Philosophy of Science, 69 (September 2002) Pp

Philosophy of Science, 69 (September 2002) pp. S83-S97. 0031-8248/2002/69supp-0008
Copyright 2002 by The Philosophy of Science Association. All rights reserved.

InterlevelExperimentsandMultilevelMechanismsintheNeuroscienceofMemory

CarlF.Craver
WashingtonUniversity

The dominant neuroscientific theoryof spatial memory is,like many theories inneuroscience, a multilevel descriptionof a mechanism. Thetheory links the activitiesof molecules, cells, brainregions, and whole organismsinto an integrated sketchof an explanation forthe ability of organismsto navigate novel environments.Here I develop ataxonomy of interlevel experimentalstrategies for integrating thelevels in such multilevelmechanisms. These experimental strategiesinclude activation strategies, interferencestrategies, and additive strategies.These strategies are mutuallyreinforcing, providing a kindof interlevel and intratheoreticrobustness that has notpreviously been recognized.

SendrequestsforreprintstoCarlF.Craver,DepartmentofPhilosophy,BuschHall,OneBrookingsDrive,WashingtonUniversity,St.Louis,St.Louis,MO63130;.

Specialthanks to William Bechteland Lindley Darden forcomments and suggestions. Figures,which also appeared inCraver and Darden (2001),are reprinted with permissionfrom the University ofPittsburgh Press.

1. Introduction.

Many theoriesin contemporary neuroscience aremultilevel descriptions of mechanisms.One aim of experimentationis to integrate thedifferent levels in suchtheories. In this paperI analyze the conceptsof "mechanism" and "level"and deploy this analysisto describe three mutuallyreinforcing kinds of interlevelexperiments used to integratethe levels in multileveltheories. My discussion isconstructed by reference torecent attempts to integrateLong-Term Potentiation (LTP), aform of synaptic plasticity,into a multilevel mechanismof memory. The imageof neuroscientific theory constructiondeveloped through this casediverges from traditional reductionisticperspectives (as advocated in,e.g., Schaffner 1993 andmore recently by Bickle1998), and does somore fully than (althoughsomewhat consistently with) recentmultilevel perspectives on neuroscientificpractice (e.g., Bechtel andRichardson 1993; Schouten andLooren de Jong 1999).What is novel inthis paper is theuse of a nonformalaccount of theory structureto ground a taxonomyof specifically interlevel experimentsand the use ofthis experimental taxonomy toexplore the process ofintegrating levels in amultilevel neural mechanism.

2. Mechanisms and Their Organization.

Mechanisms, asthey are understood incontemporary neuroscience, are collectionsof entities and activitiesorganized in the productionof regular changes fromstart or setup conditionsto finish or terminationconditions (Machamer, Darden, andCraver 2000). The entitiesin neuroscience include thingslike neurons, neurotransmitters, brainregions, and mice. Theactivities are the variousdoings in which theseentities engage: neurons fire,neurotransmitters bind to receptors,brain regions process, andmice navigate mazes. Activitiesare the things thatentities do; they arethe productive components ofa mechanism, and theyconstitute the stages ofmechanisms. When neuroscientists speakgenerally about activities, theyuse a variety ofterms; activities are oftencalled "processes," "functions," and"interactions." When they speakspecifically about activities, theyuse verbs and verbforms; they speak ofattracting and repelling, phosphorylatingand hydrolyzing, binding andbreaking, and firing andreleasing.

The entities and activitiescomposing mechanisms are organized;they are organized suchthat they do something,carry out some task orprocess, exercise some faculty,perform some function orproduce some end product.I will refer tothis activity or behaviorof the mechanism asa whole as therole to be explainedby the description ofthe mechanism. The roleis the activity atthe top of Figure 1.Below it are theentities and activities composingthe mechanism for thatrole.

/ Figure 1

The entities and activitiescomposing mechanisms have aspatial and temporal organizationthat is crucial totheir productivity. (By "crucial"I mean necessary inthe circumstances (cf. Nagel1977); this might usefullybe fleshed out withthe help of Mackie's(1974) notion of anINUS condution (for insufficientnonredundant part of anunnecessary but sufficient condition)).Spatially, the entities composingthe mechanism must beappropriately located, connected, structuredand oriented with respectto one another ifthe mechanism is towork. The activities composingmechanisms also have crucialtemporal orders, rates, anddurations. Uncovering these temporaland spatial aspects ofa mechanism's organization isa major step inthe construction of neuroscientifictheories and so isa major focus ofneuroscientific practice (see Darden and Craver, forthcoming).

