1
Dissecting the mechanisms of learning-by-doing in
Drosophila
Björn Brembs
1
& Wolfgang Plendl
2
1 Freie Universität Berlin, Institut für Biologie Neurobiologie, Königin-Luise Str.
28/30, 14195 Berlin, Germany. Email:
bjoern@brembs.net
, phone: +49 (0)30 838
55050, fax: +49 (0)30 838 55455
2 Lehrstuhl für Neurobiologie und Genetik, Biozentrum, Universität Würzburg,
Germany.
At the heart of learning-by-doing
1
lies a well-known psychological phenomenon:
information will be remembered better if it is actively generated rather than
passively read or heard
2,3
. First described in humans
2,4
, this generation effect can
also be observed in various animal models
3,5-7
. However, the neurobiological
mechanisms underlying the generation effect are unknown. Here we show that two
reciprocal interactions between its active and passive components contribute to the
generation effect in flies. One interaction consists of the active (skill-learning)
component facilitating the passive (fact-learning) component. Fact-learning, on the
other hand, inhibits skill-learning. Experiments with adenylyl cyclase I deficient
rutabaga mutant flies revealed that the fact- but not the skill-learning component
requires this evolutionarily conserved learning gene. Using mushroom-body
deficient transgenic flies we observed that the mushroom-bodies mediate the
inhibition of skill-learning. This inhibition also enables generalization and prevents
premature habit formation. Extended training in wildtype flies produced a
phenocopy of mushroom-body impaired flies, such that generalization was
abolished and goal-directed actions were transformed into habitual responses.
Thus, our results identify various neural processes underlying learning-by-doing,
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
2
delineate some of their synergisms and provide a framework for further dissecting
them in a genetically tractable model system.
In the 100 years since the term was coined
1
, "learning-by-doing" has been
recognized as a successful educational and economic strategy
8
. At its core lies a
psychological phenomenon which was described only a few years earlier: Active
engagement of the brain provides learning capabilities which are difficult or impossible
to achieve by passive observation alone
2,3
. This phenomenon is today known as the
generation effect
4
and can also be observed in animals such as monkeys
5
, cats
6
or fruit
flies
7
. Despite the impact learning-by-doing has on society and the ubiquity of the
generation effect, the mechanism by which activity enhances passive learning is
unknown.
To study its neurobiological basis, we hypothesized that the generation effect may
be brought about by an interaction of two components: an active, skill-learning
component and a passive, fact-learning component. We tested this hypothesis by
combining two simple experimental instances of fact- and skill-learning, respectively, in
the fruit-fly Drosophila melanogaster. In both tasks, the fly is tethered to a torque meter
to accomplish stationary flight. In the passive fact-learning task, the fly is presented
with one of two visual cues in alternation, independently of its own behaviour. The
presentation of one of these cues is associated with an infrared beam of light providing
instantaneous, aversive heat. The animal learns the fact that one cue is punished and
prefers the unpunished over the punished cue in a subsequent choice test without heat
9
.
In the active skill-learning task, the fly's spontaneous turning maneuvers
10
are divided
into two groups (i.e. attempts to turn left or right, respectively) and one of them is
punished by heat. During the subsequent choice test without heat, the animal generates
manoeuvres preferentially in the previously unpunished direction
11
. There are no
external cues guiding the animal during skill-training or testing. The fly has to rely
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
3
solely on its own internal representation of its movements in order to solve this task. In
the combined paradigm (Fig. 1), attempted turning will lead to either blue or green
illumination (i.e. a right turn will cause illumination of one colour while a left turn will
cause illumination of the other). During training, one of the two situations is associated
with heat. During test, the heat is permanently switched off. Extending previous results
7
,
the flies showed the generation effect also in this composite paradigm (Supplementary
Figures 6, 7). This finding is consistent with our hypothesis that an interaction of fact-
and skill-learning components may underlie the generation effect. But of what nature is
this interaction? A simple possibility is that the fact-learning component and the skill-
learning component are formed in parallel, and that the two components are summed. A
straightforward test of this `summation' hypothesis is to disable one of the two
components and then subject the animals to the composite learning task.
