Goncalves Etal 2014

ORIGINAL RESEARCH ARTICLE
published: 28 February 2014

NEURAL CIRCUITS doi: 10.3389/fncir.2014.00010
Optogenetic perturbations reveal the dynamics of an

oculomotor integrator
Pedro J. Gonçalves 1,2,3† , Aristides B. Arrenberg 4,5* † , Bastian Hablitzel 5 , Herwig Baier 4,6 and
Christian K. Machens 1,2*
1
Group for Neural Theory, Departement d’Etudes Cognitives, INSERM U960, École Normale Supérieure, Paris, France
2
Champalimaud Neuroscience Program, Centro Champalimaud – Champalimaud Centre for the Unknown, Lisbon, Portugal
3
Gatsby Computational Neuroscience Unit, University College London, London, UK
4
Neuroscience Program, Department of Physiology, University of California San Francisco, San Francisco, CA, USA
5
Faculty of Biology, Center for Biological Signaling Studies, University of Freiburg, Freiburg, Germany
6
Max Planck Institute of Neurobiology, Martinsried, Germany
Edited by: Many neural systems can store short-term information in persistently firing neurons. Such
German Sumbre, École Normale persistent activity is believed to be maintained by recurrent feedback among neurons.
Supérieure, France
This hypothesis has been fleshed out in detail for the oculomotor integrator (OI) for which
Reviewed by:
the so-called “line attractor” network model can explain a large set of observations. Here
Yoram Burak, Hebrew University,
Israel we show that there is a plethora of such models, distinguished by the relative strength
Yonatan Loewenstein, Hebrew of recurrent excitation and inhibition. In each model, the firing rates of the neurons relax
University, Israel toward the persistent activity states. The dynamics of relaxation can be quite different,
*Correspondence: however, and depend on the levels of recurrent excitation and inhibition. To identify the
Aristides B. Arrenberg, Department
correct model, we directly measure these relaxation dynamics by performing optogenetic
of Developmental Biology, Faculty of
Biology, Center for Biological perturbations in the OI of zebrafish expressing halorhodopsin or channelrhodopsin. We
Signaling Studies, Institute Biology show that instantaneous, inhibitory stimulations of the OI lead to persistent, centripetal
1, University of Freiburg, eye position changes ipsilateral to the stimulation. Excitatory stimulations similarly cause
Hauptstrasse 1, D-79104 Freiburg,
Germany
centripetal eye position changes, yet only contralateral to the stimulation. These results
e-mail: aristides.arrenberg@ show that the dynamics of the OI are organized around a central attractor state—the
biologie.uni-freiburg.de; null position of the eyes—which stabilizes the system against random perturbations. Our
Christian K. Machens, results pose new constraints on the circuit connectivity of the system and provide new
Champalimaud Neuroscience
Program, Centro Champalimaud –
insights into the mechanisms underlying persistent activity.
Champalimaud Centre for the
Unknown, Neurociências (26.34), Keywords: neural integrator, optogenetics, model, zebrafish, oculomotor system, network dynamics
Av. Brasília, Doca de Pedrouços,
1400-038 Lisbon, Portugal
e-mail: christian.machens@
neuro.fchampalimaud.org
† These authors have contributed
equally to this work.
INTRODUCTION position (Lopez-Barneo et al., 1982; Delgado-García et al., 1989;

Neural activity deep within the nervous system or close to the Fukushima et al., 1992; McFarland and Fuchs, 1992; Aksay et al.,
motor periphery is largely driven by a combination of intrin- 2000). This graded persistent activity typifies a simple form of
sic neuronal properties and recurrent feedback among neu- short-term memory and shares many similarities with the per-
rons. Such activity is almost always dynamic, changing either sistent activity found in higher-order brain areas during working
fast, as in central pattern or sequence generators (Marder and memory (Brody et al., 2003; Major and Tank, 2004; Goldman
Bucher, 2001; Hahnloser et al., 2002), or slowly, as in the neural et al., 2009). Several theoretical studies have shown how the
integrators that have been found at many levels of the ner- persistent activity in the OI can be generated through precise
vous system (Robinson, 1968; Pastor et al., 1994; Gold and recurrent synaptic feedback among neurons (Cannon et al., 1983;
Shadlen, 2001; Wong et al., 2007; Goldman et al., 2009). A key Cannon and Robinson, 1985; Seung, 1996; Seung et al., 2000;
question in neuroscience is how neural systems generate and Aksay et al., 2007). This body of work has led to a network model
control these internal dynamics through links between individual of the OI that can essentially reproduce all experimentally mea-
neurons. sured features in the goldfish, such as the distribution of tuning
One of the simplest and best-studied systems to address this curves (Seung et al., 2000; Aksay et al., 2007), the correlations
question is the oculomotor integrator (OI) for horizontal eye between simultaneously recorded neurons, the generation of sac-
movements. Neurons in the OI are persistently active with a dis- cades, or the system’s response to unilateral silencing (Aksay et al.,
charge rate that is directly proportional to the horizontal eye 2007). Moreover, several candidate mechanisms were pointed out
Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 1

Gonçalves et al. Dynamics of an oculomotor integrator
to explain the remarkable robustness of the system (Koulakov

et al., 2002; Goldman et al., 2003; Moreau and Sontag, 2003).
Here, we show that this network model can be understood
as a particular instantiation of a class of models, all of which
can explain the shared experimental features across animals. The
models are only distinguished by the specific ratio of excitatory
and inhibitory inputs to the neurons. Each network model fully
specifies the dynamics of the OI. While the dynamics are similar
in the system’s normal operating regime, they are distinct out-
side of this regime. Specifically, different network models relax
differently toward the persistent activity states. Consequently,
different instantiations of the network models make different
predictions on how the OI will react to perturbations. These pre-
dictions can be tested with the recent advances of optogenetic
tools which allow us to manipulate systems with high spatio-
temporal precision (Nagel et al., 2003; Boyden et al., 2005; Lima
and Miesenböck, 2005; Han and Boyden, 2007; Zhang et al.,
2007a,b; Douglass et al., 2008; Huber et al., 2008; Arrenberg et al.,
2009; Schoonheim et al., 2010; Fenno et al., 2011).
In systems that are strongly driven by their own internal
dynamics, the outcome of a perturbation depends on a com-
bination of the externally applied stimulation and the intrinsic
network dynamics. In many instances, neural systems are not
sufficiently well known to disentangle the two and make precise
and quantitative predictions. The modeling approach we pursue
here, however, provides the opportunity to predict the influence
of these two competing effects, and by comparison with experi-
mental data, adjust the model, and further our understanding of
the system.
Here, we test these model predictions using transgenic
FIGURE 1 | Properties of the oculomotor integrator (OI). (A) Schematic of
zebrafish expressing either halorhodopsin (NpHR), a light-driven
the oculomotor system for horizontal eye movements. The neurons carrying a
chloride pump, or channelrhodospin (ChR2), a light-activated position signal are located in the oculomotor integrator (OI) and are thought to
cation channel (Zhang et al., 2007a,b; Arrenberg et al., 2009; integrate (in the sense of calculus) the velocity signal from the burst nucleus
Fenno et al., 2011). We show that these instantaneous and small (B). Studies of pairwise neural correlations suggest recurrent excitatory
perturbations of the OI network yield crucial insights into the connections within each side of the OI and mutual-inhibition between
opposite sides. The burst nucleus (B) generates the saccadic eye
dynamics around the system’s normal operating regime. We show movements. Position and velocity signals are provided to the muscles (LR:
that only one of the network models can explain all the data. lateral rectus; MR: medial rectus) via excitation of the motoneurons in the
This model suggests a dominance of unilateral self-excitation, abducens nucleus (ABN) and the contralateral oculomotor nucleus (OMN).
concurrent with a weak coupling between OI cells in the left (B) Idealized eye position trace and concurrent position cell activity. In the
absence of external visual feedback, the fish moves its eyes spontaneously in
and right hemisphere. While the stable states still form a line
scanning cycles, side to side (left side L; right side R; central position 0),
attractor in our new model, the dynamics around the line attrac- alternating stable fixations with saccades. Stable firing rates of the position
tor differ from those of previously proposed models. Specifically, cells are linearly related to eye position during behavior. (C) Tuning curves of
the dynamics are organized around the center of the line attrac- right (blue) and left (red) position cells. (D) Overall population activity of the
tor which corresponds to the null position of the eyes. This left (pL ) and right side (pR ), defined as the sum over the activity of all individual
cells. (E) Single-cell synaptic tuning curves for right (top) and left (bottom)
network arrangement could be preferable for the animal, since
position cells, with thresholds assumed to span the whole motor range. (F)
any perturbations due to noise or synaptic mistuning will cause Synaptic population activities of right (XR ) and left (XL ) half of the integrator as
drifts toward the resting state, instead of drifting toward extreme a function of eye position. Note the staircase shape of the curves, a
population activities and eye positions. consequence of the saturation of the single-cell synaptic activity curves.
RESULTS generated by motoneurons in the abducens and oculomotor cra-

NETWORK MODELS OF THE OCULOMOTOR INTEGRATOR nial nuclei (ABN, OMN, Figure 1A), which receive a velocity
The main anatomical and physiological features of the OI are command from excitatory burst cells. The OI receives a collat-
summarized in Figures 1A–C. The OI is located in the hindbrain eral of the velocity signal and integrates it in the mathematical
and composed of two bilaterally symmetric neuronal nuclei sense to create a position signal. The firing rates of the OI neu-
(Figure 1A). In goldfish, these nuclei consist of around N = 40 rons are therefore linearly related to eye position (Figures 1B,C).
neurons (Pastor et al., 1994; Aksay et al., 2000), and are referred The slope of their tuning curves is higher for neurons with more
to as Area I. Quick horizontal eye movements (saccades) are central eye position firing thresholds and lower for neurons with

contraversive, peripheral eye position thresholds (Aksay et al., In the absence of synaptic input, neurons cease to fire given
2000), a property that has been called “recruitment order” (see the neuronal leak. Therefore, for neural activities to be sta-
Materials and Methods). Neurons from the same side are usu- ble in a network in the absence of an external input, the
ally positively correlated, whereas neurons from opposite sides are recurrent network input to each neuron has to exactly match
negatively correlated, suggesting that the two sides are coupled the neuronal leak (see Materials and Methods). An imbalance
by mutual inhibition and self-excitation (Aksay et al., 2003) as between those would cause either runaway excitation or inhibi-
shown in Figure 1A. The firing rates of these “position” neurons tion. As the eye position moves from left to right, this balance
remain stable in the absence of visual feedback and provide the of inputs and neuronal leaks needs to be maintained despite
signal that controls fixation of the horizontal eye position (Mensh the non-linear distortions (such as thresholds or saturations)
et al., 2004). introduced by the biophysics of neurons and synapses. The result-
These observations indicate that the network of position neu- ing distortions can be compensated by the successive recruit-
rons can maintain a continuum of persistent firing rates over ment of neurons from one side and the suppression of neu-
several seconds, i.e., the time scale of a typical eye fixation. Since rons from the opposite side, as prescribed by the recruitment
only these persistent firing patterns are observed, the popula- order (Figure 1C, see Materials and Methods, for details) (Seung,
tion activity of the two sides must be highly constrained. We use 1996; Seung et al., 2000; Aksay et al., 2007). It is important
these constraints to reduce the network dynamics to the dynam- to note however that to obtain a line attractor in our net-
ics of the two interacting populations, which we describe by work models, fine-tuning of the parameters is necessary, since
their summed activity (Figures 1D,F – see also Figures A1D–F). changes in the parameters as small as 1% disrupt the line attrac-
We will write XL and XR for the left- and right-side population tor dynamics.
output, measured as the resulting post-synaptic conductances Given the connectivity of the two sides of the system, the
(Figures 1E,F) and will refer to this population output as “pop- response of a neuron (on the right) is modeled as xR = H(aXR −
ulation activity.” cXL + h), where parameter “a” determines the weight of the
As further discussed below, our modeling framework requires excitatory input from the right population XR , parameter “c”
that the synaptic currents saturate to balance the progressive determines the weight of the inhibitory input from the left
recruitment of neurons. Although such synaptic non-linearities population XL , and parameter “h” models constant external
are yet to be found in the OI, we here assume their presence. inputs to the network. The function H(.) models the neuron’s
As discussed in the Materials and Methods, we choose step input-output function, and is either a threshold-linear func-
input-output functions to simplify the model tuning but can tion (in the case of a firing rate response) or a Heaviside
relax that assumption by using smoother sigmoidal functions function (in the case of a synaptic output response). The
(see Appendix). Given this choice, the synaptic output tuning threshold of this input–output function corresponds to the
curves are idealized, saturated versions of the firing rate tun- set of points for which aXR − cXL + h = 0 (orange thresh-
ing curves (Figures 1C,E). For simplicity of the model tuning, old line, Figure 2A). Any (XR , XL ) combination that is below
we also assume that the single-cell synaptic outputs have thresh- the orange line will make the right side neuron shown in
olds spanning the whole eye position range. This assumption can Figure 2A fire.
be reconciled with the data if the synaptic currents have differ- The recruitment order fixes the thresholds of the neurons on
ent thresholds (Aksay et al., 2007). In consequence, the left- and the red-blue line as illustrated by the blue circles for a few exam-
right-side synaptic population outputs XL and XR resemble a ple neurons (Figure 2A). However, the data do not specify how a
staircase function where each step is caused by the synaptic neuron would respond to population activities XR and XL out-
input-output function of a single neuron (Figures 1E,F—see also side of this line, leaving a degree of freedom that is related to
Materials and Methods). Since there are many neurons involved, the relative strength of the self-excitatory input “a” and cross-
these population outputs approximate linear functions of the eye inhibitory input “c” into each neuron (Figure 2A). Depending on
position. how this threshold line is chosen for each neuron, the dynamics
The essence of our network models is shown in Figure 2 (see of the population activities outside of the line attractor change
also Materials and Methods; a Matlab-based implementation of accordingly.
the models is available in the online supplementary informa- The dynamics of four exemplary models are illustrated in
tion). In Figure 2A, we plot the population activities of Figure 1F Figures 2C–F. Here, the arrows point in the direction in which
against each other. Given the staircase shape of the population the population activities evolve from different starting points.
outputs, the resulting relationship is composed of multiple points In Figure 2C, the line attractor is generated through mutual
on a line, which we here approximate by the red-blue line. This inhibition of the two sides. This configuration corresponds to one
line illustrates the persistent firing states of the system, which of the oldest models proposed for the OI (Cannon et al., 1983).
range from virtual silence on the right and strong activity on the In Figure 2D, the external excitatory connections are weakened,
left side (XR = 0, XL large, red color) to virtual silence on the left yet the resulting detrimental effect is compensated by weak self-
and strong activity on the right (XR large, XL = 0, blue color). excitation. This model was tested by Aksay et al. (2007), and
Different eye positions correspond to different points on this line is an extrapolation from the model in Seung et al. (2000): the
(Figure 2B; see also below), and each point corresponds to a sta- authors here included mutual inhibition, while keeping orthogo-
ble mode of firing for the network. The line is therefore often nal relaxation dynamics to the line attractor. We note that for both
called a “line attractor” (Seung, 1996). the mutual inhibition model and the weak self-excitation model,