3. Example: The Mechanism of Long Term Potentiation.

LTPis a means ofstrengthening synapses in thecentral nervous system. Manythink that LTP isa crucial activity inthe mechanisms of memory.LTP is a formof neural plasticity reminiscentof a memory mechanismproposed by D. O. Hebbin his 1949Organization of Behavior.Hebb's idea was thatmemories might be formedby strengthening synapses whenboth the presynaptic andthe postsynaptic neurons aresimultaneously active. This hypothesishas had considerable stayingpower in contemporary neuroscience;in fact, it recentlycontributed to Eric Kandel'sshare of the NobelPrize for medicine.

LTP istypically studied in thehippocampus, a brain structurelong known to exhibitthis form of synapticmodification. The hippocampus isanother of the crucialentities in the mechanismof memory. Surgical removalof the hippocampus producesprofound memory deficits inhuman and nonhuman animals.A cross section ofthe hippocampus with someof its major anatomicalregions and synaptic connectionsis shown in Figure 2.LTP can be inducedat each of thethree major excitatory synapsesin this diagram.

/ Figure 2

There isno consensus about themechanisms that produce LTP.One researcher has complainedthat the LTC (LongTerm Controversy) over LTPis threatening to becomean LTTP (a "LongTerm Tar Pit") forneurobiologists (Malinow 1998, 1226).Nonetheless, an example ofone plausible, if incomplete,sketch of the mechanismfor LTP nicely illustratesseveral aspects of themechanisms described in neuroscientifictheories.

The hippocampal synapses thatexhibit LTP use theneurotransmitter glutamate. Glutamate isreleased from the presynapticcell with each actionpotential, and binds toreceptors on the postsynapticcell. LTP can bethought of as anincrease in the effectof a single presynapticaction potential on thepostsynaptic electrical response. Thisincrease in the strengthof the synapse couldbe due, for example,to the release ofmore glutamate from thepresynaptic cell, or tothe changing receptive propertiesof the postsynaptic cell,or perhaps to both.

Onetype of postsynaptic glutamatereceptor in the hippocampusis called the NMDAreceptor (for N-Methyl D-Aspartate,a chemical agonist thathas a high affinityfor this receptor). Whenglutamate binds to NMDAreceptors on the postsynapticcell, the NMDA receptorschange their shape, exposinga pore in thecell membrane. If thepostsynaptic cell is inactive,the channel remains blockedby large Mg2+ ions.But if the postsynapticcell is depolarized, theseMg2+ ions float outof the channel, allowingCa2+ to diffuse intothe cell. The risingintracellular Ca2+ concentrations setin motion a longbiochemical cascade terminating inthe question marks ofFigure 3.

/ Figure 3

The remaining details ofthis mechanism are morespeculative, but three thingsare thought to happen.In the short term,it is thought thatthis cascade leads toan increase in thenumber or sensitivity ofso-called AMPA receptors (perhapsby phosphorylation). These changesaccount for the rapidinduction of LTP. Inthe long term, thecascade leads to theproduction of proteins inthe postsynaptic cell body.These proteins are thoughtto alter the structureof the dendritic spinesat that synapse (see,e.g., Engert and Bonhoeffer1999; Maletic-Savetic, Malinow, andSvoboda 1999). Some suspectthat there is alsoa presynaptic component ofthe LTP mechanism whereby,for example, the presynapticcell releases more glutamate.

Theentities in this mechanismare glutamate molecules, NMDAreceptors, Ca2+ ions andthe like. The activitiesinclude binding, diffusing, phosphorylating,and changing conformation. Theworking of the mechanismdepends crucially upon itsorganization. It depends uponthe order of theactivities and on theirrelative rates and durations.It also depends cruciallyupon the structures, shapes,sizes, orientations, and locationsof the component entities.LTP is a representativeexample of mechanisms incontemporary neuroscience. But noanalysis of mechanisms inneuroscience is complete withoutan analysis of theircharacteristic multilevel organization.

