06ni 1_1-burnsf - infolab - university of southern california

AbstractWithin this paper, we describe a neuroin-

formatics project (called “NeuroScholar,”http://www.neuroscholar.org/) that enablesresearchers to examine, manage, manipulate,and use the information contained within thepublished neuroscientific literature. The pro-ject is built within a multi-level, multi-compo-nent framework constructed with the use ofsoftware engineering methods that themselvesprovide code-building functionality for neu-roinformaticians. We describe the differentsoftware layers of the system. First, we presenta hypothetical usage scenario illustrating howNeuroScholar permits users to address large-scale questions in a way that would otherwisebe impossible. We do this by applyingNeuroScholar to a “real-world” neurosciencequestion: How is stress-related information

processed in the brain? We then explain howthe overall design of NeuroScholar enables thesystem to work and illustrate different compo-nents of the user interface. We then describethe knowledge management strategy we use tostore interpretations. Finally, we describe thesoftware engineering framework we havedevised (called the “View-Primitive-DataModel framework,” [VPDMf]) to provide anopen-source, accelerated software develop-ment environment for the project. We believethat NeuroScholar will be useful to experimen-tal neuroscientists by helping them interactwith the primary neuroscientific literature in ameaningful way, and to neuroinformaticiansby providing them with useful, affordable soft-ware engineering tools.Index Entries: Neuroinformatics; literature;knowledge models; bibliographic; database.

81

NeuroinformaticsCopyright © Humana Press Inc.All rights of any nature whatsoever are reserved.ISSN 1539-2791/03/01:81–110/$20.00

Original Article

Tools and Approaches for the Construction of KnowledgeModels from the Neuroscientific LiteratureGully A. P. C. Burns,*,1 Arshad M. Khan,1 Shahram Ghandeharizadeh,2 Mark A. O’Neill,3

and Yi-Shin Chen2

1 The K-Mechanics Research Group, 3641 Watt Way, Hedco Neuroscience Building, University ofSouthern California, Los Angeles, CA 90089-2520; 2 Henry Salvatori Computer Science Center, Universityof Southern California, Los Angeles, CA 90089-0781; 3 The Bee Systematics and Biology Unit, OxfordUniversity Museum of Natural History, Parks Road, Oxford, UK

*Author to whom correspondence and reprint requests should be sent. E-mail: [email protected].

82_________________________________________________________________________________ Burns et al.

Neuroinformatics _________________________________________________________________ Volume 1, 2003

“Even the most active neuroscientist spendsmore working hours in reading, reviewing andwriting scientific reports than on direct experi-mental effort”

Floyd Bloom, 1978,Trends in Neuroscience 1(1):1

“There have been few changes to the traditionalmethods of neuroscientific information gathering,sharing and analyzing: namely reading researchjournals and traveling to scientific meetings”

Floyd Bloom, 1995,Trends in Neuroscience 18(2):48-49

IntroductionA prominent and compelling justification

for the development of neuroinformatics-based approaches is that our subject isextremely complex. There are currently over50,000 working neuroscientists in the worldproducing enough data to fill more than 300journals in a wide variety of subdisciplinesranging from psychology, linguistics, and ani-mal behavior to neuroanatomy, electrophysi-ology and molecular biology (Chicurel, 2000).

A widely shared perspective within thefield is that the most effective approach toinformation management is to build a large-scale collaborative network of seamlessly inte-grated repositories of raw data (Koslow, 2000).These repositories may be linked to the pri-mary literature using modern web-based pub-lishing technology to give the data structureand form. This approach has formed the basisof collaborative projects such as CERN (theEuropean Laboratory for Particle Physics nearGeneva), the Stanford Linear Accelerator(SLAC), and the Human Genome Project andhas proven itself to be very powerful withinother disciplines. However in neuroscience,this approach is hindered by the lack of con-sensus around the data’s theoretical structure.

In some regards, the primary scientific liter-ature forms the basis for all human under-

standing of a subject; it is the crucible wherescientific discoveries are validated, tested,confirmed, or rejected. We assert that the the-oretical structure embodying the subject is anemergent property of the observations, inter-pretations, arguments, or hypotheses con-tained within the literature’s constitutive publications (Burns, 2001a). Furthermore, theliterature’s large size and scope, lack of stan-dardization, variable quality, and largely lin-guistic (i.e., qualitative, nonmathematical)nature mean that these emergent theories areoften computationally unwieldy.

It is for this reason that we directly focus onrepresenting and analyzing the contents of theliterature with knowledge management tech-niques. If we consider “data” to beunattached, unstructured values; pieces of“information” are then data with additionalstructure, and explanation; and “knowledge”would be defined as information that is con-sidered in the context of other information(Blum, 1986). While technology has developedto accelerate the delivery of published infor-mation to the modern scholar, few, if any, toolsexist to expand our understanding of it. Thetransition from paper to electronic publishingextends the structure of journal articles byproviding not only access to raw data, but alsoto embedded dynamic data viewers, and evencomputational modeling tools for readers toreally explore the data underpinning a publi-cation. The functionality of NeuroScholar con-trasts with this by permitting users to create mod-els of their own knowledge across large numbers ofsuch papers (whilst providing access to otherknowledge models from other users). We expectthat the emergent properties of such a systemwill provide a powerful new approach forgenerating neuroscientific theories.

Currently, the function of the NeuroScholarsystem is to address the following question:“What is the complete neuroanatomical cir-cuitry underlying a specific physiological phe-nomenon?” In order to address this question

Volume 1, 2003 _________________________________________________________________ Neuroinformatics

Models of Neuroscientific Knowledge ____________________________________________________________83

effectively, we first distinguish the phe-nomenon of interest (say for example, therelease of a hormone in response to stress),identify which regions of brain tissue areinvolved in this phenomenon, and finally,study all the neuroanatomical connectionslinking these regions. In principle, theNeuroScholar system may be used for anyspecies, as long as a complementary electron-ic neuroanatomical atlas is available for use bythe system. At present, we only support datawith a neuroanatomical atlas of the rat(Swanson, 1998).

NeuroScholar’s utility can be emphasizedby considering the size of the task of buildinga useful representation of a large literature.Since our stated example focuses on the stressresponse and Corticotropin ReleasingHormone (CRH; see Table 1), we performedsome broad searches on the National Library

of Medicine’s PubMed website (accessiblefrom http://www.ncbi.nlm.nih.gov/) to esti-mate the size of the literature from the follow-ing keyword combinations: “CRH,” “par-aventricular,” “CRH and paraventricular,”“CRH and stress,” and “ACTH”(see Table 1).As shown in Figure 1, the number of publica-tions conforming to the search ranged fromalmost one thousand (for the keywords “CRHand paraventricular”) to over 30,000 (for thekeyword “ACTH”). Clearly, these globalsearches are prohibitively large for an individ-ual scientist to manipulate. In most cases, the publication rate on the specified subject isincreasing. As described in the next section, itis impossible for an unaided worker toaddress questions across the whole literature.This only becomes possible within the type of neuroinformatics framework we describehere.

Fig. 1. Graph showing the normalized number of hits returned from searches of the PubMed literaturedatabase using specified search terms. [CFO]

84_________________________________________________________________________________ Burns et al.


Within this article, we describe a set ofmodular neuroinformatics tools that combineto form a prototypical application (calledNeuroScholar) that is designed to act as aknowledge management system for the neu-roscientific literature. We have previouslydescribed NeuroScholar’s underlying strategyand theoretical basis (Burns, 2001a), and itssystem’s design in detail (Burns, 2001b).

We place potential users of a knowledgemanagement system of the literature on a con-tinuum. At one end, experimental specialistsfocus on their own personal perspective of theliterature. At the other, neuroinformaticiansmay concentrate on using individual compo-nents of our system to strengthen their ownsoftware development work. Our frameworksupports the entire continuum of users. In thispaper, we begin by describing the high-levelfunctionality of the overall system, and thenprovide examples of the specific modulartools that we are implementing. NeuroScholarprovides a suite of tools for experimental spe-cialists to build and use models of their ownknowledge, and for neuroinformaticians to uti-lize the functionality of our system withintheir own. We also provide software engineer-ing tools for neuroinformaticians that assisttheir development work.

The Utility of NeuroScholar for theExperimentalist: A HypotheticalExample Concerning the Study ofthe Stress ResponseLarge-scale questions such as “How is

stress-related information processed in thebrain?” are very difficult to answer. As notedearlier, the volume and complexity of theinformation is too large for an individual toprocess effectively. Here, we attempt to pro-vide a “real-world” example of howNeuroScholar can aid the experimental spe-cialist in addressing such issues, using thequestion posed above as our example.

We first define the problem by describingsome questions currently being asked about it.We then introduce how NeuroScholar helpsaddress these issues. One point to rememberis that NeuroScholar is not providing solu-tions to the questions being asked per se, butis providing useful tools to help the experi-mental research community answer thesequestions. The promise of this work is thatindividual users will be able to collectivelybuild large-scale knowledge models of, forexample, the complete neural circuit of thesystem involved in the stress response. In thisway, the utility of NeuroScholar is that it helpsexperimentalists answer questions by helping themfirst ask the questions in a meaningful way.

Defining the QuestionWe introduce the problem with neuroscien-

tific definitions from within this specializedfield of study. We then examine some large-scale questions currently being asked byexperimental specialists within the field andhow NeuroScholar aids them in this process.

What is stress? It is important to first define what we mean

by stress (which is not, in itself, universallyagreed upon within the field). Many physiolo-gists instead try to understand stress by exam-ining the stimuli that elicit stress (“stressors”;Selye, 1974). Stressors may include those thatdisturb the homeostatic mechanisms of thebody (e.g., dehydration, infection, hemor-rhage), as well as those that threaten an indi-vidual’s state in less clear-cut ways (e.g.,restraint, footshock).