The rutabaga (rut)-mutant flies lack a type I adenylyl cyclase that is required for
most learning tasks including the instance of fact-learning tested here (Supplementary
Figure 7). If mutant rut flies are only deficient in fact-learning, the summation
hypothesis predicts two similar learning scores: reduced, but significant composite
learning and unaffected skill-learning. If the Rutabaga cyclase is required for both
learning components, the mutant flies should perform poorly in both the composite and
the skill-learning task. Surprisingly, rut mutants performed well (even exceeding
wildtype levels; Supplementary Figure 7) in the skill-learning task, but they failed in the
composite task (Fig.2a). How can this dominant-negative effect of the colours be
explained? One explanation is that the colour changes may interfere with skill-learning
in rut flies. However, colour changes unrelated to the flies' behaviour did not disrupt
performance (i.e., a yoked control; Supplementary Figure 7). Rather, in rut mutant flies,
colour changes concomitant with turning behaviour somehow inhibit skill-learning
during the composite learning task. To investigate whether this inhibition occurs during
acquisition or during retrieval of the skill-learning component, we removed the colours
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
4
after training and tested for the skill-learning component in isolation. If the inhibition
takes place at the level of acquisition, the learning score should be zero, because no skill
was ever learned. If the inhibition takes place during retrieval, the rut flies should reveal
a significant learning score, because the colours are no longer present and thus cannot
interfere with the performance of the skill which was learned during composite training.
The significant rut learning score places the inhibition firmly at the level of retrieval for
the mutant flies (Fig. 2a). The same experiment with wildtype flies did not reveal any
significant learning score. We conclude that in wildtype flies fact-learning also exerts an
inhibitory effect on skill-learning. In contrast to rut flies, this inhibition of skill-learning
acts during acquisition and not during retrieval.
These results show that the interaction between fact- and skill-learning
components is more complex than mere summation. Counter-intuitively, one factor
involved in this interaction is inhibition of skill-learning by a dominant fact-learning
component. Because the generation effect entails an overall enhancement of learning,
there must be a second, facilitating factor which more than compensates for the skill-
learning inhibition. One may assume this second factor to be reciprocal to the first, from
the skill-learning component back to the fact-learning component (Fig. 3). While on the
surface this arrangement may seem implausible, such an enhancement of fact-learning
at the expense of skill-learning allows for keeping the learned fact flexible for use with
a different behaviour than with which it was acquired. It has been shown previously that
flies can perform such a generalization
7
and that the mushroom-bodies (MB), a
prominent neuropil in the insect brain, are required for certain generalization tasks
12,13
.
Conspicuously, the general MB function has long been thought to be inhibitory in
nature
14-16
. We therefore suspected that the inhibition of skill-learning may be mediated
by the MB and enable generalization of the learned fact. To test this hypothesis, we
genetically blocked output from the MB, trained the transgenic flies in the composite
task and subsequently tested them for generalization of colour memory and for the
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
5
isolated skill-learning component (as in the rut experiments). Flies with impaired MB
function can learn visual cues and perform well in skill-learning as well as several other
learning tasks
9
. If the MB mediate the inhibition of skill-learning in order to generalize
learned facts, removal of this inhibition in the transgenic flies should lead to significant
skill-learning and no generalization. Indeed, flies with blocked MB output perform
according to these predictions (Fig. 2b). Further experiments indicate that the MB and
lobes, but not the lobes contribute to this inhibition (Supplementary Figure 8). Are
the MB also involved in the facilitation of fact-learning? There is a different composite
paradigm in which skill-learning can be prevented technically by making the behaviour-
heat association non-predictive
7
. In this experiment, a lack of fact-learning facilitation
would entail a decrement in composite performance compared to control animals. No
such decrement was observed
9
. Thus, current data are consistent with the hypothesis
that the MB mediate inhibition of skill-learning in order to generalize learned facts and
are not involved in the facilitation of fact-learning.