4. Three Kinds of Levels.

Talk of"levels" is nearly ubiquitousin neuroscience and itsphilosophy. Both philosophers ofmind and science areincreasingly recognizing the difficultiesand ambiguities attending suchtalk (see e.g., Heil1949 and Kim 1993).But it is possibleto take some initialsteps toward greater clarityby disambiguating three differentkinds of "levels": levelsof mere aggregates, functionallevels, and mechanistic levels.Each of these kindsis individuated by adifferent asymmetrical decomposition relation.The multilevel theories ofcontemporary neuroscience exhibit multiplemechanistic levels, but thecontrast with the otherkinds is revealing.

Begin withlevels of mere aggregates and the correspondingnotion of an aggregativedecomposition. An aggregative decompositioninvolves dividing some chunkof mattersome entity (itis always an entitythat is aggregatively decomposed)intosmaller chunks of matter.The ball of waxand the hippocampus caneach be sliced, diced,cubed, or spiral cutinto parts; and thesesmaller parts could then,at least in principle,be put back togetherto fill the samevolume of space occupiedbefore the decomposition (seeHaugeland 1998, chaps. 1and 9). The intendedsense of aggregativity isthat developed by Wimsatt(1986): The properties ofwholes are simple sumsof the properties ofparts (e.g., volume andmass); the wholes arestable under disaggregation andreaggregation of parts; andthe parts do notsignificantly interact with oneanother. Talk of aggregatelevels highlights relations ofspatial inclusion and sizeamong entities to thetotal neglect of activitiesand their organization.

Where aggregatelevels are relationships amongentities, functional levels are relationsamong abstract roles. Functionaldecomposition of one levelinto another involves takinga task, a routine,or a faculty andbreaking it into sub-tasks,sub-routines, or sub-faculties. Functionaldecompositions are often treatedby neuroscientists as ifthey were, at best,necessary oversimplifications in thegeneration of testable sketchesor, at worst, piein the sky speculationsthat are replaced orobviated as the detailsof a mechanism becomeavailable. This is becausedecomposition by functional rolealone does not adequatelyembody those roles inthe entities and activitiesthat the ontic storeof contemporary neuroscience hasto offer. For theneuroscientist, purely functional decompositionsare disembodied "how-possibly" descriptionsof a mechanism; theyare sometimes denigrated as"boxology."

What neuroscientists are afteris neither an aggregativedecomposition nor a purelyfunctional decomposition, but rathera mechanistic decomposition intomechanistic levelsa decomposition into entitiesand activities organized inthe performance of ahigher level role. Theactivities and properties ofthe entities in thelower level mechanism maythemselves be subject tomechanistic decomposition. In suchcases, each mechanistic decompositionadds another level towhat may become amultilevel mechanism. It istypically possible to distinguishlevels by the differententities and activities thatpopulate them and, aswe will see, bythe different techniques thatare used to investigatethose entities and activities.But how many levelsthere are and whatkinds of entities arefound at each levelare empirical questions tobe answered within agiven research program.

/ Figure 4

Consider asketch of the mechanismsof spatial memory. Thissketch has roughly fourdistinct mechanistic levels, althoughwe should expect thenumber of levels andthe descriptions at differentlevels to change overtime. At the topis a behavioral-organismic level,having to do with,for example, the varioustypes of learning andmemory, the conditions underwhich different memories maybe stored or retrieved,and the conditions underwhich storage or retrievalare likely to improveor fail. Techniques forinvestigating phenomena at thebehavioral-organismic level typically involvebehavioral tasks, such asnavigation, recognition of objects,and tests of avoidance,aversion, and preference.

Beneath thisbehavioral-organismic level is acomputational-hippocampal level, having todo roughly with therole of the hippocampusin the mechanisms ofmemory, its cytological, anatomical,and structural features, itspathology, its connectivity withother brain regions, andthe computational or processingstages it is thoughtto perform. Techniques forinvestigating phenomena at thislevel include ablation, pathologicalanatomy, multicellular recording, EEG,PET and MRI, aswell as various computationalapproaches. Claims that thecells of the hippocampus(or some part ofthe hippocampus) may functionas a "spatial map"(O'Keefe and Dostrovsky 1971),as an organ of"declarative memory" (Zola-Morgan andSquire 1993), as arelater of "an itemand its context" (Schachterand Wagner 1999), oras "self localization androute replay" (Redish andTouretzky 1998) are hypothesesabout this level.

