Opposing influence of top-down and bottom-up input on different types of excitatory layer 2/3 neurons in mouse visual cortex

Context

Predictive coding is a classic topic in neuroscience because prediction is central to the function of nervous systems. From accurately maintaining a 2-m distance with coughing people on the street to planning a week of remote work, generating predictions is a key brain process. In the past decade, Georg Keller’s lab has lead the effort in studying predictive processing in the mouse primary visual cortex (V1)¹. Their work has typically consisted of Ca²⁺ imaging of neural activity across multiple neural cell types during a visuomotor mismatch paradigm². This paradigm, also used in the preprint discussed here, entails a head-fixed mouse running in a visual virtual reality corridor where the speed of the visual flow is coupled to movement of a floating styrofoam ball, which the mouse moves as it runs. The visual flow is then suddenly stopped for 1 second, causing a mismatch between visual flow and mouse movement, which triggers robust mismatch responses in the V1 neurons.

The general framework of predictive coding with a forward model is well explained in [ref 1] and the below video. The hypothesis calls for the existence of three key nodes: representation neurons (observation of external world), prediction neurons (internal forward model) and prediction error neurons (comparison of observation and internal prediction). In the visuomotor mismatch paradigm, the internal forward model should predict expected visual flow speed based on locomotion speed, constituting an efference copy. If the hypothesis holds, the prediction error neurons should compute a difference of the input they get from representation and prediction neurons. The preprint demonstrates that this is indeed the case in L2/3 of V1.

 

 

Key findings

The preprint shows that L2/3 pyramidal neurons have heterogeneous subthreshold membrane potential responses to visuomotor mismatch. Many of them have depolarizing mismatch responses (dMM), and some have hyperpolarizing mismatch responses (hMM). While these had been indicated by previous data with Ca²⁺ imaging³, this is the first demonstration of these subtypes with whole cell recordings. The method is key here because it reveals subthreshold membrane potential (V_m) responses arising from synaptic input. It also allowed the authors to report electrophysiological properties of these cells for the first time.

The authors then record from the same neurons with uncoupled visual flow and locomotion to find out their separate influences on the cells. These recordings demonstrate that the sensory and locomotion computation in most of these neurons are consistent with the predictive coding hypothesis: in dMM neurons V_m correlation with locomotion speed is positive (prediction input) while correlation with  visual flow speed is negative (observation input), while the opposite is true of hMM neurons. This arrangement accounts for the mismatch responses of both cell types as a comparison of their sensory and locomotion responses, as expected for prediction error neurons. The dMM neurons had elevated resting membrane potential and spike threshold compared to hMM neurons, suggesting specialized electrophysiological ‘tuning’ to their functional roles.

The authors go on to record from L5/6 pyramidal neurons, showing that they have predominantly hyperpolarizing subthreshold mismatch responses. In other words, L5/6 neurons do not report prediction errors to downstream neurons. The L5/6 neurons have positive V_m correlations with locomotion and visual flow speed. So, instead of computing a prediction error response, they hyperpolarize either as a simple result of diminished visual input or through actively inhibitory input. This potential inhibition might arise from L5/6 interneurons that are excited by L2/3 dMM prediction error signals, as such connections can exist based on connectivity data⁴.

The authors end with a profound yet complex plot summarizing the differences between mismatch responses in L2/3 and L5/6 (their figure 6A,B, reprinted below). This key plot shows that mismatch responses in L2/3 are true prediction error responses because they arise from opposing signs of membrane potential correlation with locomotion speed vs visual flow speed. On the contrary, L5/6 hyperpolarizing mismatch responses are not prediction error responses because they arise from positive correlations of membrane potential with both locomotion and visual flow speeds (Figure 1).

 

Why I chose this preprint 

This study is important because it contains the first whole-cell recordings of mismatch responses from the Keller lab. In contrast to the previous Ca²⁺ imaging datasets, this new work shows subthreshold membrane potential responses in L2/3 neurons, as well as in L5/6 which has not been possible to reach with Ca²⁺ imaging. The findings show a good fit to what would be expected from prediction error computation in L2/3 pyramidal neurons. However, in L5/6 the neurons do not have the expected responses. Instead L5/6 pyramidal neurons fit the profile of representation neurons, where they would simply encode mouse speed through the world. Thus only L2/3 computes prediction errors. The work thereby demonstrates layer specific sensorimotor integration.

 

 

What next?

