Translational Science on the Cloud An Experiment in Translational Science Peter J. Tonellato Dennis Wall Center for Biomedical Informatics Harvard Medical.

Translational Science on the Cloud

An Experiment in Translational Science

Peter J. Tonellato

Dennis Wall

Center for Biomedical Informatics

Harvard Medical School

pa lav er⋅ ⋅   /pəˈlævər, ‑ˈlɑvər/ noun

1. long parley usually between persons of different cultures or levels of sophistication

2. conference, discussion

3. idle talk

4. misleading or beguiling speech

A few ambitious goals –

• Host a cross-disciplinary geographically distributed palaver using webcasting technology.

• Collaborate on a complex set of high caliber scientific and computationally complex projects.

• Provide a thematically consistent set of lectures by a world-class collection of lecturers.

• Implement and test the activities on a new technology never previously used for scientific exploration.

Project Objectives• Scientific

• Computational-BioMedical Informatic

• “Cloud”– Manage Resources, reduce complexity and costs

• “Translational”– Research -> Examination of Clinical Potential– Potential -> Clinical Efficacy– Clinical Efficacy -> Clinical Use

Gartner WarningsBest to avoid Peaks and Troughs if Possible.

Participation• ‘I like to watch’

– attend or watch recorded lectures

• ‘I like to watch - a lot’– same as above and attend (skype, webex or in person) project

discussions

• ‘I like to more than watch’– above and join active project team

– contribution to project objectives

* To Doug MacFadden for noting the “Being There” connection.

Collaborators

• Kurt Messersmith, Terry Wise, Jinesh Varia, and the AWS group

• Josh Fraser, Ed Goldberg, and the RightScale group

• Sushil Kumar, William Hodak and Oracle group

Participants (incomplete list)• Laboratory for Personalized Medicine

– Peter Tonellato, Vincent Fusaro, Prasad Patil, Rimma Pivovarov, Peter Kos

• Wall Lab– Dennis Wall, Parul Kudtarkar, Joy Poulo, Matt Hyuck

• Church Lab– Alexander Wait

• Thomson Lab

– Victor Ruotti, Ron Stewart

• University of Wisconsin – Milwaukee– Peter Kos, Dave Petering, Tom Hansen, David Stack, Joseph Bockhorst

• Tokyo Medical and Dental University– Kumiko Oohashi, Takako Takai, Yutaka Fukuoka

• Recombinant Data– Dan Housman

• Great Lakes WATER Institute– Michael Caravan, Rick Goetz

• Medical College of Wisconsin– Simon Twigger

• Marquette University

– Craig Strubble

Participants (incomplete list)

Acknowledgements

Vincent Fusaro

Prasad Patil

Peter Kos

Zhitao Wang

Dan Chen

Haiping Xia

Sumana Ramayanam

Laboratory for Personalized MedicinePeter J. Tonellato, Ph.D.

Amazon:

Tenesha Gleason

Ford Harris

Wall Lab:

Dennis Wall, Ph.D.

Tom Monaghan

Laboratory of Personalized MedicineCBMI, Harvard Medical School

Established in 2008 to Develop:

• Clinical-genetic mathematical models

• Translational science simulation paradigm and

• Personalized Medicine (PM) Web applications

and create a facilitated pathway from genetic discovery to clinical enterprise

Project Objectives• Scientific: Modeling and Prediction of Clinical Avatars and

Pharmacogenetic Dosing

• Computational-BioMedical Informatic: Accuracy of Simulations, mashup, Webapplication

• “Cloud”– Manage Resources, reduce complexity and costs

• “Translational”– Research -> Examination of Clinical Potential– Potential -> Clinical Efficacy– Clinical Efficacy -> Clinical Use

Oracle in the Cloud

Posted: May 6, 2008 10:43 AM PDT

Here at Oracle, we have been keeping track of the great strides being made by the Amazon Web Services team in enabling a Cloud Computing platform. We are looking to talk with people who are interested in utilizing Oracle technologies within the AWS platform. Please contact me directly at my email address below if you would like to share your thoughts on how Oracle technologies can help your AWS projects or if you are interested in simply sharing your experiences with AWS.

I look forward to hearing from you!

Bill Hodak

Senior Product Manager - Oracle Corporation

[email protected]

TimeLine

Amazon Web Services (AWS)Amazon Web Services (AWS)

HPCAMI

HPCAMI

AmazonS3

AmazonS3

Oracle AMI

Oracle AMI

AmazonEC2

Instances

AmazonEC2

Instances

User Application

User Application

LinuxServerLinuxServer

Fitting the Pieces Together

Math Modeling and SimulationHPC Cloud Service

Simulation as Service Options– Matlab

– Mathematica

– R

– SAS

– S-PLUS

R Benefits:– Fast computation and

statistical analysis– Large mathematical and

statistical library– Open source– Highly extensible– Supportive user

community

OpenXava

Application Ready for Production

Business Components

Controllers+ =

• Deployable on Java Application Server or any Servlet Container, or on a Portal (Liferay, JetSpeed or WebSphere)

“Clouded” Translational Science

• Web application framework is flexible

• Robust technologies

– Oracle and AWS cloud services in concert with R, OpenXava, Ruby

• Extreme Implementation: LPM team no previous collaboration

• Cloud Development Service inventory growing rapidly.