FIGURE 2 | Network models of the OI: construction and dynamics. Bottom: connectivity. Thick, thin, and dashed connectivity lines correspond
(A) State space of the network model. We assume that the state is uniquely to strong, weak, and very weak neural connections. The thickness of these
described by the value of the left and right population activities, XL and XR . connections corresponds to the absolute of the sum of the excitatory
The red-blue colored line corresponds to the stable equilibrium points (line post-synaptic potentials (EPSP) and inhibitory post-synaptic potentials (IPSP)
attractor) of the system. Left eye positions are on the red part of the line, of the neurons post-synaptic to the connection. The green lines depict
right eye positions on its blue part. Blue circles indicate the eye position external (e.g., vestibular) inputs to the integrator areas. (C) A model in which
thresholds of the tuning curves of neurons from the right population along neural activity is kept persistent due to mutual inhibition between the sides.
the line attractor. Orange lines correspond to the thresholds of firing for (D) A model in which both mutual inhibition and self-excitation provide the
other states of the system (orange shaded areas indicate the corresponding stability of the persistent states. (E) The ILA model, which can reproduce
regions above the threshold). Left: the respective slopes are given by the the goldfish inactivation results (Aksay et al., 2007). Here, the dynamics are
relative strength of the (self-) excitatory and (cross-) inhibitory input into mostly unidirectional and involve only one population on each side (arrows
each neuron. Right: as the eyes move from left to right, more and more are horizontal or vertical). In the left motor range (bottom panel, top), the left
neurons from the right population get recruited (compare Figure 1C). (B) network sustains its firing through self-excitation, whereas the right network
Mapping of position population activities onto eye position. The colored is passive, given the weakness of the recurrent inputs. Therefore, the
iso-lines correspond to different eye positions θ in the (XR , XL ) space. The dynamics are dominated by the left network. In the right range (bottom
gray dashed line indicates the line attractor and thereby the stable panel, bottom), the inputs and dynamics are reversed. (F) A model in which
equilibrium states of the system. (C–F) Example models generated within persistent activity is generated through self-excitation only. Here, both sides
our theoretical framework. Top: state space of the example models. Arrows are completely independent, and every combination of population activities
indicate the direction of the dynamics. Points indicate stationary states. is stable.
the dynamics outside of the attractor are orthogonal to the line, connections are set up so that each half of the oculomotor range
although the mutual inhibition model suggests faster dynamics is stabilized by an independent line attractor (ILA model). As a
(as indicated by the longer arrows). consequence, the population with the high activity (e.g., XR ) does
The population dynamics of the model in Figure 2E were not change its activity when the other side’s population activ-
introduced by Aksay et al. (2007) to account for unilateral ity (XL ) is reduced. This situation is given when the left half
inactivation experiments in the goldfish. To obtain such dynam- of the system is silenced, which is equivalent to setting XL = 0.
ics in our modeling framework, the inhibitory and excitatory Although the dynamics above the line attractor are unconstrained

by experimental data, this model proposes that the dynamics i.e., all points on the line attractor. Naturally, the results will
above and below the line attractor are antiparallel to each other. depend on both the length and intensity of the stimulation. We
We note that the ILA model captures the same population state used a brief stimulation (200 ms) and varying stimulation inten-
space dynamics as the model suggested in Aksay et al. (2007), sities. The net eye movement resulting from the combination
although the detailed implementation differs from the one in of stimulation and relaxation was measured as the difference
Aksay et al. (2007): the model does not incorporate input-output θ in eye positions just before the stimulation and 1 s after
functions with high synaptic thresholds and uses a different dis- the stimulation (see simulations in Figure 3C). Large θ are
tribution of tuning curves and cross-inhibition (see Materials and observed when the initial eye position is in the left range, and
Methods). negligible θ when the initial position is in the right range,
A last example model is shown in Figure 2F. Here, the (Figure 3D).
line attractor is stabilized through self-excitation only, and the With similar reasoning, we can explore the system’s response
inhibitory connections are non-existent. We note that this model to excitatory perturbations (Figures 3E–H). During left excita-
is an extension of the model in Seung et al. (2000) from one pop- tion of the ILA model (Figure 3E), the left population activity,
ulation to two populations of excitatory neurons. In this case, XL , increases. After the stimulation, the system state relaxes back
any point outside of the line will be a potential stable fixed point to the line attractor with the dynamics of the intact system
as well. The system may still be confined to the line in practice, (Figure 3F). If the initial eye position is in the left range, the
if the burst input during saccades always moves the population value of XL stays constant, and the right population activity,
activities back onto the line. XR , decreases (Figures 3F,G). Altogether, the eye makes therefore
large movements to the left. If the initial position is in the right
UNILATERAL INSTANTANEOUS PERTURBATIONS: MODEL range, the system state moves mostly in parallel to θ isolines and
PREDICTIONS the eye movements are small or null. Consequently, the pertur-
The network models allow us to predict precisely how a per- bations θ of eye position occur mostly on the ipsilateral side to
turbation would affect the system. Most importantly, these per- the stimulation, similar to the inhibitory perturbations, but with
turbations can be observed at the level of the eye position opposite sign (Figures 3G,H).
which makes the predictions experimentally accessible. To link In both cases, the perturbations θ reflect the relaxation
the population activities to the eye position, we note that the dynamics of the system, i.e., the dynamics of the intact system.
iso-eye-position curves are likely passing through a non-linear By measuring these simulated perturbations for the full range of
transform introduced in the abducens nucleus (Figure 2B; dis- initial eye positions, we can cover all points of the line attrac-
crepancy of position cell and motoneuron tuning curves—for tor. While we have illustrated these perturbations for the ILA
details see Figure A3 and Materials and Methods). This bending model, we can perform similar predictions for the whole range
of the curves also provides a simple explanation for the results of models. Conversely, we can measure the system’s response to
of unilateral silencing of the OI in which the stabilization of eye perturbations in optogenetic experiments, and then simply infer
position remains functional in only half the motor range and for the dynamics of the system around the line attractor that are
roughly half the range of population activities (Aksay et al., 2007). consistent with the experiments.
We can then simulate the response of the models to unilateral We note that we here modeled NpHR stimulations as divisive
instantaneous inhibition and excitation, mimicking optogenetic and ChR2 stimulations as additive. This distinction is based on
stimulations with NpHR and ChR2, respectively. This idea is illus- electrophysiological recordings from the caudal zebrafish hind-
trated in Figure 3, where we focus on one of the network models, brain (not limited to OI cells) which showed that NpHR stim-
the ILA model (Figure 2E). ulations induce a change in firing rate that is dependent on the
Figures 3A–D shows the effect of inhibiting the left half of initial firing rate, while for ChR2 stimulations no such effect was
the OI in the ILA model. Due to the extra inhibition, the left observed in the (small) range of firing rates tested (Figure A9).
population activity XL decreases immediately, as indicated by Therefore, for simplicity, we modeled the effect of ChR2 stimula-
the arrows in Figure 3A. If the eye position before stimulation tion as being additive. However, assuming a subtractive influence
(initial eye position) is in the left range (black point), the inhi- of NpHR on population activity, or a multiplicative influence of
bition causes the system state to cross θ isolines transversely so ChR2 yields qualitatively the same results (Gonçalves, 2012) (sim-
that the eye makes large movements to the right (orange point). ulation data not shown), and does not impede our ability to infer
If the initial eye position is in the right range, the system state the overall dynamics from measurements.
shifts mostly in parallel to θ isolines, and the eye movements
are small or null. After switching-off the inhibition stimulus, UNILATERAL OPTOGENETIC PERTURBATIONS: RESULTS OF NpHR
the system relaxes back to the line according to the dynam- EXPERIMENTS
ics of the intact system (Figure 3B). We note that if the initial To measure the effects of such instantaneous perturbations, and
eye position is in the left range, XR increases, thereby mov- in turn infer the dynamics around the line attractor, we used fiber
ing the eye further to the midline. If the initial position is in optic stimulations (Arrenberg et al., 2009) in behaving transgenic
the right range, XL increases and XR does not change after the zebrafish (Figures 4A,B, Figure A7). Zebrafish are likely to have
stimulation is turned-off, so that the system returns to its ini- the same basic oculomotor circuit architecture and physiology as
tial state and the net eye movement is null (Figures 3A,B). We adult goldfish. The zebrafish larvae (5–8 days post-fertilization,
can extend this perturbation analysis to all initial eye positions, dpf) were immobilized in agarose, and the agarose surrounding

FIGURE 3 | Instantaneous perturbations of the left OI: simulations. stimulation). The deflections of the eye position depend on the initial eye
(A) State space of the ILA model during left inhibitory stimulation. The blue position. Only initial eye positions in the left half of the motor range can be
arrows correspond to the dynamics of the model during a stimulation and affected by the stimulation. (D) ILA model predictions for the eye position
show that the left population activity is driven toward lower values, while the changes, θ, as a function of the initial eye position, θ. Color indicates the
right population activity is largely unaffected. The black dot represents one intensity of stimulation. Positive θ correspond to changes toward the right,
example eye position and population activity (XR , XL ) before stimulation. The and negative θ correspond to changes toward the left. Eye position changes
orange dot represents the population activity and eye position after are measured as a proportion of the whole motor range, and “min” and
stimulation. The gray arrow corresponds to the movement from the initial to “max” are the minimum and maximum eye positions of the motor range. For
the final position. We note that the arrows are enlarged to clarify the direction each intensity of stimulation, the respective plot corresponds to averages
of the movement, and do not correspond to the magnitude of the over 4000 stimulations in each of 21 eye positions. (E) State space of the ILA
movements observed in simulations in (C). The gray dashed line corresponds model during left excitatory stimulation. As can be seen by the dynamics, the
to the line attractor (which is overridden by the stimulation). (B) State space left population activity increases, while the right population activity is largely
of the ILA network model after stimulation offset. Population activities relax unaffected. (F) Relaxation to the line attractor after stimulation. (G) Simulated
back to the line attractor according to the dynamics of the intact system, eye position traces during left stimulation (gray shaded period). Same format
given by the arrows. The colored iso-lines are annotated with the as panel (C). Only initial eye positions in the left half of the motor range can
corresponding eye positions θ. The gray line corresponds to the line attractor. be affected by the stimulation. (H) ILA model predictions for the eye position
(C) Simulated eye position traces during left stimulation (gray shaded period) changes, θ, as a function of the initial eye position, θ. Simulations were
for the ILA model. Colors indicate the initial eye position (before the performed as in panel (D).
the eyes was removed. An optic fiber was positioned above stimulations (50 μm) to test multiple positions (Figure 4C). The
the hindbrain to stimulate halorhodopsin (NpHR) or chan- photoactivated volumes were columns of tissue approximately
nelrhodopsin (ChR2). The transgenes Tg(UAS:ChR2(H134R)- 10–15 cells wide and protruding through the entire dorsoven-
mCherry)s1986t or Tg(UAS:NpHR-mCherry)s1989t were driven tral extent of the hindbrain, as judged by Kaede photoconver-
by the enhancer trap line Et(E1b:Gal4)s1101t, resulting in broad sion experiments (Figure A7). The maximal effect was observed
expression of ChR2 or NpHR in neurons (Scott et al., 2007; around 50–150 μm caudal of the Mauthner cells, somewhat more
Arrenberg et al., 2009). To localize the zebrafish OI (Miri et al., rostral (rhombomere 5, 6, and 7) than in the previous reports
2011a) we performed unilateral light stimulations on differ- (Miri et al., 2011a,b) (Figure 4D, Supplementary Movie 1).
ent rostro-caudal positions of the hindbrain and measured the We first focused on inhibitory (NpHR) perturbations.
resulting eye drift magnitudes (Figures 4B–D). We chose short Unilateral stimulations about 100 μm caudal of the Mauthner
200 ms stimulations to use the same time scale as burst sig- cells (rhombomere 7 and 6) resulted in a drift of the eye position
nals which the position neurons typically integrate and to allow following a saccade (Figure 4E). As in the numerical simulations
the NpHR-induced hyperpolarization to saturate. Two previous protocol, we computed the changes θ in eye position from just
reports (Miri et al., 2011a,b) localized the integrator neurons before the stimulation to 1 s after the stimulation. For simplicity,
in rhombomere 7 and 8 of the larval zebrafish, based on the all unilateral stimulation results are plotted as left stimulations.
corresponding location in the goldfish, two-photon laser abla- Figures 5A,B show the results for a single fish: the eye posi-
tions and on optogenetic loss-of-function experiments. In these tions are perturbed strongly in the left range and only weakly
laser ablation and optogenetic experiments, a single location was in the right range. In the left range, the more eccentric the ini-
tested (rhombomere 7 and 8) and found to have an effect on the tial eye position, the higher the elicited change in the eye position
integrator performance. Here, we used small diameter optic fiber toward the right, as indicated by the positive θ. Additionally, the

