letters to nature

two-dimensional models, the effective mass falls steeply as wavelength decreases. In three-dimensional models, in a fairly broad
frequency range near the response peak, the effective mass becomes
substantially less sensitive to wavelength14,15. In light of this behaviour, and the intuitive argument above, the observed lack of
frequency sensitivity in the penetration depth does not seem all
that surprising.
M

Acknowledgements
I thank H. Nakajima, E. de Boer, L. Sohn, N. Cooper, S. Staggs, R. Brawer, R. Austin and
L. Page Jr. This work was supported by the National Institute on Deafness and Other
Communication Disorders.

Methods
Pressure sensor
A sensor consists of a glass capillary (inner and outer diameters 100 and 170 mm,
respectively) tipped with a gold-coated polymer diaphragm. Light from an LED is
delivered via a bre optic threaded into the capillary, and reects from the diaphragm. The
amount of light returning to the bre optic for transmission to a photodetector varies
linearly with the pressure-induced bending of the diaphragm. The acoustic impedance of
the sensors is an order of magnitude larger than that of the cochlea, and their presence does
not `load' the cochlea outright. Nevertheless, in some animals small reversible changes in
compound action potential (CAP) threshold and/or scala vestibuli pressure occurred
when the scala tympani sensor was close to the basilar membrane. However, the derived
uid velocity close to the basilar membrane was similar to that measured directly by
others11,12, suggesting that the sensor's presence does not cause large changes in cochlear
mechanics.

Experimental procedure
Animal procedures were approved by the Princeton University IACUC. The experimental
animals were young adult gerbils. A gerbil was deeply anaesthetized and its left cochlea was
exposed. Tones from a loudspeaker were delivered to the ear via a tube tted to the left ear
canal. The level of the tones was calibrated in the ear canal at the beginning of each
experiment. Basal scala tympani pressure measurements were made by inserting a pressure
sensor through the round window opening after removing the covering membrane. Scala
vestibuli and turn-one scala tympani measurements were made through small holes handdrilled in the cochlear bone. In basal experiments it was possible to see the basilar
membrane in order to position the scala tympani sensor. In turn-one experiments it was
positioned by using anatomical landmarks and referring to widely opened excised
cochleae. It was not practical to systematically check the precision of positioning in each
experiment. However, the grouped data indicate that incorrectly positioning the sensor
towards the spiral ligament caused substantial damage, and that incorrectly positioning
the sensor over the spiral lamina caused greatly diminished pressure gradients. The
distance between the basilar membrane and the sensor was determined by touching the
former with the latter, which produced a characteristically noisy signal.
Stimulus generation and recording was performed with a Tucker Davis Technologies
DA/AD system. With typical signal averaging times of 3 s, sound pressure above the level of
6070 dB SPL (2060 mPa) could be reliably measured.
As a gauge of cochlear health, an electrode at the round window measured the CAP
response of the auditory nerve to tones. The CAP threshold is the minimum sound level
required to elicit a reliable neural response. Initial thresholds in the turn one experiment
were <30 dB SPL at 15 and 20 kHz, and 50 dB SPL at 25 kHz. Close to the time of
measurements, these thresholds were 40, 40, and 5060 dB SPL. Initial thresholds in the
basal experiment were <60 dB SPL at 30 kHz, and 80 dB SPL at 40 kHz. Close to the time
of measurements, they were 60 and .80 dB SPL. At the frequencies of interest, the initial
CAP thresholds in the turn one experiment were in keeping with those of Muller6, those in
the basal experiment were elevated somewhat. The degree of nonlinearity observed in the
scala tympani pressures was consistent with the health of the cochlea as indicated by the
CAP thresholdthe basal experiment was nearly linear, and the nonlinearity in the turnone experiment was strong, but not among the strongest in the literature9.
Received 1 June; accepted 4 October 1999.
1. von Bekesy, G. Experiments in Hearing (McGraw Hill, New York, 1960).
2. Olson, E. S. Observing middle and inner ear mechanics with novel intracochlear pressure sensors.
J. Acoust. Soc. Am. 103, 34453463 (1998).
3. Brownell, W. E., Bader, C. R., Betrand, D. & de Ribaupierre, Y. Evoked mechanical responses of
isolated cochlear outer hair cells. Science 227, 195196 (1985).
4. Liberman, M. C. & Dodds, L. W. Single neuron labeling and chronic cochlear pathology. III.
Stereocilia damage and alterations of threshold tuning curves. Hear. Res. 16, 5574 (1984).
5. Spoendlin, H. Innervation densities of the cochlea. Acta Otolaryngol. (Stockh.) 73, 235248 (1972).
6. Muller, M. The cochlear place-frequency map of the adult and developing mongolian gerbil. Hear. Res.
94, 148156 (1996).
7. Dancer, A. & Franke, R. Intracochlear sound pressure measurements in guinea pigs. Hear. Res. 2, 191
205 (1980).
8. Rhode, W. S. Observations of the vibration of the basilar membrane in squirrel monkeys using the
Mossbauer Technique. J. Acoust. Soc. Am. 49, 12181231 (1971).
9. Ruggero, M., Rich, N. C., Recio, A., Narayan, S. S. & Robles, L. Basilar membrane responses to tones at
the base of the chinchilla cochlea. J. Acoust. Soc. Am. 101, 21512163 (1997).
10. Lighthill, J. Energy ow in the cochlea. J. Fluid Mech. 106, 149213 (1981).
11. Sellick, P. M., Yates, G. K. & Patuzzi, R. The inuence of Mossbauer source size and position on phase
and amplitude measurements of the guinea pig basilar membrane. Hear. Res. 10, 101108 (1983).
12. Xue, S., Mountain, D. C. & Hubbard, A. E. Electrically evoked basilar membrane motion. J. Acoust.
Soc. Am. 97, 30303041 (1995).
13. Cooper, N. P. & Rhode, W. S. Basilar membrane mechanics in the hook region of cat and guinea-pig

