AN EXTENSIBLE BIOMARKER CURATION APPROACH AND

SOFTWARE INFRASTRUCTURE FOR THE EARLY DETECTION

OF CANCER

Andrew F. Hart

, John J. Tran

1,3

, Daniel J. Crichton

, Kristen Anton

Heather Kincaid

, Sean Kelly

, J. S. Hughes

and Chris A. Mattmann

1,3

Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109, U.S.A.

Section of Biostatistics and Epidemiology, Dartmouth Medical School, Lebanon, NH 03766, U.S.A.

Computer Science Department, University of Southern California, Los Angeles, CA 90089, U.S.A.

Keywords:

Bioinformatics, Data grid, Data management, Data procurement, Ontology.

Abstract:

Modern research requires collaboration among geographically distributed scientists. This collaborative model

is transforming scientiﬁc discovery by enabling sharing and validation of data across institutions. Informatics

infrastructures are being developed to support cancer research, endowing scientists with the ability to capture

and share data with remote collegues. A critical challenge presented by such infrastructures is the develop-

ment of a curation model for the science data. While considerable emphasis has been placed on developing

grid infrastructures, few are addressing the curation aspects crucial to creating a useful scientiﬁc knowledge-

base. The United States National Cancer Institute’s (NCI) Early Detection Research Network (EDRN) is a

distributed network of research institutions focused on the discovery of cancer biomarkers. In this paper, we

describe our work building a data collection and curation infrastructure on top of the existing EDRN bioinfor-

matics data grid. The approach involves normalizing curated data through the use of a common information

model for cancer biomarker research. We argue that such a model is critical to ensuring that data can be com-

bined into an integrated knowledge system. Furthermore, we argue that human curators with backgrounds in

both informatics and science play a critical role in the overall value of the EDRN knowledge-base.

1 INTRODUCTION AND

MOTIVATION

In recent years, the cancer research community has

seen increasing dependency on complex computing

infrastructure (clusters, exotic instrument technolo-

gies with associated software and data formats, etc.)

and data modeling ‘know-how’ (common data ele-

ments, ontologies, etc.) to support its scientiﬁc re-

search needs. In contrast to the silo’ed research labs

of old, scientists are now under immense pressure

to exchange, share and disseminate data across geo-

graphically dispersed institutions. As has been shown

(Crichton et al., 2006), the rate of scientiﬁc discovery

increases as data is shared between collegues. The

need to share data has necessitated a common vocab-

ulary not only for the format of data (the organization

of its “bits”), but for annotations about the data itself

– regularly referred to as metadata.

We have borne witness ﬁrst-hand to the prolifera-

tion of data management needs speciﬁc to the cancer

research community, as have others more generally

within the ﬁeld of biology (Lynch, 2008; Howe et al.,

2008). Over the past seven years, we have helped to

construct and deploy the enabling grid infrastructure

(Foster et al., 2001) for a highly successful bioinfor-

matics grid project within the U.S. National Cancer

Institute (NCI). The Early Detection Research Net-

work (EDRN) project is a network of over forty insti-

tutions across the U.S. all collaborating with the com-

mon goal of the early detection of cancer through dis-

covery of promising cancer biomarkers – predictors

of cancer at an early stage.

To assist EDRN in the curation of its many types

387

F. Hart A., J. Tran J., J. Crichton D., Anton K., Kincaid H., Kelly S., S. Hughes J. and A. Mattmann C. (2009).

AN EXTENSIBLE BIOMARKER CURATION APPROACH AND SOFTWARE INFRASTRUCTURE FOR THE EARLY DETECTION OF CANCER.

In Proceedings of the International Conference on Health Informatics, pages 387-392

DOI: 10.5220/0001781103870392

 SciTePress

of data, we have built a suite of applications col-

lectively termed the EDRN Knowledge Environment

(EKE) (Crichton et al., 2006). EKE consists of:

a distributed specimen inventory called ERNE (for

EDRN Resource Network Exchange), study manage-

ment database applications called VSIMS and eSIS,

and two technologies that we will focus on in this pa-

per: the Biomarker Database and the EDRN Catalog

and Archive Service (or eCAS for short). Each appli-

cation is an independent software component respon-

sible for capturing and annotating its particular type

of EDRN data and/or metadata. Periodically (or at

deﬁned intervals), metadata from each of these ap-

plications can be pushed a centralized EDRN web

portal that accepts Resource Description Framework

(RDF) (Lassila and Swick, 1999) exports of the data

and metadata captured in each application, for use in

specialized search and discovery user interfaces such

as free-text and facet-based search.