The hypothalamic paraventricular nucleus(PVH) is considered to be the final outputpathway in the stress response. The PVH isdefined as a pair of densely packed wing-shaped nerve cell clusters occupying a small,dorsomedially located volume of the hypotha-lamus. This is a critical staging area for inte-grated, adaptive responses to stress (Swanson,



Table 1—A Guide to Terminology and Abbreviations Used Within This Paper

Abbreviation / Terminology Definition

‘Data’ Unstructured measurements‘Information’ Data with structure and meaning‘Knowledge’ Information in the context of other information‘Knowledge Model’ A computational representation of an individual user’s perspective of

information from the literature defined in context of either informa-tion from another source or from fragments in the literature

‘Fragments’ Individual excerpts of data, taken from published sources (i.e., journalarticles, books, databases, etc).

‘Ontology’ “An explicit formal specification of how to represent the objects, con-cepts, and other entities that are assumed to exist in some area ofinterest and the relationships that hold among them” (Howe 2001)

‘Experimental Specialist’ End-users of the NeuroScholar system who are principally concernedwith using the system to build knowledge models

‘Neuroinformatician’ Developers within the field of Neuroinformatics who may want touse the underlying infrastructure of the VPDMf and the KMC withintheir own systems

KMC The general ‘Knowledge Management Core’ infrastructure uponwhich the NeuroScholar system is built

VPDMf A software engineering paradigm called the ‘View Primitive DataModel framework’ that permits forward and reverse engineeringbased on the Unified Modeling Language (the ‘UML’)

View A hybrid composite data object defined within the VPDMf from oneor more linked classes within a data model

View Instance Instance data contained within the structure of a named ViewView Specification A description that names which classes from a data model from a

View and how they are associated (used within the VPDMf)View Graph Definition A graph-based representation of the interactions between several

views derived from the same data model within the VPDMfView Graph Instances A graph-based representation of interrelated View InstancesUML The Universal Modeling Language, an widely-used object-oriented

design methodologyXML The eXtensible Markup Language.ORT The Objective Relational Transformation methodology for translating

data between parcellation schemes (used in the CoCoMac system).PVH The paraventricular nucleus of the hypothalamus PVHmpd The medial parvocellular division of the paraventricular nucleus of

the hypothalamus‘HPA Axis’ The pathway of activation from cells in the PVH, to cells in the anteri-

or pituitary gland to cells in the adrenal cortex cells implicated in thestress response

Stressor Stimuli that elicit stressCRH Corticotropin releasing hormoneCORT CorticosteroneACTH Adrenocorticotropic hormone

86_________________________________________________________________________________ Burns et al.


1986; Sawchenko and Swanson, 1989;Swanson, 1991; Herman and Cullinan, 1997;Sawchenko et al., 2000). Activation of thisregion ultimately results in the release of pitu-itary and adrenal hormones in the blood-stream, which can then exert a variety of effectson both central and peripheral target tissues.

The PVH consists of many distinct sub-groups of cells (Swanson, 1991), only one ofwhich will be discussed here. Specifically,neurons within the medial parvocellular divi-sion of the PVH (PVHmpd) respond to stress-related inputs by synthesizing the hormonecorticotropin-releasing hormone (CRH) andreleasing it into the bloodstream. Oncereleased, CRH activates the anterior pituitarygland, causing it to release the adrenocorti-cotropic hormone (ACTH), which in turn,travels down to activate the adrenal cortex,causing it to release glucocorticoids, such ascorticosterone (CORT). CORT then acts uponmultiple tissues both centrally and peripheral-ly to mobilize the body’s energy stores inresponse to stress.

This pathway of activation, from PVH cellsto anterior pituitary cells to adrenal cortexcells, is an example of the “hypothalamo-pitu-itary-adrenal” axis (de Groot and Harris,1950). The output of the HPA axis, for examplethe CRH-ACTH-CORT response, is a majorindicator that physiological responses to stresshave been triggered. It is no surprise then, thatthe synthesis and release of CRH, ACTH, andCORT, in the HPA axis are under exquisitecontrol at a number of levels (e.g., Axelrodand Reisine 1984; Dallman et al., 1987; Wattsand Swanson, 1989; Tanimura et al., 1998;Tanimura and Watts, 1998; Sapolsky et al.,2000). Activation of the HPA axis is generallyconsidered to be the hallmark of the stressresponse, with the PVH serving as “the finalcommon pathway for all types of stressresponse mediated by the central nervous sys-tem” (Swanson, 2000).

From this, we emphasize the followingpoints:

a) It is useful to study stress in terms of its phys-iological causes (stressors).

b) A stressor can preferentially activate certainbrain regions to produce a characteristic “pat-tern of activation.”

c) The PVH is a critical brain region involved inthe brain’s response to stress and is composedof many subgroups of cells.

d) One PVH subgroup, the PVHmpd, respondsto stress-related signals by synthesizing CRH.CRH can then trigger a cascade of hormonerelease, first involving ACTH from the pitu-itary, and then CORT from the adrenal cortex.

e) The activation of PVH cell groups (such as thePVHmpd), which often results in the produc-tion and release of hormones in the blood-stream, is a hallmark of the brain’s response tostress.

We now address some likely questionsposed by experimental specialists.

Questions of Interest for ExperimentalSpecialists

How does the brain discriminate betweenstressors? As noted earlier, it is generallythought that part of the way that the brain candifferentiate between different types of stres-sor is from the selective activation of the neu-ral circuitry that processes the informationencoding each type. These “patterns of activa-tion” differ for each type of stressor. A concep-tual challenge facing experimentalists is toprecisely document the regional and connec-tivity relationships for the brain subjected todifferent types of stress.

Can stressors act similarly on the brain? Arelated question is whether the patterns ofactivation caused by one stressor are alwaysmutually exclusive from the patterns elicitedby another. It is now well known that somebrain regions are activated by multiple types



of stressor, while others only respond to stres-sors of a specific type.

How does a single cell population such asthe PVH integrate the various signals encod-ing information about different stressors? Aspreviously noted, multiple stressors may act,in part, upon common sets of target areas inthe brain. For these common areas, somemechanisms must exist that allow for signalsencoding information from multiple stressorsto somehow be integrated at the level of thesingle cell or cell population. How is thisachieved? For example, PVH neurons becomeactivated in response to cold stress as well asmetabolic stress, such as hyperglycemia(excess blood glucose). How do these neuronsact when both stressors occur at the same time(i.e., when signals conveying both types ofstressor arrive at a PVH neuron essentiallysimultaneously)? Does the hormone releasethat typically forms the output response ofPVHmpd cells increase in amplitude or fre-quency in any way?

Addressing these questions requires aseamless integration of relevant data from avariety of primary literature sources within acoherent, shared structure to which multipleusers can contribute. We now describe indetail ways in which the NeuroScholar systemcan contribute to helping solve this problem.

The Usefulness of NeuroScholarKnowledge Models Concerning the Brain’s Response to Stress

Organizing the Primary LiteraturePerhaps the most logical starting point for

any experimentalist attempting to make senseof data within this subject is having a means toorganize the primary literature in a usefulway. Traditionally, experimental specialistshave used reference management programs tohelp create databases of the literature in whichthey are most interested. This method contin-ues to be a useful, albeit simple, way to orga-

nize publications. However, few tools havebeen developed that allow the user to takeonly relevant portions of the actual publishedliterature. Such “parts of papers,” or fragments(see “The Neuroscholar System as a Whole”)may include data tables, photomicrographs,electrophysiological traces, and, in the case ofpapers published electronically, even supple-mental data (including published sets of rawdata) or digital animation. It would be usefulto store such fragments within a convenientenvironment that allows the user to makesense of them.

NeuroScholar’s user interface (the compo-nents of which are described in“NeuroScholar Components”) provides suchan environment. As shown in Figure 2, theuser interface provides a basic “workspace”within which a user may link multiple frag-ments of information. Fragments from thispaper can then be linked to the document,including textual and graphical fragments.NeuroScholar contains tools to create frag-ments efficiently and store them in a database(see “Fragmenter”) for subsequent retrievaland viewing within a user’s workspace envi-ronment. Figure 3, for example, illustrateshow we may delineate Figure 4 from Rho andSwanson (1989) as a graphical fragment. Withthis tool, the user may use the fragment as thebasis for definitions and assertions within hisor her knowledge model. The utility providedby the Fragmenter is to represent portions ofthe actual contents of the papers themselvesand is clearly not commonly available withinconventional bibliographic software.

Through the use of fragments, the literaturerequired to answer the questions posed in thepreceding section can be readily broken downand organized into fragments that may thenbe associated with other information accord-ing to the preferences of the individual user.For example, electrophysiologists may collectonly fragments of data represented by stimu-

88_________________________________________________________________________________ Burns et al.


lus and recording traces (or text fragmentsdescribing such graphical data), whereasanatomists may only collect fragments depict-ing and/or textually describing structure-function relationships within stress-relatedcircuitry. This latter task is made easier byother component tools of NeuroScholar, asdescribed next.

Patterns of Physiological Intervention and Activation Related to Stress As described in “Questions of Interest for

Experimental Specialists,” one question posedby experimental specialists is how the braindiscriminates between stressors. Stressors pro-duce characteristic “patterns of activation.”Many studies have described such patterns,which involve regions both common andunique to multiple stressors. Moreover, whiletwo stressors may exert their actions upon acommon brain region, a single stressor canalso have opposite effects on two differentregions. For example, CRH synthesis in twobrain regions, the PVH and the central nucle-us of the amygdala (CeA), are markedlydecreased and increased, respectively, inresponse to elevated glucocorticoids (Watts,1996; Makino et al., 2002). Also, a single stres-sor can activate more than one subgroup of abrain region. For example, metabolic stresscan increase the activation of enzymes in mul-tiple divisions of the PVH (A.M. Khan,unpublished observations), each of whichmay be mediating different aspects of theresponse to this stressor.