Encouraged by these results, we developed our hypothesis one step further. When
the colours were removed after composite training, flies with impaired MB function
stereotypically continued their attempts to turn in the unpunished direction, despite the
change in the environment. Accumulating evidence suggests that the inhibitory nature
of the MB allows them to serve a gating function
12,13,16
, preventing all but the most
important events from forming memories. In our case, one may interpret the results as
the MB preventing the formation of a motor memory or habit. Skills and habits are
important for efficiently carrying out often-repeated behaviours by limiting the amount
of behavioural variability. If our simple skill-learning paradigm indeed were an
adequate model for studying habit formation in flies, extended composite training
should overcome the MB-mediated inhibition and lead to stereotyped turning attempts
and abolished generalization, much as in the transgenic flies. Remarkably, wildtype flies
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
6
trained for twice the regular amount of time in the composite task indeed perform as
phenocopies of the flies with impaired MB function (Fig. 2c).
Our results allow for the first time to establish a mechanistic model of how active
and passive learning systems interact in composite learning situations and which
biological substrates mediate the processes resulting in the generation effect (Fig. 3).
Acquisition of the rut-dependent fact-learning component suppresses acquisition of the
rut-independent skill-learning component via the MB. The skill-learning component
facilitates fact-learning via still unknown, non-MB pathways. This interaction leads to
efficient learning, enables generalization and prevents premature habit-formation. Habit
formation after extended training reveals the gate-keeping role of the MB, allowing only
well-rehearsed behaviours to consolidate into habits. Despite these advances, we still do
not know what specific mechanisms lead to the enhancement of learning in the
generation effect. However, with this new set of behavioural tools and the molecular
genetic arsenal of Drosophila, it is only a matter of time until we see progress in this
area as well. Lacking any direct evidence, an attractive speculation is that the operant
behaviour serves to focus the fly's attention, to more quickly detect coincidences
between stimuli. Recent work shows that flies modulate their attention, can focus it to
different areas of their visual field and that these attention-like processes require short-
term memory genes
17-19
. It is also still unknown what molecular cascades mediate skill-
learning. There is preliminary but converging evidence from mice and the marine snail
Aplysia that one critical molecular component of skill-learning is a calcium-independent
but dopamine-dependent adenylyl cyclase acting upstream of protein kinases A and
C
20,21
.
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
7
Methods summary:
Wild-type strain Berlin (WT) and rutabaga mutant strain rut
2080
were used for this
study. Experimental transgenic flies were obtained by crossing a MB-specific GAL4
driver strain (mb247) to an effector strain expressing the catalytic subunit of bacterial
tetanus toxin (UAS
GAL4
-TNT). This cross results in a block of synaptic output in the
MB neurons targeted by the driver strain. The heterozygote offspring from crossing
driver and reporter strain, respectively, to Canton-S wildtype flies served as genetic
controls for these experiments. As both crosses were tested simultaneously and their
results did not differ, both control groups were pooled. After briefly immobilizing 24-
48h old female flies by cold-anesthesia, the flies were glued with head and thorax to a
triangle-shaped copper hook the day before the experiment. The animals were then kept
individually overnight in small moist chambers containing a few grains of sucrose. The
apparatus for dissecting learning-by-doing in Drosophila is shown in Fig. 1a, b. The
tethered fly, suspended at a torque meter, is flying stationarily in the centre of an arena
that is illuminated from behind. The torque meter records the attempts of the fly to turn
around its vertical body axis (yaw torque). For green and blue illumination of the arena,
the light is passed through monochromatic broad band filters. Filters can be exchanged
by a fast solenoid within 0.1s. Alternatively, the arena is illuminated with `daylight' by
passing it through a blue-green filter. Yaw torque is recorded every 50 ms and direction
preferences are calculated for nine (extended 15) consecutive 2-min periods
(performance index (PI) 19; Fig. 1c). During training, one yaw torque/colour
combination is paired with ambient temperature and the other with heat from an infrared
laser diode. If t
A
is the time the fly spends in one of the two situations, and t
B
the time in
the other, the performance index is calculated as PI = (t
A
- t
B
)/(t
A
+ t
B
). Error bars in the
figures are s.e.m.; asterisks indicate levels of significance against zero (one-sample t-
test; two-sided P-value). For details see Supplementary Methods.