NATURE | VOL 402 | 2 DECEMBER 1999 | www.nature.com

cochlea: Sharp tuning and nonlinearity in the absence of baseline position shifts. Hear. Res. 63, 163
190 (1992).
14. de Boer, E. Auditory physics. Physical principles in hearing theory. II. Phys. Rep. 105, 141226 (1984).
15. Steele, C. R. & Taber, L. A. Comparison of WKB calculations and experimental results for threedimensional cochlear models. J. Acoust. Soc. Am. 65, 10071018 (1979).
16. Plassman, W., Peetz, W. & Schmidt, M. The cochlea in gerbilline rodents. Brain Behav. Evol. 30, 82
101 (1987).

Correspondence and requests for materials should be addressed to E.O.

(e-mail: eolson@princeton.edu).

.................................................................
Stable propagation of synchronous
spiking in cortical neural networks
Markus Diesmann*, Marc-Oliver Gewaltig* & Ad Aertsen
Department of Neurobiology and Biophysics, Institute of Biology III,
Albert-Ludwigs-University, Schanzlestrae 1, D-79104 Freiburg, Germany
..............................................................................................................................................

The classical view of neural coding has emphasized the importance of information carried by the rate at which neurons
discharge action potentials. More recent proposals that information may be carried by precise spike timing15 have been challenged by the assumption that these neurons operate in a noisy
fashionpresumably reecting uctuations in synaptic input6
and, thus, incapable of transmitting signals with millisecond
delity. Here we show that precisely synchronized action potentials can propagate within a model of cortical network activity
that recapitulates many of the features of biological systems. An
attractor, yielding a stable spiking precision in the (sub)millisecond range, governs the dynamics of synchronization. Our results
indicate that a combinatorial neural code, based on rapid associations of groups of neurons co-ordinating their activity at the
single spike level, is possible within a cortical-like network.
Evidence is accumulating that cortical neurons in vivo are capable
of producing action potentials with high temporal accuracy. In
recordings of multiple single-neuron activity in behaving monkeys,
precisely timed action potentials have been systematically related to
stimuli and behavioural events, indicating that these instances of
precise spike timing play a functional role13. Independent evidence
for precise spike timing in cortical neurons came from intracellular
recordings in vitro4,5. But can an instance of synchronous spiking,
once it has occurred, be successfully propagated by subsequent
groups of cortical neurons? Under which input conditions can a
group of cortical neurons engage in precisely coordinated spike
timing, and are such conditions feasible in the cortical network?
How can we clarify and quantify the notions of `well timed' and
`reliable', which gained such a prominent role in the on-going debate
on temporal coding in the brain?710 Clearly, these questions must be
resolved to determine whether cortical computation on the basis of
precise spike timing is possible. Preliminary results have been
presented in abstract form11,12.
To address these questions, we have studied the ne-grained
temporal response properties of the `integrate-and-re' neuron, a
widely used class of model neurons capturing essential properties of
* Present addresses: Department of Nonlinear Dynamics, Max-Planck-Institut fur Stomungsforschung,
Gottingen, Germany (M.D.); Future Technology Research, Honda R&D Europe, Offenbach, Germany
(M.-O.G.)