FIGURE 4 | Halorhodopsin (NpHR) stimulation in the hindbrain induces A maximum projection of selected optical slices is shown. Note: The
drifts in the eye position. (A) An optic fiber (red) is placed above the broadly distributed red speckles are NpHR-mCherry fluorescent protein
zebrafish head (6 dpf) to perturb neuronal activity in the brain (green). hb, aggregates. (D) Inverse of the time constants of drift τ (in (1/s), see
hidbrain; sc, spinal cord; mb, midbrain; dc, diencephalon. (B) Bilateral Materials and Methods—Data Analysis) induced by NpHR stimulations at
NpHR stimulation in the hindbrain induces eye drifts toward the null different rostro-caudal levels, relative to the Mauthner cell position (n = 7
position. Saccades to the left side are shown for both the left eye (black animals). The data is split according to the eye position before stimulation.
trace) and the right eye (gray trace). The red shade indicates the Eye positions (averaged across the two eyes) on the same side as the
stimulation period (6 s or 200 ms). Top: Wildtype animal. Middle: NpHR stimulation (left) are plotted in red (ipsiversive) and eye positions on the
expressing animal, 6 s stimulation. Bottom: NpHR expressor, 200 ms opposite side (contraversive) are plotted in green. Control trials (open
stimulation. θ (blue) is the change in eye position from before the symbols, no light stimulation) are included. The data points are fitted by
stimulation to 1 s after the stimulation. (C,D) An optic fiber (50 μm Gaussian curves. Approximate locations of rhombomeres r4–r8 are
diameter) was placed at different rostro-caudal positions of the hindbrain indicated in gray. (E) Two minutes of recording in a fish with frequent
and the resulting eye drift magnitudes during unilateral stimulation were spontaneous eye movements. In this experiment, NpHR was stimulated
measured. (C) A 5 dpf hindbrain transgenic for Et(E1b:Gal4-VP16)s1101t, four times on the left side after the fish made a saccade. Each stimulation
Tg(UAS:NpHR-mCherry)s1989t, and Tg(UAS:Kaede)s1999t. The green is marked as a vertical red line in the recording and the red color intensity
channel contains non-converted Kaede signal (upper left), while the red of the line corresponds to the light power used for the stimulation. The
channel contains converted Kaede and mCherry fluorescence. Kaede was relative stimulation intensity (black line) is plotted in addition below the eye
photoconverted from green to red at three fiber positions (yellow circles). traces. Panel (A) is modified from Figure 4 of Baier and Scott (2009).

magnitude of this change increased with the stimulation intensity. UNILATERAL OPTOGENETIC PERTURBATIONS: RESULTS OF ChR2
These results become more distinct when averaging over all fish EXPERIMENTS
(n = 24) (see Figure 5C). Following the same protocol as in the NpHR experiments, we
The results of the halorhodopsin activation experiments there- next performed unilateral instantaneous stimulations (100 ms) in
fore agree with the ILA model that we used as an example network animals expressing ChR2 so that one side of the OI was excited
model in Figures 3A–D. We note two small differences to the pre- (Figures 5D–F). Again, all unilateral stimulations were pooled
dictions of the ILA model. First, even in the absence of NpHR and plotted as left stimulations. When the initial eye position
stimulation, the eyes move slowly toward the null position from was in the right range, the perturbations generally caused an
the whole motor range. Accordingly, the zebrafish eyes are slightly eye movement toward the left, i.e., toward the null position,
mistuned, whereas the ILA model is not. Second, the eyes show as indicated by the negative θ. In this case, we furthermore
positive perturbation θ at the null position, unlike the model observed that more eccentric initial eye positions or higher inten-
predictions (with θ = 0 for θ = 0). This difference could be due sity stimulations induced stronger changes in the eye position. On
to uncertainties in the eye position: since the model only predicts the other hand, when the initial eye position was in the left range,
positive perturbations, θ ≥ 0, any ambiguity in eye position, stimulations elicited very small or no changes toward the right
be it due to uncertainty of the null position (see Materials and after 1 s, when compared to the control case (Figures 5D–F).
Methods), measurement errors, or hysteresis in the system itself, For very high ChR2 stimulation intensities, we additionally
will shift the average perturbation at θ = 0 to a positive value. noted eye movements toward the null position when the initial
Altogether, the dynamics of the OI below the line attractor are eye position was on the same side as the stimulation (ipsiver-
therefore well-captured by the ILA model (compare Figure 2E). sive eye positions; Figure A6). Since we are interested in inferring
The halorhodopsin inactivation experiments thereby confirm the the dynamics in the immediate vicinity of the line attractor,
pharmacological inactivation experiments of Aksay et al. (2007). we linearly regressed the magnitude of the drift at very low
FIGURE 5 | Halorhodopsin and channelrhodopsin instantaneous Figure 3D. The motor ranges of all fish were normalized before averaging,
stimulations of the left OI: experimental results. (A) Experimentally and the y-axis indicates the eye position changes θ as a proportion of the
recorded eye position traces, averaged across eyes, during NpHR stimulation motor range. Legends indicate the light power averages in mW for each bin
(0.5 mW). Averages over 96 eye traces (for each eye) of 1.5 s each, distributed and respective standard errors across recordings. The shaded envelopes
across 10 initial eye position bins, in one fish. Same format as in Figure 3C. indicate standard errors. (D) Experimentally recorded eye position traces
Stimulation of the left OI elicited movements toward the right in the left during ChR2 stimulation (between 0.37 and 0.56 mW). Averages over 76
range, and small or no movements toward the left in the right range, in stimulations in one fish. Same format as in Figure 3C. The stimulation
agreement with the ILA model predictions from Figure 3C. (B) elicited movements toward the left in the right range, and small or no
Experimentally recorded eye position changes, θ, as a function of initial eye movements toward the right in the left range, in contrast with ILA model
position, θ, averaged across both eyes (see Figure A6, for individual eyes). predictions from Figure 3C. (E) Experimentally recorded eye position
Data are from the same single fish recording as in (A) (for each eye, 372 changes, θ, as a function of initial eye position, θ, averaged across both
stimulations, 258 control points, 4 intensity bins, over the course of 3 h). eyes (see Figure A6 for individual eyes). Data are from the same single fish
Same format as in Figure 3D. The stimulations were binned and colored recording as in (D) (for each eye, 523 stimulations, 382 control points, over
according to the average stimulation light power in mW for each bin. Spline the course of 3 h). Same format and analysis protocol as in panel (B).
fits were performed on the binned data to highlight the data trend. (C) Spline (F) Spline fit averages across all tested fish (n = 19 recordings), same format
fit averages across all tested fish (n = 24 recordings), same format as in and analysis protocol as in panel (C).

stimulation intensities (Figure 6A). In agreement with the data reflection-symmetric with respect to θ = 0, as proposed by the
for medium intensities (Figure 5F), low intensity stimulation ILA model.
induces eye movements mostly during contraversive eye posi-
tions (when the initial eye position is on the other side as the THE OI DYNAMICS AROUND THE LINE ATTRACTOR
stimulation; Figures 6B,C). The results from the NpHR and ChR2 experiments provide us
The results of the channelrhodopsin activation experiments with the means to infer the OI dynamics around the line attrac-
therefore invalidate the ILA model, as we find that weak exci- tor. Since the inhibitory stimulations yield results similar to those
tatory unilateral perturbations of the left OI induce centripetal predicted by the ILA model (compare Figure 3D and Figure 5C),
(toward the null position) and not centrifugal eye movements. we conclude that the dynamics below the line attractor are similar
Furthermore, we find the strongest effect on the side contralat- to those of the ILA model (Figure 2E). To obtain the observed
eral to the perturbation, and almost no effect on the ipsilateral point-symmetry of the perturbations (Figures 5C,F), however,
side. We note that the perturbations resulting from experimen- we need to assume that the dynamics above the line attractor
tal excitation and inhibition are approximately point-symmetric are roughly orthogonal to the dynamics below. Consequently, our
with respect to the central eye position, (θ = 0, θ = 0), and not experimental results suggest a model with dynamics as shown in
Figure 7A. Here, the dynamics outside the line attractor are orga-
nized around a central point, corresponding to the null position
of the eyes (the “null position” or NP model). Although Figure 7A
illustrates the overall flow of the trajectories toward the line, we
note that our line attractors are composed of individual fixed
points, such that the fine-scale dynamics change in the vicinity
of the line (see Figure A4C). Although the fine-scale dynamics
predict small changes toward the left in the left range after ChR2
stimulation, these small changes do not appear in the predictions
if we assume that the system is noisy.
These conclusions are further supported when we analyze the
transient dynamics of the experimental eye movements around
the NpHR or ChR2 stimulation period (Figures 7C–F). If the
initial eye positions are in the left range, the NpHR stimula-
tion generates monotonic eye movements toward the right, as
indicated by the transient, positive derivatives of the eye move-
ments (green traces in Figure 7C). If the initial eye positions
are in the right range, the NpHR stimulation leaves them essen-
tially unaffected. This contrasts with the ChR2 stimulations which
transiently affect all eye positions. For initial eye positions in the
left range, the stimulation causes transiently biphasic eye move-
ments, as indicated by the negative, then positive derivatives of the
eye movements (green traces in Figure 7E). In other words, the
eye is briefly moved to a more eccentric position, before relaxing
back to its original position. For initial eye positions in the right
range, we observe an overall negative derivative of the eye move-
ment, generating a centripetal eye position drift (yellow traces in
Figure 7E).
These transient dynamics match with the population activity
dynamics predicted by the NP model (Figures 7D,F). Let us focus
on the excitatory perturbations. When the left OI is excited, the
FIGURE 6 | Effect of ChR2 stimulations at very low light intensities. (A)
left population activity XL increases in the whole motor range
Eye position changes θ caused by left side ChR2 stimulation vs. eye
position previous to stimulation, for one recording, averaged across both (Figure 7F). In the relaxation phase, if the initial eye position is in
eyes. Data were binned and colored according to light power. Legends the left range, the system counterbalances the left stimulation by
indicate the stimulation light power average in mW for each bin. The two decreasing XL without changing XR . Accordingly, eye movements
gray shaded areas highlight the ranges of initial eye positions θ for which are biphasic and altogether eye position changes only marginally.
the function between θ and intensity of stimulation, I, is analyzed in (B).
(B) θ vs. intensity of stimulation, I, for the two highlighted eye position
In the right motor range, the right population activity decreases in
ranges in (A). In blue are the data points, and in red a linear regression fit the relaxation phase, while the left population activity stays intact,
for the data in the range 0 mW < I < 0.28 mW. The derivative of the thus causing centripetal (to the null position) eye movements
function between θ and I at I = 0 is assumed to be the slope of this linear (Figure 7F).
regression fit. (C) Averages of the derivatives dθ/dI (I = 0) vs. eye position
Besides explaining the perturbation data well (Figures 7G,H),
previous to stimulation across all tested fish (n = 19 recordings). The plot
envelopes indicate standard errors.
the resulting NP model is also in better agreement with the nat-
ural leakiness of the OI, i.e., the slow drift in eye positions that