The sample questions described above aretypical of an individualized enquiry conduct-ed by a single researcher and, as such, may beof little or no interest for other scientists out-side or even within the same field. It is impor-tant to state that these questions are all inter-linked, since they all involve shared concepts;namely the pattern of activity within the brainunder a specific physiological condition andthe structure of the neuroanatomical circuit

serving as a physical substrate for that activa-tion. Thus, the specialized knowledge modelsthat individual researchers use to answer spe-cific questions are based on components thatwill be naturally useful to other researchers.

NeuroScholar helps address these issues byproviding users with tools that can help themdelineate “brain volumes” according to theirown choosing and represent these volumeswith reference to a brain atlas. As shown inFigure 4, NeuroScholar’s AtlasMapper plug-in (described in detail in “The Atlas Mapper”)can allow a user to delineate an enclosed vol-ume on a template of the brain region of inter-est, obtained from an electronic atlas file.Figure 4 specifically depicts a brain volumedelineated by a user and superimposed uponmultiple subdivisions of the PVH. This, forexample, might be useful if one wishes to notea pattern of stress-induced cellular activationwithin neuronal populations that do not nec-essarily conform to the boundaries of pub-lished atlases. This method of delineation maybe used to describe lesion sites, injection sites,regions of labeling, sites of activation, or anyother data located physically within the brain.

It should also be mentioned that the litera-ture reporting such data (i.e., activation pat-terns from different stress patterns) is growingrapidly, and this method immediately placeseach newly published set of results within aframework that straightforwardly permits dif-ferent data to be compared. Tools such asthose provided by NeuroScholar will be indis-pensable for users who need to sort throughsuch large numbers of papers to begin formu-lating general hypotheses about the data at asystems level.

Integration of Stress-Related Information at the Level of the PVHmpdThe Fragmenter and AtlasMapper plug-in

can also be used to organize information con-cerning the many inputs arriving at PVH neu-roendocrine neurons mediating the final out-



put pathway of the stress response. The prin-cipal question asked in the preceding sectionin relation to these inputs was how signalsconveying stress are received and integratedby the cells within the PVHmpd. This ques-tion has received a fair amount of attention inthe literature (see reviews by Swanson, 1986,Swanson et al., 1987, Herman and Cullinan,1997, Sawchenko et al., 2000) and remains anactive area of experimental investigation.

Roughly fifty different sources of neuralinputs to the PVH have been identified usingaxonal transport tracing methods, making theafferent control of PVH function extraordinar-ily complex (Swanson, 2000). PVHmpd neu-rons are known to receive four major sourcesof neural input (reviewed by Swanson, 1986):(1) catecholaminergic inputs from the brain-stem that convey primarily viscerosensoryinformation; (2) subfornical inputs, some ofwhich contain angiotensin II as a transmitter;(3) inputs from the bed nucleus of the stria ter-minalis (BST), a limbic region that is believedto be a principal conduit of information arriv-ing to the PVH from the neocortex; and (4) avariety of inputs from the hypothalamusitself. Without computational support, mak-ing sense of the data describing these regionsand their inputs to PVH would be highly chal-lenging. Within the NeuroScholar system, itbecomes a matter of using the Fragmenter andAtlas Mapper tools to link the original figuresand text of the input regions to delineations astandard atlas and then manage the accountsof connections with this neuroanatomicalorganization within the NeuroScholar userInterface.

NeuroScholar as an Aid for DesigningExperiments for Stress-Related ResearchIn addition to providing experimental spe-

cialists with a systematic means to keep trackof the primary literature, their own experi-mental results, and the relationship betweenthese two sets of information, NeuroScholar

also provides users with a tool to help designthe experiments themselves. Experimentalresearch plans that include iterative steps, inparticular, can be readily outlined usingNeuroScholar’s experimental flowchart plug-in (see “The Experimental Flowchart” for moreelaboration on this topic; see Fig. 5).

As NeuroScholar is still under develop-ment, efforts are ongoing to tailor the compo-nents of the system to the needs of the experi-mental specialist (as well as neuroinformatician).The preceding section provided a small sam-pling of what can potentially be achievedusing the NeuroScholar system. We welcomethe input of our colleagues who come acrossthis article and find a specific need for theNeuroScholar system that has not been explic-itly described here. Indeed, the promise of thissystem ultimately rests in its operation by asmany users as possible.

Computational Implementation of the NeuroScholar System

The development of new approaches in thefield of neuroinformatics is supported by best-practices and approaches from computer sci-ence. Three key computational concepts formthe foundation for our work: the separation ofprocesses, data schemata, and data instances,modularization (within designated names-paces), and the use of sound software engi-neering practices.

If processes are the mechanisms that imple-ment a specific functionality, then the dataschemata and instances provide the contextfor that functionality. By distinguishingbetween these three components explicitly, wegreatly expand the applicability of our toolsoutside of the scope of this project. If wedesign the processes to work under data withdifferent data schemata, then the context maybe adapted for other tasks.

The systems we have built are modular andmulti-level so that a subsystem provides a

90_________________________________________________________________________________ Burns et al.


specific functionality (either as a contributingweb service or a plug-in component for theoverall application). Web services exist whereboth software and data are provided as a com-putational service implemented as remoteprocedure calls. These then may act as theconstitutive components of cooperative appli-cations that may be dispersed across a net-work such as an intranet, the Internet, or both.

In this section, we describe some of the sub-systems that combine to provide the function-ality of NeuroScholar. Within this section, wedescribe this functionality in a top-down man-ner. We begin with components ofNeuroScholar’s user interface that permitusers to manipulate specialized data types

such as fragments from papers, delineatedvolumes of brain tissue, and representationsof experimental design (see “The Neuro-Scholar System as a Whole” for the system asa whole and then “NeuroScholarComponents” for each individual subsystem).The next subheading describes how weembed these data types into a framework thatlinks scientific descriptions, interpretations,and rules to the primary literature and per-mits users to annotate and discuss the con-tents of the system (seethe “KnowledgeManagement Core”). The lowest level of thesystem is called the View-Primitive DataModel framework (VPDMf), and provides adata-management methodology to support

Fig. 2. A screen shot from the user interface of the NeuroScholar Knowledge Management system. [CFO]



Fig. 3. Screenshot illustrating the Global View Graph Instance and the Fragmenter. [CFO]

Fig. 4. Detail from the AtlasMapper plugin showing the delineation of labeled CRH cells shownin fragment screenshot in Figure 3. [CFO]

92_________________________________________________________________________________ Burns et al.


the other components (see ”The View-Primitive-Data Model Framework”). It isimportant to note that while the different sub-systems interoperate cooperatively within theframework of NeuroScholar, they may alsofunction outside of that framework as inde-pendent tools.

All software developed within this projectis open-source except where developmentwork involves third-party commercial soft-ware (from http://www.thebrain.com andhttp://www.tomsawyer.com) where we donot publish our code. This is to protect ourcommercial partners’ intellectual property. Allavailable software is documented athttp://www.neuroscholar.org/ and may beaccessed by navigating from that site tohttp://www.sourceforge.net/.

The NeuroScholar System as a WholeIn this section we consider the system as a

whole, rather than focus on an individual sub-set of the system’s design or implementation.The way the NeuroScholar system works isillustrated in Figure 6. The top tier of the dia-gram represents the heterogeneous scatteringof information that occurs in the world as“publications.” At the present time, we onlywork with online papers that are expressed inthe format known as portable document for-mat (PDF) from Adobe (http://www.adobe.com/). We provide tools to extract textual“fragments” from a paper (see “Fragmenter”)and save a pointer to the location of that frag-ment in the top level of the NeuroScholar sys-tem. Within NeuroScholar, each user has a“workspace” designated to them, where theymay construct a computational representationof their knowledge. These “knowledge mod-els” are considered the user’s intellectualproperty and may be published within thesystem, or external to the system viaNeuroScholar’s web interface. In this section,we discuss the high level applications of inter-est to experimental specialists before entering

into the low-level components of interest toneuroinformaticians.

The fragments serve as our image of the pri-mary literature and as such they support thedefinition of knowledge models by the user.Every single entity within an individual’s userspace must be supported by links to frag-ments. It is possible to link knowledge modelsto entities from other users’ spaces as well inorder to build a consolidated view, but theentities themselves must be linked to the pri-mary literature.

The symbols on the top tier of Figure 6 referto non-pdf document types and databases. Infuture iterations of the NeuroScholar system,we will be able to treat any source of data onthe web as a publication as long as we cannavigate to intelligible fragments within it andhave confidence that the fragments will per-sist in the same online location over time. Thismay include raw data from other neuro- orbioinformatics systems such as graphs,images, neuroimaging, or time-series datafiles. At present, we have deliberately restrict-ed ourselves to peer-reviewed articles toensure that the fragments being used in thesystem arise from reviewed sources. In orderto extend this mechanism to non-reviewedsources of data, we permit users to attach theirown personal “reliability score” to individualsources (based on structures from theKnowledge Management Core, see “Know-ledge Management Core”).

According to the definitions of the UnifiedModeling Language (“the UML” Rational1997), if an object is “an entity with a well-defined boundary and identity that encapsu-lates state (attribute and relationship values),and behavior (operations and methods),” thena class is a description of a set of objects thatshare the same attributes, operations, meth-ods, relationships, and semantics.