Our recent work involves developing a stan-

dard procedure for curating information in both the

Biomarker Database and the eCAS. Curation of high

quality, peer-reviewed information for each of these

applications has numerous beneﬁts within the EDRN,

including: (1) providing a means for tracking research

publications associated with discovered biomarkers

(an indicator of viability of promising biomarkers

identiﬁed by EDRN investigators); (2) for track-

ing biomarker validation (EDRN has a ﬁve phase

biomarker validation process ranging from early in-

vestigation to clinically validated results); and (3) for

sharing and disseminating both raw and processed

science data of use to researchers in the biomedical

community.

In this paper, we discuss our experience in devel-

oping a data model and curation approach and asso-

ciated software infrastructure for curating biomarker

information from the Biomarker Database and eCAS.

Though in its nascent stages, the curation approach

and software are proving worthwhile as we have been

successful in capturing publication, study, organ and

sensitivity/speciﬁcity information for a number of

promising biomarkers in the EDRN, and also for a

number of popular science data sets.

The rest of this paper is organized as follows. Sec-

tion 2 discusses background and related work in the

areas of bioinformatics grid infrastructures and data

curation for biomedicine. Section 3 presents our data

model, approach, and software infrastructure for cu-

rating biomarker research information, and important

cancer science data sets. Section 4 describes the use

of our curation system and approach in the context of

real EDRN use cases. Section 5 rounds out paper by

pointing the reader to future work and conclusions.

2 RELATED WORK

In this section we discuss relevant related work in the

areas of large-scale bioinformatics grid infrastructure

and data curation for cancer research, highlighting

similar projects and underpinning the unique contri-

bution of our own work – the assertion that, unlike

related projects, data curation has been made a ﬁrst-

class process within our work within the context of

curating biomarker information in the EDRN.

2.1 Bioinformatics Grid Infrastructure

There are several related projects building bioinfor-

matics grid infrastructure. We discuss a representative

subset of them here.

2.1.1 BIRN

The Biomedical Informatics Research Network

(BIRN) is a comparable initiative being undertaken by

the National Center for Research Resources (NCRR)

and the National Institutes of Health (NIH) for the

purpose of sharing neuroscience and brain imaging

data between collaborating scientists (Keator et al.,

2008). Like the EDRN, BIRN participants conduct

research at their own facilities and retain control over

the data they generate. While clearly a successful

project, BIRN allows scientists to submit data via the

BIRN portal, according to a speciﬁcation documented

in an associated PDF ﬁle (BIR, 2008). Our experience

indicates that allowing scientists to curate their own

data, unmoderated, can at times introduce data qual-

ity and overlap issues that can be difﬁcult to resolve.

Within EDRN, we are focused on curation by highly

trained informatics personnel, with science expertise

in a particular organ-related cancer.

2.1.2 caBIG

The Cancer Biomedical Informatics Grid (caBIG)

(von Eschenbach and Buetow, 2006) is a large, ini-

tiative funded by the U.S. National Cancer Institute

(NCI) whose mission includes building out an all-

inclusive bioinformatics grid infrastructure for shar-

ing data among the whole of the cancer research com-

munity. Like BIRN, caBIG embraces the idea that its

participating sites maintain control of the data they

contribute.

A primary difference between the EDRN and

caBIG initiatives is their respective scope. Rather

than endeavoring towards a broad-reaching connec-

tive web for cancer research as a whole, the EDRN

has taken a more reﬁned approach, focusing narrowly

in on the speciﬁc issues facing researchers involved

HEALTHINF 2009 - International Conference on Health Informatics

388

in the early detection of cancer. We have recently in-

tegrated EDRN with a caBIG enabled specimen cu-

ration tool, caTissue, which we will discuss in detail

below.

