A neural mechanism for looming-evoked escape behavior

By Yatang Li

Neural pathways for defensive behaviors elicited by non-visual cues include hypothalamus, amygdala and periaqueductal gray (PAG) (Silva et al., 2016). Although visual cues play an important role in the defensive behaviors of animals, the underlying neural circuit mechanism remain poorly understood. In the mouse, an overhead expanding dark disc, which mimics an aerial approaching predator, elicits robust freezing or flight behaviors (Yilmaz and Meister, 2013). Previous studies indicate that the superior colliculus (SC) occupies a key node in the neural pathway controlling looming-evoked behaviors. Freezing behaviors appear to be mediated via the retina-SC-thalamus-amygdala pathway (Wei et al., 2015) or the retina-SC-PBGN-amygdala pathway (Shang et al., 2015). In a recent journal club, we discussed a paper from Tiago Branco’s lab at UCL about the neural circuit for the flight behavior (Evans et al., 2018). This paper, titled “A synaptic threshold mechanism for computing escape decisions”, was posted on February 27, 2018 in BioRxiv.


In this study, Evans et al. demonstrated a monosynaptic excitatory connection from the deep layers of medial superior colliculus (dmSC) to dorsal PAG (dPAG) neurons that is necessary for initiating escape behavior via a threshold mechanism. Although the connection is weak and unreliable, the synaptic drive to dPAG can be strongly amplified by recurrent excitation and short-term synaptic facilitation in the dmSC.

The authors started with an exploration of how the behaviors of mice were modulated by looming stimuli with different contrasts. They showed that both the response time and running speed were correlated with the contrast. This relationship was modeled using a variant of drift-diffusion model, a popular model in perceptual decision making in primates. The key variable of the model is the threat level which increases with sensory evidence and decays over time. The authors found this model fit well with their behavioral data. Nevertheless, the interpretation deserves more thought. How is the internal threat reflected by behavioral measurements? Does the running speed reflect the threat level above the threshold? If the authors monitor the animal’s autonomous responses or muscle activity, they may have a better estimation about the quality of their model below the threshold for flight. In addition, an alternative model should be provided for fitting the behavioral data. One such hypothesis is that the mouse escapes simply because it recognizes an overhead looming stimulus. To test this hypothesis, the authors need also include the posture and orientation of the mouse, which determine the visual field of the looming stimulus.


Then, the authors aimed to define the critical nodes for the behavior. They focused on mSC, V1, amygdala, and dPAG; but they did not provide rationales for why they only chose these four regions. They found that inactivation of V1 or amygdala with muscimol had a modulatory effect on the behavior. Interestingly, inactivation of mSC and dPAG chemically or optogenetically affected the behavior in different manners. Specifically, inactivation of the dPAG led to a switch from escape to freezing behavior while inactivation of the mSC eliminated the defensive responses. However, this is not well supported by their data as shown below. Clearly, inactivation of either region decreased the running speed. However, it’s not clear whether the mouse entered a freezing state or just took a rest. This ambiguity may be primarily attributed to the unclear definition of freezing.


Calcium imaging revealed that dmSC neurons encode both visual and behavioral information while dPAG neurons encode behaviors only. We think it would be interesting if the authors can show the neuronal responses to stimuli of different contrasts during escape behavior and fit that to the drift-diffusion model. If a dmSC neuron’s activity can be fitted as consistently as the behavioral data, this neuron is likely to encode the threat level. In addition, the different roles of dmSC and dPAG were demonstrated by optogenetic stimulation of mSC and dPAG with different laser power. When the light intensity was increased gradually, dmSC neurons showed a progressive increase in the probability of escape and a decrease of the variability of the responses. In contrast, dPAG neurons showed a steep increase in the probability of escape and stereotyped responses for each intensity.