1999 Macmillan Magazines Ltd

529

letters to nature
a

in

a in

Time (ms)

70

Time (ms)

the ring behaviour of cortical neurons13,14. We focused on spike

responses to transient membrane-potential excursions, implied by
the physiological ndings15. As a rule, such transients are explained
by convergent inputs from simultaneously spiking neurons onto a
target neuron (Fig. 1a). These transients, in turn, result in well timed
response spikes in target neurons. Neurons that share a large enough
pool of simultaneously discharging input cells tend to align their
action potentials15 (Fig. 1b). By repeating this arrangement, a group
of neurons can reproduce its synchronous input activity and act as
the source of synchronous shared input to the following group (Fig.
1c). This idea of connecting groups of neurons into feed-forward
arrangements16 was formalized by Abeles in the `synre chain'17. The
degree of temporal accuracy of spike times among the group's
members determines whether subsequent groups can reproduce
(or even improve) this accuracy (Fig. 1d), or whether synchronous
excitation disperses and eventually dies out (Fig. 1e). Thus, in the
context of cortical network function, the timing precision of a
neuron's action potentials is measured vis-a-vis the timing of those
of its companion neurons; the quality of timing is judged on the
effect the group's activity has in the network, that is, whether
synchronous spiking is sustained or whether it dies out.
Existing measures of short-term dynamics in neural transmission
focus on two extreme cases of input activity: full synchrony and
random arrival of spikes1820. Intermediate cases with limited degree
of temporal dispersion are generally not addressed. `Pulse
packets'11,12 were introduced to overcome this restriction and to
quantify the degree of synchrony in a propagating spike volley. A
pulse packet characterizes a spike volley by two parameters: activity,
a, and temporal dispersion, j. Activity is dened as the number of
spikes in the volley; their temporal dispersion is measured by the
standard deviation of the underlying pulse density (Fig. 2a). Thus,
in simulations we measured the response of a cortical model neuron
(see Methods) to different pulse-packet inputs in the presence of
background activity (Fig. 2b). The neuronal transmission function
for transient input activity is dened by the transformation of the
input pair (ain,jin) into the output pair (a,jout), where a is the single
neuron response probability. The ring probability curves (Fig. 2c)

out

5
4

0.8

Figure 1 Propagation of synchronous spiking in cortical networks. Effect (a) and

generation (b) of tight spike correlation by convergence of synchronous shared input onto
target neurons (simultaneously recorded neurons are highlighted by black circles).
c, Combination of a and b (no intragroup connections) generates synchronous spike
volley. Propagation across neuron groups by repeating the arrangement (`synre chain'17)
can be stable (d) or unstable (e). d, e, Raster displays of propagating spike volley along
fully connected synre chain. Panels show the spikes in 10 successive groups of 100
neurons each (synaptic delays arbitrarily set to 5 ms). Initial spike volley (not shown) was
fully synchronized (j 0 ms), containing a 50 (d) or a 48 (e) spikes; background
activity identical in both cases.

530

out (ms)

70

1
2
3
4
5
6
7
8
9
10

1
2
3
4
5
6
7
8
9
10

Group

Group

0.6
0.4
0 ms
1 ms
3 ms
5 ms

0.2
0
0

50
100
a in (spikes)

3
2
45
65
105
145

1
0
150

in (ms)

Figure 2 Neural transmission function for pulse-packet input. a, Pulse-packet

representation of propagating spike volley. b, Cortical model neuron, stimulated with
pulse packets by repeatedly drawing ain spike times from Gaussian pulse density with
standard deviation jin (green). Pulse packets were embedded in random spikes,
mimicking background activity. The neuron responded with one spike at most, its timing
varying trial by trial. The temporal distribution was measured by trial averaging (red). Net
area under curve, spike probability a; standard deviation, response jitter jout. c, Spike
probability versus input spike number for different degrees of input synchrony.
d, Temporal accuracy of spike response versus input synchrony for different input spike
numbers. The diagonal (dashed grey line) marks points with equal temporal precision of
input and response.

resemble the well known sigmoid activation function, the slope is

determined by the degree of input synchrony (the saturation slightly
below 1 reects the chance that the neuron was refractory on
stimulus arrival). As expected, the spread of the response distribution (Fig. 2d) increases with the input spread; however, the slope is
less than 1, hence output spread increases more slowly than input
spread. In addition, the curves show an offset; even for fully
synchronized volleys (jin 0 ms), some residual jitter of the
response spike remains, reecting the inuence of background
activity. Hence, each dispersion curve crosses the diagonal at
some critical value of input synchrony. Up to this intersection,
the neuron's response is less precise than the input, that is,
synchronous input is desynchronized. Beyond the intersection,
however, the neuron's spike response is more precise than the
input, that is, the neuron exhibits a synchronizing behaviour.
Duplication of the experiment with identical input pulse-packet
realizations across trials yielded essentially the same results, conrming that trial-by-trial response variability is due to uctuations
in background activity21.
We used this neuronal transmission function to test whether the
cortical network is capable of sustaining synchronous spiking
activity. As each neuron responds to an incoming pulse packet
with at most one spike, stable propagation of synchronous spike
volleys inevitably requires the activation of successive, large enough
groups of neurons (Fig. 1c). For a group of identical independent
neurons, the distribution of response spikes to an input pulse packet
is identical to the response distribution for a single neuron (Fig. 2).
Thus, the spread of the group's response equals the single neuron's
response dispersion jout (Fig. 2d). The expected number of response
spikes aout in a group equals a (Fig. 2c) multiplied by the group size
w. Figure 3a, b (grey curves) shows the inputoutput relation for a
group of w 100 neurons. Assuming that the group's response to a