This work shows robust evidence for prediction error computation in V1 L2/3 and raises many interesting questions. Obviously the mechanisms generating the responses are of interest, and they are schematized by the authors in their Figure 6D. A previous study from the Keller lab concluded that somatostatin (SOM) and vasoactive intestinal peptide (VIP) interneurons are involved³. The delta oscillation seen in whole-cell traces during visual flow surely suggests involvement of a slowish excitatory/inhibitory oscillatory circuit somewhere in the mix. This could arise from burst firing of mutually inhibitory VIP and SOM interneuron populations. Top-down projections from the cingulate cortex are known to disinhibit V1 through primarily targeting VIP interneurons⁵. Previous work from the Keller lab indicates that this projection is also important for generating mismatch responses and has topographical organization on a gross level⁶. While it has not been quantified whether this projection is organized into finely clustered axonal fields in L1 of V1 (though Figure3A in Leinweber et al. 2017⁶ suggests this may be the case), inputs from thalamus and secondary visual areas certainly are. Thalamic and secondary visual axons form an interleaved patchwork of ~60um diameter patches in V1, and these correspond to muscarinic receptor distributions and differences in interneuron circuitry^7,8. Various other strands of evidence suggest existence of local interneuron teams at a similar spatial scale⁹. Might there be a topographical distribution of hyperpolarizing and depolarizing mismatch neurons associated with these patches?

How does V1 learn to generate these experience dependent responses?³ Are the responses ‘calibrated’ to match the visual flow conditions? For example, if a mouse were to move from the tall grass to arid plains, its visual flow might require retuning. For many eye movements, the cerebellum is thought to perform the role of a ‘teacher’, which is not required for the movement itself, but enables its tuning. Might loops between the cerebellum and prefrontal cortex, through thalamus and pontine nuclei^10–12, also be involved in modifying V1 prediction error responses? It has been demonstrated that V1 SOM interneurons do not undergo marked plasticity while parvalbumin (PV) and neurogliaform neuropeptide Y (NPY) interneurons do^1,3,13. PV and NPY interneurons inhibit SOM interneurons, pyramidal cells and VIP interneurons, which positions them as potential mediators of prediction error modifications.

How do V1 mismatch responses differ from those in other sensory cortices with similar circuitry, like somatosensory or auditory cortex? For example, is the hMM response in L5/6 somehow related to sensory suppression that occurs during saccades or self-generated sounds? The potential logic being that after reporting an observation, sensory representation should be suppressed during mismatch in order to prevent it from participating in further processing loops or modification.

Finally, how do the various mismatch neurons interact with each other? As the authors discuss, the predictive coding framework suggests the prediction error neurons and representative neurons would communicate to update predictions.¹ As PV and NPY interneurons have strong, experience dependent mismatch responses³, perhaps dMMs drive them during mismatch causing hMM responses in L5/6. Would this inhibition be an update of the representation?

 

References:

1. Keller, G. B. & Mrsic-Flogel, T. D. Predictive Processing: A Canonical Cortical Computation. Neuron 100, 424–435 (2018).
2. Keller, G. B., Bonhoeffer, T. & Hübener, M. Sensorimotor mismatch signals in primary visual cortex of the behaving mouse. Neuron 74, 809–815 (2012).
3. Attinger, A., Wang, B. & Keller, G. B. Visuomotor Coupling Shapes the Functional Development of Mouse Visual Cortex. Cell 169, 1291-1302.e14 (2017).
4. Jiang, X. et al. Principles of connectivity among morphologically defined cell types in adult neocortex. Science 350, aac9462–aac9462 (2015).
5. Zhang, S. et al. Selective attention. Long-range and local circuits for top-down modulation of visual cortex processing. Science 345, 660–5 (2014).
6. Leinweber, M., Ward, D. R., Sobczak, J. M., Attinger, A. & Keller, G. B. A Sensorimotor Circuit in Mouse Cortex for Visual Flow Predictions. Neuron 95, 1420-1432.e5 (2017).
7. D’Souza, R. D., Bista, P., Meier, A. M., Ji, W. & Burkhalter, A. Spatial Clustering of Inhibition in Mouse Primary Visual Cortex. Neuron 104, 588-600.e5 (2019).
8. Ji, W. et al. Modularity in the Organization of Mouse Primary Visual Cortex. Neuron 87, 632–643 (2015).
9. Karnani, M. M. & Jackson, J. Interneuron Cooperativity in Cortical Circuits. Neuroscientist 24, 329–341 (2018).
10. Gao, Z. et al. A cortico-cerebellar loop for motor planning. Nature 563, 113–116 (2018).
11. Chabrol, F., Blot, A. & Mrsic-Flogel, T. D. Cerebellar contribution to preparatory activity in motor neocortex. bioRxiv 335703 (2018) doi:10.1101/335703.
12. Proville, R. D. et al. Cerebellum involvement in cortical sensorimotor circuits for the control of voluntary movements. Nat. Neurosci. 17, 1233–9 (2014).
13. Xue, M., Atallah, B. V. & Scanziani, M. Equalizing excitation–inhibition ratios across visual cortical neurons. Nature 511, 596–600 (2014).

Tags: in vivo patch clamp, mismatch, mouse, pyramidal neuron, v1, visual cortex, whole cell recording

Posted on: 7 April 2020

doi: https://doi.org/10.1242/prelights.18184

Read preprint (No Ratings Yet)