How does the SC pass the signal to the PAG? Retrograde monosynaptic rabies tracing showed that the dPAG receives inputs mainly from deep and intermediate layers of the SC with a 11:1 convergence ratio. In vitro whole-cell patch-clamp data showed that optogenetic stimulation of the mSC with a single light pulse resulted in very small postsynaptic currents in the dPAG and that the statistics of neurotransmitter release were not different from a Poisson model. However, the postsynaptic currents were significantly increased with 5 light pulses at 20Hz, which are attributed to short-term facilitation among the light pulses and a long lasting increase in sEPSCs frequency after the pulses. Further experiments indicated that the light-train-evoked neuronal responses were amplified by the strong recurrent excitatory connection found in the dmSC. The authors also confirmed that this 20 Hz-light-evoked response is in the range of sensory-evoked response. Lastly, the authors tested the necessity of the dPAG by co-expressing the synaptically targeted inhibitory designer receptor hM4D-neurexin5 and ChR2 in VGluT2+ dmSC neurons. When the synaptic inputs to dPAG were suppressed, the animals stopped showing escape behaviors. In contrast, inhibition of the projections to the lateral posterior nucleus of the thalamus (LP) slightly increased the escape probability. The authors provide a circuit model in which evidence about threats is integrated in the dmSC network and passed through a threshold at the dPAG level to initiate defensive escape.


In sum, this is a nice study on an interesting question using a variety of techniques. It demonstrated a novel neural mechanism for looming-evoked escape behavior in which the SC, not the visual cortex or the amygdala, plays an important role.



Evans, D.A., Stempel, V., Vale, R., Ruehle, S., Lefler, Y., and Branco, T. (2018). A synaptic threshold mechanism for computing escape decisions. BioRxiv 272492.

Shang, C., Liu, Z., Chen, Z., Shi, Y., Wang, Q., Liu, S., Li, D., and Cao, P. (2015). A parvalbumin-positive excitatory visual pathway to trigger fear responses in mice. Science 348, 1472–1477.

Silva, B.A., Gross, C.T., and Gräff, J. (2016). The neural circuits of innate fear: detection, integration, action, and memorization. Learn. Mem. 23, 544–555.

Wei, P., Liu, N., Zhang, Z., Liu, X., Tang, Y., He, X., Wu, B., Zhou, Z., Liu, Y., Li, J., et al. (2015). Processing of visually evoked innate fear by a non-canonical thalamic pathway. Nat. Commun. 6, 6756.

Yilmaz, M., and Meister, M. (2013). Rapid Innate Defensive Responses of Mice to Looming Visual Stimuli. Curr. Biol. 23, 2011–2015.



Superior Colliculus for Rapid Sensorimotor mapping

By Mu Qiao

We recently discussed two papers from Carlos Brody’s lab about the role of the superior colliculus in rapid executive control of rats’ sensorimotor behaviors. The first paper, “Requirement of Prefrontal and Midbrain Regions for Rapid Executive Control of Behavior in the Rat” was published in Neuron in June 2015 [1]. The second paper, titled “Collicular circuits for flexible sensorimotor routing,” was posted on Biorxiv in January 2018 [2].

Context-dependent fast sensorimotor routing, an important aspect of cognition and executive control, requires an animal to quickly adapt to different rules under different contexts. This rapid sensorimotor mapping has been studied extensively in primates, using tools such as the pro/anti-saccade paradigm [3]. In this paradigm, the subject fixates on the center of a screen, where a colored dot appears to instruct the subject to either saccade towards a peripheral stimulus (pro-saccade) or saccade in the opposite direction (anti-saccade). The subject has to change their saccade strategy based on the color of the central dot (context). This paradigm has rarely been applied to rodents. Can rodents perform a similar task? If so, which brain regions are responsible for sensorimotor routing, and how is the task encoded in these brain regions? The authors tried to address these questions in these two papers.

In the first paper, Duan and colleagues developed a cued pro/anti-orienting task, which is similar to the pro/anti-saccade paradigm in primates. Briefly, a rat is facing three ports in front of it. The light from the central port indicates the start of a trial. The animal has to nose-poke the central port and hold there for 1.5s. During the first 1s, one of two clearly distinguishable sounds is played to the animal. Of the two, one indicates that this is a pro-task, in which the animal will poke the port with the light on to get the reward, and the other sound is for an anti-task, in which the rat will poke the port without the light to get the reward. Incorrect choices will result in punishment with a loud sound and a short time-out.