FIGURE 7 | Inference of the OI dynamics. (A) State space of the model after stimulation. In the right range, the system relaxes to positions close
inferred from the NpHR and ChR2 instantaneous stimulations: the NP to the position previous to stimulation. (E) Average across all ChR2 trials
model. In the NP model, the dynamics below the line attractor are (light power between 0 and 0.28 mW) of the derivatives of the eye traces.
orthogonal to the dynamics above the line attractor. (B) Effective For initial eye positions in the right half of the motor range (yellow traces),
connectivity of the NP model. In the left range (top), the left network the left stimulation leads to negative deflections of the eye position, i.e.,
receives stronger self-excitatory and weaker cross-inhibitory inputs than the toward the left. In contrast, for initial eye positions in the left half of the
right network. For low activity of the right network, the left network motor range (green traces), the left stimulation leads to biphasic eye
sustains its firing, whereas for high activity of the right network, the movements, first negative derivatives (i.e., leftwards) then positive
activity of the left network is decreased. In the right range (bottom), the derivatives (i.e., rightwards). (F) State space interpretation of the transient
inputs and dynamics are reversed. Overall, cross-inhibition is very weak. (C) dynamics in (E). Left: dynamics during left excitatory stimulation. The left
Average across all experimental recordings of the derivatives of the eye population activity increases, while the right population activity is largely
position traces during left side NpHR stimulation (light power between 0 unaffected. Right: relaxation to line attractor after stimulation. (G) NpHR
and 0.8 mW). Eye position derivatives are measured as a proportion of the stimulation of the NP model. Left: simulated eye position traces during left
motor range. Colors indicate the initial eye position (before the stimulation). stimulation (gray shaded period), with same format as Figure 5A. Only
The shaded envelopes indicate standard errors. For initial eye positions in initial eye positions in the left half of the motor range are affected by the
the right half of the motor range (yellow traces), the left side stimulation stimulation. Right: NP model predictions for the eye position changes, θ,
leads to either small positive deflections of the eye position (i.e., toward as a function of the initial eye position, θ. (H) ChR2 stimulation of the NP
the right), or null deflections. In contrast, for initial eye positions in the left model. Left: simulated eye position traces during left stimulation (gray
half of the motor range (green traces), the left stimulation leads to positive shaded period), with same format as Figure 5D. Stimulation causes eye
derivatives, i.e., strong rightward eye movements. (D) State space position changes only in the right half of the motor range. Right: NP model
interpretation of the transient dynamics in (C). Left: dynamics during left predictions for the eye position changes, θ, as a function of the initial eye
inhibitory stimulation. The left population activity decreases, while the right position, θ. In the right panels of (G,H), averages were computed over 4000
population activity is largely unaffected. Right: relaxation to line attractor stimulations in each of 21 eye positions, for each intensity of stimulation.
we observe even in the absence of perturbations (Figures 5C,F). to the integrator cells remains a possibility. Using the modeling
First, we note that many (random) mistunings of the synaptic framework, we show below that accidental stimulation of these
parameters will automatically lead to leaky eye positions. Second, afferent inputs does not invalidate our conclusions.
even in the perfectly tuned NP model, the slow centripetal drift We distinguish four possible scenarios for afferent stimula-
will emerge due to the presence of noise in the neuron’s firing. tion (Figure 8A): in the first two scenarios, afferent stimulation
Since any internally generated noise in the OI will have similar (if leads to either ipsilateral excitation or inhibition of the integra-
smaller) effects than our perturbations, such noise will cause slow tor cells (violet and green arrows), and in that case our overall
centripetal, and thereby leaky eye movements (see Figure A4D). stimulation direction does not change in the synaptic population
output space. In the other two scenarios, the afferent stimulation
STIMULATION OF AFFERENTS TO THE INTEGRATOR leads to either contralateral excitation or inhibition of integrator
While the dynamics that we inferred through the experiments cells (red and blue arrows), therefore introducing a component
seem perfectly reasonable, the applicability of the NP model to the to the stimulation which is orthogonal to the pure integrator
OI hinges on the correctness of our experiments. One particular stimulation.
concern is that the limits of the integrator cells in the larval For a cross-inhibitory architecture, the dynamics above the
zebrafish are not well-established. Hence, we cannot rule out that line attractor are either as in the ILA model (Figure 2E), as
we may be stimulating efferent motor neurons or afferent inputs in the NP model (Figure 7A), or somewhere in between (in
(for instance, vestibular neurons) to the integrator. Since stimula- that case, the dynamics are roughly perpendicular to the line
tion of motor neurons should not lead to persistent eye position attractor, as in Figure 2D). We simulated all four afferent stim-
changes, we ignore the potential stimulation of those neurons. ulation scenarios for each of the three models, and assumed
However, stimulating afferent inputs to the integrator, in addition that afferent stimulation has half the magnitude of the direct

FIGURE 8 | Stimulation of afferents to the integrator (model). Potential according to the respective afferent stimulation. (C–E) Eye position changes
ChR2 stimulation of afferents to the left and right half of the integrator in θ caused by direct and afferent stimulation (afferent stimulation has half the
addition to direct ChR2 stimulation of the left OI. (A) Stimulation of afferents magnitude of the direct integrator stimulation) vs. eye position previous to
to the left OI changes the overall magnitude of the stimulation. When stimulation for the three different models, where color indicates the direction
affecting excitatory afferent neurons, the stimulation magnitude is increased of afferent stimulation in the state space. Neither the ILA model nor the
(green arrows), when affecting inhibitory afferent neurons, it is decreased intermediate model lead to results similar to the data: the ILA model mainly
(violet arrows). When afferents to the right OI are stimulated, an orthogonal leads to centrifugal movements on the ipsiversive side regardless of the type
component is added to the overall stimulation effect. Additional stimulation of of afferent stimulation; the intermediate model leads to changes toward the
inhibitory afferent neurons pulls the perturbation to smaller values of XR (blue left in most of the motor range, for all afferent stimulations. Only the NP
arrow) whereas additional stimulation of excitatory afferent neurons model leads to results similar to the data, and especially so when the net
increases the right population activity (red arrow). (B) Relaxation after afferent stimulation leads to contralateral excitation (red curve). For each
stimulation for three different models: ILA model, intermediate model, and intensity of stimulation, the respective plot corresponds to averages over
NP model. The final state of the system is represented by dots, colored 4000 stimulations in each of 21 eye positions.
integrator stimulation. As can be seen in Figures 8B–E, nei- and a self-excitation whose effective strength depends on eye posi-
ther the ILA model nor the intermediate model produce results tion, suggesting that both sides are almost, albeit not completely,
similar to the data (regardless of the type of afferent stimu- decoupled, in line with previous findings in the goldfish (Debowy
lation). Only the NP model leads to the results observed in and Baker, 2011). To keep the self-excitation in check, external
the data. (e.g. vestibular) inputs to this model are inhibitory. A detailed
Furthermore, the NP model provides an explanation for the description of the implementation of the network models can be
effect of the high intensity stimulation mentioned earlier (com- found in the Materials and Methods.
pare Figure A6): if at high intensity, unilateral ChR2 stimulation
also excited the contralateral side via light scattering, then the DISCUSSION
NP model would indeed predict centripetal eye drifts on the SUMMARY
ipsilateral side (Figure 8E, red curve). We showed that a whole range of network models could
account for the set of electrophysiological features that have been
CONNECTIVITY IN THE INTEGRATOR measured in OIs across animals of different species. These mod-
In conclusion, the NP model is the only candidate model within els differ from each other by the relative strength of self-excitation
our theoretical framework that agrees with all aspects of the and mutual inhibition, and the respective dynamics prevailing
data. Most importantly, these dynamics suggest a different con- in the population activity state space outside of the line attrac-
nectivity than implied by the previously proposed models. The tor. To test these models and experimentally unveil the dynamics
respective, effective connectivities for the NP model are shown in of the OI, we performed optogenetic perturbations in the lar-
Figure 7B. The dynamics are mediated by weak mutual inhibition val zebrafish. Using the silencer NpHR, we found that unilateral

light stimulation induced eye movements back to the midpoint, if stimulation in this line led to marked behavioral changes, the
the eye position prior to the stimulation was on the same side manipulation was not specific to the neural integrator. The neural
as the stimulation. The experimental results for NpHR in lar- integrator in larval zebrafish is distributed across approximately
val zebrafish were in accordance with results from goldfish using 150 μm in the hindbrain of the larval zebrafish (Miri et al.,
inactivation with lidocaine (Aksay et al., 2007), therefore cor- 2011a), with non-integrator neurons interspersed between the
roborating the previously proposed ILA population dynamics. cells with position signals. However, our results are not easily
However, we found that unilateral ChR2 stimulations did not explained by a stimulation of these non-integrator neurons. First,
have the opposite effect to NpHR stimulations predicted by the both NpHR and ChR2 stimulation induce stable and persistent
ILA dynamics, i.e., centrifugal (away from the null position) eye changes in eye position. This persistent change makes an influ-
movements on the stimulation side (ipsiversive eye positions). ence of the motoneurons that lie in close rostral proximity to
Instead, ChR2 stimulations had a centripetal (toward the null the integrator unlikely. Exclusive motoneuron stimulation should
positions) effect on eye positions on the side opposite to the cause the eyes to move back to the original position immediately
stimulation (contraversive eye positions). after stimulation offset, an effect we did not observe. Second, we
Hence, perturbations always tend to drive the eye positions may have stimulated cells which project to the integrator such
toward the midpoint, indicating that this point is the focus of the as the saccade-generating neurons. However, saccade generating
OI dynamics. Consequently, we inferred the dynamics around the neurons are only active during saccades and unilateral stimula-
line attractor from these experiments, and named the resulting tion is therefore expected to only change the saccade frequency
model the null-position or NP model. This model suggests an OI (Schoonheim et al., 2010) and have no effect in-between saccades.
architecture with strong self-excitation and weak cross-inhibition. Nevertheless, a low level stimulation of the saccade-generating
Only a specific combination of excitation on one side and inhibi- neurons could have occurred without the generation of a measur-
tion on the other will lead to actual changes in the position signal able saccade: in that case, given that saccade-generating neurons
as required in saccadic eye movements (e.g., movements from one excite the ipsilateral OI and inhibit the contralateral OI, a per-
range to the opposite range) (Gonçalves, 2012). Interestingly, our turbation of these neurons would cause an indirect stimulation
results can at least in part explain the findings in pharmacologi- of the integrator neurons roughly in the same direction as the
cal experiments where glutamate or GABA agonists were injected direct integrator stimulation, therefore not invalidating the inter-
in the OI. Both an increase in excitation as well as an increase in pretation of our results. Third, we may have affected some of the
inhibition resulted in centripetal eye movements, much as in our vestibular inputs to the OI. However, even in this scenario, our
experiments (Arnold et al., 1999). conclusions about the integrator dynamics hold up. Since vestibu-
lar inputs are included in the network models, we can simply
EXPERIMENTAL FEATURES AND LIMITATIONS simulate their accidental stimulation. As shown in Figure 8, acci-
In the presented study we show how optogenetic experiments can dental stimulation offsets the magnitude of induced eye drifts,
be combined with modeling to infer the dynamics of a neural but overall does not alter their eye position dependence. Within
circuit module for integration. In the last few years, the field of the range of models considered, the data can therefore only be
optogenetics has provided a powerful set of techniques to per- explained by the NP model, but not by the other models.
form gain- and loss-of-function experiments (reviewed in Zhang We have stimulated excitatory and inhibitory integrator cells
et al., 2007a; Luo et al., 2008; Fenno et al., 2011) and has been at the same time, which could potentially lead to unexpected
applied to zebrafish (Szobota et al., 2007; Douglass et al., 2008; network effects, e.g., due to induced imbalances of excitation
Arrenberg et al., 2009; Baier and Scott, 2009; Zhu et al., 2009; and inhibition within the network. However, several observations
Schoonheim et al., 2010). A fundamental problem in interpreting support our interpretation of the data. First, electrophysiological
the effects of optogenetic stimulations is that a system’s response is recordings in the hindbrain of the same zebrafish lines used in this
a combination of the stimulation magnitude and the intrinsic net- study suggest that more than 80% of NpHR expressing cells were
work dynamics. Indeed, the eye movements induced through the significantly silenced during illumination and more than 90%
optogenetic perturbations depended on both the light-intensity, of ChR2 expressing cells showed an increase in firing rate upon
i.e., the strength of stimulation, and on the eye position, i.e., the illumination (Arrenberg et al., 2009). Second, NpHR stimulation
internal state of the system prior to stimulation. To understand results are in agreement with the pharmacological inactivations in
these interdependencies, we relied on network modeling (Seung, the goldfish integrator (Aksay et al., 2007), therefore confirming
1996; Seung et al., 2000; Goldman et al., 2003; Eliasmith, 2005; the inhibitory nature of the NpHR stimulations on the integra-
Aksay et al., 2007). In turn, the mismatches between the model tor. Third, ChR2 stimulation leads to different results than NpHR
predictions and the experimental results allowed us to constrain stimulation, which is proper inhibition. Fourth, both ChR2 and
the class of feasible network models and thereby improve our NpHR experiments were performed with the same Gal4 driver
understanding of the OI. This general approach illustrates the line. Therefore, if ChR2 stimulation led to net inhibition of the
importance of the internal state of a system during a pertur- integrator, then NpHR stimulation would lead to excitation of the
bation. Wherever this internal state is at least partially known, integrator which is ruled out based on the second observation.
optogenetic perturbations can provide useful clues toward the While these results suggest that optogenetic manipulations of the
underlying network dynamics. neural integrator changed the network activity in the expected
In our experiments we made use of a Gal4 driver line direction, future zebrafish lines, e.g., with specificity for excitatory
that drove strong expression broadly in neurons. While local or inhibitory neurons, will facilitate the dissection of this circuit.