1999 Macmillan Magazines Ltd

NATURE | VOL 402 | 2 DECEMBER 1999 | www.nature.com

letters to nature
the state space into two regimes. In the basin of attraction, all
trajectories converge into the attractor. A spike volley starting
anywhere inside this regime rapidly (that is, after only few stages)
reaches a stable conguration (,95 spikes) with submillisecond
dispersion. Volleys starting outside the stable regime decay after
only few stages; too weak or too dispersed activity rapidly dies out.
Note that neither the relationship between input and output activity
(Fig. 3a) nor that between input and output jitter10 (Fig. 3b) alone
determines whether synchronous activity survives. The nonmonotonic evolution of these variables along a trajectory (Fig. 3c,
coloured curves) demonstrates this fact. An initial increase in
temporal spread may still support stable propagation, provided
that the number of spikes in the volley is large enough (blue). If,
however, this number is too small, the volley dies out, in spite of its
initial increase (purple). Conversely, synchronous activity may still
vanish with an initial decrease in dispersion (red), unless the volley
is large enough (green). Thus, the system dynamics are governed by
the interaction of the two state variables.
Evidently, the number of neurons per group inuences the
evolution of the activity (see Fig. 3a). The analysis in Fig. 3 was
made using the group size w 100. To determine how many
simultaneously ring neurons are needed to guarantee that synchronous activity survives in the network, we examined how the
structure of the state space depends on the groups size (Fig. 4). For
increasing numbers of neurons per group, the two xpoints move
apart, thereby increasing the basin of attraction (yellow). The
regime over which synchronous spiking survives in the network
increases accordingly. By contrast, for decreasing groups size, the
two xpoints approach each other until, at some critical value (here
w 89), they merge into a single saddle node. Below this critical
value, no xpoint exists and, hence, all trajectories lead to extinction
(Fig. 4a). Thus, a minimum of some 90 neurons per group is needed
to maintain precise spike synchrony. This lower bound is essentially
determined by the ratio of the distance from mean membrane
potential to spike threshold and the postsynaptic potential (PSP)
amplitude; stronger intergroup synapses reduce this number (see
Methods: Fluctuations).
Our results show that for a wide range of amplitudes and
dispersions of spike input distributions, the response of successively
activated groups of cortical neurons is governed by an attractor,
which describes a stationary conguration of activity in (a,j) space.
However, unlike the Hopeld attractor22, this attractor describes a
dynamic activity conguration in neuron space, that is, different
neuron groups, one after the other, contribute single spikes to the
propagating synchronous wave. The basin of attraction guarantees
robustness of the propagating synchrony against perturbations
exceeding the response variability accounted for by the transmission
function (Fig. 2). Thus, temporal dispersion due to differences in
axonal or dendritic delays, uctuations in synaptic transfer properties or correlated background uctuations will not destroy the
synchronous transmission, as long as they do not push the network
outside the basin of attraction.

b 3

100
80

out (ms)

aout (spikes)

60
40

20
0

0
0

20

40 60 80
a in (spikes)

100

in (ms)

100

a (spikes)

80

60

40

20

0
0

(ms)

Figure 3 State-space analysis of propagating spike synchrony. a, b, Grey curves,

transmission function for pulse-packet input for a group of w 100 neurons. Coloured
curves, evolution of synchronous spike volley along a chain of such groups (Fig. 1c) for
different input congurations. Arrows, group to group transformation. Two cases of stable
propagation are shown, one starting with few, fully synchronized spikes (blue; see Fig.
1d), the other with many, weakly synchronized spikes (green), and two unstable cases
where synchronous activity rapidly dies out (red, purple; see Fig. 1e). c, State space
portrait of dynamic variables a and j. Coloured trajectories portray the four cases in a, b.
The stable (upper) and unstable (lower) regimes are separated by the separatrix (dashed
line).

spike volley is adequately described by a and j (see Methods:

Fluctuations), we can capture the evolution of synchronous spiking
activity by repeatedly applying the (a,j)-transformation while
moving from one group to the next (Fig. 3a, b, coloured curves).
Thus, the evolution of synchronous activity on its course through
the network is described by a trajectory in the 2-dimensional state
space, spanned by a and j (Fig. 3c).
The state-space portrait exhibits two xpoints: an attractor
(a 95, j 0:5 ms), and a saddle point (a 65, j 1:2 ms). A
separatrix (dashed line) running through the saddle point divides

a (spikes)

100
80
60
40
20

0
0

w = 80
1
(ms)

3 0

w = 90
1

(ms)

Figure 4 Dependence of the spike synchrony state space on the size of the neuron
groups. ad, State-space portraits of propagating spike synchrony for four different
group sizes w, increasing from w 80 (a) in steps of 10 up to w 110 (d). In each
NATURE | VOL 402 | 2 DECEMBER 1999 | www.nature.com

3 0

w = 100
1

(ms)

3 0

w = 110
1

(ms)

panel, the continuous red line indicates the a-isocline, the red dashed line indicates the jisocline (see Methods: Isoclines). The intersections of the two isoclines dene the
xpoints. The basin of attraction is indicated in yellow.

