Dialogue DataGrid Motivating Applications

Dialogue DataGridMotivating Applications

Joel Saltz, MD, PhDChair Biomedical Informatics

College of MedicineThe Ohio State University

Dialogue DataGrid

• Relational databases, files, XML databases, object stores

• Strongly typed• Multi-tiered metadata management

system• Incorporates elements from OGSA-DAI,

Mobius, caGrid, STORM, DataCutter, GT4 …

• Scales to very large data, high end platforms

Requirements

• Support or interoperate with caGrid, eScience infrastructure

• Interoperate with or replace SRB• Well defined relationship to Globus

Alliance• Services to support high end large scale

data applications • Design should include semantic metadata

management• Well thought out relationship to

commercial products (e.g. Information Integrator, Oracle)

Motivating Application Class I: Phenotype characterization

• Information heterogeneity, data coordination and data size

• Synthesize information from many high throughput information sources

• Sources can include multiple types of high throughput molecular data and multiple imaging modalities.

• Coordinated efforts at multiple sites

• Detailed understanding of biomedical phenomena will increasingly involve the need to analyze very large high resolution spatio-temporal datasets

Structural Complexity

Example Questions (Phenotypes associated with Rb knockouts)

1. What are the mechanisms of fetal death in mutant mice?

2. What structural changes occur in the placenta?

3. How different are the structural changes between the wild and mutant types?

4. …

Rb-Rb+

Dataset Size: Systems Biology

Future big science animal experiments on cancer, heart disease, pathogen host responseBasic small mouse is 3 cm3

1 μ resolution – very roughly 1013 bytes/mouse

Molecular data (spatial location) multiply by 102

Vary genetic composition, environmental manipulation, systematic mechanisms for varying genetic expression; multiply by 103

Total: 1018 bytes per big science animal experiment

Now: Virtual Slides(roughly 25TB/cm2 tissue)

Compare phenotypes of normal vs Rb deficient mice

Placenta

Slides/Slices Alignment

SegmentationVisualization

Computational Phenotyping Challenges

• Very large datasets• Automated image analysis • Three dimensional reconstruction• Motion• Integration of multiple data sources• Data indexing and retrieval

Large Scale Data Middleware Requirements

• Spatio-temporal datasets• Very large datasets

– Tens of gigabytes to 100+ TB data

• Lots of datasets– Up to thousands of runs for a study are possible

• Data can be stored in distributed collection of files

• Distributed datasets– Data may be captured at multiple locations by multiple groups– Simulations are carried out at multiple sites

• Common operations: subsetting, filtering, interpolations, projections, comparisons, frequency counts

Our Example: Ohio Supercomputing Center Mass Storage

Testbed

D V D

D V D

D V D

D V D

D V D

D V D

MetaData Servers

Core Storage Pool (35/50 TB) with SAN.FS

(4)772 MB/s throughput

(4)772 MB/s throughput

(4)772 MB/s throughput

(4)772 MB/s throughput

SAN Volume Controller(4 servers)

FAStT900 (4)

Backup Storage3584 Tape

1 L32 2 D32Actual: 640 cartridges @ 200

GB for a total of 128 TB4 drives

max drive data rate is 35 MB/s

Cisco Directors 9509

(4)

(2)890 MB/s throughput

(2)890 MB/s Throughput

(2) (2) (2)

(2)

(2) (2)

(16 -

4 per serve

r)

890 MB/s

throughput

Scratch / Archive Storage Pool (310/420 TB)FAStT600 Turbo (20)

LinTel boxes (PvFS/Active Disk Archive) (20)

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

D V D

(40 - 2 per T600)384 MB/s throughput

(40 - 2 per xSeries)10 GB/s

• 50 TB of performance storage

– home directories, project storage space, and long-term frequently accessed files.

• 420 TB of performance/capacity storage

– Active Disk Cache - compute jobs that require directly connected storage

– parallel file systems, and scratch space.

– Large temporary holding area

• 128 TB tape library– Backups and long-term

"offline" storage

Very Large Dataset Hardware is Proliferating

STORM Services• Query• Meta-data• Indexing• Data Source• Filtering• Partition

Generation• Data Mover

STORM Results

STORM I/O Performance

0

500

1000

1500

2000

2500

3000

3500

4000

4500

1 2 4 8 16

# XIO nodes

Ban

dw

idth

(M

B/s

)

2 Threads

4 Threads

Max

Seismic Datasets10-25GB per file. About 30-35TB of Data.

Motivating Application II: caBIG In vivo Imaging Workspace

Testbed• Study the effects of image acquisition and reconstruction

parameters (i.e. slice thickness, reconstruction kernel and dose) on CAD and on human ROC. – use multiple datasets and several CAD algorithms to

investigate the relationship between radiation dose and nodule detection ROC.

• Cooperative Group Imaging Study Support– Children’s Oncology Group: quantify whether perfusion

study results add any additional predictive value to the current battery of histopathological and molecular studies

– CALGB: Grid based analysis of PET/CT data to support phase I, II studies

– NTROI: Grid based OMIRAD -- registration, fusion and analysis of MR and Diffusive Optical Tomography (DOT).

CAD Testbed Project RSNA 2005 (joint with Eliot Siegel et al at Univ.

Maryland)• Expose algorithms and data management as Grid Services

• Remote execution of multiple CAD algorithms using multiple image databases

• CAD algorithm validation with larger set of images

• Better research support — recruit from multiple institutions, demographic relationships, outcome management etc.