One could hypothesize that the effect of ChR2 stimulation sat- one could interpret the NP model as suggesting that the OI oper-
urates or reverses (depolarization block, Kleinlogel et al., 2011) ates like a single fixed point, and not a line attractor, as extensively
with increasing stimulation intensities or for highly active cells. suggested in previous literature. However, for large noise levels,
Our previous electrophysiological recordings provided no evi- we can re-tune the NP model to recover the NP dynamics in the
dence for such an effect (Arrenberg et al., 2009, Figure A9). proximity to the line by implementing stronger cross-inhibition
More importantly, this possibility is not supported by the eye (simulations not shown). In any case, random perturbations of
movement data in the range of intensities analyzed, since the the NP model (such as noise) are unlikely to cause a centrifugal
eye movements scale linearly with the stimulation light intensity drift of the eye position.
(Figure 6B). Also, the effect does not reverse at high light intensi- Given the centripetal drift suggested by the NP model, we
ties (Figure A6). As a final note, the modeling framework already hypothesize that the OI features dynamics with a higher degree
assumes that the synapses of highly active neurons are saturated, of built-in “safety” than previously thought. The OI has been
so that, at least within the model, ChR2 stimulation does not observed to be leaky on longer time scales, both in goldfish and
affect these neurons. zebrafish (Mensh et al., 2004; Miri et al., 2011a). This leakiness
In this study, we deliberately focused on the dynamics in the may be a behaviorally advantageous feature, since, by bringing the
neighborhood of the stable eye position states. Consequently, system to the central position by default, it enables the relaxation
our analysis was restricted to stimulation with low or medium of the eye muscles. Yet even higher brain systems may rely on such
light intensities. The NP model, however, makes predictions for a built-in leakiness. In working memory tasks that employ graded
any stimulation intensity, opening the question of what hap- persistent activity (Machens et al., 2005), for instance, a tendency
pens when the stimulation intensity is increased. As shown in the to drift toward the central point while memorizing a sensory stim-
Figure A6, for high NpHR stimulation intensities, we addition- ulus could explain the psychophysical errors that are known as
ally found small centripetal movements when the eye position was contraction bias (Ashourian and Loewenstein, 2011).
on the side opposite to the stimulation. For high ChR2 stimula- While this built-in “safety” may help against noise in the sys-
tions, we found centripetal eye movements when the eyes were tem, it does not solve the fine-tuning problem, i.e., the instability
on the same side as the stimulation. In the NP model, this would of the line attractor against perturbations in the synaptic weights
require a change in the dynamics far away from the line attrac- in the network. Indeed, this fine-tuning problem is a separate
tor, requiring the arrows to bend further toward the midline. problem, somewhat orthogonal to the problems that we have
While it seems unlikely that the effects at higher light intensities investigated here, for which several solutions have been proposed
can be explained through scattering of light into the other hemi- (Koulakov et al., 2002; Goldman et al., 2003; Moreau and Sontag,
sphere (data not shown), the effect could potentially be explained 2003).
through strong stimulation of vestibular inputs. Furthermore, While we here have assumed that neural integration in the ocu-
we notice that strong ChR2 stimulation could synchronize the lomotor system is generated through precise recurrent feedback
activities of cells, which may have a range of effects, including in a neural circuit, in previous literature single-cell mechanisms
complete shutting down of persistent activity (Dipoppa, 2012). have been put forward to explain neural integration observed
We therefore refrained from including these observations in the in multiple areas in the brain. In particular, following an exper-
model. imental demonstration of integration in individual cells from
the entorhinal cortex (Egorov et al., 2002), a body of theoret-
MODEL FEATURES, LIMITATIONS, AND PREDICTIONS ical work has proposed several biophysical mechanisms which
Integrators are ubiquitous in the brain and are involved in several could underlie single cell integration, dispensing synaptic feed-
important computations. For instance, in decision-making tasks back (Loewenstein and Sompolinsky, 2003; Fransén et al., 2006).
requiring sensory integrations, neurons in the lateral intrapari- In the OI, unilateral disruption of the connectivity leads to neu-
etal cortex behave similar to integrators (Gold and Shadlen, 2001; ral activity drifts with time constants which are typically above
Wong et al., 2007). In working memory tasks, neurons in the pre- 1 s (Pastor et al., 1994; Aksay et al., 2007), suggesting that single-
frontal cortex exhibit almost linear dynamics during the times cell mechanisms possibly play a role in the process of integration.
in which an animal needs to remember a stimulus, similar to However, single-cell mechanisms remain largely uncharacterized
integrators operating in several dimensions (Singh and Eliasmith, in the integrator, and therefore we here have followed the net-
2006; Machens et al., 2010). In the head direction system, a head work mechanisms hypothesis as in previous studies of this system
velocity signal is integrated into head position (Zhang, 1996). (Seung, 1996; Aksay et al., 2007). The contribution of single-cell
In previous line attractor models, it has generally been mechanisms to the slow dynamics in the integrator is a challenge
assumed that noise causes random drift along the line (Seung, for future research.
1996). While this is true in models with orthogonal dynamics We also note that our network model is a rate-model, in which
around the line (such as the model illustrated in Figure 2D), in the activities of individual cells are described by rates rather than
the case of the NP model the relaxation to the line has a pre- precise spike times. Although we lose biophysical realism with this
ferred direction, therefore causing a systematic drift toward the type of model, we gain analytical tractability, a very useful asset
null position. The term “line attractor” for the NP model is there- in interpreting experimental results within a theoretical frame-
fore strictly only valid in the limit of vanishing noise. For large work, and in constructing models in accordance with data. Since
noise levels, the model shows flow toward the central eye position the position cells exhibit persistent activity with regular firing
with equivalent speed from every point in the state space. Hence, (Aksay et al., 2003), temporal averages of spiking events are a good

qualitative description of the system. Nevertheless, an equivalent are sufficiently above that threshold (when they run into satura-
model with spiking neurons could be built as shown in Seung tion). This prediction could e.g., be tested with single-cell ChR2
et al. (2000), Eliasmith (2005). stimulation. Excitatory stimulation of different cells would then
Given the weak mutual-inhibition, the NP model behaves close lead to movements in different eye ranges, enhancing the fact
to a system with a plane of stable fixed points (see Figure 2F), and that different neurons are responsible for different stable activities
shows slow dynamics around the line. We note that the relaxation on the motor range (Gonçalves, 2012). Consequently, combin-
of the eye positions after stimulation is indeed slow (on the order ing single-cell optogenetics with the framework here designed has
of 200 ms). However, these slow dynamics could be reflecting the the potential to provide even deeper insights into the detailed
dynamics of the muscle physics rather than the slow dynamics structure of the OI in the future.
of the integrator. Future work should show whether such slow
dynamics can indeed be observed. MATERIALS AND METHODS
Our network model features multiple stable fixed points, EXPERIMENTS
which suggests that the eye positions corresponding to these fixed Animals
points should be held comparatively longer than eye positions in- For all experiments, we used zebrafish larvae between the age of 5
between the fixed points. However, in our data, the system seems and 8 dpf. Animals were transgenic for a combination of the fol-
to visit a range of eye positions in a homogeneous way (both dur- lowing transgenes: Et(E1b:Gal4)s1101t, Tg(UAS:NpHR-mCherry)
ing NpHR stimulation and during spontaneous, slow eye position s1989t, Tg(UAS:Kaede)s1999t, Tg(UAS:ChR2-mCherry)s1986t. In
decay), which contrasts with the prediction of our model. Such addition, the larvae were mutant for the mitfa/nacre gene
homogeneity could be due to external factors to the integrator, (mitfas170 , mitfas184 , or mitfab692 alleles), which rendered the skin
such as small saccadic commands causing smooth eye movement transparent and facilitated fiber optic stimulation as well as eye
fluctuations, or the dynamics of the motor neurons and muscles. position detection. Siblings that did not express NpHR or ChR2
Nevertheless, the homogeneity found in the data challenges our served as control groups and are labeled wt here. Adult fish
hypothesis of discrete fixed points in the integrator and suggests were either transgenic for Et(E1b:Gal4)s1101t or for the opto-
further studies to elucidate this question. genetic responders, since keeping optogenetic expressors in the
Our modeling framework assumes a specific mapping from s1101t line would have resulted in variegation of the expression.
network activity to eye position, based on the difference between Embryos/larvae were raised in the dark and not fed. For each
the tuning curves of position neurons and motor neurons (see experiment, about 4 clutches were screened and the strongest
Materials and Methods). Although this assumption is essential in expressors were kept.
our framework to account for the unilateral inactivation results Many Gal4s1101t/UAS:NpHR expressing larvae had non-
in Aksay et al. (2007), one could relax it and assume linearity inflated swim bladders or showed only infrequent eye move-
between population synaptic outputs and eye position by intro- ments. For this reason, each mounted larva was observed for
ducing high synaptic thresholds in the same fashion as in Aksay 1 min under a stereoscope and only larvae that showed saccades
et al. (2007). However, we believe that the non-linear mapping is in both directions and good peripheral eye fixations were used
most likely present in the system and should be included in future for the experiments. This way, only the best behaving 20% of the
modeling studies of the oculomotor system. mounted larvae were used. Control larvae were screened the same
In a recent study (Miri et al., 2011a), zebrafish position neu- way, although a higher percentage of larvae could be used for
rons were shown to have variable timescales of integration, so experiments. The screened NpHR expressing and non-expressing
that the associated relaxation time constants varied across neu- larvae had similar eye drift rates in the absence of stimulation (see
rons over one order of magnitude. This suggests that the dynamics Figure 4 in Miri et al., 2011a). The magnitudes of the induced eye
of the OI could be high-dimensional (on the order of the number position drifts were somewhat variable between animals stimu-
of neurons), in contrast with our line attractor model, which is lated at the same position, which we attribute to the expression
implemented with homogeneous time constants across neurons level/variegation variability between animals. For example, in one
and has low (2D) dimensional dynamics. Given that in this study animal in Figure A8, we noted a patch of cells in which NpHR
we are interested in the dynamics of the population activities, the expression was absent, which resulted in a much reduced effect
details of single-cell time constants do not affect our conclusions. on eye position (see points [0.08, 86 μm] and [0.15, 86 μm] for
In the future, it will be interesting to perform optogenetic stimu- the left and the right eye in the ipsi stim. condition). We excluded
lations and at the same time measure the activities of the position 5 animals from the analysis in Figure 5F, since the induced eye
cells to explore the full state space, and realize the dimensionality position changes were much smaller than in the majority of
of the system’s dynamics. animals.
Our models belong to a series of works suggesting that the OI
builds up a line attractor by a balance between neural saturation Mounting
and progressive recruitment of neurons to compensate such satu- Larvae were mounted in a drop of low-melting agarose (1.6%)
ration (Seung, 1996; Seung et al., 2000; Aksay et al., 2007). Future in a petri dish (35 mm diameter). A platinum wire (100 μm in
work should specifically target the validity of this assumption as diameter) glued to a pasteur pipette was used to flatten the liq-
it is crucial for the whole modeling framework. Specifically, this uid agarose drop by moving the wire at the perimeter of the drop
assumption predicts that neurons have no responsibility or influ- and thus increasing the agarose-covered area in the dish, so that
ence on eye positions that are below their firing threshold or that the height of the liquid approximately matched the height of the

larva. This step took 10 s and ensured that the optic fiber could dark, because wildtype animals showed a behavioral response
be placed close to the skin in the experiment. Next, we used the to the stimulation light in the dark, which was much reduced
wire to position the larva dorsal side up. The agarose solidified or absent in the presence of the backlight (Miri et al., 2011a)
for 5–10 min and egg water was added. A second platinum wire (the larvae probably see some of the scattered stimulation light).
(100 μm) that was flattened at the tip and bend 70◦ about 2 mm We note that larval post-saccadic eye fixations were somewhat
from the tip was used to make agarose incisions, moving the plat- leakier in the dark (when compared to low light conditions),
inum wire sideways so that the flattened part acted like a knife. suggesting that larvae were using visual feedback to improve ocu-
Sometimes the agarose would lose its adhesion and we found lar stability. However, this effect was small in comparison to eye
that using a fresh petri dish for every fish ameliorated this prob- position changes induced by medium light intensities, so that
lem. Also, the fish sometimes managed to escape and we found we are confident to be measuring the integrator performance in
that inserting the platinum wire directly at the eye and mov- our experiments. A custom-written LabVIEW program (National
ing it outwards worked better than the other way round. Two Instruments, Austin, TX) was used to image, record angular
blocks surrounding the eyes were cut and the flat side of the eye positions, and to trigger stimulations via a NI USB DAQ
platinum wire was now used to scoop the blocks out. The stere- box. Images were acquired at 30–45 Hz and particle analysis (NI
oscope backlight was set at an angle (dark field) to visualize the Vision) was used to detect eye positions and write them to a text
cut agarose and make sure that the agarose surrounding the eyes file (together with the stimulation time points) for later analy-
was completely removed. After 3 min of rest, the frequency of eye sis. Eye positions were measured and plotted in real time and
movements was observed (see Animals). To minimize water evap- fiber optic stimulations were triggered based on them. In most
oration, a dish lid was placed on the petri dish in some of the experiments a single 200 ms stimulation of constant intensity was
recordings. The dish lid had a 1 cm diameter hole to allow for the delivered automatically 1 s after the eyes had made a saccade. In
placement of the optic fiber. some experiments four consecutive stimulations were delivered
for each saccade (1, 3, 5, and 7 s after the saccade). The stimula-
Laser stimulation tion intensities were chosen randomly: we either used five fixed
For laser stimulations, we used an AOTF to couple lasers of three intensities or the computer picked values from an exponential
different wavelengths (633, 561, 488 nm) into a multimode optic distribution of intensities that biased toward smaller intensities.
fiber. We made use of the laser system of a disassembled confo- No stimulation was applied after every third saccade, and these
cal microscope. For UV photoconversions, we manually switched periods served as our control trials. Recordings typically lasted
the stimulation fiber to a fiber coupled UV laser (355 nm). Fiber for at least 30 min and up to 12 h. The LabVIEW program can be
preparations were described previously (Arrenberg et al., 2009). requested from the authors.
The AOTF was used to modulate the laser intensity by providing
an analog voltage from the DAQ device connected to the com- DATA ANALYSIS
puter. For some experiments we used a multi-laser system from The files containing eye position trace and stimulation data were
Ikecool (Anaheim, CA), however the analog modulation was not analyzed with custom algorithms using MATLAB. We only ana-
good enough to precisely control light intensity on a millisecond- lyzed the first 3 h of each experiment, since in some recordings
basis (significant baseline light at 0 Volts and the intensity was not (without dish lid) water evaporation affected the camera focus
stable enough after a switch from 0 Volts to almost full power). and thereby the correct detection of eye position angles. The
We recommend the inclined scientist to rather buy a system from induced eye movements were measured as the change in eye posi-
Toptica (Munich, Germany) or Omicron (Rodgau-Dudenhofen, tion from stimulation start to 1 s after stimulation start. Most
Germany). The optic fiber (including Thorlabs’ FT030 protec- induced eye movements consisted of a single monotonic drift
tive tubing) was placed in a glass pipette with an angled tip. that was completed within 400 ms after stimulation. However, a
The pipette was mounted on a fine micromanipulator and the 1 s interval was chosen to account for potential variations of the
fiber was positioned over the hindbrain. While x-y position duration of the post-stimulation drift. The eye position change,
could be judged easily by looking through the stereoscope, the averaged across left and right eye, was then plotted against the ini-
z-positioning was more difficult. In cases in which the z-position tial eye position, averaged across eyes. The results from each fish
was unclear, we used the fine micro-manipulation to lower the were fitted by a cubic smoothing spline with boundary second
fiber tip slowly until (a) the mechanic strain on the agarose moved derivatives equal to zero, using the spline fit-package developed
the skin of the fish slightly or (b) until the fish startled. We then for MATLAB by Jonas Lundgren. This fitting procedure was
moved the fiber back up by a small distance (≈20 μm). repeated for several stimulation intensity bins. A population plot
was generated by averaging the spline fits of all fish, for each inten-
Experimental setup sity bin. We observed some variability of motor ranges across
The larva was placed under a stereoscope and a custom-built LED fish, especially in animals expressing ChR2, which tended to
array was used as a backlight. White LED light was used to posi- have a smaller motor range (average motor range and respec-
tion the animals and IR LED light (850 nm) to image the head tive standard deviations in NpHR animals: [−20.0◦ ± 5◦ ; 18.7◦ ±
with a CCD camera (up to 60 Hz, TheImagingSource, Bremen, 4.4◦ ]; in ChR2 animals [−13.7◦ ± 3.9◦ ; 13◦ ± 3.5◦ ]). Therefore,
Germany). During the experiment, dim room lights were kept before averaging across animals, we normalized all motor ranges
on and sometimes a weak white backlight was used in addi- to generate a smooth population plot. In some fish, there was
tion so that the experiments were performed under low light an undersampling of events at the eccentric eye positions, and
intensity and low contrast conditions. We did not record in the therefore we removed the 5% most eccentric events on the right