We use a uniform approach to the differenttypes of data being processed that uses andextends the basic object-oriented class struc-



ture of the UML. Each publication, fragment,entity, annotation, and rule (the differentspecies of computational item defined withinthe knowledge management framework ofNeuroScholar, see “Knowledge ManagementCore”) is considered to simply be a “View”within the data-management framework soft-ware at the system’s lowest level (see “TheView-Primitive-Data Model Framework”).Each view is essentially a composite object(similar to the concept of “materializedviews” in relational databases) made up of

combinations of interlinked classes from thesystem’s underlying data model. Each viewmay be represented as an encapsulated nodein a so-called “View Graph” where the associ-ations, overlapping relationships, and relativeenclosure of different types of items can berepresented as edges in the same graph (i.e., ifView B’s constituent classes were a subset ofthe classes of View A, then an edge from B toA would exist in the View).

A screenshot of the basic user interface ofthe NeuroScholar system is shown in Figure 2,

Fig. 5. The experimental flowchart control with the local View Graph Instance viewer fromhttp://www.theBrain.com.The flowchart refers to the first ethological study of the classical conditioning in therabbit eyeblink response (Gormezano, 1962; #606) where animals are “adapted” to their housings over twodays, they acquire the conditioned response over 8 days and the response is made “extinct” over the final 8days. [CFO]

94_________________________________________________________________________________ Burns et al.


illustrating the View Graph Definition (agraph that shows the possible relationshipsbetween Views) and a simple form generatedfor data in a “Publication View.” The defini-tions of views are related to each other byvirtue of the fact that the set of classes in oneview may be related to the set of classes inanother; they may also be connected via asso-ciation in the UML representation. These rela-tionships may be calculated to form the basisof the definition of the View Graph (see Figure7 for the underlying schema navigated in thisprocess). This system utilizes View Graphs inthree forms. Figure 2 shows a View GraphDefinition with the possible links between theUser, Publication, Graphical Fragment, andTextual Fragment Views derived from theschema shown in Figure 3. The left hand paneof Figure 3 shows the “Global View GraphInstance” which illustrates all the ViewInstances that are currently loaded and avail-able to the user. Users may reformat and rear-range this graph by hand or by using layouttools. Thus, in Figure 2, we illustrate the orga-nization of the software displaying a view.The left hand panel shows the relationshipsbetween three views defined over our simpleexample from Figures 2, 3, and 5. The viewsshown here are denoted by the User,Publication, Textual Fragment, and GraphicalFragment (each one derived from one or morelinked classes from the schema shown inFigure 7).

Within the NeuroScholar system, we usecommercial software packages to manipulateand navigate our graph representations. Thegraph-based representation shown in Figure 2uses the commercial graph drawing softwarefrom Tom Sawyer Software (http://www.tomsawyer.com/). This software provides severalfunctionality including automatic layout func-tions, subgraph representations (see “TheExperimental Flowchart”), zoom, and graphediting.

NeuroScholar ComponentsNeuroscholar’s primary goal is geared to

cater to the experimental specialist. We wish toprovide computational tools for scientists whootherwise would not consider usingapproaches from modeling or computationalneuroscience. We describe some of the userinterface methodologies that were built on topof the data model described previously(Burns, 2001b). These user-interface method-ologies deal with some of the issues that neu-roscientists are forced to consider almostevery time they read a paper: these includeselecting the excerpts of online papers thatform fragments within the system(“Fragmenter”), tools to assist the delineationof structures on a brain atlas in a way that dis-plays the users’ uncertainty concerning thedelineation (“The Atlas Server” and “TheAtlas Mapper”), and graphically describingthe design of an experiment (“TheExperimental Flowchart”). The developmentof these tools was designed to make the pro-cess of interacting with the literature as effort-less as possible for the user whilst empower-ing users to disseminate knowledge.

Fragmenter

All interpretations in this system are basedon excerpts from the primary literature itself.While building a system to represent workdescribing the neural connectivity of the rat,we discovered that copying the text of eachexcerpt into the database was the rate-deter-mining step of data entry (Burns, 1997; Burnsand Young, 2000). In this application, wesought to simplify this process based on neu-roinformatics techniques of annotating pas-sages of text (Ovsiannikov and Arbib, 2001).We simply wanted to be able to select excerptsfrom a journal article, store them in our sys-tem, and then manipulate them as we wouldany other view within the system as a whole.



Fig.

6.T

he h

igh

leve

l str

uctu

re o

f the

Neu

roSc

hola

r sy

stem

.[C

FO

]

96_________________________________________________________________________________ Burns et al.


The interface is referred to as the“Fragmenter” and is illustrated below.

The Fragmenter serves as a form for textualand graphical fragments for online docu-ments in pdf format. There are some copy-right issues that we have to circumnavigate; inmost cases we would not be permitted to storea reproduction of the text and figures of thedocument in our document without permis-

sion. We address this issue by only storing thelocation of the delineation around the relevanttext or figure on the page, the page number,and a unique index citation to the article thatpoints the user directly to the article (such asthe PMID identifier on the PubMed system).This means that users must have externalaccess to the publication from which they wishto extract fragments (as files on their local

Fig. 7.The relationship of the data model,Views, and View Graph for a simple example. [CFO]



machine, for example). If permission is givenby the publisher, users may be able to accessonline articles directly. The right hand pane inFigure 3 shows the Fragmenter in use, illus-trating a graphical fragment View Instance(Rho and Swanson, 1989).

It is hoped that the use of the Fragmenterwill have a significant impact on the way thatscientists perform their research. The act ofreferring to the original fragment from theoriginal publication necessitates a rigorousapproach to the literature to minimize misin-terpretation and miscommunication. Not onlywill users be able to track what precise frag-ments support or refute their interpretations,authors will be able to track exactly how peo-ple are interpreting their work.

The Atlas ServerWeb-based atlases are a widely used, effec-

tive web service (see http://www.mapquest.com for the current leader in this field).Other websites routinely use maps generatedby these web services as plug-ins within theirown application. In this way, we propose aweb service based on neuroanatomical atlasesfor neuroscientists. In this system, parametersare delivered to the application via aUniversal Resource Locator (URL) that speci-fies the atlas level number, the desired zoomfactor, and the coordinates of the desired viewto receive a scaled, cropped, and annotatedbitmap image of the desired region of theatlas. The atlas server is used directly by theAtlas Mapper project (see “The AtlasMapper”) to provide the images on which theAtlas Mapper’s delineations appear. It is alsopossible to search for named landmarks (i.e.,areas and nuclei) so that the label of the struc-ture in question appears in the center of view.Our current version is based on the Swansonatlas of the rat brain (Swanson, 1998) butcould use any electronic atlas that is expressedas a set of Adobe Illustrator or PDF files so

that the system could be used for any specieswith a sufficiently detailed electronic atlas.This project, like any other that involves copy-righted information, will require approvalfrom the publishing house that owns the atlas.

The Atlas MapperThe unification and standardization of neu-

roanatomical nomenclature in order to allevi-ate much of the past and present confusionsurrounding the naming of brain regions hasbeen an ongoing effort for over 100 years(Wilder, 1896; Bowden and Martin, 1995;Bowden and Dubach, 2003, this issue).Enforcing a set of standardized definitionslacks flexibility, however, and may not allowresearchers to describe neuroanatomicaldelineations in the detail they would like(Swanson, 1998). Rather than devise newtreatments for the neuroanatomical nomencla-ture (i.e., the names researchers use to denotebrain structures), our approach (called the“Atlas Mapper”) allows users to delineate vol-umes of brain tissue explicitly on a standardatlas using mapping techniques from stan-dard drawing applications. They may thensave those delineations in the knowledgemanagement system.

Figure 4 shows a typical screenshot (corre-sponding to a region from the graphical frag-ment shown in Figure 4) of the Atlas Mappercontrol in “Insert” mode.

These delineations are based on the percep-tions from users derived from images inpapers. As such, they are unlikely to be veryaccurate for several reasons (size of printedimage, plane of section of the image, etc.), andso our system permits users to estimate theerror in their delineation through the use ofFuzzy Bezier Splines (FBS). The graphical con-trols of FBS are shown in Figure 5. Essentially,the FBS handles define the extent of a “borderzone” where inclusion in the structure may bedefined probabilistically so that the central

98_________________________________________________________________________________ Burns et al.


anchor point has a 0.5 probability of beinginside the structure and each of the FuzzySpline Handles lie one standard error eitherinside or outside the structure. FBS curves aredrawn on more than one section to provide astack that delineates a volume. Within theNeuroScholar system, many such volumescomprise a map that may describe how dif-ferent regions of the brain possess differentcharacteristics (such as histological labelingpatterns generated from neuroanatomicalexperiments).

The Experimental FlowchartThe NeuroCore project seeks to provide a

“base ontology” for neuroscientific data bydefining a data model with base classes thatmay be extended for different laboratories anddata sets (Grethe et al., 2001). One section ofNeuroCore that was carefully emphasizedwas a generic table to capture the experimen-tal method for individual papers. We haveelaborated this idea by representing the work-flow of the experimental method in a paper asa modified UML activity diagram. This isshown in Figure 5 where the user is adding anactivity to the workflow (in this case, the“extinction day” procedure).

The experimental flowchart plug-in allowsusers to build descriptions of the organiza-tional flow in an experiment in a methoddescribed in detail elsewhere (Burns, 2001b).Figure 6 illustrates the organization of a semi-nal eyeblink classical conditioning study pub-lished in 1962 (Gormezano et al., 1962, alsodiscussed in Burns, 2001b) where each “activ-ity state node” in the flowchart on the right-hand side corresponds to an experimental day(each node contains subnodes that denote pro-cedures performed that day). Each edge thatpasses between nodes has a weight corre-sponding to the number of animals involvedin that transition (18 in Figure 5). When anedge connects a node to itself, its weightshows how many times that step is repeated.

This tool permits experimental procedures tobe depicted graphically in a straightforwardmanner. Within the finalized system, weintend to link activity states that are con-cerned with specific measurements to theactual data themselves. The structure of thisconnection has been described previously (seeFigure 3 in Burns, 2001b) and will appear aslinked Views within the View Graph.