Duan and colleagues trained rats to perform the task in blocks: task blocks switched only when the rats’ local performance was above 70%. Under this training procedure, the performance of the animals came close to 80%. The authors also proceeded with blocks of random length (with a mean of 15 trials and a standard deviation of 5 trials), in which the animals’ performance dropped to ~77%, and random interleaved trials, in which the animals’ performance was ~72-75%. The authors further analyzed data from the blocked sessions. They found behavioral asymmetries between the pro- and anti-tasks. The rats had a lower error rate and a shorter reaction time in the pro-task than in the anti-task. In addition, the authors noticed that when the animals switched from the pro- to the anti-task, the accuracy drop and the reaction time change were smaller. They called this switch-cost asymmetry.

The authors then inactivated different brain regions to see how it affected animals’ performance. Two brain regions were selected: the superior colliculus (SC), which is known to be important in orienting behaviors [4], and the prefrontal cortex (PFC), which is suggested to be involved in primates’ pro/anti-saccade behaviors [3]. Unilateral inactivation of SC using muscimol led to orienting contralaterally to the inactivated side, consistent with the role of the SC in orienting behaviors. When the authors inactivated the SC bilaterally, surprisingly, they found a large decrease in performance in the anti-task while the decrease in performance effect on the pro-task was reduced, suggesting that the SC is not only important for orienting but may also be involved in executive control. Both unilateral and bilateral inactivation of the PFC led to a larger impairment on the anti-task than on the pro-task. In addition, bilateral SC and PFC inactivation reduced switch-cost from the anti- to the pro-task but left intact switch-cost from the pro- to the anti-task.


Results from the first paper suggested the role of the SC in executive control of rapid sensorimotor routing. But how is the information from the behavioral task processed in SC? To answer this question, in the second paper, Duan and colleagues performed tetrode recordings from SC when rats were performing the tasks.

The authors recorded neural signals from 193 neurons in the intermediate and deep layers of the SC and 331 neurons in the PFC. Encoding of these neurons was heterogeneous and multiplexed. The authors then applied a linear decoder to the SC neurons and the PFC neurons and see how their responses decoded the tasks and choices. Surprisingly, the decoding of task information was much better and the decoding of the choice information was significantly earlier in the SC than in the PFC.


The authors then separated these neurons into two groups based on the temporal profiles of their task selectivity: SC cue neurons, which predict the pro- or anti-tasks most strongly during the auditory cue period; and SC delay/choice neurons, which maintain the task information after the cue is presented. The authors thought the delay/choice neurons were important for the behavioral output because their task decoding was close to chance (~50%) in error trials, and their decoding of the choice was faster and more accurate. The authors then focused on the delay/choice neurons and found that their task selectivity was usually correlated with their choice selectivity. The pro-encoding neurons were usually contra-preferring and the anti-encoding neurons were often ipsi-preferring.

In addition, the authors optogenetically inactivated the SC during different phases of the tasks and they found that only inactivation during the delay period led to a significant impairment to the anti-task compared with the pro-task. The authors then modeled an SC circuit outputting animals’ behavioral choices based on results from the neural recording and the optogenetic inactivation. For a model with 12 free parameters, there could be 354 solutions of the circuit, suggesting a variety of configurations of the circuits in the SC.


Overall, these two papers provide evidence for the involvement of SC in rapid sensorimotor mapping during the pro/anti-switching task: in the first paper, the authors showed the necessity of the SC in the task by pharmacological inactivation; in the second paper, the authors presented neural correlations of the SC neurons to the tasks. What’s surprising to me from the second paper is that the neural decoding of the choice is faster in the SC than in the PFC, suggesting that against our intuition, choice information may not be computed first in the PFC and then delivered to the SC.

However, we also came up with a few questions and issues during the discussion. The first one is about the behavioral task. The center light is on and the animal approaches and pokes the central port to start a trial. This approaching behavior may add an additional bias to the pro-task. Maybe that’s why the animals’ performance on the anti-trials is not as good as that on the pro-trials. A better design might be to replace the central light with a sound.

A second question has to do with the behavioral asymmetries. The rats have better performance in the pro-tasks than in the anti-tasks, suggesting that the animals prefer approaching the visual target. Is this stimulus-specific? Our lab shows that mice will escape from a looming stimulus [5]. If we replace the light stimulus with a looming spot, will the animal learn the anti-task easily?