1999 Macmillan Magazines Ltd

531

letters to nature
In the stable state, essentially all response spikes in a volley fall
within 61 ms. This temporal precision is consistent with the
accuracy of observed spike patterns in cortical recordings13.
Thus, in contrast to other studies using this neuron model7,8, we
conclude that precise synchronous ring of cortical neurons is
indeed feasible, in spite of the membrane time constant of 10 ms
or more (see Methods: Membrane time constant). The notion of
pulse packets yields a natural solution to the question of whether the
cortical neuron acts as `integrator' or `coincidence detector'a
question raised many years ago18 and revived recently710 by
embedding these two conceptions in a single framework. The
temporal structure of the input determines which of the two aspects
is emphasized.
For each neuron in our network model, the synapses delivering
activity from the preceding group and the synapses delivering
background activity were set at equal strength. Our ndings show
that stable transmission of synchronous spiking does not require
dedicated strong synapses, provided enough neurons can be
recruited in successive groups. Each neuron in a group contributes
a single spike to the passing volley. Once the neuron recovers from
refractoriness, it is ready to engage in another group. Thus, each
neuron may participatespike by spike, not limited by a specic
arrangement of synaptic weightsin multiple volleys with different
neuron compositions, provided its engagements differ by more than
the refractory period. Hence, the network only needs to be locally
feed forward. Several such volleys may propagate through the
network simultaneously, allowing multiple synchronous processes
to coexist while maintaining their identities23,24. Their degree of
temporal coupling may differ, depending on overlap (numbers and
arrangement of shared neurons and inputs)23,25. This scheme indicates that a combinatorial neural code, based on the continual
reconguration of the cortical network into short-lived functional
groups depending on the immediate computational demands, is
compatible with biological constraints.
M

Methods

Fluctuations
As conrmed by network simulations (M.-O.G., manuscript in preparation), the statespace portrait derived here describes the evolution of synchronous activity in the mean,
that is, by subsequent values of the expectation (a,j) across trials with different background activity realizations. Around each point of a trajectory, these realizations form a
distribution with width determined by a, j, w and intergroup connectivity. This width
becomes more important near the separatrix because of the increased probabilityeven
for trajectories stable in the meanthat individual realizations leave the basin of
attraction (and vice versa). Upscaling the synaptic weights by a factor up to 10 while
downscaling the groups size accordingly does not alter the structure of the state space,
except close to the separatrix where the probability to leave the basin of attraction
increases. The contribution of individual input spikes grows; consequently, uctuations in
membrane-potential response to pulse-packet realizations with identical parameters and,
hence, trial-by-trial variability, increases.

Background activity
Background activity in different neurons was considered independent, stationary Poisson.
However, on-going cortical activity is known to exhibit coherent spatio-temporal
structure30. Hence, anatomically nearby neurons within a group tend to be excited
(inhibited) together. This affects the pulse-packet properties needed to make these
neurons re simultaneously. The impact of such coherence in background activity is
currently being studied.

Membrane time constant

The temporal precision of spike response is not constrained by membrane time constant
itself; the limiting factor is the up slope of the excitatory PSP. The larger this slope, the
faster the membrane-potential response to a pulse packet traverses the threshold region17.
This reduces the chance of interference with background uctuations, which degrade
response spike precision. The membrane time constant does, however, limit transmission
of synchronous spikes in the opposite way. It determines the integration time window of
the receiving neuron, limiting the extent over which a spike volley is `seen' as a single
packet, rather than as individual spikes. For a too-small membrane time constant, PSPs no
longer overlap and cannot add up to threshold. Reliable transmission of incompletely
synchronized spike volleys, therefore, requires a minimal (rather than maximal) membrane time constant, in the order of the volley duration.
Received 17 June; accepted 23 August 1999.

Model neuron
Simulations were performed using a leaky-integrator with voltage-threshold model13,14,
with physiological and anatomical parameters taken from experimental literature. The
model neuron (membrane time constant 10 ms, resting potential -70 mV, spike threshold
-55 mV, absolute refractoriness 1 ms, relative refractoriness (t < 15 ms) modelled by Kconductances) was supplied with synaptic noise input, reecting on-going activity in the
cortical network (20,000 synapses: 88% excitatory, 12% inhibitory)26. Postsynaptic
currents (PSCs) were modelled by an a-function to yield realistic PSPs (peak amplitude
0.14 mV, time-to-peak 1.7 ms, half-width 8.5 ms)27. Identical values were used for intergroup and background connections; excitatory and inhibitory PSPs only differed in sign.
Background ring rates (excitatory, 2 Hz; inhibitory, 12.5 Hz; all uncorrelated stationary
Poisson) were chosen to yield an output rate of 2 Hz. At this consistency condition, output
statistics were approximately Poisson, membrane potential shot noise (mean 8.25 mV, s.d.
2.85 mV) was close to `balanced' excitation/inhibition28. It can be shown that details of the
construction of background uctuations are not essential. Simulations were performed in
0.1 ms time steps using the simulation tool SYNOD29.

