaxone+Lansoprazole Lorberbaum, et al. JACC (In press)

Ceftriaxone+Lansoprazole Cefuroxime+Lansoprazole Lorberbaum, et al. JACC (In press)

Ceftriaxone+Lansoprazole Cefuroxime+Lansoprazole Ceft+Lanso effect on hERG current Cefu+Lanso effect on hERG current 1.2 1.2 1.0 1.0 Change from Control Change from Control 0.8 0.8 0.6 0.6 0.4 0.4 Ceftriaxone + 10 μM Lansoprazole 0.2 0.2 Ceftriaxone + 1 μM Lansoprazole Cefuroxime + 1 μM Lansoprazole Ceftriaxone alone Cefuroxime alone 0.0 0.0 0 μM 0.1 μM 1 μM 10 μM 50 μM 100 μM 0 μM 0.1 μM 1 μM 10 μM 50 μM 100 μM Ceftriaxone Concentration ( μM) Cefuroxime Concentration ( μM) Lorberbaum, et al. JACC (In press)

Computational model of human ventricular myocyte Wildtype channel 1 μM Lansoprazole + 100 μM Ceftriaxone (10% block) 10 μM Lansoprazole + 100 μM Ceftriaxone ( 55 % block) 0mV 50mV 100ms Lorberbaum, et al. JACC (In press)

Computational model of human ventricular myocyte Wildtype channel 1 μM Lansoprazole + 100 μM Ceftriaxone (10% block) 10 μM Lansoprazole + 100 μM Ceftriaxone ( 55 % block) most common at CUMC 0mV 10ms longer 50mV 100ms Lorberbaum, et al. JACC (In press)