Multiscale informatics for understanding drug response

Russ B. Altman, MD, PhD

Departments of Bioengineering, Genetics, Medicine & (by courtesy) Computer Science

Stanford University

PharmGKB, http://www.pharmgkb.org/

Outline of comments today

Our focus: drug action at molecular, cellular and organismal level (for pharmacogenetics, mechanism, repurposing, drug interactions)

Our toolkit: informatics tools and datasets for discovery and for leveraging the investment.

1. Structural molecular data for drug repurposing

2. Expression data for understanding cellular responses to drugs

3. Text information for predicting drug interactions

4. Population and clinical data to discover drug interactions

Future prospects for systems pharmacology.

http://www.pharmgkb.org/

Summarizing clinical implications of molecular changes

PMID: 20435227

Beta 2 Adrenergic Receptor, a G-protein coupled receptor (GPCR), frequent pharmacological target.

Beta 2 Adrenergic Receptor in lipid bilayer with surrounding intra- and extracellular aqueous solution.

Molecular Data

• Structural genomics initiative: great increase in “solved” 3D structures

• Includes many complexes of proteins with small molecule ligands

Can we identify use this information to predict new interactions between proteins and small molecule drugs?

The FEATURE system for describing molecular environments

Builds a statistical model of microenvironments in proteins, using occurrence of biophysical and biochemical properties.

Can detect remote similarities in environments from different proteins.

Atom TypeAtom ElementResidue NameResidue ClassPartial Charge

HydrophobicityAromaticEtc.

FEATURE1

FEATURE microenvironments detect remote flexible FAD binding sites

FEATURE “sees” similarities in diverse kinase structures (PIK3CG and SRC) = repurposing opportunity (Liu & Altman, PLoS Comp Bio, 2011).

Pairs with high similarity

Validating kinase predictions from 40 tested predictions (11 positive & 29 negative)

ValidatedHighly interestedMammalian

20

Liu & Altman, CPT: PSP, 2014

Moving from molecular to cellular data for drug action

• Expression data universally available at Gene Expression Omnibus

• Measures human gene expression in thousands of conditions, cell types

• Recent examples of drug repurposing using this information.

Can we understand drug action using this data?

Independent Component Analysis (ICA) on all human GEO data (~9K arrays)

component = hidden signal in data

Independent Component Analysis

Independent Component Analysis of GEO human datayields ~420 fundamental components that explain most variability in expression experiments.(Engreitz et al, J Biomed Inform. 2010)

Analyzing response to a drug: gene expression in response to parthenolide

• CD34+ cells • 12 AML patients • Samples before/after treatment

Key components expressed at different levels in treated vs. untreated

373 = component with biggest difference

Highly weightedGenes:

ICOSGNA15SOCS2GCH1ID3NR4A3PELI2ID1ZNF394KHDRBS3

Diseases with high level of component 373:

Pediatric T-ALLChildhood ALLAMLALLLLeukemiaInteferon treatment of Hep C

Parthenolide effects NFKB signaling. Pathways suggests new diseases and new targets.

Parthenolide alters expression of TNF, NFKB, kinases.

Immediately suggests a way to decide if Parthenolide is appropriate for other cancers, based on their gene expression.

Augmenting the network of molecular and cellular data with textual information.

• PubMed now holds more than 20 million abstracts

• Most biomedical knowledge stored in the published literature

Can we have computers “read” PubMed abstracts and reason over them to predict new drug-drug interactions?

Advances in natural language parsing enable high fidelity extraction of relations

Dependency graph

We can identify genes, drugs, phenotypes in text using these technologies.… an ontology to normalize

Coulet, Shah, Garten, Musen, Altman, JBI, 2010.

KEY: we map extracted entities to standard terminologies

Coulet, Shah, Garten, Musen, Altman, JBI, 2010.

Semantic network of 170,598 normalized relations from all PubMed abstracts.

CYP3A4

CYP2D6

ACE

ABCB1 gene and verapamil drug

We chain together two gene-drug relationships to create a drug-gene-drug relationship = potential drug interaction!

Percha et al, Pacific Symposium on Biocomputing, 2012.