- Subversion - i2b2 - R/S/Splus - Research Data

- Development Platform:

- OpenXava and dependecies

- Ruby-on-Rails and dependencies

- Clinical Trial simulation service,

Oracle in the Cloud

Posted: May 6, 2008 10:43 AM PDT

From: Tonellato, Peter

Sent: Tuesday, June 24, 2008 12:09 PM

We have successfully launched the personalized medicine translational research platform on AWS. …

P

Peter J. Tonellato, Ph.D.

Center for Biomedical Informatics

Harvard Medical School

Children's Hospital of Boston

617.432.7185 866.771.2566 (fax)

TimeLine

Footnote:The team never met together and more than half had never worked together.

Warfarin Pharmacogenetic Simulation Service Application

Goals– Predict dosage to achieve rapid therapeutic dosing

– Create clinical ‘avatar’ patient-base – reflects real data

– Identify patients-types or sub-populations who may experience difficulty achieving therapeutic Warfarin level

– Create flexible and extensible modular framework as the basis for future translational science studies

Dosage/INR Prediction OverviewModels used for generating initial dosage:Anderson et. Al.1 :

Dose = 1.64 + exp[3.984 + c(x) + v(x) + g(x) - age*(0.009) + weight*(0.003)]

{ 0 if genotype = CYP2C9*1/*1 {-0.197 if genotype = CYP2C9*1/*2c(x) = {-0.360 if genotype = CYP2C9*1/*3 {-0.947 if genotype = CYP2C9*2/*3 {-0.265 if genotype = CYP2C9*2/*2 {-1.892 if genotype = CYP2C9*3/*3

{ 0 if VKORC1 1173 genotype = C/Cv(x) = {-0.304 if VKORC1 1173 genotype = C/T {-0.569 if VKORC1 1173 genotype = T/T

g(x) = { 0 if gender = female { 0.094 if gender = male

1. Anderson JL, Horne BD, Stevens SM, Grove AS, Barton S, Nicholas ZP, et al. Randomized trial of genotype-guided versus standard warfarin dosing in patients initiating oral anticoagulation. Circulation 2007 Nov 27;116(22):2563-2570.

CYP2C9 genotype elements in this algorithm are derived from the CYP2C9 gene/allele generic hash map

Gage et. Al 2: Dose = exp[0.9751 − 0.3238 × v(y) + (0.4317 × BSA) - 0.4008 ×

c_3(y) − (0.00745 × age) − 0.2066 × c_2(y) + (0.2029 × target INR) − (0.2538 x amiodarone) + (0.0922 ×smokes) - (0.0901 × African-American race) + (0.0664 × DVT/PE)]

{ 0 if VKORC1 -1639 genotype = G/Gv(y) = { 1 if VKORC1 -1639 genotype = G/A { 2 if VKORC1 -1639 genotype = A/A { 0 if CYP2C9*2 genotype = C/Cc_2(y) = { 1 if CYP2C9*2 genotype = C/T { 2 if CYP2C9*2 genotype = T/T { 0 if CYP2C9*3 genotype = A/Ac_3(y) = { 1 if CYP2C9*3 genotype = A/C { 2 if CYP2C9*3 genotype = C/C

2. Gage B, Eby C, Johnson J, Deych E, Rieder M, Ridker P, et al. Use of Pharmacogenetic and Clinical Factors to Predict the Therapeutic Dose of Warfarin. Clin.Pharmacol.Ther. 2008 Feb 27.

Variation of CYP2C9 Genotype (Gage Model)

A/A G/A G/G

02

46

810

12*1/*1

VKORC1 Genotype

Dos

age

(mg)

A/A G/A G/G

02

46

810

12

*1/*2

VKORC1 Genotype

Dos

age

(mg)

A/A G/A G/G

02

46

810

12

*1/*3

VKORC1 Genotype

Dos

age

(mg)

A/A G/A G/G

02

46

810

12

*2/*2

VKORC1 Genotype

Dos

age

(mg)

A/A G/A G/G

02

46

810

12*2/*3

VKORC1 Genotype

Dos

age

(mg)

A/A G/A G/G

02

46

810

12

*3/*3

VKORC1 Genotype

Dos

age

(mg)

Dosage vs. WSI by CYP2C9 Genotype (20,000 patients)

Current Results

• LPM Warfarin Web App Completed in two months

• 100 Million clinical avatar and dosing simulations

• Translational Science paradigm supports clinical trial simulation, incidentalome testing, and leads to new metrics for clinical efficacy

• New Metrics for Clinical Efficacy e.g. Warfarin ‘Sensitive’ Participants

We have demonstrated the value and flexibility of Cloud Services and Framework for future projects.

Acknowledgements

Vincent Fusaro

Rimma Pivovarov

Prasad Patil

Peter Kos

Zhitao Wang

Dan Chen

Haiping Xia

Sumana Ramayanam

Laboratory for Personalized MedicinePeter J. Tonellato, Ph.D.