The final variety of View Graphs is illus-trated in Figure 5. The example of a “LocalView Graph Instance” only shows the currentView Instance on display in the center sur-rounded by its immediate neighbors. We usecommercial software specifically designed forknowledge navigation (http://www.the-brain.com) to move between Views so thatonly the local knowledge is displayed. Byusing graph-based approaches that may begenerated automatically from data modelsand view specifications we hope that usersmay straightforwardly organize and navigatethrough large knowledge models.

Knowledge Management Core (KMC)The intentions and the design of the knowl-

edge management core have been described indetail previously (Burns, 2001a; Burns, 2001b)and will be briefly reiterated here. There arefive key capabilities that the KMC delivers.

1) Users may interact directly with the contentsof the primary literature. Rather than sup-porting an interpretation of data by a refer-ence to the paper, users may support theirideas by linking to the relevant paper, page,and passage of interest as fragments.

2) Users may build models to represent papers’data, methodology, and conclusions. Usersmay build models that accurately capture theknowledge content of papers; we providespecialized tools to accomplish this based onmalleable data models. Justification for thedefinition of these models is derived fromlinks to supporting fragments.

3) Users may build computational models of



their own knowledge using human reason-ing. Users may model their own knowledgebased on their representation of papers’ con-tents.

4) Users may argue, refute, support, and ques-tion the knowledge models of other users inthe system. Users may allow their knowledgemodels to be accessed by other users on thesystem and the KMC provides toolsets thatpermit that conversation to be constructedsurrounding the knowledge models in thesystem. Users may also query the data basedon their preferences and opinions, tracingwork that they have said they believe to bereliable.

5) Knowledge models may be aggregated andanalyzed to address the specific questionunder study. We provide tools to summarizethe contents of knowledge models into a larg-er model. We also provide data analysis tech-niques to map the organization of these datasummaries in order to provide a completedescription of the subject under study.

Essentially, the KMC is an implementationof our underlying data management system(see “The View-Primitive-Data ModelFramework”) concerned with the defininghigh-level entities and rules that are extendedfurther by the definition of the NeuroScholarsystem’s data model and toolsets within thedomain of neuroscience (see Burns, 2001b).The KMC might, in principle, be extendedinto other domains as well. This section is con-cerned with some of the technology we use atthis level to evaluate users’ opinions and tointerpret the rule sets generated within thesystem.

There are four constructs we use to recordhuman opinions within the system: Comments(where users may annotate the system’s con-tents however they would like), Justifications(where users are required to justify the defini-tion of an item in the system by linking it toanother piece of data and to explain the link),Viewpoints (where users score their confidencelevels in the attached piece of information),

and finally a Judgment (where users may selectbetween two items that have been shown tocontradict one another). These constructs areincluded to provide users with a clear way ofcapturing and reconstructing their own rea-soning rather than attempting to use compu-tational inference to automate it.

The usage of the NeuroScholar system willdepend on the opinions and preferences of itsusers. To accommodate this, we have incorpo-rated support for “soft queries,” whichemploys a fuzzy-logic based aggregation tech-nique, to permit users to retrieve customizedinformation from the system based on theirpreferences. Our proposed soft query tech-nique can further use genetic algorithms toextract users’ confidence values for differentusers in the system based on their usagebehaviors. This methodology has beenapplied in the field of e-commerce applica-tions (Chen and Shahabi, 2002).

The soft query method aggregates high-level entity data and the corresponding humanperceptions (such as confidence values toother users, authors, journal, experimentalmethods) in order to provide customizedresults that are appropriate to users’ prefer-ences (Chen and Shahabi, 2001). This methodalso allows users to consult and adopt otherusers’ opinions. In this way, we assert that “ifuser x believes user y, the concepts that satisfythe query criteria based on y’s judgments canbe retrieved for x.” For example, assume user xdoes not provide any confidence values con-cerning which methodology he prefers. If userx believes user y, who has specified confidenceweights for different methodologies, the softquery method would also take user y’s opin-ions into consideration during aggregationprocesses for user x. Within the system, a usercould assign confidence weights for many ofthe different computational entities within theNeuroScholar system (individual users,authors, specific journals, experimental meth-ods, etc.), so that when querying the system,

100________________________________________________________________________________ Burns et al.


each object in the result set will be prioritizedaccording to the weighted aggregation data.

The Portable UNIX Programming System(PUPS) uses homeostatic computational pro-cesses to run robust, adjustable computations(O’Neill and Hilgetag, 2001). PUPS has beenused to build optimized maps showing order-ing or clustering of complex data sets in neu-roinformatics (Hilgetag et al., 1996a; Hilgetaget al., 1996b). We will use PUPS in conjunctionwith users’ opinions to map the current con-tents of each user’s knowledge modelingspace dynamically, so that each time a userupdates the model, the map will accommo-date changes. We are initially focused on eval-uating each user’s account of the neural con-nectivity between the structures of interest tothem.

The PUPS system is coded in ANSI-C andwas developed under the LINUX operatingsystem using the Free Software FoundationGNU compiler tools. PUPS has been ported toa number of POSIX.1b compliant operatingsystems including OSF1, Solaris, SunOS 4.1and BSD4.4. PUPS is supported as a open-source project on Sourceforge (http://www.sourceforge.net/projects/pups.

The View-Primitive-Data Model framework (VPDMf)The VPDMf forms the base of the technolo-

gy described in this article and as such formsthe foundation of almost all the softwaredescribed here. This section is geared to theinterests of dedicated neuroinformaticians:specialists whose primary interest lies in theconstruction and evolution of computationalsystems themselves. Here, we describe thefunctionality and design of the VPDMf as a setof computer aided software engineering(CASE) tools that neuroinformaticians mayhave direct access to at zero cost. CASE toolspermit the standardization of tool sets,straightforward communication of softwaredesign, and the acceleration of code develop-

ment with forward/reverse engineeringmethods. Although common in industry,broad adoption of CASE software in academiais slow (due mainly to the high cost of theseproducts such as Rational Rose). The VPDMfdoes reproduce some of the automated meth-ods of commercial tools, but it also providesan entirely novel approach to representing asystem by superimposing a formalized frame-work over the populated data model to encap-sulate content into “Views.” Within this sec-tion, we discuss a simplified example basedon a small section of the design of the KMC(see “Knowledge Management Core (KMC)”Burns, 2001a) to illustrate how views are builtwithin the VPDMf and these definitions maybe used to build navigable models of the datamodel that support the Graph-basedapproaches shown in Figures 2, 3, and 5.

Figure 7 shows a class diagram in the UMLthat illustrates the static characteristics of sev-eral classes defined in the KMC. For example,instances of the “Publication” class are cita-tions to the literature. The “n-to-n” associationbetween the Publication and Person classessignifies that each cited paper must have “oneor more” authors, and each person may be anauthor of “one or more” publications. Eachpublication refers to the controlled vocabulary(CV) class in two attributes, the “publica-tion_type,” and the “language.” The princi-ples of object-oriented design and the use ofthe UML are well documented, and will not bedescribed here (see Rumbaugh et al., 1999).

The VPDMf provides an abstractionmethod that may be superimposed over adata model to encapsulate related classes into“Views” made up of “Primitives” spanning asmall portion of the data model. As shown inthe central section of Figure 7, every instanceof the Publication class would be linked to oneor more authors as instances of the Personclass, a CV object denoting the language of thepaper and another CV object denoting thetype of the publication. It is straightforward to


Models of Neuroscientific Knowledge ___________________________________________________________101

specify the structure of the view by (a)describing the class-composition of each prim-itive, (b) naming which primitive is consid-ered “primary” (i.e., this primitive always hasa cardinality of one and forms the core of theview definition), and finally (c) describing theassociations that link the primitives together.All other data that must be taken into accountwhen dealing with the representation (such asthe type of each attribute, the cardinality ofeach association, etc.) is explicitly described inthe data model. These three pieces of informa-tion are referred to in this article as a “ViewSpecification.”

Naturally, it is possible to define a largenumber of views by superimposing differenttiled, enclosed or overlapping definitions ontothe data model. This permits us to construct aformal methodology for navigating from viewto view by traversing these relationships. Forexample, if we defined a “Textual Fragment”view based on a single primitive made up ofthe Fragment and Textual_Fragment classes,we can calculate that there is one route in thedata model linking the two views: the 1-to-naggregation association represented shownwith a diamond at one end (the apparent over-

lap would not be considered since eachPrimitive has non-overlapping conditionalityplaced on their attribute values). By capturingthese relationships, we permit a graph-basedrepresentation for navigating between Viewscalled (somewhat unimaginatively) a “View-Graph.” This forms the basis of the graph-based data navigational systems in Figures 2,3, and 5.

The data model diagram in Figure 7 couldeasily serve as the conceptual design of a data-set in several different applications: a databasescheme, object-oriented classes in a user inter-face or a schema description of web-basedresources. In Figure 8, we illustrate the use ofsoftware engineering techniques to automatethe generation of a working system (made upof these three specific applications). The inputdata for this process consists of the specifica-tion of a data model combined with an appro-priate set of View specifications (as definedabove). Importantly, the process of communi-cating between the three components is sim-plified enormously since they all derive fromthe same design.

The schema of the physical system basedupon the conceptual design may undergo lan-

Fig. 8. Program flow underlying forward engineering in the VPDMf. (Burns, G.A. P. C., pp. 46, 7/24/2002). [CFO]

102________________________________________________________________________________ Burns et al.


guage- or system-specific design changes (forexample, adapting object-oriented models to arelational design require that n-to-n relation-ships are represented by the insertion of anintermediate link class and the addition ofattributes for primary and foreign keys,Ullman and Widom, 1997). Within theVPDMf, this transformation process is entire-ly automated.

