CNS*2020 Online

Neurostars discussion link

One of the central questions in neuroscience is how structure of brain circuits determines their activity and function. To explore such structure-function relations systematically, we integrate information from large-scale experimental surveys into data-driven, bio-realistic models of brain circuits, with the current focus on the mouse cortex.

Our 230,000-neuron models of the mouse cortical area V1 [1] were constructed at two levels of granularity – using either biophysically-detailed or point-neurons. These models systematically integrated a broad array of experimental data [1–3]: the information about distribution and morpho-electric properties of different neuron types in V1; connection probabilities, synaptic weights, axonal delays, and dendritic targeting rules inferred from a thorough survey of the literature; and a sophisticated representation of visual inputs into V1 from the Lateral Geniculate Nucleus, fit to in vivo recordings. The model activity has been tested against large-scale in vivo recordings of neural activity [4]. We found a good agreement between these experimental data and the V1 models for a variety of metrics, such as direction selectivity, as well as less good agreement for other metrics, suggesting avenues for future improvements. In the process of building and testing models, we also made predictions about the logic of recurrent connectivity with respect to functional properties of the neurons, some of which have been verified experimentally [1].

In this presentation, we will focus on the model’s successes in quantitative matching of multiple experimental measures, as well as failures in matching other metrics. Both successes and failures shed light on the potential structure-function relations in cortical circuits, leading to experimentally testable hypotheses. Our models are shared freely with the community: https://portal.brain-map.org/explore/models/mv1-all-layers. We also freely share our software tools – the Brain Modeling ToolKit (BMTK; alleninstitute.github.io/bmtk/), which is a software suite for model building/simulation [5], and the SONATA file format [6] (github.com/allenInstitute/sonata).

References
1. Billeh, Y. N. et al. Systematic Integration of Structural and Functional Data into Multi-scale Models of Mouse Primary Visual Cortex. Neuron 106, 388-403.e18 (2020).
2. Gouwens, N. W. et al. Classification of electrophysiological and morphological neuron types in the mouse visual cortex. Nat. Neurosci. 22, 1182–1195 (2019).
3. Gouwens, N. W. et al. Systematic generation of biophysically detailed models for diverse cortical neuron types. Nat. Commun. 9, 710 (2018).
4. Siegle, J. H. et al. A survey of spiking activity reveals a functional hierarchy of mouse corticothalamic visual areas. bioRxiv 805010 (2019) doi:10.1101/805010.
5. Gratiy, S. L. et al. BioNet: A Python interface to NEURON for modeling large-scale networks. PLoS One 13, e0201630 (2018).
6. Dai, K. et al. The SONATA data format for efficient description of large-scale network models. PLOS Comput. Biol. 16, e1007696 (2020).

Ali Almasi, Hamish Meffin, Shi Sun, Michael R Ibbotson

Neurostars discussion link

Our understanding of sensory coding in the visual system is largely derived from parametrizing neuronal responses to basic stimuli. Recently, mathematical tools have developed that allow estimating the parameters of a receptive field (RF) model, which are typically a cascade of linear filters on the stimulus, followed by static nonlinearities that map the output of the filters to the neuronal spike rates. However, how much do these characterizations depend on the choice of the stimulus type?

We studied the changes that neuronal RF models undergo due to the change in the statistics of the visual stimulus. We applied the nonlinear input model (NIM) [1] to the recordings of single units in cat primary visual cortex (V1) in response to white Gaussian noise (WGN) and natural scenes (NS). These two stimulus types were matched in their global RMS contrast; however, they are fundamentally different in terms of second- and higher-order statistics, which are abundant in natural scenes but do not exist in white noise. We estimated for each cell the spatial filters constituting the neuronal RF and their corresponding nonlinear pooling mechanism, while making minimal assumptions about the underlying neuronal processing.