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
8
References:
1.
Baden-Powell, R. Scouting for Boys. (C. Arthur Pearson Ltd,, London, 1908).
2.
James, W. The Principles of Psychology. (Holt, New York, 1890).
3.
Thorndike, E. Animal Intelligence. An Experimental Study of the Associative
Processes in Animals. (Macmillan, New York, 1898).
4.
Slamecka, N. J. & Graf, P. Generation Effect - Delineation of a Phenomenon. J.
Exp. Psychol. [Hum. Learn]. 4, 592-604 (1978).
5.
Kornell, N. & Terrace, H. S. The Generation Effect in Monkeys. Psychol. Sci.
18, 682-685 (2007).
6.
McVea, D. A. & Pearson, K. G. Stepping of the forelegs over obstacles
establishes long-lasting memories in cats. Curr. Biol. 17, R621-R623 (2007).
7.
Brembs, B. & Heisenberg, M. The operant and the classical in conditioned
orientation in Drosophila melanogaster at the flight simulator. Learn. Mem. 7, 104-115
(2000).
8.
Arrow, K. J. The Economic Implications of Learning by Doing. Rev. Econ.
Stud. 29, 155-173 (1962).
9.
Wolf, R. et al. Drosophila mushroom bodies are dispensable for visual, tactile
and motor learning. Learn. Mem. 5, 166-178 (1998).
10.
Maye, A., Hsie, C.-h., Sugihara, G., & Brembs, B. Order in spontaneous
behavior. PLoS One 2, e443 (2007).
11.
Wolf, R. & Heisenberg, M. Basic organization of operant behavior as revealed
in Drosophila flight orientation. J. Comp. Physiol. A Neuroethol. Sens. Neural. Behav.
Physiol. 169, 699-705 (1991).
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
9
12.
Liu, L., Wolf, R., Ernst, R., & Heisenberg, M. Context generalization in
Drosophila visual learning requires the mushroom bodies. Nature 400, 753-756 (1999).
13.
Brembs, B. & Wiener, J. Context generalization and occasion setting in
Drosophila visual learning. Learn. Mem. 13, 618-628 (2006).
14.
Martin, J.-R., Ernst, R., & Heisenberg, M. Mushroom Bodies Suppress
Locomotor Activity in Drosophila melanogaster. Learn. Mem. 5, 179-191 (1998).
15.
Huber, F. in Acoustic Behavior of Animals, edited by RG Busnel (Elsevier,
Amsterdam, 1963), pp. 440-488.
16.
Zhang, K. et al. Dopamine-Mushroom Body Circuit Regulates Saliency-Based
Decision-Making in Drosophila. Science 316, 1901-1904 (2007).
17.
van Swinderen, B. & Greenspan, R. J. Salience modulates 20-30 Hz brain
activity in Drosophila. Nat. Neurosci. 6, 579-586 (2003).
18.
Tang, S. M., Wolf, R., Xu, S. P., & Heisenberg, M. Visual pattern recognition in
Drosophila is invariant for retinal position. Science 305, 1020-1022 (2004).
19.
van Swinderen, B. Attention-Like Processes in Drosophila Require Short-Term
Memory Genes. Science 315, 1590-1593 (2007).
20.
Kheirbek, M. A., Beeler, J. A., Ishikawa, Y., & Zhuang, X. in Annual meeting
of the Society for Neuroscience (San Diego, Ca., USA, 2007).
21.
Lorenzetti, F. D., Baxter, D. A., & Byrne, J. H. in Annual meeting of the Society
for Neuroscience (Atlanta, Ga. USA, 2006).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
10
Acknowledgments: Fly strains were provided by Martin Heisenberg, Hiromu Tanimoto and Scott
Waddell. Tilman Franke created the 3D renderings of the experimental setup using Moray and Pov-Ray.
Graduate students Gretel Wittenburg, Michael Vogt and Joshua Lilvis ensured that the manuscript was
understandable by non-specialists.
Author contributions: B.B. designed and performed the experiments, analyzed the data and wrote the
manuscript. Undergraduate student W.P. discovered that rut flies are not defective in yt-learning and is
honorary co-author in recognition of his discovery. W.P. has never seen the manuscript before
acceptance.