side and on the left side before averaging across animals. In several Assumption 1 (Two opposing populations). We assume that the
fish, some intensity bins did not span the full motor range, and we dynamics are controlled by two population variables, XR and
extrapolated the corresponding splines. To account for the uncer- XL , one for each side. These variables shall represent the effect
tainty of the extrapolation, we performed a weighted average over one population has on the postsynaptic currents of neurons in
fish for eccentric eye positions, with weights that decreased as the the other population; a specific definition will follow below. We
extrapolation distance increased. assume that in its normal working regime, these two “synaptic”
The null-position of the eyes was defined as the average of the population activities oppose each other, so that, as the activity in
extreme eye positions. Note that this definition only works well one population grows, that of the other drops. For simplicity we
for eyes which explore eye positions in both directions equally therefore assume that
well, which was not the case for every fish. However, in our hands
β = XR + XL (1)
this was the best definition since other definitions (median, aver-
age, fixed at +8◦ /−8◦ ) were more subjective or decreased the
where β is a constant value.
across-fish consistency of the results. An animal’s distribution of
Assumption 2 (Eye position). We furthermore assume that
post-saccadic eye positions was often bimodal, since the animal
within this working regime (but not outside of it) the eye position
explored the peripheral positions more frequently. In some exper-
can be read out as
iments, we stimulated multiple times (1, 3, 5, and 7 s) after a
single post-saccadic fixation in order to access the full scope of
θ = XR − XL (2)
eye positions.
The location of stimulation was in caudal proximity to the where right range positions are defined as positive eye positions,
previously identified region containing saccade generating neu- and left range positions as negative.
rons (Schoonheim et al., 2010), and in a fraction of the events, Assumption 3 (Firing rates). We furthermore assume that each
the stimulation induced a saccade (more frequently for ChR2 neuron’s firing rate in the stationary state is determined by its
stimulations than for NpHR). Since this study focuses on the excitatory input from the same population, scaled by a weight
performance of the integrator, we did not include these events ai , the inhibitory input from the opposite population, scaled by
in our analysis. In Figure 5, these were excluded by a velocity a weight ci , and some external (e.g., vestibular) input, hi , so that
threshold (>20◦ /s). In Figure 4D, the data points result from
eye velocity vs. eye position fits (linear regression through the rR,i = [ai XR − ci XL + hi ]+ (3)
origin, see Figure A8), and correspond to the inverse of the eye
position drift time constant at a specific hindbrain location. For rL,i = [ai XL − ci XR + hi ]+ (4)
each fit, only the middle 70% of the eye velocity/eye position data
points were fitted, thereby excluding outliers caused by saccades. where i = 1, . . . , n indexes the neuron, and [·]+ is a threshold-
For each eye, the eye drift was normalized across hindbrain linear function (Figure A1A). For simplicity, we assume complete
positions. Each data point in Figure 4D corresponds to one symmetry of the two systems, so that neurons in the two popula-
fish (average of the normalized left and right eye, for individual tions have exactly the same set of parameters values.
eyes see Figure A8) and hindbrain position. The data points for
Figure 4D (eye drift vs. hindbrain position) were then fitted by Tuning curve constraints
gaussian functions. For the Gaussian fit functions in Figure 4D, Using the (abstract) threshold-linear tuning curves, we can
the mean was restricted to the interval [0; 200]. This modification describe the firing rates of the right and left position neurons, rR,i
was made because for approximately flat distributions (control and rL,i , as a function of the eye position θ (see Figure 1C), so
data points ipsi ctrl, contra stim, contra ctrl) peripheral data that (for i = 1, . . . , n)
points sometimes caused Gaussian fits with means far away from
the tested hindbrain region. rR,i = [si (θ − ti )]+ (5)
rL,i = [−si (θ − ti )]+ (6)
THEORETICAL FRAMEWORK: NETWORK MODEL
In previous work (Gonçalves, 2012), we show how to build a line where si and ti are the tuning curves’ slope and threshold,
attractor model accounting for the main findings in the oculomo- which are assumed to obey the recruitment order (Aksay et al.,
tor integrator, using basic design principles that have previously 2000) (see Figure A2). Since the parameters si and ti are given
been suggested in the literature (Seung, 1996; Eliasmith, 2005; by the data, they constrain the possible choices of ai , ci , hi
Machens and Brody, 2008). We use these principles to clarify how from Equations (3,4). More specifically, on the curve defined by
the network connectivity is related to the tuning curve properties. Equation (1), the following relations need to hold:
Here, we summarize the main features of the model developed in
Gonçalves (2012). ai + ci
si = (7)
We start by describing the assumptions and simplifications 2
underlying the network model. Most of these assumptions are (ai − ci ) β + 2hi
adapted from the previous literature and are based either on ti = − . (8)
ai + ci
arguments of plausibility and simplicity or on observations about
the oculomotor integrator (Seung, 1996; Aksay et al., 2000; Seung To obtain the constraints on the excitatory and inhibitory inputs,
et al., 2000; Aksay et al., 2007; Machens and Brody, 2008). ai and ci , as well as the external inputs, hi , we need to invert

this relationship. Since there are three parameters on one side Input-output functions
(ai , ci , hi ) and two on the other (si , ti ), this inversion is not Before fitting the parameters bi , we need to make a specific choice
unique. We therefore introduce a free parameter λi to obtain for the input-output function H(·). Our choice will be driven by
a quest for simplicity:
ai = 2λi si (9) Assumption 5 (Input-output function). We assume that the pos-
sible synaptic saturation can be modeled by a heaviside function
ci = 2(1 − λi ) si (10)
so that
hi = si (1 − 2λi ) β − si ti . (11)
0 if x ≤ 0
H(x) = (14)
The free parameter λi ∈ [0, 1] controls the relative role of self- 1 if x > 0
excitation and mutual inhibition. For λi = 0, a neuron receives
only inhibitory inputs from the other side, and for λi = 1, it In this case, the summation of a set of Heaviside functions H(·)
receives only excitatory inputs from the same side. No matter how in Equation (13) results in a function that resembles a staircase
λi is chosen, as long as ai , ci , and hi follow the above equations, (see Figure A1H). In turn, a particularly simple solution is given
the measured recruitment order will be obeyed (see Figure A2). if we assume that all bi = b are the same, and that the ti are
equally spaced across the whole range of eye positions. In this
Stationarity constraints case, the staircase approximates a linear function whose slope is
Next, we need to make sure that the firing rates of the neurons are determined by b, and we only need to set this single parame-
in a stationary state during fixations of the eye. To do so, we have ter to its proper value. This solution assumes, though, that the
to define how the individual neural firing rates combine to give thresholds ti extend over the whole eye position range which has
rise to the synaptic population activities XL and XR . not been observed experimentally. Rather, the thresholds cluster
Assumption 4 (Synaptic Population Activity). We assume that in only half the eye position range (as indicated in Figure 1C).
the synaptic population activities are generated through a linear However, the framework can easily incorporate this constraint by
combination of the individual neural activities such that (for the making the assumption that a) H(·) does not saturate completely
right population) and is a concave function for large inputs, and b) H(·) features
high-synaptic thresholds. This has indeed been the assumption in

n
the past (Aksay et al., 2007). In this case, the parameters bi have to
XR = bi g rR,i (12) be fit by linear regression. Use of the Heaviside function, however,
i=1 simplifies the mathematics and makes the underlying architecture
more transparent (Machens and Brody, 2008).
where bi is a weighting factor, that determines the contribution of
the i-th neuron to the synaptic population activity and g(·) is a
sigmoidal function that captures possible non-linearities. These Dynamics of the network
non-linearities can capture saturations in the contribution of To develop a dynamical equation for the network and clarify the
individual neural firing rates to the synaptic population activity. network connectivity, we define the effective (or synaptic) output
Plugging Equations (2,5) into the above equation, and using of a neuron as
the relation XL = β − XR , we obtain the following condition for
stationarity: xR,i = g rR,i (15)

n
so that Equation (12) becomes
XR = bi H(si (2XR − β − ti )) with H (·) = g [·]+

i=1 XR = bi xR,i . (16)
(13)
i
Here, H(·) is a neural input-output function that combines the
effect of the neuron’s threshold-linear tuning curves and other Identical equations hold for the left population. Using
non-linearities modeled through the sigmoidal function g(·), Equations (3,4), we can now reformulate the stationarity
such as synaptic saturation (see Figure A1B). For instance, H(·) condition on the level of single cells, using the effective outputs
could be a function similar to the one shown in Figure A1C. of the neurons,
To obtain a continuum of stationary solutions, i.e., solutions to
the above equation for many values of XR , the parameters bi
xR,i = H (ai XR − ci XL + hi ) (17)
can be fit so that the above relation holds approximately (see ⎛ ⎞
Figures A1G,H) (Seung, 1996; Eliasmith, 2005; Machens and n
n
Brody, 2008). Since the input-output functions H(·) are shifted = H⎝ ai bj xR,j − ci bj xL,j + hi ⎠. (18)
with respect to the XR -axis due to the differently distributed j=1 j=1
threshold values ti , a workable solution usually exists. We note
that with this fitting, we recover the desired line of synaptic pop- Just as the population equations, these equations have solutions
ulation activities in the normal working regime (see Equation (1) for a continuum of values of xR,i as long as the parameters bi are
and Figure A1I). fitted as described above. To equip the network with dynamics,

we need to assume how the neural activities relax to this station- more weaker (and orthogonal) modes that would only impact the
ary state once the system is perturbed. Let us define the following transient dynamics of single neurons, while leaving the dynam-
abbreviations, ics of the summed population intact. These weaker modes can
change the (biophysical) connectivity in many ways, while leaving
wij,E = ai bj , (19) the effective connectivity the same.
wij,I = ci bj , (20) THEORETICAL FRAMEWORK: FROM NETWORK ACTIVITY TO EYE
POSITIONS
where wij,E denotes the weight of an excitatory connection from Neural integrator mapping to eye positions
neuron j to neuron i and wij,I denotes the weight of an inhibitory Oculomotor plant dynamics are the result of the innervation
connection from neuron j to neuron i. of two antagonist muscles (medial rectus and lateral rectus)
Assumption 6 (Exponential relaxation). We assume that the by motor neurons delivering the position signal (Figure 1A).
neural activities relax exponentially to the stationary state. This Consequently, we assume that the eye position θ is related to
leads to a network with standard Wilson-Cowan dynamics the difference between right and left motor population activi-
(Dayan and Abbott, 2001), ties, making the function fx between synaptic population activ-
ities (XR , XL ) and eye position non-linear (see Appendix for
⎛ ⎞ details).
n
n
τẋR,i (t) = −xR,i (t) + H ⎝ wij,E xR,j − wij,I xL,j + hi ⎠ (21) Oculomotor plant dynamics
j=1 j=1 To model the observed inertia or sluggishness of the whole sys-
tem in response to the perturbations (most likely a consequence
where now each neuron i receives excitatory and inhibitory inputs of muscle inertia and eye dynamics) we implemented a simple
from the two populations. An identical equation holds for the exponential decay toward the network stable activities with a time
neurons in the left population and is obtained by switching the constant τθ , so that
L and R subscripts.
Firing rate dynamics are assumed to be very fast as compared τθ θ̇ = −θ + fx (XR , XL ) . (23)
to synaptic dynamics, and are therefore always in equilibrium:
⎡ ⎤ By visual inspection of zebrafish eye traces, we chose τθ = 200 ms.
n
n
rR,i =⎣ wij,E xR,j − wij,I xL,j + hi ⎦ (22) UNILATERAL INACTIVATION
j=1 j=1 + Now that we have fully determined the relationship between the
integrator neural activities and the eye position, we can appropri-
As previously, an identical equation holds for the neurons in ately interpret the pharmacological inactivation results of (Aksay
the left population and is obtained by switching the L and R et al., 2007). After unilateral silencing of the oculomotor integra-
subscripts. tor, the animals manage to fixate the eyes on the contralateral half
In this model, neurons are excitatory to their own popula- of the motor range, and the activities of the intact neurons remain
tion and inhibitory to the opposing population. However, we can stable within this range. In our framework, this suggests a range
build equivalent models obeying Dale’s law, i.e., where neurons of stable points on the upper half of the activity axis XR and XL .
are either excitatory or inhibitory, but not both at the same time These stable ranges provide an important constraint for the ILA
(unpublished results). A systematic way of mapping networks and NP models, limiting the possible choices of λi that we can
with mixed excitatory and inhibitory neurons to networks that make in Equations (9–11).
obey Dale’s law was recently proposed in Parisien et al. (2008).
As can be seen in Equations (19,20), we here assume low-rank PARAMETER TUNING OF NETWORK MODELS
connectivity. This connectivity has been a standard assumption We have clarified the dynamics along the axes required by the
of all previously proposed models of the oculomotor integrator pharmacological inactivation results of Aksay et al. (2007), so that
(Seung, 1996; Aksay et al., 2007) for two reasons: (1) the exper- we can proceed to the final tuning of the network model.
imentally observed stable activities across cells in the integrator Independent of the choice of λi , the proposed network con-
along the eye position range suggest that the system has low nectivities feature a line attractor in the (XR , XL ) space that agrees
dimensional dynamics, characteristic of a low-rank connectivity with the recruitment order. However, the desired dynamics out-
system; (2) this assumption simplifies the theoretical treatment side of the attractor, e.g., the dynamics along the axes required
of the problem. However, the assumed connectivity should only by the pharmacological inactivation results, further constrain the
be viewed as an effective connectivity rather than a direct map- tuning of the parameter λi . Given the low number of neurons and
ping onto biophysical synapses. Mathematically, one can relax the simplicity of the tuning, we manually tuned this parameter to
low-rank assumption: the connectivity matrix can be expanded change the dynamics around the line attractor. However, a least
into “modes” (using singular value decomposition, e.g.), and the squares optimization procedure could also be built to solve the
dynamics of the oculomotor integrator are simply governed by problem.
the first and strongest mode which governs the dynamics in the In Figure A4, we present two model solutions presented in the
plane of population activities. However, one could add many main section, the independent line attractor model (ILA model)