The “view encapsulation” component is anXML-based wrapper around the database toprovide a standardized view-based web inter-face to the contents of the system. This mayprovide a way of mediating knowledgebetween systems with dissimilar data models(Burns et al., 2001). This may also provide thebasis for publishing the system as a web ser-vice since web services use XML formatters tocommunicate (see work in the W3C consor-tium concerning SOAP messages athttp://www.w3.org/TR/SOAP/).

The most powerful aspect of this frame-work is that it incorporates a model of theUML itself (similar to the reflection capabili-ties of Java, see Campione et al., 2002). Thisfeature allows the VPDMf to be superim-posed over any software that may itself berepresented by the UML. The applicability ofthis is very widespread since the UML isdesigned to be generally applicable withinthe software engineering industry (Rum-baugh et al., 1999). Data models may bedescribed in the UML or the XML Schemalanguage (Duckett et al., 2001).

DiscussionNeuroScholar strives to provide an online

environment for the comprehensive evalua-tion and interpretation of the neuroscientificliterature in order to answer specific, stated,high-level questions. This will permit neuro-scientists to expand their theoretical view ofthe subject by removing ambiguity concerningthe large-scale information sets that describecharacteristics for the wider system. The main

obstacles to designing and building this sys-tem may be distilled into three main issues: (1)Neuroscience is heterogeneous and nonstan-dardized (with regard to techniques, data,interpretations or the nomenclature); (2) theinformation in the literature is subjective andcontextual; (3) most experimental neuroscien-tists are not inclined to adopt new computa-tional methods if the methods are technicallyproblematic or unreliable. Here, we briefly dis-cuss our strategies for overcoming these obsta-cles and how they relate to existing work.

Consider that the process of “Knowledgeengineering” involves three interlinked sub-disciplines: logic, ontology, and computation(Sowa, 2000) where the term “ontology” refersto “an explicit formal specification of how torepresent the objects, concepts, and other enti-ties that are assumed to exist in some area ofinterest and the relationships that hold amongthem” (Howe, 2001). We assert that the lack ofstandardization across neuroscience’s variousdomains can really only be addressed bydefining explicit, unambiguous ontologies foreach of the domains in turn and then relate thecomponents of different ontologies to oneanother to permit interaction and translation.If users feel that it is scientifically inappropri-ate to adopt any given set of standards withina specific context, then they should be encour-aged to define their own as long as theydescribe how to map from their descriptionsto the standard set.

For example, the ambiguity within the neu-roanatomical nomenclature was mentionedbriefly in a previous section (see “The AtlasMapper”). The objective relational transfor-mation (ORT) project defines a methodologyto use set-theory to track the relationshipsbetween different brain structures and thentranslate the data embedded in those regionsbetween different parcellation schemes(Stephan et al., 2000). This method provides apractical mechanism of standardization with-out relying on individual researchers to agree



to adopt a standard reference scheme. Withinthis approach, standards will emerge overtime as the most widely used solutions and ifnew discoveries force a change of approach orterminology the system can naturally evolve.The KMC supports this functionality withinits use of rules and set theory (see “KnowledgeManagement Core” and Burns, 2001b).

An important characteristic of the neurosci-entific literature is that it is too large for oneindividual to definitively understand every-thing about a given subject when consideredfor the whole brain. Every neuroscientist’sunderstanding of the literature can be consid-ered subjective. We specifically target this con-cern within NeuroScholar by providing aunique “workspace” for users so that theymay examine the literature, build their ownknowledge models, and then use the knowl-edge models in analyses or publish them sothat other users may adopt (or refute) them.

It is the authors’ experience that when pre-sented with this idea, some researchers feelthat neuroscientists as a community would bereluctant to express their ideas in this way,preferring still to publish new theories withinthe narrative of text-based articles andreviews. We assert that a trend already existswithin publishing that will naturally evolveinto a system similar to NeuroScholar. Forexample, the Signal Transduction KnowledgeEnvironment is an online journal (http://stke.sciencemag.org/) with the following statedpurpose: “[to] maximize the efficiency withwhich the reader gathers, assimilates, andunderstands information about cell regulatoryprocesses.” The development of the Neuro-Scholar system enables neuroscientists toaddress questions that hitherto, have beenimpossible to formulate effectively.

Central to this concept is the usability of thesystem. The success of the project is complete-ly dependent on whether the system is simpleenough for noncomputational neuroscientiststo understand, beneficial to users in their

work, and reliable within its functionality on aday-to-day basis.

The origins of information science itselfderived from early attempts to systematizepublished knowledge. The pioneering work ofPaul Otlet was the driving force behind theInternational Institute of Bibliography from1895 to 1935. He constructed the “Classif-ication Décimale Universelle” (in English, the“UDC”) as an extension of the Dewey Decimalsystem. This methodology acted as “animmense map of the domains of knowledge”(Otlet, 1918; Rayward, 1998) and was used forthe institute’s efforts to catalog large volumesof bibliographic citations (the “RépertoireBibliographique Universel” or “RBU” con-tained 16 million records), graphical records(containing as many as 250,000 entries) andalso “full-text documents” (containing asmany as one million items in 10,000 subjectfiles). The UDC was based on a complex num-bering scheme that was designed to providemultiple routes of access to an individual doc-ument, and could be considered as an early,fully formed practical database managementsystem.

From these beginnings, the technology sur-rounding library-based databases has grownmassively. The National Library of Medicinehouses several web-accessible databases thatare used by the scientific community world-wide. The usage statistics for PubMed exceed-ed 329 million individual searches in 2001,which corresponds to a rate of roughly 10 persecond (National Library of Medicine, internaldocumentation). PubMed contains over 11million records, which, incredibly, is less thanthe maximum content of the RBU at its great-est extent in 1930 (Boyd-Rayward, 1998). Fulltext for an increasing number of journals isavailable online, some as an archival resourceat no cost (http://www.jstor.org). Therefore,indexes of published scholarly informationnot only constituted the earliest form of cen-tralized databases, but also form one of theessential tools of scholarly work.

104________________________________________________________________________________ Burns et al.


The Brain Browser is an atlas-enabled cross-domain computational encyclopedia that wasoriginally distributed as a MacintoshHypercard application, and was probably thefirst serious attempt to generate a neuroinfor-matics database that was based on published(or “public”) neuroscientific data (Bloom etal., 1990). This was a commercial product,designed to act as a form of computationaltextbook that could be annotated and expand-ed by the user according to individual require-ments rather than a direct interface to the pri-mary literature.

One of the necessary precursors of earlystudies of neural connectivity was the devel-opment of databases capable of storing thedata (Nicolelis et al., 1990). A large-scale colla-tion study of the hierarchical organization ofprimary visual cortex was performed withoutcomputational support by Felleman and VanEssen in 1991, which then prompted a sub-stantial research effort into the analysis andrepresentation of connectivity data for themacaque monkey (Young, 1992; Young, 1993;Young et al., 1995), the cat (Scannell et al.,1995; Scannell et al., 1999), and the rat (Burns,1997; Burns and Young, 2000). Thus by themid-to-late nineties, several independent neu-ral connectivity applications existed (with lit-tle or no interoperability between them).

These were notably followed by the devel-opment of a literature-based database for thestudy of Macaque Cortical Connectivity(called “CoCoMac”; Stephan et al., 2001). Thissystem utilizes a well-defined methodologyfor translating connectivity data between dif-ferent neuroanatomical nomenclatures usingset-theoretical rules to track inferences thatpermit data to be translated from one neu-roanatomical schema to another (Stephan etal., 2000). CoCoMac is an example of a well-populated, fully functional relational databasesystem based on information from the litera-ture that permits data exchange with othersystems via an XML-enabled web interface.

Other projects include the NeuroHomologydatabase, designed to evaluate the validity ofhomologies between brain structures in differ-ent species (Bota and Arbib, 2001).

An important feature of the earlier neuralconnectivity database work in the rat (Burns,1997; Burns and Young, 2001) was that eachrecord in the database included an abbreviat-ed copy of the original text that described thedata. This concept was identified as pivotalwithin development work into “summarydatabases” (Arbib, 2001). Within Neuro-Scholar, we define “fragments” to denote the“raw data” that forms the substrate ontowhich the interpretations of NeuroScholarmay be overlaid. The Annotator project in theUniversity of Southern California’s BrainProject was formative in the conceptualizationof the Fragmenter component of Neuro-Scholar (Ovsiannikov and Arbib, 2001).

This perspective, of superimposing an inter-pretative framework onto fragments extractedfrom any data source (as long as it is web-accessible) is unique to the NeuroScholar sys-tem, and may provide a powerful capabilityfor expansion of the system’s capabilities inthe future. Notably, this perspective dovetailswith other neuroinformatics developers whoare designing their systems so that their con-tent might serve as publications themselves(Gardner et al., 2001). Thus, NeuroScholarmight be useful as an interpretive methodolo-gy working within the conventional literatureas well as other emerging technologies fromwithin neuroinformatics.

When faced with a real-world application,systems designers parameterize and describetheir view of the world as a “data model” (or“ontology” according to the definition above),and importantly, their world-view is directlyinfluenced by the application that they arebuilding. If the designers are forwardthinkers, they will attempt to maximize theutility of their system by designing it for morethan one purpose, or they will make it con-



form to standard methods that permit sharingand reuse of the knowledge in the system(Musen, 1992). It is important to note that thesoftware development process is iterative,based on a cycle of design, implementation,and testing strategies. From a developmentviewpoint, it is highly desirable to iteratethrough this cycle as rapidly as possible. Thismeans that effective software engineeringstrategies for adapting and redesigning datamodels are potentially very important tools(Muller, 1999).