Threshold packet
The (ain,aout) curves (grey) in Fig. 3a cross the diagonal (dashed) in two points. Here, the
number of output spikes equals the number of input spikes. The lowest intersection (at the
left-most vertical line) occurs with the curve for fully synchronized input (jin 0),
dening a lower boundary on the size of the threshold packet, that is, the minimum input
spike number needed to reach the attractor. Smaller packets cannot survive, since for other
curves, the output spike number is even smaller. Decreasing the group's size rotates the
diagonal relative to the (ain,aout) curves counterclockwise around the origin; the intersection points approach each other and the threshold packet size increases. Thus, for
decreasing w, more curves fall below the diagonal until even the curve for fully
synchronized input only touches (short-dashed oblique line intersection with right-most
vertical line). For smaller groups, all curves run below the diagonal; the attractor vanished,
and stable propagation of synchronous spiking is no longer possible.

Isoclines
The a-isocline is the collection of states for which the spike number in a volley does not
change from stage to stage, irrespective of j (Fig. 4bd, solid red curves). The j-isocline
contains all states maintaining temporal spread, irrespective of a (Fig. 4ad, dashed red
curves). Thus, the isoclines are the loci of horizontal/vertical ow. The xpoints (neither a

532

nor j changes) are the intersections of the isoclines. The j-isocline is independent of w
(Fig. 3b). The spike number, however, is proportional to w (Fig. 3a). Hence, with
decreasing w (Fig. 4d to 4a), the a-isocline shrinks and moves left until it ceases to exist.
Thus, the group's size w acts as bifurcation parameter, controlling the existence and
separation of xpoints.

1. Abeles, M., Bergman, H., Margalit, E. & Vaadia, E. Spatiotemporal ring patterns in the frontal cortex
of behaving monkeys. J. Neurophysiol. 70, 16291638 (1993).
2. Prut, Y. et al. Spatiotemporal structure of cortical activity: Properties and behavioral relevance. J.
Neurophysiol. 79, 28572874 (1998).
3. Riehle, A., Grun, S., Diesmann, M. & Aertsen, A. Spike synchronization and rate modulation
differentially involved in motor cortical function. Science 278, 19501953 (1997).
4. Mainen, Z. F. & Sejnowski, T. J. Reliability of spike timing in neocortical neurons. Science 268, 1503
1506 (1995).
5. Nowak, L. G., Sanchez-Vives, M. V. & McCormick, D. A. Inuence of low and high frequency inputs
on spike timing in visual cortical neurons. Cereb. Cortex 7, 487501 (1997).
6. Calvin, W. H. & Stevens, C. F. Synaptic noise and other sources of randomness in motoneuron
interspike intervals. J. Neurophysiol. 31, 574587 (1968).
7. Shadlen, M. N. & Newsome, W. T. Noise, neural codes and cortical organization. Curr. Opin.
Neurobiol. 4, 569579 (1994).
8. Shadlen, M. N. & Newsome, W. T. The variable discharge of cortical neurons: Implications for
connectivity, computation, and information coding. J. Neurosci. 18, 38703896 (1998).
9. Konig, P., Engel, A. K. & Singer, W. Integrator or coincidence detector? The role of the cortical neuron
revisited. Trends Neurosci. 19, 130137 (1996).
10. Marsalek, P., Koch, C. & Maunsell, J. On the relationship between synaptic input and spike output
jitter in individual neurons. Proc. Natl Acad. Sci. USA 94, 735740 (1997).
11. Diesmann, M., Gewaltig, M.-O. & Aertsen, A. in Computational NeuroscienceTrends in Research
1995 (ed. Bower, J.) 5964 (Academic, San Diego, 1996).
12. Diesmann, M., Gewaltig, M.-O. & Aertsen, A. in From Membrane to Mind (eds Elsner, N. & Wassle,
H.) 62 (Thieme, Stuttgart, 1997).
13. Gerstein, G. L. & Mandelbrot, B. Random walk models for the spike activity of a single neuron.
Biophys. J. 4, 4168 (1964).
14. Tuckwell, H. C. Introduction to Theoretical Neurobiology (Cambridge Univ. Press, Cambridge, 1988).
15. Perkel, D. H., Gerstein, G. L. & Moore, G. P. Neuronal spike trains and stochastic point processes. II.
Simultaneous spike trains. Biophys. J. 7, 419440 (1967).
16. Grifth, J. S. On the stability of brain-like structures. Biophys. J. 3, 299308 (1963).
17. Abeles, M. Corticonics (Cambridge Univ. Press, Cambridge, 1991).
18. Abeles, M. Role of cortical neuron: integrator or coincidence detector? Isr. J. Med. Sci. 18, 8392
(1982).
19. Bernander, O., Koch, C. & Usher, M. The effect of synchronized inputs at the single neuron level.
Neural Comp. 6, 622641 (1994).
20. Murthy, V. N. & Fetz, E. E. Effects of input synchrony on the ring rate of a three-conductance cortical
neuron model. Neural Comp. 6, 11111126 (1994).
21. Arieli, A., Sterkin, A., Grinvald, A. & Aertsen, A. Dynamics of ongoing activity: explanation of the
large variability in evoked cortical responses. Science 273, 18681871 (1996).