and the null position model (NP model) (see Appendix for the similar results in both models, with minor differences (results not
detailed parameter tuning procedure). In general, the nullcline of shown).
each half OI is the set of points in space where the dynamics in Due to the use of Heaviside input-output functions and the
the corresponding direction are zero. In both ILA and NP mod- resulting staircase character of the nullclines, the network imple-
els, we tuned the nullclines of either side to consist of densely mentation of the NP model features additional fine-scale dynam-
arranged parallel lines, such that the vectors in the flow field ics that are not illustrated in Figures 7A, A4B (see detail of the
are bound to be either horizontal or vertical. For instance, in dynamics in Figure A4C). These dynamics do not affect the tra-
the NP model, left range, vertical nullclines of the right side (in jectories in the noise-less case, and disappear when more realistic
blue) ensure that the system has weak dynamics on the horizontal (e.g., sigmoidal) input-output functions are used for tuning the
direction, and therefore the dynamics are mostly on the vertical model (simulations not shown).
direction.
We note that in the NP model, we have tuned the nullclines to THEORETICAL FRAMEWORK: INSTANTANEOUS PERTURBATIONS
end close to the line of fixed points, to ensure that each piece of As shown in the results section, we experimentally test the
nullcline intersects only one piece of the opposite nullcline and range of models by performing instantaneous perturbations with
therefore obtain a line of fixed points (Figure A4B, left). If we optogenetic tools: inactivation with halorhodopsin (NpHR) and
relax that assumption and allow for a band attractor for instance, excitation with channelrhodopsin (ChR2). To model the pertur-
then the nullclines do not have to end as close to the diagonal. bations, it is crucial to understand the effects of optogenetic tools
in the activities of stimulated cells.
ILA MODEL DYNAMICS It has previously been shown (Arrenberg et al., 2009) that inac-
The ILA model with 36 neurons on each half of the integra- tivation by NpHR stimulation leads to stronger inactivation of
tor is illustrated in Figure A4A. As can be seen, the nullclines high activity cells than low activity cells, so the inactivation effects
resulting from the summation of the input-output functions H(.) are approximately divisive. ChR2 stimulation, on the other hand,
of the individual neurons intersect in a line of stable points causes an increase in activity independent of the original activities
(Figure A4A, left). Furthermore, after complete left inactivation, (Arrenberg et al., 2009). To incorporate optogenetic stimulations
the system is able to maintain stable positions in the right range in our dynamical systems model, we take these experimental
(results not shown), as obtained in the pharmacological experi- observations into account.
ments (Aksay et al., 2007). Finally, this model shows the experi-
Perturbations with halorhodopsin (NpHR)
mentally observed recruitment order feature, i.e., the higher the
Given the divisive nature of the inhibitory perturbations, we
threshold the higher the slope of the tuning curves (Figure A2),
model these as multiplicative interactions of the input-output
since it was imposed into the model. The dynamics of the ILA
function. Assume that a neuron is perturbed with a brief stimu-
model are qualitatively similar to the dynamics of the model
lation of magnitude αi (0 < αi < 1) and duration T. We simulate
developed in Aksay et al. (2007) (see Figure A4A, right). The
the synaptic output of such a perturbed neuron (from the left) as
respective network parameters are listed in Table A1.
τẋL,i (t) = −xL,i (t) + (1 − αi D(t)) H(ai XL (t) − ci XR (t) + hi ) ,
NP MODEL DYNAMICS
In Figure A4B, we illustrate the NP model. Just as the ILA i = 1, . . . , n (24)
model, the system is able to maintain stable eye positions in
half the original range after unilateral silencing. The model dif- where D(t) = H(t − t0 )H(t0 + T − t) is a unit perturbation
fers from the ILA model in one important aspect, however. The pulse, starting at time t0 , and lasting for T seconds. Consequently,
upper half of the nullcline (Figure A4B, left) is now maintained at the population level, the system is described by the following
largely by self-excitation. In fact, the two halves of the oculo- population equations during an inactivation of the left half of the
motor integrator are almost independent, and the threshold of integrator:
the neurons is mostly determined by the self-excitatory inputs.
However, due to the weak, mutual cross-inhibition, this inde-
n
τẊR (t) = −XR (t) + H(ai XR (t) − ci XL (t) + hi ) (25)
pendence is slightly disturbed, leading to the dynamics shown
i=1
in Figure A4B, right. Altogether, the NP model has qualitatively
the same dynamics below the line attractor as the ILA model
n
but different dynamics above the line attractor. However, the τẊL (t) = −XL (t) + (1 − αi D(t)) H(ai XL (t)−
NP model dynamics are slower than the ILA model dynam- i=1
ics in the whole state space, given the low inter-dependence ci XR (t) + hi ) (26)
of the two areas in the NP model (the arrow lengths in
Figure A4A, right panel, and Figure A4B, right panel, are in dif- The parameters αi are chosen at random from a Gaussian distri-
ferent scales). The respective network parameters are listed in bution whose mean μ is equal across neurons, but grows linearly
Table A1. with the stimulation intensity. The standard deviation of the dis-
Despite the connectivity differences between ILA model and tribution of αi was given as σ = μmin /2 where μmin indicates the
NP model, since these models share the same qualitative dynam- smallest perturbation intensity tested. To ensure that the stimu-
ics below the line attractor, NpHR left stimulation leads to very lation strength is always positive, negative αi were rectified. Since

the mapping of the physical stimulation intensities (in mW) onto Aristides B. Arrenberg, Herwig Baier, and Christian K. Machens
the network was unknown, we hand-tuned the mean stimulation wrote the paper. Pedro J. Gonçalves and Aristides B. Arrenberg
effect μ to give results in a reasonable range for the ILA and NP contributed equally to this work.
models.
ACKNOWLEDGMENTS
Perturbations with channelrhodopsin (ChR2) Pedro J. Gonçalves was part of the PhD Program in
Given the assumed additive nature of the excitatory perturba- Computational Biology of Instituto Gulbenkian de Ciência,
tions, we model these as external excitatory inputs into the supported by Fundação Calouste Gulbenkian, Siemens SA and
cells: Fundação para a Ciência e Tecnologia (Portugal) through grant

n
τẊR (t) = −XR (t) + H(ai XR (t) − ci XL (t) + hi ) (27) SFRH/BD/33210/2007. Aristides B. Arrenberg was supported
i=1 by a Krevans fellowship and Aristides B. Arrenberg and Bastian
Hablitzel are supported by the Excellence Initiative of the German

n
τẊL (t) = −XL (t) + H(ai XL (t) − ci XR (t) + hi Federal and State Governments (Centre for Biological Signaling
i=1
Studies EXC 294-B5-Driever). Herwig Baier was funded by
NIH R01 NS053358, the NIH Nanomedicine Development
+ αi D(t)) (28) Center “Optical Control of Biological Functions” and the Max
Planck Society. Christian K. Machens was funded by a “Chaire
where D(t) = H(t − t0 )H(t0 + T − t) is once more a unit per- d’excellence” of the French Research Agency (ANR), a “Chaire
turbation pulse, starting at time t0 , and lasting for T seconds. The d’excellence” of INSERM and the Emmy-Noether Program of the
parameters αi are chosen from Gaussian distributions, as above, German Research Foundation. We thank Jan Huisken for his help
with mean μ and fixed standard deviation, equivalent to one half regarding the laser setup.
of the mean of the smallest tested intensity, σ = μ/2. As above,
the range of intensities tested was hand-tuned for both models, to
SUPPLEMENTARY MATERIAL
fall in the experimentally observed range.
The Supplementary Material for this article can be found
Perturbations protocol online at: http://www.frontiersin.org/journal/10.3389/fncir.
To simulate the perturbations, we roughly follow the outline of 2014.00010/abstract
the experimental perturbations. The model is simulated numeri-
cally in MATLAB, using the Euler method. Each trial lasts 1.5 s, REFERENCES
Aksay, E., Baker, R., Seung, H. S., and Tank, D. W. (2000). Anatomy and discharge
and is initialized at one stable eye position, and after 0.5 s, is
properties of pre-motor neurons in the goldfish medulla that have eye-position
provided a perturbation of the same duration as in experiments signals during fixations. J. Neurophysiol. 84, 1035–1049.
(200 and 100 ms for NpHR and ChR2 trials, respectively). This Aksay, E., Baker, R., Seung, H. S., and Tank, D. W. (2003). Correlated dis-
procedure is repeated for perturbations of different stimulation charge among cell pairs within the oculomotor horizontal velocity-to-position
intensities, and several eye positions across the motor range, and integrator. J. Neurosci. 23, 10852–10858.
Aksay, E., Olasagasti, I., Mensh, B. D., Baker, R., Goldman, M. S., and Tank, D. W.
typically 4000 times for each condition (initial eye position, per- (2007). Functional dissection of circuitry in a neural integrator. Nat. Neurosci.
turbation intensity). The population activities in the network are 10, 494–504. doi: 10.1038/nn1877
simulated according to Equation (25) and Equation (26) for the Arnold, D. B., Robinson, D. A., and Leigh, R. J. (1999). Nystagmus induced
NpHR and Equation (27) and Equation (28) for the ChR2 per- by pharmacological inactivation of the brainstem ocular motor inte-
turbations. Eye positions are assigned to the resulting population grator in monkey. Vision Res 39, 4286–4295. doi: 10.1016/S0042-6989
(99)00142-X
activities according to the mappings outlined in Appendix. Just Arrenberg, A. B., Del Bene, F., and Baier, H. (2009). Optical control of zebrafish
as for the experimental data, we measure the resulting changes behavior with halorhodopsin. Proc. Natl. Acad. Sci. U.S.A. 106, 17968–17973.
in eye position, θ, as a function of the eye position θ just doi: 10.1073/pnas.0906252106
before perturbation. For each perturbation intensity, all neurons Ashourian, P. and Loewenstein, Y. (2011). Bayesian inference underlies the contrac-
tion bias in delayed comparison tasks. PLoS ONE 6:e19551. doi: 10.1371/jour-
are assumed to receive a perturbation magnitude taken from a
nal.pone.0019551
normal distribution, as explained above. Baier, H. and Scott, E. K. (2009). Genetic and optical targeting of neural circuits
and behavior–zebrafish in the spotlight. Curr. Opin. Neurobiol. 19, 553–560. doi:
MATLAB PACKAGE 10.1016/j.conb.2009.08.001
A Matlab package in which the ILA model and the NP model Boyden, E. S., Zhang, F., Bamberg, E., Nagel, G., and Deisseroth, K. (2005).
are implemented is provided in the online supplementary Millisecond-timescale, genetically targeted optical control of neural activity.
Nat. Neurosci. 8, 1263–1268. doi: 10.1038/nn1525
information. Brody, C. D., Romo, R., and Kepecs, A. (2003). Basic mechanisms for graded
persistent activity: discrete attractors, continuous attractors, and dynamic
AUTHOR CONTRIBUTIONS representations. Curr. Opin. Neurobiol. 13, 204–211. doi: 10.1016/S0959-
Pedro J. Gonçalves and Christian K. Machens designed the 4388(03)00050-3
model. Pedro J. Gonçalves performed the simulations. Aristides B. Cannon, S. C., and Robinson, D. A. (1985). An improved neural-network model for
Arrenberg and Herwig Baier conceived the experiments. Aristides the neural integrator of the oculomotor system: more realistic neuron behavior.
Biol. Cybern. 53, 93–108. doi: 10.1007/BF00337026
B. Arrenberg and Bastian Hablitzel performed the experiments. Cannon, S. C., Robinson, D. A., and Shamma, S. (1983). A proposed neural net-
Pedro J. Gonçalves, Aristides B. Arrenberg, Bastian Hablitzel, work for the integrator of the oculomotor system. Biol. Cybern. 49, 127–136.
and Christian K. Machens analyzed the data. Pedro J. Gonçalves, doi: 10.1007/BF00320393