Different groups in neuroinformaticsapproach the problem of positioning theirdata models in different ways, with someworkers emphasizing markup languages inorder to facilitate communication betweensystems (Goddard et al., 2001), while othershave accomplished interoperability by defin-ing a common data model for different sys-tems, (Gardner et al., 2001; Grethe et al., 2001).Other developers have simply described thedata models of their systems diagrammatical-ly in order to describe their approach asexplicitly as possible (Stephan et al., 2001;Burns, 2001a). The KIND architecture (Guptaet al., 2000), adopts the ontology of the UnifiedMedical Language System of the NationalLibrary of Medicine (UMLS) within a FLOG-IC-based system. This approach may alsoform the basis for mediation between databas-es (Burns et al., 2001).

The Protégé 2000 project at Stanford is con-cerned with building practical modeling toolsfor knowledge sharing and reuse (Noy et al.,2000). Users may download the Protégé appli-cation to their local machine and build anontology for a specific domain. The project isopen-source and is supported by an interna-tional community of developers. The VPDMfproject and Protégé both use data modeling attheir core but have some key differences.Protégé is a frame-based knowledge manage-ment system with emphasis placed on thedevelopment of ontologies in widely dis-

persed domains. The VPDMf is principally asoftware engineering paradigm based on theUML to allow accelerated software develop-ment. Protégé supports interfaces to otherknowledge representations such asOntolingua (Gruber, 1993), the KnowledgeInterchange Format (KIF; Genesereth, 1991),the open knowledge base connectivity specifi-cation (OKBC; http://www.ksl.stanford.edu/software/OKBC/), the resource descriptionframework (RDF) as part of the developmentwork within the semantic web (http://www.w3.org/2001/sw/), and the XMLSchema language (which is itself supportedby the VPDMf).

The field of bioinformatics has many highlyengineered tools for various well-definedtasks such as displaying molecular structure,performing data analysis, etc. (e.g., see theProtein Databank’s website for a cross sectionof tools: http://www.pdb.org/). An in-depthanalysis between the bio- and neuroinformat-ics tools is beyond the scope of this article butmany of the research problems in bioinfor-matics work within a problem space definedby a quantitative, mathematically-tractableontology (as defined above). As we have dis-cussed in this paper and elsewhere (Burns,2001b), the problem space of neuroinformaticsis defined by a qualitative, nonstandardizedset of ontologies. We are therefore compelledto address our technical problems in a differ-ent way (by building subjective, multi-usersystems; by addressing one specified highlevel question, etc.) whereas bioinformaticssolutions can be large-scale without having toimplement these measures.

NeuroScholar is a natural extension of thetechnology described in many of these otherprojects (such as work in the analysis of neuralconnectivity). Many of the tool sets we aredeveloping are similar to other developers’work (e.g., the Atlas Mapper was derived fromthe NeuARt project from the USC BrainProject, Dashti et al., 1997), however the com-

106________________________________________________________________________________ Burns et al.


bination of all of these techniques into our the-oretical framework is unique (Burns, 2001a).

At present, the development of ontologiesand data models is not a continuous process.If the data model needs to be redesigned,changes are typically made offline and thensubsidiary systems would need to be updatedand the data ported to the new system.Cellular Database Management Systems(CDBMS) present a very new technology forrobust, dynamic data-handling without theneed for predefined data models (Gelfand,2002 a,b).

Cellular Database Management Systems area new technology. The process of building adatabase starts by having to fully define itsdata model, which is essentially how thedatabase “sees the world.” In general, thisviewpoint is static and relatively simple (con-sisting of, in the absolute largest cases, severalthousand typed object classes in DataWarehouse solutions). Within this databasesystem, any and all facts present in the worldmay only ever be expressed within this datamodel and all users of the system are restrict-ed to the same absolute vocabulary. If thevocabulary becomes obsolete, or the informa-tion in the world does not exactly conform tothe rigidly defined rules of the system, at best,the system loses accuracy, and at worst, itbreaks. Within the CDBMS model, it is possi-ble to insert and retrieve ambiguous data atany level. The system works with a large net-work of interconnected “cells” to store data(rather than tables, or objects). Each cell is avector of three integers and a data value. Theconnections between cells may be formed andunformed dynamically to reconfigure the sys-tem’s data model intelligently while the sys-tem is in use. If the stated objectives of thesystem’s inventors are realized, this will be anextremely important development, permittingdata models to be changed and updated onthe fly, with no loss of functionality.

The fuzzy relationships that are often found

between individual data items within neu-roinformatics datasets, especially thosederived from knowledge management sys-tems, are usually analyzed with numericallyintensive techniques (for example, stochasticoptimization, fuzzy template matching andanalysis of large graphs). Thus, even with par-allel or clustered computing facilities, applica-tions may have to run for a time period ofdays to weeks, given the relatively low speedsof current hardware. This has a number ofimplications. Applications should berestartable in the event that the host systemsoftware and hardware fail. The use ofdynamic load balancing in network-clustercomputer environments may permit the opti-mal use of available resources and could facil-itate the reliability of the system, since pro-cesses may migrate away from malfunction-ing nodes.

The Portable UNIX Programming System(see “Knowledge Management Core”) isdesigned to perform homeostatic computa-tions. Namely, the computational processesperforming the work of the calculation man-age their environment through a number ofmechanisms. These include automated pro-cess migration via MOSIX (Barak et al., 1993);support for recoverable processes via theTennessee checkpointing protocol (Plank etal., 1995); homeostatic protection of dataitems; peer-to-peer and user-to-peer dynamicinteraction with running processes; supportfor dynamic goal reassignment and steeringparameter update in running applications;and support for manipulation and storage ofcomplex datasets via a practical multihost/mulitprocess implementation of a persistentobject store.

For the NeuroScholar project, this is of par-ticular interest, since it is expected that thetheoretical landscape of the subject of neuro-science will be dramatically reshaped bydevelopments in neuroinformatics. As newdiscoveries arise, our successes will be deter-



mined by the agility with which we can shiftour theoretical perspective to meet new data.

Acknowledgments This work was supported under an RO1

from the National Library of Medicine (num-ber LM07061). We thank the collective mem-bers of the Brainwalk discussion group in theNeurobiology department at USC for theirvaluable input and suggestions. Thanks toAlan Watts for his conceptual input concern-ing application of NeuroScholar to the CRHsystem. Thanks to Shyam Kapadia, KevinChan, Wei-Cheng Chen, Atousa Golapayegani,Shanshan Song, Yan Zhou and Ning Zhang fortheir contributions to the development of theapplication. Thanks to Cyrus Shahabi for hisvaluable feedback and support during thewriting of this manuscript.

ReferencesArbib, M. (2001) NeuroInformatics: the issues, in:

Computing the Brain, 1 (Arbib, M. and J. Grethe,J., eds.) Academic Press Inc., San Diego, CA, pp.3–28.

Axelrod, J. and Reisine, T. D. (1984) Stress hor-mones: their interaction and regulation. Science224:452–459.

Bloom, F. E., Young, W. G., and Kim, Y. M. (1990)Brain Browser, Hypercard application for theMacintosh, Academic Press Inc., San Diego, CA.

Blum, B. (1986) Clinical Information Systems,Springer, New York.

Bota, M. and Arbib, M. (2001) The NeuroHomologyDatabase, in: Computing the Brain, 1, (Arbib, M.and Grethe, J., eds.) Academic Press Inc., SanDiego, CA, pp. 337–354.

Bowden, D. M. and Martin, R. F. (1995) NeuroNames Brain Hierarchy. Neuroimage 2:63–83.

Boyd-Rayward, W. (1998) The Origins ofInformation Science and the InternationalInstitute of Bibliography/InternatinoalFederation for Information and Documentation(FID), in: Historical Studies in Information Science,(Hahn, T. and Buckland, M. eds.) InformationToday, Inc. Medford, NJ.

Burns, G. (2001a) Knowledge Mechanics and theNeuroscholar project: a new approach to neuro-

scientific theory, in: Computing the Brain, 1(Arbib, M. and Grethe, J. eds.) Academic Press,Inc., San Diego, CA, pp. 319–336.

Burns, G. A. (2001b). Knowledge management ofthe neuroscientific literature: the data modeland underlying strategy of the NeuroScholarsystem. Philos Trans R Soc Lond B Biol Sci356:1187–1208.

Burns, G. A. and M. P. Young (2000). Analysis ofthe connectional organization of neural systemsassociated with the hippocampus in rats. PhilosTrans R Soc Lond B Biol Sci 355:55–70.

Burns, G. A. P. C. (1997) Neural connectivity of therat: theory, methods, and applications,Laboratory of Physiology, University of Oxford,Oxford, UK.

Burns, G. A. P. C., Stephan, K. E., Ludäscher, B.,Gupta, A., and Kotter, R. (2001) Towards a fed-erated neuroscientific knowledge system usingbrain atlases. Neurocomputing 38–40:1633–1641.

Campione, M., Walrath, K., and Huml, A. (2002)The Java Tutorial, Third Edition.

Chen, Y.-S. and Shahabi, C. (2001) Automaticallyimproving the accuracy of user profiles withgenetic algorithm. IASTED InternationalConference on Artificial Intelligence and SoftComputing, Cancun, Mexico.

Chen, Y.-S. and Shahabi, C. (2002). Improving userprofiles for e-commerce by genetic algorithms,in: E-Commerce and Intelligent Methods in studiesin Fuzziness and Soft Computing, (Segovia, J.,Szczepaniak, P. S., and Niedzwiedzinski, M.,eds.), Springer-Verlag. Heidelberg, Germany.

Chicurel, M. (2000) Databasing the brain. Nature406:822–825.

Dallman, M. F., Akana, S. F., Cascio, C. S.,Darlington, D. N., Jacobson, L., and Levin, N.(1987) Regulation of ACTH secretion: variationson a theme B. Rec Prog Horm Res 43:113–173.