Toppredicted DDIsfromtext processingafter training on known DDIs

drug pair interact npaths avgvotes p.glm LAB

fluoxetine-diazepam 1* 975 59.68% 0.608tramadol-dextromethorphan 1 536 95.27% 0.533venlafaxine-dextromethorphan 1 528 94.99% 0.523naproxen-diazepam 0 806 65.60% 0.506 1paroxetine-dextromethorphan 1 489 95.32% 0.489metoprolol-dextromethorphan 0 490 90.17% 0.441 0lisinopril-enalapril 1-- 986 37.46% 0.410verapamil-omeprazole 0 469 87.30% 0.395 1dextromethorphan-codeine 1- 440 86.98% 0.367omeprazole-naproxen 1* 911 40.09% 0.366sertraline-dextromethorphan 1** 357 95.02% 0.365omeprazole-diazepam 1 872 38.84% 0.322verapamil-fluconazole 1 252 97.34% 0.296verapamil-fexofenadine 0 248 93.28% 0.262 1verapamil-atorvastatin 1 296 88.32% 0.261naproxen-fluoxetine 1* 626 52.50% 0.240warfarin-glibenclamide 1* 221 92.29% 0.236warfarin-fluoxetine 1 272 86.93% 0.234verapamil-carvedilol 1 200 92.99% 0.227venlafaxine-tramadol 1 145 98.45% 0.227verapamil-clopidogrel 1- 184 94.40% 0.226 1venlafaxine-paroxetine 1 132 99.18% 0.223verapamil-simvastatin 1 263 86.04% 0.222tramadol-paroxetine 1 133 98.42% 0.219warfarin-ibuprofen 1 172 94.41% 0.218omeprazole-dextromethorphan 0 150 96.36% 0.216 0fluoxetine-dextromethorphan 1 160 95.16% 0.215venlafaxine-metoprolol 1* 132 94.89% 0.196venlafaxine-sertraline 1 97 97.97% 0.194tramadol-sertraline 1 97 97.84% 0.193dextromethorphan-citalopram 1** 104 96.35% 0.188paroxetine-metoprolol 1 126 93.79% 0.186tramadol-metoprolol 0 132 93.13% 0.186 0verapamil-diazepam 1- 300 76.30% 0.185 1dextromethorphan-aripiprazole 0 132 92.67% 0.183 0trazodone-dextromethorphan 1 88 96.78% 0.181donepezil-dextromethorphan 0 88 96.68% 0.181 0verapamil-amlodipine 1- 184 87.04% 0.180tolterodine-dextromethorphan 0 92 95.71% 0.178 0enalapril-diazepam 1* 641 40.72% 0.176verapamil-naproxen 1* 282 76.05% 0.174verapamil-lovastatin 1 177 86.27% 0.172omeprazole-folicacid 0 600 44.01% 0.171 0verapamil-famotidine 0 46 98.80% 0.170 1venlafaxine-codeine 1* 120 91.24% 0.169sertraline-metoprolol 1 91 93.70% 0.166memantine-dextromethorphan 1 45 98.14% 0.166tramadol-codeine 1** 120 90.28% 0.163venlafaxine-tolterodine 1- 24 99.56% 0.162

Metoprolol

Class Beta blocker

Brand Lopressor, Toprol

Treats High blood pressureAngina (chest pain)Heart attack (improves survival)Heart failure

Effects Relaxes blood vesselsSlows heart rate

Mech Blocks beta receptors on sympathetic nerves

Dextromethorphan

Class Non-opioid antitussive

Brand Robitussin (+ tons of others)

Treats Cough

Effects Acts on central nervous systemto elevate threshold forcoughing

Mech NMDA and glutamate antagonistBlocks dopamine reuptake site(?)

Metoprolol Dextromethorphan

metoprolol Drug isMetabolizedBy Gene CYP2D6…

Metoprolol Dextromethorphan

… CYP2D6 Gene metabolizes Drug Dextromethorphan

DEEP DIVE for gene-drug-phenotype (Chris Re, CS)

Using population-based and clinical data to generate hypotheses

• FDA releases adverse event reports regularly, but very complex data.

• Electronic Medical Records have information about patient drug exposures?

Can we mine FDA and clinical EMR records to discover new drug-drug interactions?