1999 Macmillan Magazines Ltd

NATURE | VOL 402 | 2 DECEMBER 1999 | www.nature.com

letters to nature
22. Hopeld, J. J. Neural networks and physical systems with emergent collective computational abilities.
Proc. Natl Acad. Sci. USA 79, 25542558 (1982).
23. Bienenstock, E. A model of neocortex. Network 6, 179224 (1995).
24. Herrmann, M., Hertz, J. A. & Prugel-Bennett, A. Analysis of synre chains. Network 6, 403414
(1995).
25. Arnoldi, H.-M. R. & Brauer, W. Synchronization without oscillatory neurons. Biol. Cybern. 74, 209
223 (1996).
26. Braitenberg, V. & Schuz, A. Anatomy of the Cortex (Springer, Berlin, 1991).
27. Fetz, E., Toyama, K. & Smith, W. in Cerebral Cortex Vol. 9 (eds Peters, A. & Jones, E. G.) 147 (Plenum,
New York, 1991).
28. van Vreeswijk, C. & Sompolinsky, H. Chaos in neuronal networks with balanced excitatory and
inhibitory activity. Science 274, 17241726 (1996).
29. Diesmann, M., Gewaltig, M.-O. & Aertsen, A. SYNOD: An Environment for Neural Systems
Simulations Technical report GC-AA/95-3 (The Weizmann Institute of Science, Rohovot, Israel,
1995). (http://www.synod.uni-freiburg.de).
30. Arieli, A., Shoham, D., Hildesheim, R. & Grinvald, A. Coherent spatiotemporal patterns of ongoing
activity revealed by real-time optical imaging coupled with single-unit recording in the cat visual
cortex. J. Neurophysiol. 73, 20722093 (1995).

Acknowledgements
We thank M. Abeles, E. Bienenstock, S. Grun, I. Nelken, A. Riehle, S. Rotter and C. von der
Malsburg for their constructive comments. Supported in part by grants for the Deutsche
Forschungsgemeinschaft, the GermanIsraeli Foundation for Scientic Research and
Development, and Human Frontier Science Program..
Correspondence and requests for materials should be addressed to A.A.
(e-mail: aertsen@biologie.uni-freiburg.de).

.................................................................
Membrane-anchored aspartyl
protease with Alzheimer's
disease b-secretase activity

Riqiang Yan*, Michael J. Bienkowski, Mary E. Shuck, Huiyi Miao*,

Monica C. Tory, Adele M. Pauley, John R. Brashlerk,
Nancy C. Stratman, W. Rodney Mathews, Allen E. Buhlk,
Donald B. Carter, Alfredo G. Tomasselli#, Luis A. ParodiI,
Robert L. Heinrikson# & Mark E. Gurney
* Cell & Molecular Biology, Genomics, # Protein Sciences, k Pharmacology,
Structural, Analytical & Medicinal Chemistry and Neurobiology, Pharmacia
& Upjohn, Inc., 301 Henrietta Street, Kalamazoo, MI 49007, USA
I
Bioinformatics, Pharmacia & Upjohn, Inc., Lindhagensgatan 133,
S-11287 Stockholm, Sweden
These authors contributed equally to this work
..............................................................................................................................................

Mutations in the gene encoding the amyloid protein precursor

(APP) cause autosomal dominant Alzheimer's disease13. Cleavage
of APP by unidentied proteases, referred to as b- and gsecretases47, generates the amyloid b-peptide, the main component of the amyloid plaques found in Alzheimer's disease
patients8. The disease-causing mutations ank the protease cleavage sites in APP and facilitate its cleavage. Here we identify a new
membrane-bound aspartyl protease (Asp2) with b-secretase
activity. The Asp2 gene is expressed widely in brain and other
tissues. Decreasing the expression of Asp2 in cells reduces amyloid
b-peptide production and blocks the accumulation of the carboxy-terminal APP fragment that is created by b-secretase cleavage. Solubilized Asp2 protein cleaves a synthetic APP peptide
substrate at the b-secretase site, and the rate of cleavage is
increased tenfold by a mutation associated with early-onset
Alzheimer's disease in Sweden3. Thus, Asp2 is a new protein
target for drugs that are designed to block the production of
amyloid b-peptide peptide and the consequent formation of
amyloid plaque in Alzheimer's disease.
Visual inspection suggests that the b- and g-secretase cleavage
sites in APP might be substrates for cleavage by aspartyl proteases,
NATURE | VOL 402 | 2 DECEMBER 1999 | www.nature.com

and indeed, cathepsin D cleaves synthetic b-secretase substrates9.