Dashti, A. E., Ghandeharizadeh, S., Stone, J.,Swanson, L. W., and Thompson, R. H. (1997)Database challenges and solutions in neurosci-entific applications. Neuroimage 5:97–115.

de Groot, J. and Harris, G. W. (1950) Hypothalamiccontrol of the anterior pituitary gland and bloodlymphocytes. J Physiol (London) 111:335–346.

Duckett, J., Griffen, O., Mohr, S., Norton, F., Stokes-Rees, I., Williams, K., Cagle, K., Ozu, N., andTennison, J. (2001) Professional XML Schemas,Wrox Press, Inc, Birmingham, UK.

Gardner, D., Abato, M., Knuth, K. H., DeBellis, R.,and Erde, S. M. (2001) Dynamic publicationmodel for neurophysiology databases. Philos

108________________________________________________________________________________ Burns et al.


Trans R Soc Lond B Biol Sci 356:1229–1247.Gelfand, B. (2002a) Data cells and data cell genera-

tions. US Patent Office. USA.Gelfand, B. (2002b) Data cells, and a system and

method for accessing data in a data cell. USPatent Office. USA.

Genesereth, M. R. (1991) Knowledge InterchangeFormat. Proceedings of the Second InternationalConference on the Principles of KnowledgeRepresentation and Reasoning, MorganKaufmann, Cambridge, MA.

Goddard, N. H., Hucka, M., Howell, F., Cornelis,H., Shankar, K., and Beeman, D. (2001) TowardsNeuroML: model description methods for col-laborative modelling in neuroscience. PhilosTrans R Soc Lond B Biol Sci 356:1209–1228.

Gormezano, I., Schneiderman, N., Deaux, E., andFuentes, I. (1962) Nictitating membrane: classi-cal conditioning and extinction in the Rat.Science 138:33–34.

Grethe, J., Mureika, J., and Merchant, E. (2001)Design concepts for NeuroCore and neuro-science databases, in: Computing the Brain, 1,(Arbib, M. and Grethe, J., eds.) Academic Press,Inc., San Diego, CA, pp. 135–150.

Gruber, T. R. (1993) A translation approach toportable ontology specifications. KnowledgeAcquisition 5:199–220.

Gupta, A., Lüdascher, B., and Martone, M. (2000)Knowledge-based Integration of NeuroscienceData Sources. Proceedings of 12th Int. Conf.Scientific and Statistical Database ManagementSystems (SSDBM’00), Berlin, IEEE.

Herman, J. P. and Cullinan, W. E. (1997)Neurocircuitry of stress: central control of thehypothalamo-pituitary-adrenocortical axis.Trends Neurosci 20:78–84.

Hilgetag, C. C., O’Neill, M. A., and Young, M. P.(1996a) Indeterminate organization of the visualsystem. Science 271:776–777.

Hilgetag, C. C., O’Neill, M. A., and Young, M. P.(1996b) Optimization analysis of complex neu-roanatomical data, in: ComputationalNeuroscience, Plenum Press, Boston, MA.

Howe, D. (2001) The Free On-Line Dictionary ofComputing (http://www.foldoc.org/).

Koslow, S. H. (2000) Should the neuroscience com-munity make a paradigm shift to sharing pri-mary data? Nat Neurosci 3:863–865.

Muller, R. J. (1999) Database design for smarties,using UML for data modeling, MorganFreeman, San Francisco, CA.

Musen, M. A. (1992) Dimensions of knowledgesharing and reuse. Comput Biomed Res 25:435–467.

Nicolelis, M. A., Tinone, G., Sameshima, K., Timo-Iaria, C., Yu, C. H., and Van de Bilt, M. T. (1990)Connection, a microcomputer program for stor-ing and analyzing structural properties of neu-ral circuits. Comput Biomed Res 23:64–81.

Noy, N. F., Fergerson, R. W., and Musen, M. A.(2000) The knowledge model of Protege-2000:Combining interoperability and flexibility.Proceedings of 2nd International Conference onKnowledge Engineering and KnowledgeManagement, Juan-les-Pins, France.

Otlet, P. (1918) Transformations in the bibliograph-ic apparatus of the sciences, in The internationalorganization and dissemination of knowledge:Selected essays of Paul Otlet (1990), W. B.Rayward ed., Elsevier. Amsterdam.

Ovsiannikov, I. and Arbib, M. (2001) Annotator:Annotation Technology for the WWW, in:Computing the Brain, 1, (Arbib, M. and Grethe, J.,eds.) Academic Press Inc., San Diego, CA, pp.255–264.

Plank, J. S., Beck, M. and Kingsley, G. (1995)Libckpt: transparent checkpointing underUNIX. Proc. USENIX Winter Tech. Conf., NewOrleans, LA.

Rational (1997) UML Semantics Version 1.1. SantaClara, CA, Rational Software Corp.

Rayward, W. B. (1998) The origins of informationscience and the International Institute ofBibliography/International Federation forInformation and Documentation (FID), in:Historical Studies in Information Science, (Hahn, T.and Buckland, M., eds.) Information Today, Inc.Medford, NJ.

Rho, J. H. and Swanson, L. W. (1989) A morpho-metric analysis of functionally defined subpop-ulations of neurons in the paraventricular nucle-us of the rat with observations on the effects ofcolchicine. J Neurosci 9:1375–1388.

Rumbaugh, J., I. Jacobson and G. Booch (1999) TheUnified Modeling Language Reference Manual,Addison-Wesley, Reading MA.

Sapolsky, R. M., Romero, L. M., and Munck, A. U.(2000) How do glucocorticoids influence stressresponses? Integrating permissive, suppressive,stimulatory and preparative actions. EndocrineRev 21:55–89.

Sawchenko, P. E., Li, H. Y., and Ericsson, A. (2000)Circuits and mechanisms governing hypothala-mic responses to stress: a tale of two paradigms.Prog Brain Res 122:61–78.



Sawchenko, P. E. and Swanson, L. W. (1989)Organization of CRF immunoreactive cells andfibers in the rat brain: immunohistochemicalstudies. In: Corticotropin-Releasing Factor: Basicand Clinical Studies of a Neuropeptide, (Souza, E.D. and Nemeroff, C., eds.) CRC Press. BocaRaton, FL, pp. 29–51.

Scannell, J. W., Blakemore, C., and Young, M. P.(1995) Analysis of connectivity in the cat cere-bral cortex. Journal of Neuroscience 15:1463–1483.

Scannell, J. W., Burns, G. A., Hilgetag, C. C.,O’Neill, M. A., and Young, M. P. (1999) The con-nectional organization of the cortico-thalamicsystem of the cat. Cereb Cortex 9:277–299.

Sowa, J. (2000) Knowledge Representation.Logical, Philosophical and ComputationalFoundations, Brooks/Cole, Pacofoc Grove, CA.

Stephan, K. E., Kamper, L., Bozkurt, A., Burns, G.A., Young, M. P., and Kotter, R. (2001) Advanceddatabase methodology for the Collation ofConnectivity data on the Macaque brain(CoCoMac). Philos Trans R Soc Lond B Biol Sci356:1159–1186.

Stephan, K. E., Zilles, K., and Kotter, R. (2000)Coordinate-independent mapping of structuraland functional data by objective relational trans-formation (ORT). Philos Trans R Soc Lond B BiolSci 355:37–54.

Swanson, L. (1986) Organization of mammalianneuroendocrine system. In: Handbook ofPhysiology, The Nervous System, IV, (Bloom, F.,ed.) Waverly Press, Baltimore, MD, pp. 317–363.

Swanson, L. (2000) Paraventricular nucleus. In:Encyclopedia of Stress, Vol. 3, Academic Press Inc.,San Diego, CA, pp. 130–133.

Swanson, L., Sawchenko, P., Lind, R., and Rho, J.-H. (1987) The CRH motoneuron: differentialpeptide regulation in neurons with possiblesynaptic, paracrine and endocrine outputs. Ann

NY Acad Sci 512:12–23.Swanson, L. W. (1991) Biochemical switching in

hypothalamic circuits mediating responses tostress. Prog Brain Res 87:181–200.

Swanson, L. W. (1998) Brain Maps: Structure of theRat Brain, Elsevier Science Publishers B. V.,Amsterdam, The Netherlands.

Tanimura, S., Sanchez-Watts, G., and Watts, A. G.(1998) Peptide gene activation, secretion, andsteroid feedback during stimulation of rat neu-roendocrine corticotropin-releasing hormoneneurons. Endocrinol 139:3822–3829.

Tanimura, S. and Watts, A. (1998) Corticosteronecan facilitate as well as inhibit corticotropin-releasing hormone gene expression in the rathypothalamic paraventricular nucleus.Endocrinol 139:3830–3836.

Ullman, J. and Widom, J. (1997) A first course indatabase systems, Prentice Hall, Inc., UpperSaddle River, NJ.

Watts, A. and Swanson, L. (1989) Diurnal varia-tions in the content of preprocorticotropin-releasing hormone messenger ribonucleic acidsin the hypothalamic paraventricular nucleus ofrats of both sexes as measured by in situhybridization. Endocrinol 125:1734–1738.

Wilder, B. G. (1896) Neural terms, internationaland national. J Comp Neurol 6:216–352.

Young, M. P. (1992) Objective analysis of the topo-logical organization of the primate cortical visu-al system. Nature 358:152–155.

Young, M. P. (1993) The organization of neural sys-tems in the primate cerebral cortex. Pro Royal SocLondon B Biol Sci 252:13–18.

Young, M. P., Scannell, J. W., and Burns, G. A. P.C.(1995) The Analysis of Cortical Connectivity,Springer/RG Landes.

06ni 1_1-burnsf - infolab - university of southern california

Documents