2.2 Curation of Biomarker Information

While there are several tools and efforts already un-

derway to capture biomarker related medical data

(e.g., see (Keerthi et al., 2002; Baral et al., 2005)),

due to space limitations we will focus on two re-

lated approaches for curating specimen information

called caTissue and the National Biospecimen Net-

work (NBN), respectively.

2.2.1 caTissue

caTissue (caT, 2008), is a specimen capture system

developed at the NCI. Focusing largely on gene ex-

pression and sequence data as it relates to cancer re-

search, the system operates as a ’plug-in’ to caBIG.

Patient tissue sample is combined with metadata an-

notations to form a comprehensive specimen bank

that can be queried by a caBIG user.

It should be noted that ERNE has successfully

plugged into the caTissue suite, and can thus EDRN

can integrate with caBIG grids.

2.3 NBN

The National Biospecimen Network (NBN) (Birm-

ingham, 2004) is an NCI-funded initiative arising out

of a general consensus during the March 2002 Na-

tional Dialogue on Cancer that access to annotated tis-

sue collections was a critical enabling factor in the ap-

plication of recent technological advances to the ﬁght

against cancer. NBN has faced a similar need for a

framework in which to provide meaningful annota-

tions of the tissue samples. While the scope of the

NBN is much more broadly deﬁned than the EDRN,

there is some degree of overlap with regards to data

curation issues.

3 CURATION OF BIOMARKER

DATA AND RAW SCIENCE

RESULTS

While it is undeniable that a robust, extensible infras-

tructure is an integral part of a complete informatics

solution, we feel that the issues surrounding the cu-

ration of that data need to be placed on at least equal

footing. The specialized nature of the data generated

by cancer biomarker research, combined with issues

of coordination and control that arise from the inher-

ent geographical distribution of the work being done,

place an emphasis on the importance of data curation.

In this section we will discuss our approach to build-

ing data models and curation tools for the BMDB and

eCAS.

3.1 Biomarker Database Data Model

The raw input data for the Biomarker Database is pro-

vided by various distributed applications and services.

Interoperability with these data sources depends on

the collective ability of all participants to ‘speak the

same language.’ To that end, the Biomarker Database

relies heavily on a common information model, the

EDRN Ontology, used throughout EDRN informat-

ics. The Ontology provides all applications with

common reference for naming conventions using the

EDRN registry of Common Data Elements (CDE), as

well as relationships that add meaning between the

various biomarker research data objects. The EDRN

Ontology has been an organizing presence at every

step in the development of the BMDB, from the de-

sign of the initial ingestion and storage routines, to

speciﬁcations for the export of the curated data. Most

critically, however, the Ontology provides a formal

foundation for the design of the BMDB data model.

The EDRN Ontology has been implemented using

Stanford’s Prot

e toolkit (Noy et al., 2000)

Because the data is so highly specialized, the data

model for the BMDB needs to be expressive and ﬂex-

ible, providing a curator with the ability to indicate

associations between data objects with as much lati-

tude as possible. Research on a particular biomarker

is comprised of many individual elements, many of

which may be related on sometimes overlapping lev-

els. As an example, a given study may make a passing

reference to a biomarker as a member of a panel of

biomarkers. Alternatively, a study might consider the

biomarker’s relationship to a particular organ-site in

great detail. Each of these studies may have relevant

publications, external resources, and other important

relationships to the biomarker that need to be curated

in order to truly capture a comprehensive representa-

tion of the research.

The challenge has been to develop a data model

ﬂexible enough to handle these nuances, yet reﬁned

and focused enough to retain the semantic meaning

of the underlying information. Figure 1a indicates

the quantity of supporting data that can be associated

with a biomarker by a curator. It is important to note

that a given type of data (a publication, for example)

can be associated at a number of locations, reﬂecting

AN EXTENSIBLE BIOMARKER CURATION APPROACH AND SOFTWARE INFRASTRUCTURE FOR THE

EARLY DETECTION OF CANCER

389

Biomarker

Biomarker-Study

Data

Study

Publication

Resource

Biomarker-Organ

Data

Biomarker-Organ-Study

Data

Study

Publication

Resource

Publication

Resource

Publication

Resource

Publication

Resource

Publication

Resource

CAS File

EDRN

Data Set

SELDI

product id

filename

product type

file location

...

study

specimen

...

productType elements

type-element

map

instance of

(a) (b)

Figure 1: (a) Biomarker Database data model diagram and (b) example eCAS data model.

