Recurrent interactions in local cortical circuits

Simon Peron, Ravi Pancholi, Bettina Voelcker, Jason D. Wittenbach, H. Freyja Ólafsdóttir, Jeremy Freeman, Karel Svoboda

TL;DR

By combining computational modeling with precision laser "photoablation" in living mice, researchers discovered that local recurrent circuits in the cortex act as specialized amplifiers for sensory signals. Surprisingly, these amplifiers are highly sensitive: losing fewer than 5% of the neurons in a specific functional group can significantly degrade the sensory encoding of the remaining survivors.

What Problem Does This Paper Address?

The cerebral cortex is a dense thicket of connections, the majority of which are local and excitatory. For decades, neuroscientists have debated the purpose of this "recurrent excitation." Does it exist to amplify weak sensory inputs, to allow the brain to "fill in the blanks" (pattern completion), or perhaps to maintain memories?

A major hurdle in answering this is that the cortex is messy. Neurons that respond to different things—like the touch of an object versus the movement of a whisker—are spatially intermingled. This makes it incredibly difficult to determine how a specific "subnetwork" of similarly tuned neurons influences its peers. The authors sought to determine if these subnetworks drive input-specific amplification and how robust these circuits are to the loss of individual members.

Methods

The study employed a "model-predict, experiment-verify" approach:

  1. Computational Modeling: The team built a network model of cortical Layer 2/3 (L2/3) featuring 1,700 excitatory and 300 inhibitory integrate-and-fire neurons. They varied the "subnetwork connectivity"—the probability that neurons with similar tuning are connected to each other.
  2. In Vivo Volumetric Imaging: Using two-photon microscopy and the calcium indicator GCaMP6s, they recorded the activity of thousands of neurons (~8,000 per mouse) in the vibrissal somatosensory cortex (vS1) while mice performed a tactile discrimination task.
  3. Targeted Photoablation: This is the study's "killer app." After identifying which neurons responded to "touch" and which to "whisking," the researchers used high-intensity multiphoton excitation to precisely kill a small, specific subset of neurons (e.g., only touch-sensitive ones) and then observed the functional impact on the remaining, spared neurons.

Datasets

The study relies on original experimental data collected from 16 mice trained on a pole-localization task.

  • Neural Data: Volumetric imaging across 18 planes in the barrel cortex, capturing nearly 10,000 neurons per animal.
  • Behavioral Data: High-speed video tracking of whisker movement and curvature (to define "touch" vs. "whisking" events).
  • Simulated Data: Results from 30 randomized network simulations per connectivity regime.

Key Insights and Ideas

  • The Amplification-Fragility Tradeoff: The model revealed three regimes. Low-connectivity networks are robust to damage but don't amplify signals. High-connectivity networks amplify signals beautifully but are "brittle"—losing just a few members causes the whole subnetwork's performance to tank.
  • Input-Specific Recurrence: The cortex isn't just one big amplifier; it's a collection of overlapping, specific amplifiers. Ablating "touch" neurons only degraded the "touch" representation, leaving "whisking" representations untouched.
  • The Similarity Rule: The impact of ablation wasn't random. Spared neurons that were most "similar" in their tuning to the killed neurons suffered the greatest loss in encoding quality. This provides strong evidence that these neurons were part of the same functional recurrent subnetwork.
  • Spatial Decay: The effect of losing a neighbor decayed over a distance of ~87 µm, which matches the known physical scale of local excitatory connections in the rodent cortex.

Key Results

  • Dramatic Impact from Tiny Losses: Ablating a tiny fraction of neurons (10–20 cells, or <5% of the total) representing touch caused a significant decline in the touch encoding scores ($R_{touch}$) of the remaining neurons (dropping from ~0.123 to ~0.100).
  • Specificity of Damage: While touch representation was degraded by touch-cell ablation, whisking representation remained stable ($P = 0.820$). Conversely, ablating whisking neurons or "silent" neurons had no significant effect on the touch representation.
  • Model Validation: The experimental results closely matched the "elevated subnetwork connectivity" (0.4) regime in their model, suggesting that real cortical circuits are tuned for high amplification at the cost of some redundancy.

Limitations and Future Work

  • The Whisking Paradox: Interestingly, ablating whisking neurons didn't seem to degrade the whisking representation. The authors hypothesize this is because whisking input is "asynchronous" (neurons fire at different phases of a whisk), whereas touch is a highly synchronized "pop" that is better suited to trigger recurrent amplification.
  • Homeostatic Recovery: The study focused on the immediate aftermath (the day after ablation). It remains unknown if the cortex can rewire itself over longer periods to recover the lost amplification.
  • Cell-Type Simplicity: The computational model used a generalized inhibitory population. Future work could incorporate the known diversity of inhibitory interneurons (like PV, SOM, and VIP cells) to see how they modulate this recurrent amplification.

This research suggests that our brains prioritize signal "loudness" and clarity through specific subnetwork amplification, even if it makes the system more vulnerable to the loss of individual neurons than previously assumed.

Summary updated

Abstract

The majority of cortical synapses are local and excitatory. Local recurrent circuits could implement amplification, allowing for pattern completion and other computations 1 . Cortical circuits contain subnetworks, consisting of neurons with similar receptive fields and elevated connectivity relative to the network average 2,3 . Understanding the computations performed by these subnetworks during behavior has been hampered by the fact that cortical neurons encoding different types of information are spatially intermingled and distributed over large brain volumes 4,5 . We used computational modeling, optical recordings and manipulations to probe the function of recurrent coupling in layer (L) 2/3 of the somatosensory cortex during tactile discrimination. A model of L2/3 dynamics revealed that recurrent excitation enhances sensory signals via amplification, but only for subnetwork with elevated connectivity. Networks with high amplification were sensitive to damage: loss of a few subnetwork members degraded stimulus encoding. We tested this prediction experimentally by mapping neuronal selectivity 5 and photoablating 6,7 neurons with specific selectivity. In L2/3 of the somatosensory cortex, ablating a small proportion (10-20, &lt; 5 % of the total) of neurons representing touch dramatically reduced responses in the spared touch representation, but not other representations. Network models further predicted that degradation following ablation should be greatest among spared neurons with stimulus responses that were most similar to the ablated population. Consistent with this prediction, ablations most strongly impacted neurons with selectivity similar to the ablated population. Our data shows that recurrence among cortical neurons with similar selectivity can drive input-specific amplification during behavior.
BibTeX
@inproceedings{peron2020recurrent,
  author = {Simon Peron and Ravi Pancholi and Bettina Voelcker and Jason D. Wittenbach and H. Freyja Ólafsdóttir and Jeremy Freeman and Karel Svoboda},
  title = {{Recurrent interactions in local cortical circuits}},
  booktitle = {Nature},
  year = {2020},
  doi = {10.1038/s41586-020-2062-x},
}
Figure