We built a statistical model able to recognize glucose-altering drugs based on their “adverse event signature”

Fij = frequency of AEj for Drugi

DiabetesDrugs

CorrelatedIndications

Fij = frequency of AEj for Indicationi

Adverse Events reported by FDA

The adverse event signature is able to recall glucose-altering drugs

Model trained on single drug reports can be used to classify drug pairs. EXAMPLE: Glucose-homeostasis adverse effects

Fij = frequency of AEj for Drugi

DrugsPairs

Pairs of drugs that match signature:

Paroxetine and Pravastatin each taken by approximately 15 million Americans.We estimate 500,000 to 800,000 take both.

Pravastatin and paroxetine significantly increase blood glucose by 20 mg/dl

VariablePravastatin and Paroxetine

Combination

N 10

Demographics

Age (mean ± SD) 59.9 ± 11.09

Gender (% Female) 90.0

Race (% of group)White 50

African American 20

Hispanic 0

Other 30

Glucose (mg/dl mean ± SD)

Baseline (base) 114.08 ± 14.79

After treatment(s) (post) 133.96 ± 19.54

paired t-test (t: post - base) 0.020

Change (base to post) 19.88 ± 21.04

N patients with increase 8 (80)

Combination therapy increases glucose by 20 mg/dl; pravastatin and paroxetine effects alone are modest

Site 2(10 patients)

Replication sites validate combination findings

Vanderbilt(18 patients)

Harvard Partners(106 patients)

Combining all three sites

Not a “class affect” of statins and SSRIs

Effect larger in diabetics

N = 177

Mice on two drugs also show increase(Tatonetti et al, Clin Pharm Ther. 2011)

Web searches? Patient search for pravastatin & paroxetine and DM-related words more frequently.

White et al, J Am Med Inform Assoc. 2013 May 1;20(3):404-8

Conclusions

• Biomedical informatics provides powerful tools for discovery at the molecular, cellular, and organismal levels.

• They operate on both primary data and processed knowledge.

• They can individually contribute to understanding drug action, effects and interactions.

• The combination & integration of these methods is likely to be even more powerful going forward…

Combining text mining + FDA datadrug pair interact npaths avgvotes p.glm LAB

fluoxetine-diazepam 1* 975 59.68% 0.608tramadol-dextromethorphan 1 536 95.27% 0.533venlafaxine-dextromethorphan 1 528 94.99% 0.523naproxen-diazepam 0 806 65.60% 0.506 1paroxetine-dextromethorphan 1 489 95.32% 0.489metoprolol-dextromethorphan 0 490 90.17% 0.441 0lisinopril-enalapril 1-- 986 37.46% 0.410verapamil-omeprazole 0 469 87.30% 0.395 1dextromethorphan-codeine 1- 440 86.98% 0.367omeprazole-naproxen 1* 911 40.09% 0.366sertraline-dextromethorphan 1** 357 95.02% 0.365omeprazole-diazepam 1 872 38.84% 0.322verapamil-fluconazole 1 252 97.34% 0.296verapamil-fexofenadine 0 248 93.28% 0.262 1verapamil-atorvastatin 1 296 88.32% 0.261naproxen-fluoxetine 1* 626 52.50% 0.240warfarin-glibenclamide 1* 221 92.29% 0.236warfarin-fluoxetine 1 272 86.93% 0.234verapamil-carvedilol 1 200 92.99% 0.227venlafaxine-tramadol 1 145 98.45% 0.227verapamil-clopidogrel 1- 184 94.40% 0.226 1venlafaxine-paroxetine 1 132 99.18% 0.223verapamil-simvastatin 1 263 86.04% 0.222tramadol-paroxetine 1 133 98.42% 0.219warfarin-ibuprofen 1 172 94.41% 0.218omeprazole-dextromethorphan 0 150 96.36% 0.216 0fluoxetine-dextromethorphan 1 160 95.16% 0.215venlafaxine-metoprolol 1* 132 94.89% 0.196venlafaxine-sertraline 1 97 97.97% 0.194tramadol-sertraline 1 97 97.84% 0.193dextromethorphan-citalopram 1** 104 96.35% 0.188paroxetine-metoprolol 1 126 93.79% 0.186tramadol-metoprolol 0 132 93.13% 0.186 0verapamil-diazepam 1- 300 76.30% 0.185 1dextromethorphan-aripiprazole 0 132 92.67% 0.183 0trazodone-dextromethorphan 1 88 96.78% 0.181donepezil-dextromethorphan 0 88 96.68% 0.181 0verapamil-amlodipine 1- 184 87.04% 0.180tolterodine-dextromethorphan 0 92 95.71% 0.178 0enalapril-diazepam 1* 641 40.72% 0.176verapamil-naproxen 1* 282 76.05% 0.174verapamil-lovastatin 1 177 86.27% 0.172omeprazole-folicacid 0 600 44.01% 0.171 0verapamil-famotidine 0 46 98.80% 0.170 1venlafaxine-codeine 1* 120 91.24% 0.169sertraline-metoprolol 1 91 93.70% 0.166memantine-dextromethorphan 1 45 98.14% 0.166tramadol-codeine 1** 120 90.28% 0.163venlafaxine-tolterodine 1- 24 99.56% 0.162