Biomarker

Database

EDRN Public

Portal

DMCC Protocol

Database

PubMed

Publication

Repository

EDRN P.I.s

Peer-Review Board

Curator

Public

RDF

XML

HTTP

RDF

HTTP

users: read only

curator: read & write

eCAS system

high fidelity data

ingestion

meta data

(a) (b)

Figure 2: (a) Biomarker Database process ﬂow and (b) eCAS data curation ﬂow.

the ﬂexibility a curator has in indicating data relation-

ships. Likewise, it is also possible to link a particular

study with a biomarker in meaningful ways: (1) re-

lated directly to the biomarker without mentioning a

speciﬁc organ (or, alternatively, mentioning multiple

organs); or (2) as speciﬁcally relating the study to a

given biomarker-organ pair. We have found that pro-

viding such a meaningful variety of curation options

greatly enhances the utility of the curated data.

3.2 Biomarker Database Curation

Process

The Biomarker Database curation process is realized

through a combination of efforts of one or more cura-

tors with a background in both informatics and biol-

ogy, and supporting software tools (as shown in Fig-

ure 2a). Active communication between researchers

and a dedicated curator serves to mitigate the difﬁcul-

ties posed by the ad-hoc submission of research data

from multiple independent sites. The curator acts as

a liaison between the research scientists and the sys-

tem, providing an additional layer of assurance on the

quality of the data.

Once a biomarker has been selected for curation,

the curator works with the research team to identify

the relevant details (including EDRN studies, sensi-

tivity, speciﬁcity, and predictive value data, publica-

tions and resources) that would contribute to a greater

understanding of the current research state. Using a

streamlined, web-based tool, a curator is able to indi-

cate associations between the data, directly importing

and linking resources from across the EDRN enter-

prise as needed to create a uniﬁed view of the research

that is ready for the peer review process.

3.3 eCAS Data Model

Whereas the Biomarker Database data model has

been designed to speciﬁcally store research data

related to the discovery and validation of cancer

biomarkers, the eCAS data model is generalizable to

all scientiﬁc research communities. The model it-

self places no restrictions on the type of data being

processed, and follows an inheritance pattern which

allows for unlimited extensibility and specialization.

The core meta-data representation consists of three

HEALTHINF 2009 - International Conference on Health Informatics

390

basic building blocks: (1) elements – basic key/pair

nucleus; (2) productType – a composite set of ele-

ments; and (3) typeElementMap – mapping represen-

tation between a product type deﬁnition and its corre-

sponding elements list.

All meta-data representations are deﬁned using

these three core data structures. More complex data

deﬁnitions speciﬁc to the particular science domain

inherit and extend their properties from the base-

line deﬁnition. This built-in extensibility allows an

eCAS instance to be tailored to suit even very com-

plex data warehousing scenarios. Figure 1b shows

how this hiearchical organization of the eCAS data

model is extended to a speciﬁc implementation of

e.g., a Surface-enhanced laser desorption/ionization

(SELDI) dataset deﬁnition. It is possible to trace the

inheritance graph by starting with the casFile pol-

icy entity, a generic meta-data deﬁnition, and travers-

ing backwards, each entity expanding and specializ-

ing the previous one, until arriving at a policy speciﬁc

to the SELDI data set.

3.4 eCAS Curation Process

As mentioned earlier, the eCAS data model is de-

signed to be domain agnostic. As a consequence, the

web-based curation interface, modeled in Figure 2b,

does not speak to any speciﬁc data ﬂow and its users’

roles and delegation of authority are more generic.

While the curation requirements of a particular do-

main may vary to some extent, all curation efforts in-

volve, at a minimum, (1) high-ﬁdelity data ingestion,

and (2) meta-data manipulation.