Reprints and permissions information is available at npg.nature.com/reprintsandpermissions
Correspondence and requests for materials should be addressed to B.B, (bjoern@brembs.net).
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
11
Fig. 1: Drosophila composite learning at the torque meter.
a-b, Experimental setup. The fly is attached to a torque meter and its yaw
torque, generated by attempted left and right turns, controls the colour of the
panorama around it as well as a punishing beam of infrared light. The coloration
of panorama illumination is accomplished by a fast solenoid moving two colour
filters at the light source such that only light of a specific wavelength (either
green or blue) illuminates the panorama at any given time. For instance, right
turning may lead to green illumination of the panorama and heat off (a), while
left turning may lead to blue illumination and heat on (b). c, Course of
experiment. Bars show performance indices (PI) of successive 2-min intervals
of pre-test (yellow bars; PI
1
, PI
2
), training (orange bars; PI
3
, PI
4
, PI
6
, PI
7
) and
memory test (yellow bars; PI
5
, PI
8
, PI
9
). A PI of 1 means the fly spent the entire
period in the unpunished situation, whereas a PI of -1 indicates that the fly
spent the entire period in the situation associated with heat. Accordingly, a PI of
zero indicates that the fly distributed the time evenly between heated and non-
heated situations. Therefore, PIs were tested against zero for statistical
significance. The following bar graphs all show the first PI after the last training
period (hatched bar). Error bars are s.e.m. throughout.
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
12
Fig. 2: Experiments with wildtype, mutant and transgenic flies reveal
hierarchical interactions between fact- and skill-learning.
a1, Abolished composite and unaffected skill-learning
rut mutant flies (red,
composite: t
16
=0.7, p<0.5; light-green, skill-learning: t
16
=4.3, p<0.001). After
composite training, the skill-learning component is significant (dark green:
t
29
=2.9, p<0.007) indicating skill-learning inhibition at the level of retrieval. a2,
Significant composite and skill-learning in wildtype (WT) flies (composite:
t
31
=5.1, p<0.001; skill-learning: t
29
=3.0, p<0.006). After composite training, the
skill-learning score is not significant (t
24
=-0.3, p<0.8) indicating inhibition of skill-
learning during acquisition. b1, Flies expressing the bacterial tetanus toxin light
chain in most mushroom-body intrinsic Kenyon cells perform well in composite
learning (red: t
19
=3.1, p<0.01), but do not inhibit the skill-learning component
during composite training (green: t
18
=2.6, p<0.05). Without inhibition of skill-
learning, these transgenic flies are unable to generalize the colour memory
(blue: t
20
=-0.5, p<0.6.). b2. The genetic control flies (the two heterozygote
strains did not differ and were pooled) reproduce the wild-type results:
significant composite learning (t
26
=3.8, p<0.001), inhibition of skill-learning
(t
31
=0.7, p<0.5) and successful generalization (t
14
=2.7, p<0.05). c, Extended
training overcomes the inhibition of skill-learning in wildtype flies. The results
constitute a phenocopy of the transgenic animals (b1). Extended composite
learning does not lead to an overtraining decrement (t
16
=2.8, p<0.013). Testing
for the skill-learning component after extended composite training shows a
release from the inhibition of skill learning (t
16
=2.6, p<0.02). Without inhibition of
skill-learning, the flies are unable to generalize (t
19
=0.1, p<0.91).
Grey shading mutant or experimental animals. No shading WT or control
animals. Numbers at bars number of animals. * statistical significance.
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
13
Fig. 3: Composite learning consists of two components with reciprocal,
hierarchical interactions.
The rut-independent skill-learning component facilitates acquisition of the rut-
dependent fact-learning component (generation effect) via unknown, non-
mushroom-body pathways. This facilitated fact-learning inhibits acquisition of
skill-learning via the mushroom-bodies. These interactions lead to efficient
learning, generalisation and prevent premature habit-formation.
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007
Nature Precedings : hdl:10101/npre.2007.1354.1 : Posted 20 Nov 2007