We found that cells respond differently to these two stimulus types, with mostly higher spike rates and shorter response latencies to NS than to WGN. The most striking finding was that NS stimuli resulted in around twice as many uncovered RF filters compared to using WGN stimuli. Via careful analysis of the data, we discovered that this difference between the number of identified RF filters is not related to the higher spike rates of cells to NS stimuli. Instead, we found it to be attributed to the difference in the contrast levels of specific features that exhibit different prevalence in NS versus WGN. These features correspond to the V1 RF filters recovered in the model. This specific feature-contrast attains much higher values in NS compared to WGN stimuli. When the feature-contrast is controlled for, it explains the differences in the number of RF filters obtained. Our findings imply that a greater extent of nonlinear processing in V1 neurons can be uncovered using natural scene stimulation.

Acknowledgements The authors acknowledge the support the Australian Research Council Centre of Excellence for Integrative Brain function (CE140100007), the National Health and Medical Research Council (GNT1106390), and Lions Club of Victoria.

References

[1] McFarland JM, Cui Y, Butts DA. Inferring nonlinear neuronal computation based on physiologically plausible inputs. PLoS Comput Biol. 2013, 9(7).

Yoram Ben-Shaul, Rohini Bansal, Romana Stopkova, Maximilian Nagel, Pavel Stopka, Marc Spehr

Neurostars discussion link

The broad goal of this work is to understand how consistency on a macroscopic scale can be achieved despite random connectivity at the level of individual neurons.

A central aspect of any sensory system is the manner by which features of the external world are represented by neurons at various processing stages. Yet, it is not always clear what these features are, how they are represented, and how they emerge mechanistically. Here, we investigate this issue in the context of the vomeronasal system (VNS), a vertebrate chemosensory system specialized for processing of cues from other organisms. We focus on the accessory olfactory bulb AOB, which receives all vomeronasal sensory neuron inputs. Unlike the main olfactory system, where MTCs sample information from a single receptor type, AOB MTCs sample information from a variable number of glomeruli, in a manner that seems largely random. This apparently random connectivity is puzzling given the presumed role of this system in processing cues with innately relevant significance.

We use multisite extracellular recordings to measure the responses of mouse AOB MTCs to controlled presentation of natural urine stimuli from male and female mice from various strains, including from wild mice. Crucially, we also measured the levels of both volatile and peptide chemical components in the very same stimulus samples that were presented to the mice. As subjects, we used two genetically distinct mouse strains, allowing us to test if macroscopic similarity can emerge despite variability at the level of receptor expression.

First, we then explored neuronal receptive fields, and found that neurons selective for specific strains (regardless of sex), or a specific sex (regardless of strain), are less common than expected by chance. This is consistent with our previous findings indicating that high level stimulus features are represented in a distributed manner in the AOB. We then compared various aspects of neuronal responses across the two strains, and found a high degree of correlation among them, suggesting that despite apparent randomness and strain specific genetic aspects, consistent features emerge at the level of the AOB.

Next, we set out to model the responses of AOB neurons. Briefly, AOB responses to a given stimulus are modelled as dot products of random tuning profiles to specific chemicals and the actual level of those chemicals in the stimulus. In this manner we derive a population of AOB responses, which we can then compare to the measured responses. Our analysis thus far reveals several important insights. First, neuronal response properties are best accounted for by sampling of protein/peptide components, but not by volatile urinary components. This is consistent with the known physiology of the VNS. Second, several response features (population level neuronal distances, sparseness, distribution of receptive field types) are best reproduced in the model with random sampling of multiple, rather than single molecules per neuron. This suggests that the sampling mode of AOB neurons may mitigate some of the consequences of random sampling. Finally, we note that random sampling of molecules provides a reasonable fit for some, but not all metrics of the observed responses. Our ongoing work aims to identify which changes must be made to our initial simplistic model in order to account for these features.

This work is funded by GIF and DFG grants to Marc Spehr and Yoram Ben-Shaul