Addressing the ﬁrst concern, eCAS provides an

interface to support data ingestion. Because the eCAS

curation application is built on top of the underlying

OODT CAS layer (Mattmann et al., 2008), it is able

to support voluminous and high ﬁdelity data inges-

tion. For the second concern, meta-data manipulation,

eCAS provides a web-based interface (similar to the

BMDB curation web interface described earlier) that

presents curators with an efﬁcient method for entering

and editing meta-data. The tool provides an option to

update existing data (and meta-data) or initiate new

product type entry from scratch using a “wizard” in-

terface.

4 APPLICATION WITHIN EDRN

4.1 An Emphasis on Data Quality

The efforts of the EDRN center around the collec-

tion, storage, annotation and presentation of cancer

biomarker research. The BMDB and eCAS facili-

tate these efforts by providing a complete data storage

and curation infrastructure. As EDRN aims to pro-

vide authoritative, comprehensive coverage of can-

cer biomarker discovery and validation research, the

quality of the data is a paramount concern.

As discussed in Section 3, the majority of the de-

velopment effort on the Biomarker Database has been

directed towards working with the NCI, EDRN Prin-

cipal Investigators, and curators to develop a ﬂexible

web-based curation interface to support curation ef-

forts. The collaborative effort has at times unearthed

signiﬁcant model differences among the participants

which, if left undiscovered, might lead to conﬂicting

interpretations and ambiguities and threaten to under-

mine the value of the data.

The Biomarker Database has provided us with a

sandbox of sorts in which to iteratively reﬁne our data

model until a consensus was reached. At present, we

have approximately 15 highly curated biomarkers and

11 data sets (including 1000s of data ﬁles and meta-

data ﬁles) representing a range of sub-disiplines in-

cluding liver, lung and ovary. Each of the biomark-

ers is meaningfully linked to studies, organ-site data,

publications, and external resources and thus forms a

largely complete picture of the existing research.

Curation is an essential ingredient in our efforts to

construct the foundations of a comprehensive knowl-

ege environment of value to the scientiﬁc community.

The presence of a knowledgeable curator capable of

acting as liaison and proverbial “trafﬁc-cop”, has been

pivotal in maintaining the integrity of the captured

data.

4.2 Discussion

Curation is the linchpin in a process that extends from

initial ingestion of data from various external sources,

including the EDRN Data Management and Coor-

dinating Center (DMCC) Protocol Database at the

Fred Hutchinson Cancer Research Center (FHCRC)

in Seattle, and publication repositories like PubMed

(both shown in the upper left portion of Fig. 1a) to

the eventual release of peer-reviewed biomarker re-

search through the EDRN Public Portal (shown in the

middle right portion of Fig. 1a). A similar process

is shown for the eCAS as depicted in Fig. 1b. The

Biomarker Database and the eCAS system can per-

haps be thought of as data reﬁneries, taking in vol-

umes of raw, uncorrelated data. The resulting curated,

peer-reviewed data that is ultimately made available

for public consumption (shown in the lower right por-

tion of Fig. 1a and upper left version of Fig. 1b) is of

high quality (as it has been peer-reviewed) and highly

AN EXTENSIBLE BIOMARKER CURATION APPROACH AND SOFTWARE INFRASTRUCTURE FOR THE

EARLY DETECTION OF CANCER

391

valuable (as it has been curated) as an authoritative

information resource.

The key factor distinguishing our work from that

of others building bioinformatics grids is that many

of the other bioinformatics grid efforts are pursuing

technology research and have, we believe, not given

curation sufﬁcient prominence, even as the data man-

agement problems in science have continued to grow.

Our key lesson learned is that scientists need to be in-

volved in the planning and curation process and the

bioinformatics grid software needs to be able to grow

and evolve in unison with the data model during cu-

ration activities.

5 CONCLUSIONS AND FUTURE

WORK

In this paper we discussed our efforts to deﬁne a cura-

tion process for biomarker information collected with

the National Cancer Institute (NCI)’s Early Detection

Research Network (EDRN) project. Biomarker re-

search data is curated and stored within two appli-

cations running on top of EDRN’s data grid infras-

tructure: (1) the Biomarker Database (BMDB), and

(2) the EDRN Catalog and Archive Service (eCAS).