The contributionof the hippocampus tothe phenomena of memoryis thought to involveLTP and various synapticcomponents. This electrical-synaptic levelincludes such entities asneurons, synapses, and dendriticspines and such activitiesas vesicular release andthe generation and propagationof action potentials. Phenomenaat this level aretypically investigated with pharmacologicaland electrophysiological techniques.

Bottoming outthis hierarchy are entitiesand activities at amolecular-kinetic level. At this levelentities like the NMDAand AMPA receptors, glutamate,Ca2+ ions, and Mg2+ ions engage in activitieslike attracting and repelling,binding and breaking, phosphorylatingand hydrolyzing. These componentsare investigated with ahost of biochemical, andincreasingly, molecular biological techniques.

Tosummarize, the mechanism sketchfor memory is multilevel;its current description includesmice learning and remembering,hippocampi generating spatial maps,synapses inducing LTP, andmacromolecules binding and changingconformation. These levels aremechanistic levels in thatthey describe parts andwholes related as componentsto mechanisms or, moreappropriately, as the activitiesof parts to theactivities of the mechanismas a whole. Itis a "sketch" becausecertain of its levelsare poorly understood andbecause gaps exist ineven the most wellunderstood levels. This isoften the case inthe process of mechanismdiscovery.

The elaboration and refinementof such multilevel descriptionstypically proceeds piecemeal withthe goal of integratingthe entities and activitiesat different levels (Craverand Darden 2001; Craver2001). Integrating a componentof a mechanism intosuch a hierarchy involves,first, contextualizing the itemwithin the mechanism forthe role to beexplained. This involves "lookingup" a level andidentifying a mechanism thathas the item asa component. Integrating involves,second, "looking down" alevel and showing thatthe properties or activitiesof an entity canbe explicated in termsof a lower levelmechanism. For example, LTPmight be integrated intoa memory mechanism bylooking up to seeit as a componentin a computational-hippocampal mechanismand by looking downto explain it interms of its molecularlevel mechanisms. This understandingof mechanisms, levels, andintegration yields a tidytaxonomy of interlevel experimentsin neuroscience.

5. Interlevel Experimental Strategies.

Interlevel experiments aretools for integrating thelevels in hierarchical descriptionsof mechanisms. Interlevel experimentstell us what therelevant entities and activitiesare, how they arenested in component/sub-component relations,and how the activitiesof the component entitiesfit into their mechanisticcontext.

As a first pass,experiments for testing mechanismshave three basic elements:(i) an experimental model(e.g., a strain ofmouse), (ii) an interventiontechnique (e.g., electrical stimulation),and (iii) a detectiontechnique (e.g., whole-cell recording).These elements are depictedin the abstract experimentalprotocol in Figure 5, whichshows an experiment fora single mechanistic level.The connected circles andarrows represent a hypothesizedmechanism putatively instantiated inan experimental model. Onthe left hand sideof the figure arearrows standing for anintervention technique (I). (Theintended sense of interventionmight be explicated alongthe lines of Woodward2000). The perturbation thatis produced by Iin the experimental preparationhas "downstream" results whichare detected or amplifiedusing a detection technique(D).

/ Figure 5

It is easy toextend this view ofexperiments to interlevel experiments.Interlevel experiments are experimentsin which the techniquesfor intervening and detectingare targeted at differentlevels in the mechanistichierarchy. For simplicity, startwith experiments spanning twolevels. The left handside of Figure 6 exhibitsa case of interveningto perturb a componentin the lower levelmechanism and detecting theconsequences for a higherlevel role; these arebottom-up experiments. The righthand side of Figure6 shows the opposite:intervention to perturb thehigher level role anddetection of the activitiesor properties of componentsin the lower levelmechanism. These can bethought of as top-downexperiments. I now wantto consider three prevalentinterlevel experimental strategies incontemporary neuroscience: activation strategies,interference strategies, and additivestrategies.