Another issue is about the inactivation experiments. In the first paper, the authors performed muscimol inactivation. This pharmacological inactivation was quite unnatural and could change the circuits severely. In the second paper, the authors performed optogenetic perturbation and they could control the time window of the inactivation. However, the results seem to be controversial. The SC cue neurons decode the task information perfectly, but inactivating the SC during the cue period didn’t affect performance. Also, the SC neurons encode choices faster but inactivation of the SC during the choice period didn’t affect the choice accuracy. The results will be more convincing if more sanity controls are included, such as testing whether unilateral optogenetic inactivation leads to a deficit in orienting behaviors. If true, does the result suggest redundant representations of the task and choice information in other brain regions? And does this depend on training?

Finally, we had a discussion on the model. Given that there are many parameter sets that can generate the model and there is not much in common between these models, it is hard to gain information from these models. Is there a way to use these candidate models to make predictions and then narrow down the model set based on some experimental tests?

To sum up, these two papers are interesting to us in the way that it was long believed that the cortex is where the executive control happens, and these studies suggest that subcortical regions such as the SC might also be an essential part.


[1]        Duan, C. A., Erlich, J. C. & Brody, C. D. Requirement of Prefrontal and Midbrain Regions for Rapid Executive Control of Behavior in the Rat. Neuron 86, 1491–1503 (2015).

[2]        Duan, C. A. et al. Collicular circuits for flexible sensorimotor routing. BioRxiv preprint. http://dx.doi.org/10.1101/245613. (2018)

[3]        Munoz, D.P., and Everling, S. (2004). Look away: the anti-saccade task and the voluntary control of eye movement. Nat. Rev. Neurosci. 5, 218–228.

[4]        Sparks, D.L. (1999). Conceptual issues related to the role of the superior colliculus in the control of gaze. Curr. Opin. Neurobiol. 9, 698–707.

[5]        Yilmaz, M, Meister, M (2013) Rapid innate defensive responses of mice to looming visual stimuli. Curr. Biol. 23:2011–2015.

Image recurrence sensitive retinal ganglion cells

By Kyu Hyun Lee

We recently discussed “Sensitivity to image recurrence across eye-movement-like image transitions through local serial inhibition in the retina” by Vidhyasankar Krishnamoorthy, Michael Weick, and Tim Gollisch [1]. It’s a nice paper about a novel response property in the mouse retina that the authors call image recurrence sensitivity (IRS).

Vision under natural conditions involves constant eye movements, such as large saccades, smooth pursuits for tracking moving objects, and smaller fixational movements. The last kind is the reason why we are able to see while focusing our gaze on an object – without it the image will soon fade due to neural adaptation (as can be demonstrated in the lab with retinal stabilization). Even though these eye movements likely pose a significant problem to the early visual system, little is known about how the retinal circuits deal with them. Krishnamoorthy and colleagues investigate this question in their paper by recording from the retinal ganglion cells (RGCs) of the isolated mouse retina using multi-electrode arrays. They present stimuli designed to mimic rapid eye movements: square wave gratings whose phase changes (by translation) to one of four possible locations. The transition period lasts about 100 ms, after which the grating remains stationary for another 800 ms before the next transition begins. The authors summarize the neural response to these stimuli with a matrix like the one below, whose (i,j)th element is the PSTH following a transition period (shaded) from position i to position j:

Curiously, they noticed that 5-10% of the RGCs in their recordings fired a burst of action potentials shortly after the transition period – but only if the starting and target positions were the same (noted by arrows in the figure above). In other words, these neurons were highly sensitive to the recurrence of the image that they just saw.

The authors call them IRS cells and suggest that they form a particular cell type: most of them have large cell bodies, non-overlapping receptive fields, similar temporal filter waveform, and transient OFF responses. Based on these observations, the authors think that they are the OFF transient α-RGCs [2].

This intriguing response property was then further studied along a number of dimensions. IRS was remarkably invariant to several kinds of changes. For example, it didn’t matter if the transition period was replaced with a gray illumination, or if the contrast between the starting and target positions were different. Spatially limiting the stimulus to cover only the receptive field center and replacing white bars with gray ones also had little effect. Finally, changing the stimulus from gratings to naturalistic images preserved the IRS (though the effect is weaker), indicating that this may generalize to real stimuli that a mouse would see.