Amazon:

Terry Wise

Kurt Messinger

Tenesha Gleason

Ford Harris

Projects• Network Analysis for Disease Genetics

• The Translational Variome

• Next Generation Sequence Analysis– DNA– RNA

• i2b2

• Pharmacogenetics - with Clinical Avatars

• Cloud Computational Center

About i2b2 and Recombinant

Recombinant Data Corp. (http://www.recomdata.com)

–Translational Research Open Source implementation and support

–i2b2 deployments: UMass, Johnson and Johnson, Wash U./UCSF/UC Davis collaboration

–Clinical data warehousing & integration services

“The i2b2 Center is developing a scalable informatics framework that will bridge clinical research data and the vast data banks arising from basic science research in order to better understand the genetic bases of complex diseases.”http://www.i2b2.org

i2b2: Informatics for Integrating Biology and the Bedside

Service based “i2b2 Hive” open source framework

i2b2 Running on Amazon Cloud Objectives

• Establish an i2b2 AMI• Test the AMI with clinical avatar data sets• Create a model/QA environment for federated queries

using SHRINE• Benchmark query performance with large SNP and gene

expression data sets• Define a security model/requirements for deploying

sensitive clinical data in the cloud• Investigate relevant implementation of high-compute

“cloud” models for correlation analysis

The HiveMind of Mechanical Turks(The Translational Variome)

Laboratory of Personalized Medicine

Can crowdsourcing be used to solve common biomedical information processing dilemmas?

Rimma Pivovarov February 22, 2009

What is Mechanical Turk?

Database Annotation

DNA Change

Amino Acid Change

Accession Number

How much understanding of biology is necessary?How long will this take?

How accurate will they be?

Can the Turks extract variant data from dbSNP?

10 RS Numbers = 10 tasks3 individual Turks perform each task

10 x 3 = 30 Human Intelligence Tasks (HITs)

HIT Design

ResultsTo

tal N

umbe

r of H

ITs

Com

plet

ed

Number Correct

% Correct

DNA Change 100%

Accession ID 100%

Amino Acid Change 90%

Average 96.6%• Time elapsed: 11.5 hours

• Total Cost: 33 cents

• 7 Individual Turks Participated

Time

Number of HITs Completed Over Time

Abstract Interpretation

SHP2 HSP70

"The Src homology phosphotyrosyl phosphatase, SHP2, is a positive effector of EGFR signaling. However, the molecular mechanism and biological functions of SHP2 regulation are still not completely known. To better understand the cellular processes in which SHP2 participates, we carried out mass spectrometry to find SHP2 binding proteins. FLAG-SHP2 complexes were isolated by affinity purification, and associated proteins were identified by in-gel trypsin digestion followed by LC/MS/MS mass spectrometry. Among the identified proteins, we focus in this report on the heat shock protein 70 (HSP70). Physical interactions of SHP2 with HSP70 were confirmed in vivo. Further experiments demonstrate that EGF does not activate binding of SHP2 with HSP70 rather the binding appears to be constitutive. However, the formation of an HSP70/SHP2 complex affected the binding of SHP2 with EGFR and (or) GAB1. These data suggest that binding of HSP70 with SHP2 regulates to some extent the EGF signaling pathway. In addition, immunostaining experiments indicated that SHP2 and HSP70 co-localized in the cell membrane region after EGF treatment. Our findings propose a possible involvement of HSP70 in the regulation of EGF signaling pathway by SHP2."

Turkers Response

Cloud Computing CenterAIM: Understand how to properly launch and configure AWS servers,

monitor performance and cost, and manage large volumes of data on the cloud for a mixture of simultaneous start-up projects.

Lead by an Individual who does not know what he is doing.

Maintaining 'virtual' computing centers for each of the Palaver project teams. – Typical setting, launching 6-10 significant computing projects with

diverse hardware, software, flexibility and resource needs would take some time (months?).

– We will attempt to manage startup needs in a matter of days and manage them going forward with minimal effort ('minimal' to be determined!).

– Resource requirements implemented on AWS using RightScale

Cloud Computing Center• Amazon is sponsoring resources

• Vince Fusaro will wrangle resources

• Each Project lead will predict needs and coordinate through Vince

• “Special” requests will be managed directly with AWS – and must be justified, ….

• William Crawford will conduct meta-analysis of use and implementation. Please interact with him as needed.

• RightScale is interested in our experiences

RightScale

• Manage virtual servers

• Monitor usage statistics

Palaver WebSite

Website created and managed by Rimma Pivovarov

Website designed by Kristian St. Gabriel

Logistics• Monday 3-5 pm from now on….• Monitor the Web site for updates

• Review the Project sites in the next week or so and confirm your level of participation

• Technical glitches … • Rimma on lecture/ Website issues• Vince on Project Computational Center issues.

• Project Team Leaders –• Will refine Project statement, Coordinate participants, project (skype) meetings, project logistics

• Palaver Day – May 6th

• Other??

Translational Science on the Cloud

Amazon Web Services: A Clouded Architecture

Jinesh Varia

Amazon