This cleavage is facilitated by the KM ! NL mutation, referred to as
the `Swedish' mutation, found in patients with early-onset
Alzheimer's disease10; however, APP processing to amyloid b (Ab)
peptides occurs normally in hippocampal neurons cultured from
cathepsin-D-null mice11. Nevertheless, it seemed plausible that the
APP b- or g-secretases could be as yet uncharacterized aspartyl
proteases; therefore, we searched for new human enzymes of this
mechanistic set. Sequencing of the Caenorhabditis elegans genome
was nearing completion, which offered the possibility of enumerating the complete set of aspartyl proteases encoded in a simple
metazoan genome, and using these as a bridge to human sequence
databases.
Simple AWK scripts scanning for the D(S/T)G active-site motif,
PROSITE and hidden Markov models were used to search the
WormPep database of predicted C. elegans proteins. This revealed at
least 10 candidate aspartyl proteases. Seven of these ten were found
on a single chromosome, chromosome V (F21F8.3, F21F8.4,
F21F8.7, Y39B6B.G, Y39B6B.J, Y39B6B.H and T18H9.2), and three
each of these were found in the same cosmid clones (F21F8 and
Y39B6B), suggesting that they represent a recently evolved family of
proteins that arose by ancestral gene duplication. Other homologous predicted genes were found in the same cluster (F21F8.2,
F21F8.6 and Y39B6B.I); however, these contain only a single DTG or
DSG motif. Additional predicted aspartyl protease genes were
found on chromosomes IV (C11D2.2) and X (R12H7.2 and
H22K11.1). Searches of vertebrate expressed sequence tag (EST)
databases with the 10 C. elegans sequences identied 7 known and 4
new candidate aspartyl proteases. The new human sequences were
numbered in order of their discovery (Asp14). R12H7.2 and
H22K11.1 appear to be C. elegans homologues of cathepsin D.
Most of the chromosome V aspartyl proteases had no clear vertebrate orthologues; however, one of these (T18H9.2) bridged to two
unusual sequences (Asp1 and Asp2) which contained the less
common DSG motif in the second active site. In turn, C11D2.2
identied two additional sequences (Asp3 and Asp4) which have
since been reported in the literature as napsins A and B12.
The two predicted aspartyl protease sequences identied by
T18H9.2 were of greatest interest. Completion of their sequences
by a combination of EST sequencing, 59 rapid amplication of
complementary DNA ends by the polymerase chain reaction, and
library screening showed that both Asp1 and Asp2 had an unusual
C-terminal extension containing a single predicted transmembrane
domain (Fig. 1). Asp1 maps to human chromosome 21q22 within
the Down's syndrome critical region, and Asp2 to chromosome
11q2324. Northern hybridization to human tissue blots showed
widespread expression of both Asp1 and Asp2. Both are expressed at
the highest levels in pancreas. Asp2 is also expressed at high levels in
brain, whereas Asp1 is expressed in brain at somewhat lower levels.
In situ hybridization showed expression of Asp2 primarily in acinar
cells of the exocrine pancreas, whereas faint hybridization was seen
over neurons in hippocampus; however, we identied two Asp2 EST
in a human astrocyte cDNA library indicating that Asp2 may be
expressed in both neurons and glial cells. Transcripts for both Asp1
and Asp2 were expressed in human embryonic kidney 293 cells,
human IMR-32 neuroblastoma cells and mouse Neuro-2a neuroblastoma cells, three commonly used cellular models of APP
processing.
We used a panel of antisense oligomers to test the involvement of
each of the four predicted aspartyl proteases in APP processing by a
stable clone of HEK293 cells that had been engineered to process
APP to Ab peptides at high levels. These cells were transformed with
a modied human APP695 cDNA containing the Swedish
KM ! NL mutation to which two lysine residues had been added
to the C terminus (HEK/APP-Sw-KK cells). The KK motif greatly
increases the processing and release of Ab peptides but does not
inuence the ratio of Ab42=Ab42 Ab40, nor alter the effect of

1999 Macmillan Magazines Ltd

533

Copyright of Nature is the property of Nature Publishing Group and its content may not be copied or emailed to
multiple sites or posted to a listserv without the copyright holder's express written permission. However, users
may print, download, or email articles for individual use.