Dayan, P., and Abbott, L. (2001). Theoretical Neuroscience: Computational and Machens, C. K., and Brody, C. D. (2008). Design of continuous attractor net-
Mathematical Modeling of Neural Systems, 1st Edn. Cambridge, MA: The MIT works with monotonic tuning using a symmetry principle. Neural Comput. 20,
Press. 452–485. doi: 10.1162/neco.2007.07-06-297
Debowy, O., and Baker, R. (2011). Encoding of eye position in the goldfish Machens, C. K., Romo, R., and Brody, C. D. (2005). Flexible control of mutual inhi-
horizontal oculomotor neural integrator. J. Neurophysiol. 105, 896–909. doi: bition: a neural model of two-interval discrimination. Science 307, 1121–1124.
10.1152/jn.00313.2010 doi: 10.1126/science.1104171
Delgado-García, J. M., Vidal, P. P., Gómez, C., and Berthoz, A. (1989). A neuro- Machens, C. K., Romo, R., and Brody, C. D. (2010). Functional, but not anatomical,
physiological study of prepositus hypoglossi neurons projecting to oculomotor separation of “what” and “when” in prefrontal cortex. J. Neurosci. 30, 350–360.
and preoculomotor nuclei in the alert cat. Neuroscience 29, 291–307. doi: doi: 10.1523/JNEUROSCI.3276-09.2010
10.1016/0306-4522(89)90058-4 Major, G., and Tank, D. (2004). Persistent neural activity: prevalence and mecha-
Dipoppa, M. (2012). The Role of Correlations and Oscillations as a Unified nisms. Curr. Opin. Neurobiol. 14, 675–684. doi: 10.1016/j.conb.2004.10.017
Mechanism Controlling Persistent Neural Activity and Working Memory. PhD Marder, E., and Bucher, D. (2001). Central pattern generators and the con-
thesis, Universite Pierre et Marie Curie. trol of rhythmic movements. Curr. Biol. 11, R986–R996. doi: 10.1016/S0960-
Douglass, A. D., Kraves, S., Deisseroth, K., Schier, A. F., and Engert, 9822(01)00581-4
F. (2008). Escape behavior elicited by single, channelrhodopsin-2-evoked McFarland, J. L., and Fuchs, A. F. (1992). Discharge patterns in nucleus preposi-
spikes in zebrafish somatosensory neurons. Curr. Biol. 18, 1133–1137. doi: tus hypoglossi and adjacent medial vestibular nucleus during horizontal eye
10.1016/j.cub.2008.06.077 movement in behaving macaques. J. Neurophysiol. 68, 319–332.
Egorov, A. V., Hamam, B. N., Fransen, E., Hasselmo, M. E., and Alonso, A. A. Mensh, B. D., Aksay, E., Lee, D. D., Seung, H. S., and Tank, D. W. (2004).
(2002). Graded persistent activity in entorhinal cortex neurons. Nature 420, Spontaneous eye movements in goldfish: oculomotor integrator performance,
173–178. doi: 10.1038/nature01171 plasticity, and dependence on visual feedback. Vision Res. 44, 711–726. doi:
Eliasmith, C. (2005). A unified approach to building and controlling spiking attrac- 10.1016/j.visres.2003.10.015
tor networks. Neural Comput. 17, 1276–1314. doi: 10.1162/0899766053630332 Miri, A., Daie, K., Arrenberg, A. B., Baier, H., Aksay, E., and Tank, D. W. (2011a).
Fenno, L., Yizhar, O., and Deisseroth, K. (2011). The development and application Spatial gradients and multidimensional dynamics in a neural integrator circuit.
of optogenetics. Annu. Rev. Neurosci. 34, 389–412. doi: 10.1146/annurev-neuro- Nat. Neurosci. 14, 1150–1159. doi: 10.1038/nn.2888
061010-113817 Miri, A., Daie, K., Burdine, R. D., Aksay, E., and Tank, D. W. (2011b). Regression-
Fransén, E., Tahvildari, B., Egorov, A. V., Hasselmo, M. E., and Alonso, A. A. (2006). based identification of behavior-encoding neurons during large-scale optical
Mechanism of graded persistent cellular activity of entorhinal cortex layer v imaging of neural activity at cellular resolution. J. Neurophysiol. 105, 964–980.
neurons. Neuron 49, 735–746. doi: 10.1016/j.neuron.2006.01.036 doi: 10.1152/jn.00702.2010
Fukushima, K., Kaneko, C. R., and Fuchs, A. F. (1992). The neuronal substrate Moreau, L., and Sontag, E. (2003). Balancing at the border of instability. Phys. Rev.
of integration in the oculomotor system. Prog. Neurobiol. 39, 609–639. doi: E Stat. Nonlin. Soft Matter Phys. 68:020901. doi: 10.1103/PhysRevE.68.020901
10.1016/0301-0082(92)90016-8 Nagel, G., Szellas, T., Huhn, W., Kateriya, S., Adeishvili, N., Berthold, P.,
Gold, J. I. and Shadlen, M. N. (2001). Neural computations that underlie deci- et al. (2003). Channelrhodopsin-2, a directly light-gated cation-selective
sions about sensory stimuli. Trends Cogn. Sci. 5, 10–16. doi: 10.1016/S1364- membrane channel. Proc. Natl. Acad. Sci. U.S.A. 100, 13940–13945. doi:
6613(00)01567-9 10.1073/pnas.1936192100
Goldman, M., Compte, A., and Wang, X. (2009). “Neural integrator models,” in Parisien, C., Anderson, C. H., and Eliasmith, C. (2008). Solving the problem of
Encyclopedia of Neuroscience, Vol. 6, ed L. R. Squire (Oxford: Academic Press). negative synaptic weights in cortical models. Neural Comput. 20, 1473–1494.
165–178. doi: 10.1162/neco.2008.07-06-295
Goldman, M. S., Levine, J. H., Major, G., Tank, D. W., and Seung, H. S. (2003). Pastor, A. M., la Cruz, R. R. D., and Baker, R. (1994). Eye position and eye velocity
Robust persistent neural activity in a model integrator with multiple hys- integrators reside in separate brainstem nuclei. Proc. Natl. Acad. Sci. U.S.A. 91,
teretic dendrites per neuron. Cereb. Cortex 13, 1185–1195. doi: 10.1093/cercor/ 807–811. doi: 10.1073/pnas.91.2.807
bhg095 Robinson, D. A. (1968). Eye movement control in primates. the oculomotor system
Gonçalves, P. J. (2012). A Neural Circuit Model of the Oculomotor Integrator: Theory contains specialized subsystems for acquiring and tracking visual targets. Science
for Optogenetic Dissection. PhD thesis, Universite Pierre et Marie Curie. 161, 1219–1224. doi: 10.1126/science.161.3847.1219
Hahnloser, R. H. R., Kozhevnikov, A. A., and Fee, M. S. (2002). An ultra-sparse Schoonheim, P. J., Arrenberg, A. B., Del Bene, F., and Baier, H. (2010).
code underlies the generation of neural sequences in a songbird. Nature 419, Optogenetic localization and genetic perturbation of saccade-generating neu-
65–70. doi: 10.1038/nature00974 rons in zebrafish. J. Neurosci. 30, 7111–7120. doi: 10.1523/JNEUROSCI.5193-
Han, X., and Boyden, E. S. (2007). Multiple-color optical activation, silencing, 09.2010
and desynchronization of neural activity, with single-spike temporal resolution. Scott, E. K., Mason, L., Arrenberg, A. B., Ziv, L., Gosse, N. J., Xiao, T., et al.
PLoS ONE 2:e299. doi: 10.1371/journal.pone.0000299 (2007). Targeting neural circuitry in zebrafish using gal4 enhancer trapping.
Huber, D., Petreanu, L., Ghitani, N., Ranade, S., Hromádka, T., Mainen, Z., Nat. Methods 4, 323–326. doi: 10.1038/nmeth1033
et al. (2008). Sparse optical microstimulation in barrel cortex drives learned Seung, H. S. (1996). How the brain keeps the eyes still. Proc. Natl. Acad. Sci. U.S.A.
behaviour in freely moving mice. Nature 451, 61–64. doi: 10.1038/nature06445 93, 13339–13344. doi: 10.1073/pnas.93.23.13339
Kleinlogel, S., Feldbauer, K., Dempski, R. E., Fotis, H., Wood, P. G., Bamann, Seung, H. S., Lee, D. D., Reis, B. Y., and Tank, D. W. (2000). Stability of the memory
C., et al. (2011). Ultra light-sensitive and fast neuronal activation with the of eye position in a recurrent network of conductance-based model neurons.
Ca2+ -permeable channelrhodopsin catch. Nat. Neurosci. 14, 513–518. doi: Neuron 26, 259–271. doi: 10.1016/S0896-6273(00)81155-1
10.1038/nn.2776 Singh, R., and Eliasmith, C. (2006). Higher-dimensional neurons explain the tun-
Koulakov, A. A., Raghavachari, S., Kepecs, A., and Lisman, J. E. (2002). Model for a ing and dynamics of working memory cells. J. Neurosci. 26, 3667–3678. doi:
robust neural integrator. Nat. Neurosci. 5, 775–782. doi: 10.1038/nn893 10.1523/JNEUROSCI.4864-05.2006
Lima, S. Q., and Miesenböck, G. (2005). Remote control of behavior through Szobota, S., Gorostiza, P., Del Bene, F., Wyart, C., Fortin, D. L., Kolstad, K. D.,
genetically targeted photostimulation of neurons. Cell 121, 141–152. doi: et al. (2007). Remote control of neuronal activity with a light-gated glutamate
10.1016/j.cell.2005.02.004 receptor. Neuron 54, 535–545. doi: 10.1016/j.neuron.2007.05.010
Loewenstein, Y., and Sompolinsky, H. (2003). Temporal integration by calcium Wong, K.-F., Huk, A. C., Shadlen, M. N., and Wang, X.-J. (2007). Neural circuit
dynamics in a model neuron. Nat. Neurosci. 6, 961–967. doi: 10.1038/nn1109 dynamics underlying accumulation of time-varying evidence during percep-
Lopez-Barneo, J., Darlot, C., Berthoz, A., and Baker, R. (1982). Neuronal activ- tual decision making. Front. Comput. Neurosci. 1:6. doi: 10.3389/neuro.10.
ity in prepositus nucleus correlated with eye movement in the alert cat. 006.2007
J. Neurophysiol. 47, 329–352. Zhang, F., Aravanis, A. M., Adamantidis, A., de Lecea, L., and Deisseroth, K.
Luo, L., Callaway, E. M., and Svoboda, K. (2008). Genetic dissection of neural (2007a). Circuit-breakers: optical technologies for probing neural signals and
circuits. Neuron 57, 634–660. doi: 10.1016/j.neuron.2008.01.002 systems. Nat. Rev. Neurosci. 8, 577–581. doi: 10.1038/nrn2192

Zhang, F., Wang, L.-P., Brauner, M., Liewald, J. F., Kay, K., Watzke, N., et al. (2007b). Received: 26 February 2013; accepted: 30 January 2014; published online: 28 February
Multimodal fast optical interrogation of neural circuitry. Nature 446, 633–639. 2014.
doi: 10.1038/nature05744 Citation: Gonçalves PJ, Arrenberg AB, Hablitzel B, Baier H and Machens CK (2014)
Zhang, K. (1996). Representation of spatial orientation by the intrinsic dynamics Optogenetic perturbations reveal the dynamics of an oculomotor integrator. Front.
of the head-direction cell ensemble: a theory. J. Neurosci. 16, 2112–2126. Neural Circuits 8:10. doi: 10.3389/fncir.2014.00010
Zhu, P., Narita, Y., Bundschuh, S. T., Fajardo, O., Schärer, Y.-P. Z., Chattopadhyaya, This article was submitted to the journal Frontiers in Neural Circuits.
B., et al. (2009). Optogenetic dissection of neuronal circuits in zebrafish Copyright © 2014 Gonçalves, Arrenberg, Hablitzel, Baier and Machens. This is an
using viral gene transfer and the tet system. Front. Neural Circuits 3:21. doi: open-access article distributed under the terms of the Creative Commons Attribution
10.3389/neuro.04.021.2009 License (CC BY). The use, distribution or reproduction in other forums is permit-
ted, provided the original author(s) or licensor are credited and that the original
Conflict of Interest Statement: The authors declare that the research was con- publication in this journal is cited, in accordance with accepted academic practice.
ducted in the absence of any commercial or financial relationships that could be No use, distribution or reproduction is permitted which does not comply with these
construed as a potential conflict of interest. terms.

Goncalves Etal 2014

Uploaded by

Copyright:

Available Formats

Goncalves Etal 2014

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Goncalves Etal 2014

Uploaded by

Copyright:

Available Formats

ORIGINAL RESEARCH ARTICLE

published: 28 February 2014

Optogenetic perturbations reveal the dynamics of an

equally to this work.

INTRODUCTION position (Lopez-Barneo et al., 1982; Delgado-García et al., 1989;

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 1

to explain the remarkable robustness of the system (Koulakov

RESULTS generated by motoneurons in the abducens and oculomotor cra-

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 2

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 3

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 4

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 5

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 6

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 7

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 8

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 9

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 10

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 11

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 12

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 13

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 14

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 15

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 16

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 17

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 18

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 19

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 20

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 21

Frontiers in Neural Circuits www.frontiersin.org February 2014 | Volume 8 | Article 10 | 22

You might also like