• Remote CAD execution - reduced data transfer & avoid need to transmit PHI

• CAD compute farms that reduce the turnover time

• Scalable and open source — caBIG standards

Architecture

Image Data Service

•Expose data in DICOM PACS with grid service wrappers

•An open source DICOM server — Pixelmed

•XML based data transfer

1x Los Angeles

3x Chicago

1x Columbus

5 Participating Data Services

CAD Algorithm Service• Grid services for algorithm invocation and image retrieval service• caGRID middleware to wrap CAD applications with grid services• Report results to a result storage service

caGrid Introduce facilitates service creation

GUMS/CAMS is used to provide secure data communication and service invocation

CAD algorithms provided by iCAD and Siemens Medical Solutions.Prototypes for investigational use only; not commercially available

Framework Support Services

• Result storage server — A distributed XML database for caching CAD results

• GME — Manage communication and data protocols

16

15

1714

18

5

12

Slice = 127 W/L = 2200/-500

User Interface

Available data services

Queried results

DICOM image viewer

Click to browse images, submit CAD analysis, and view results

Motivating Application III: Integrative Cancer Biology Center

on Epigenetics (PI Tim Huang, OSU)

• TGFβ/Smad4 targets are regulated via epigenetic mechanisms. Promoter hypermethylation is a hallmark of this epigenetically mediated silencing. In this study, we have combined both chromatin immunoprecipitation microarray (ChIP-chip) and differential methylation hybridization (DMH) to investigate this epigenetic mechanism in ovarian cancer

Translating a goal into workflow

Construct KbTSMAD(Knowledgebase of the

TGF-B/SMAD signaling pathway)

Chip-on-chip Experiment DMH Experiment

Data Mining(ex: clustering

Description of Experiment

Description of Experiment

KbTSMAD

DMH Results(Microarray

Data)

Chip-chip Results(Microarray data)

Normilization with statistical tools

Data Collection-Genome-UCSC Genome -BLAT alignment

Literature Survey- Experimentally verified TGF-B target genes- Housekeeping Genes

Chip design

Custom chip design info (e.g. from Agilent)

Candidate cutting Enzyme Information

Clinical data

Experimental ResultsFrom Other Groups

Data source A

Data source B and C

Analytical service F

Data source D and E

Data source G

Analytical service I

Analytical service J Data source K

Data source L

Data source H

A

B C D E

G

H

K

L

n1

n2 n3

n4n5

n6

n7

n8

n9

n10

ArrayAnnotator

Application of caGrid to the workflow

• Application needs to support access to a diverse set of databases and execution of different analysis methods

• Data services– KbSMAD– Chip information from chip company– Enzyme data– Clinical data– Experimental results– Experimental design

• Analytical services– Designing a custom array– Normalization– Data mining (ex: clustering)

Example: Prototype of Clone Annotation Analytical Service

• Analytical Service: ArrayAnnotator Goal: Provide a annotation for each clone to select a subset of clones

among 400,000 candidate clones to design a custom array for DMH experiment

Clone selection criteria Clones within a promoter region Clones with proper internal and external cuts Clones within CpG island region and/or high CG contents Clones with Transcription Factor binding sites

Input: CloneType information extended sequence, enzyme info, genomic location, etc

Functions• Determine external cut locations around a clone region (e.g., cut-site by

BfaI)• Examine the internal cut around a clone region (e.g., cut-site by HapII,

HinpII, and MCrBc)• Identify the location of clone in genome• Show ether it is within promoter region or not• Calculate CG content and overlapping with CpG islands• Identify which Transcription Factor binding sites are overlapped with clones

Chip design application

Clone InfoData Services

Genome SequenceData Source

Annotation Analytical ServiceQ

uer

y (g

eneI

d)

Query

Example caGrid Usage in P50 chip design application

21

3

5

4

Res

ult

: L

ist

of c

lon

es

Result: extended genome sequence of clone

6

Request (cloneInfo)

Result: annotation (CpG, cutsite, promoter region, etc)

ArrayAnnotator output (Hao Sun, Ramana Davuluri)

Multiscale Laboratory Research Group

Ohio State UniversityJoel SaltzGagan AgrawalUmit CatalyurekDan CowdenMike GrayTahsin KurcShannon HastingsSteve LangellaScott OsterTony PanDK PandaSrini Parthasarathy

P. SadayappanSivaramakrishnan (K2)Michael Zhang

The Ohio Supercomputer Center

Stan AhaltJason BryanDennis SessannaDon StredneyPete Wycoff

Microscopy Image Analysis

• Biomedical Informatics– Tony Pan– Alexandra Gulacy – Dr. Metin Gurcan– Dr. Ashish Sharma– Dr. Kun Huang– Dr. Joel Saltz

• Computer Science and Engineering– Kishore Mosaliganti– Randall Ridgway– Richard Sharp

● Pathology–Dr. Dan Cowden

● Human Cancer Genetics

–Pamela Wenzel–Dr. Gustavo Leone–Dr. Alain deBruin

caGrid Team

• National Cancer Institute– Peter Covitz – Krishnakant

Shanbhag

• Ohio State University– Shannon Hastings – Tahsin Kurc– Stephen Langella – Scott Oster – Joel Saltz

• SAIC– Tara Akhavan– Manav Kher – William Sanchez – Ruowei Wu – Jijin Yan

• Booze | Allen | Hamilton– Manisundaram Arumani

• Panther Informatics Inc.– Nick Encina– Brian Gilman

Tony Pan, Stephen Langella, Shannon Hastings, Scott Oster, Ashish Sharma, Metin Gurcan, Tahsin

Kurc, Joel SaltzDepartment of Biomedical Informatics

The Ohio State University Medical Center, Columbus OH

Eliot Siegel, Khan M. SiddiquiUniversity of Maryland School of Medicine, Baltimore, MD

Thanks to Siemens, ICAD for supplying CAD algorithms

RSNA 2005 Team