/ Figure 6

5.1. Activation Strategies.

Experiments exemplifying activation strategieshave a top-down structure;one activates, engages, triggers,or stimulates the roleof interest and thendetects the properties oractivities of one ormore putative components ofthe mechanism instantiating thatrole. (Activation strategies havebeen discussed in thecontext of functional PETand MRI techniques byStufflebeam and Bechtel 1996and Bogen 2001).

In theearly 1970's, O'Keefe andDostrovsky (1971) recorded theelectrical activity of neuronsin the rat hippocampuswhile the rats navigateda maze. The interventionin this case involvesactivating the spatial memorysystem by putting therat in a maze.The detection technique isthe electrical recording fromhippocampal cells. They foundthat certain of theseneurons generate bursts ofaction potentials whenever therat enters a particularlocation while facing ina particular direction. Theseneurons have come tobe called "place cells,"and the region ofspace occupied by therat when the placecell increases its activityis known as thecell's "place field." Theseplace cells have slightlyoverlapping place fields thatcover the animal's immediatespatial environment, and manybelieve for this reasonthat the hippocampus (orportions of it) couldserve as a spatialmap. These findings haverecently been confirmed withmultiunit electrodes that allowone to record from70150 pyramidal cells atonce. Astonishingly, it ispossible to predict thepath taken by therat on the basisof these recordings (Wilsonand McNaughton 1993).

As compellingas these results are,they are, taken alone,far from establishing themechanistic relevance of thehippocampus to navigation. Forexample, the electrical activityof the hippocampus andnormal navigation may eachbe effects of acommon cause. In thatcase, the activity ofhippocampal cells would notbe crucial for themechanism and would infact be incidental tothe mechanism. Objections ofthis sort can bepartly redressed by thesecond form of inter-levelexperimental strategy.

5.2. Interference Strategies.

Interference experiments arebottom-up experiments in whichone intervenes to diminish,retard, eliminate, disable, ordestroy some component entityor activity in alower level mechanism andthen detects the resultsof this intervention forsome higher level role.(The epistemic status ofinterference strategies has beendiscussed by Glymour 1994and Bub 1994; seealso Pearl's 2000 discussionof "surgery" on acausal graph). Consider anexample from recent geneknockout experiments on LTPand memory.

In late 1996,researchers at MIT, Columbia,and Cal Tech publisheda series of papersdescribing the effects ofhighly specific genetic deletionson the other levelsin the memory hierarchy(McHugh et al. 1996;Rotenberg et al. 1996;Tsien et al. 1996a;Tsien et al. 1996b).The researchers invented amolecular scalpel for deletingthe NMDAR1 gene andfor deleting it onlyin the CA1 regionof the hippocampus. Theythen performed detection techniquesat each of thefour levels in themultilevel theory. Knockout micehad difficulty learning thelocation of a submergedplatform in a Morriswater maze, a commonbehavioral-organismic experimental technique. Knockoutmice swim randomly aroundthe pool; controls quicklylearn to swim directlyto the platform. Multiunitrecordings from knockout hippocampirevealed significant impairments inspatial map formation; theplace fields were muchlarger and much lesssharply defined. These deficitsin spatial map formationare arguably the resultof the absence ofLTP since knockout synapsesdid not exhibit LTPunder normal conditions.

This experimentis a bottom-up interferenceexperiment with detection atmultiple levels. The interventiontechnique interferes with theactivities of the NMDAreceptor by deleting theNMDAR1 gene. The detectiontechniques register the resultsof this intervention onLTP, spatial map formation,and spatial memory.

Like activationstrategies, interference strategies havetheir characteristic weaknesses. Inthe case at hand,for example, the interventiontechnique is, in effect,a form of cellulardamage which may haveunintended implications for cellfunction independent of theeffect on LTP and,further, this cell damagemay have unintended consequencesfor the organization ofthe hippocampus and thebrain. One would bemore certain of arole for NMDA-dependent hippocampalLTP in the mechanismsof memory if onecould enhance memory bychanging NMDA receptors andthereby enhancing hippocampal LTP.This is an additivestrategy; it is thelast I will discuss.