Orange highlighted predictions from textmining are also predictions in the FDA mining effort!

Combining molecular & cellular data

Proteins with similar pockets can be associated with disease pathways for repurposing.

Thus, the emerging network for drugs….

Target structure & dynamics

Drug recognition &binding

Cellular response &pathways

Text mining of gene, drug, phenotype associations

Clinical response datamining

Population effect reporting

Teri Klein Michelle Roxana Li Mei Beth Percha Jenn

Jesse Grace Tang Nick Tatonetti Caroline Mark Jesse Engreitz

Guy Joan Sara Fen Tianyun Liu Katrin

Jenelle Yael Garten Konrad Ellie Blanca Ryan

But the database biases can kill you.

Corrects for *unmeasured* biasees

Tatonetti et al, Science STM, 2012

Thanks to PharmGKB Team

Teri KleinMichelle CarrilloCurators: Li Gong, Joan Hebert, Katrin Sangkuhl, Laura Hodges, Connie Oshiro, Caroline ThornDevelopers: Mark Woon, Ryan Whaley, Feng Liu, Mei Gong, Rebecca TangData Center: Tina Zhou, TC Truong

NIH: GM-61374

Key Papers

PharmGKB: a logical home for knowledge relating genotype to drug response phenotype. Altman RB. Nat Genet. 2007 Apr;39(4):426. No abstract available. PMID: 17392795 [PubMed - indexed for MEDLINE]

Independent component analysis: mining microarray data for fundamental human gene expression modules. Engreitz JM, Daigle BJ Jr, Marshall JJ, Altman RB. J Biomed Inform. 2010 Dec;43(6):932-44. Epub 2010 Jul 7. PMID: 20619355

Detecting drug interactions from adverse-event reports: interaction between paroxetine and pravastatin increases blood glucose levels. Tatonetti NP, Denny JC, Murphy SN, Fernald GH, Krishnan G, Castro V, Yue P, Tsao PS, Kohane I, Roden DM, Altman RB. Clin Pharmacol Ther. 2011 Jul;90(1):133-42. doi: 10.1038/clpt.2011.83. Epub 2011 May 25. Erratum in: Clin Pharmacol Ther. 2011 Sep;90(3):480. Tsau, P S [corrected to Tsao, P S]. PMID: 21613990

Using text to build semantic networks for pharmacogenomics.Coulet A, Shah NH, Garten Y, Musen M, Altman RB. J Biomed Inform. 2010 Dec;43(6):1009-19. Epub 2010 Aug 17. PMID: 20723615 [PubMed - indexed for MEDLINE]

DISCOVERY AND EXPLANATION OF DRUG-DRUG INTERACTIONS VIA TEXT MINING. Percha B, Garten Y, Altman RB. Pacific Symposium on Biocomputing 2012, in press. Available at http://psb.stanford.edu/psbonline/

A novel signal detection algorithm for identifying hidden drug-drug interactions in adverse event reports. Tatonetti NP, Fernald GH, Altman RB. J Am Med Inform Assoc. 2011 Jun 14. [Epub ahead of print] PMID: 21676938

The FEATURE framework for protein function annotation: modeling new functions, improving performance, and extending to novel applications. Halperin I, Glazer DS, Wu S, Altman RB. BMC Genomics. 2008 Sep 16;9 Suppl 2:S2. PMID: 18831785