We described the data model and curation process for

each of these applications and described real EDRN

use cases for each application. We have experi-

enced ﬁrsthand some of the difﬁculties in transform-

ing raw research data from geographically diverse

sources into a comprehensive query-driven knowlege-

base. These difﬁculties reinforce the notion that (1)

data models must be developed with evolvability as

a cornerstone; (2) scientists need to be actively in-

volved in the model development process to the great-

est extent possible; and (3) data management (cura-

tion) policy development is at least as important to

address as decisions about the underlying technology

infrastructure.

ACKNOWLEDGEMENTS

This effort was supported by the Jet Propulsion Lab-

oratory, managed by the California Institute of Tech-

nology under a contract with the National Aeronautics

and Space Administration. The authors would like to

thank Donald Johnsey, Christos Patriotis, and Sudhir

Srivastava and the NCI leadership as a whole for their

collaborative guidance and support.

REFERENCES

(2008). Birn - describing your data,

http://nbirn.net/bdr/study information.shtm.

(2008). catissue core,

https://cabig.nci.nih.gov/tools/catissuecore.

Baral, C., Davulcu, H., Nakamura, M., Singh, P., Tari, L.,

and Yu, L. (2005). Collaborative curation of data from

bio-medical texts and abstracts and its integration. In

Data Integration in the Life Sciences, pages 309–312.

Birmingham, K. (2004). An inauspicious start for the

us national biospecimen network. J. Clin. Invest.,

113(3):320–320.

Crichton, D., Kelly, S., Mattmann, C., Xiao, Q., Hughes,

J. S., Oh, J., Thornquist, M., Johnsey, D., Srivastava,

S., Essermann, L., and Bigbee, W. (2006). A dis-

tributed information services architecture to support

biomarker discovery in early detection of cancer. In

e-Science, page 44.

Foster, I., Kesselman, C., and Tuecke, S. (2001). The

anatomy of the grid: Enabling scalable virtual organi-

zations. J. Supercomputing Applications., pages 1–25.

Howe, D., Costanzo, M., Fey, P., Gojobori, T., Hannick, L.,

Hide, W., Hill, D. P., Kania, R., Schaeffer, M., Pieer,

S. S., Twigger, S., White, O., and Rhee, S. Y. (2008).

Big data: The future of biocuration. Nature, 455:47–

50.

Keator, D., Grethe, J., Marcus, D., Ozyurt, B., Gadde,

S., Murphy, S., Pieper, S., Greve, D., Notestine, R.,

Bockholt, H., and Papadopoulos, P. (2008). A na-

tional human neuroimaging collaboratory enabled by

the biomedical informatics research network (birn).

IEEE Trans. Information Technology in Biomedicine,

12(2):162–172.

Keerthi, S. S., Ong, C. J., Siah, K. B., Lim, D. B. L., Chu,

W., Shi, M., Edwin, D. S., Menon, R., Shen, L., Lim,

J. Y. K., and Loh, H. T. (2002). A machine learn-

ing approach for the curation of biomedical litera-

ture: Kdd cup 2002 (task 1). SIGKDD Explor. Newsl.,

4(2):93–94.

Lassila, O. and Swick, R. (1999). Resource description

framework (rdf) model and syntax speciﬁcation. Tech-

nical report, W3C.

Lynch, C. (2008). Big data: How do your data grow? Na-

ture, 455:28–29.

Mattmann, C., Freeborn, D., Crichton, D., Hughes, J. S.,

Ramirez, P., Hardman, S., Woollard, D., and Kelly, S.

(2008). Transformation of oodt cas to perform larger

tasks. NASA Tech Briefs., 32(6):44.

Noy, N. F., Fergerson, R. W., and Musen, M. A. (2000).

The knowledge model of protege-2000: Combining

interoperability and ﬂexibility. In Knowledge Engi-

neering and Knowledge Management Methods, Mod-

els and Tools, pages 69–82.

von Eschenbach, A. C. and Buetow, K. (2006). Cancer in-

formatics vision: cabig. Cancer Informatics, 2:22–24.

HEALTHINF 2009 - International Conference on Health Informatics

392