The IRS cells were, however, sensitive to small spatial displacements. For example, if the target position was 45 µm removed from the starting position (measured on the retinal surface; about the size of retinal bipolar cell receptive field), the IRS would decline by about 50 %. Sensitivity to such small changes hints that the role of IRS may be to counteract small ocular drifts.

The authors then tried to understand the circuit-level mechanism of IRS by infusing pharmacological agents during recording. Blocking inhibition with strychnine (glycine antagonist) and gabazine (GABA antagonist) led to loss of IRS, but interestingly, by two distinct mechanisms: strychnine decreased the specificity (post-transition activity now occurs even if the starting and target position are different), whereas gabazine decreased the amplitude. This led to the idea that two types of amacrine cells are involved in this circuit, which is reflected in the model proposed by the authors: the IRS cell receives excitatory input from an OFF bipolar cell directly and inhibitory input from an ON bipolar cell sign-inverted via an ON amacrine cell; in addition, the OFF bipolar cell contacts an OFF amacrine cell, which inhibits the ON amacrine cell. This serial feedforward inhibition (similar to the switching circuit in [3]) is key to truncating the inhibition from the ON amacrine cell (from the transition period) and preserving the excitatory input from the OFF bipolar cell in the recurrence of dark patches. The authors were able to identify a set of parameters that qualitatively generates IRS response, as well as the other experimental observations.


Overall, I thought this was a very nice paper. It clearly presents an interesting response property uncovered by a careful stimulus design. The model is grounded on pharmacological experiments and intracellular recordings, and ties together many of the results. Though the experiments are done in the isolated retina, the authors make serious attempts to make their results relevant to the behaving animal by using realistic stimulus conditions (rapid image transitions) and naturalistic images (this should make those of us who still show static sinusoidal gratings to intact animals blush). Finally, I was surprised by the rich selectivity and invariance of the IRS cells – concepts usually reserved for neurons in the higher order cortical areas.

However, we also identified a number of lingering questions. The first has to do with the classification of IRS cells as α-RGCs. The authors’ observations lend some support to this, but it may not be enough. Since we now have genetic markers that unambiguously identify α-RGCs, it would have been good to see some molecular evidence, e.g. by filling the IRS cells after intracellular recording and staining them for α-RGC markers such as osteopontin [4].

Another question has to do with the putative function of these cells. As the authors acknowledge, it is unclear if the mouse makes small fixational eye movements to counteract gaze shifts. Given that the mouse retina lacks a fovea, eye movements in general may be less important for mice than e.g., head movements, and they may have developed a different strategy to combat adaptation-induced blindness. IRS is also agnostic of low-level visual features: it simply remembers a scene (with some negative contrast) and indicates its recurrence. Could a downstream neuron decode this for visual processing? And how long does this visual memory last? Figure 8 supplement 1 shows that the model loses IRS when transition periods becomes longer than 100 ms, but it would be good to see experimental evidence for this. This may be important for understanding the relevance of this response property to behavior, which happens on the order of seconds.

All in all, this study reminds us that a brain area that has been extensively studied (such as the retina) can still yield novel insights and open up new avenues of research when one cleverly matches the experimental condition to the natural ethology of the animal.

[1]       V. Krishnamoorthy, M. Weick, and T. Gollisch, “Sensitivity to image recurrence across eye-movement-like image transitions through local serial inhibition in the retina,” Elife, vol. 6, pp. 1–25, 2017.

[2]       B. Krieger, M. Qiao, D. L. Rousso, J. R. Sanes, and M. Meister, “Four alpha ganglion cell types in mouse retina: Function, structure, and molecular signatures,” PLoS One, vol. 12, no. 7, pp. 1–21, 2017.

[3]       M. N. Geffen, S. E. J. de Vries, and M. Meister, “Retinal Ganglion Cells Can Rapidly Change Polarity from Off to On,” PLoS Biol., vol. 5, no. 3, p. e65, 2007.

[4]       J. R. Sanes and R. H. Masland, “The types of retinal ganglion cells: current status and implications for neuronal classification,” Annu. Rev. Neurosci., vol. 38, pp. 221–246, 2015.