Sequential Transmission of Task-Relevant Information in Cortical Neuronal Networks

Nikolas A. Francis, Shoutik Mukherjee, Loren Koçillari, Stefano Panzeri, Behtash Babadi, Patrick O. KanoldbioRxiv 2021 · DOI

TL;DR

This study investigates how the mouse auditory cortex processes and transmits information during a sound discrimination task. Researchers found that specific neuronal networks encode task-relevant information through sequential communication, with the structure of these networks dynamically changing based on sensory input and behavioral choice. This work sheds light on the neural mechanisms underlying decision-making in the brain.

What Problem Does This Paper Address?

Understanding how the brain makes decisions based on sensory input is a fundamental challenge in neuroscience. While it's known that neuronal networks are involved, the precise mechanisms by which task-relevant information is encoded, transmitted, and used to guide behavior remain unclear, especially in complex discrimination tasks. This paper specifically addresses how the auditory cortex processes information to distinguish between similar sounds (targets vs. non-targets) and how the communication patterns within neuronal networks reflect this decision-making process.

Methods

The researchers employed a combination of advanced techniques:

  • In vivo 2-photon calcium imaging: This allowed them to record the activity of large populations of neurons in the primary auditory cortex (A1) of awake, behaving mice with high spatial and temporal resolution.
  • Auditory discrimination task: Mice were trained to differentiate between low-frequency target tones and high-frequency non-target tones, requiring them to make a behavioral choice (lick or no-lick).
  • Information theory: Techniques like Shannon's mutual information were used to quantify how much information individual neurons carried about the stimulus (SI), behavioral choice (CI), and the intersection of both (II - sensory information used for choice).
  • Granger Causality (GC) analysis: This statistical method was used to infer directed functional connections between neurons and analyze the structure of neuronal networks. They extended this by examining interactions across different timescales (short and long).
  • Support Vector Machine (SVM) classification: Used to determine if network structure features could predict behavioral choices.

Datasets

The study utilized data collected from 9 transgenic CBA x Thy1-GCaMP6s F1 mice. No new datasets were introduced; the data was generated through the experiments described.

Key Insights and Ideas

  • Specialized Neurons for Task-Relevant Information: A subset of neurons in A1 was identified as encoding "intersection information" (II), meaning they carry sensory information that is directly used for making behavioral choices. This highlights a specialized role for these neurons in decision-making.
  • Sequential Information Transmission: Task-relevant information isn't just present; it's actively transmitted sequentially between neurons within functional networks. This suggests a dynamic flow of information rather than a static representation.
  • Network Structure Encodes Behavioral Choice: The way neurons are connected and communicate (network structure) is not fixed but changes based on the sensory input and the animal's behavioral choice. This implies that the brain dynamically reconfigures its internal wiring to adapt to task demands.
  • Timescale Matters: Interactions between neurons occur on different timescales (short and long). The study found that short-timescale interactions are crucial for transmitting behaviorally relevant sensory information, while longer timescales might be involved in more complex or top-down processing.
  • Redundancy for Robustness: Neurons share redundant information, particularly for behaviorally relevant signals. This redundancy, rather than being inefficient, may contribute to a more robust and persistent representation of information, aiding in behavioral readout.

Key Results

  • Information Encoding: 42.37% of recorded neurons carried significant stimulus (SI) or choice (CI) information. A subset (40.15%) integrated both sensory and behavioral information (II).
  • Transient Information: Individual neurons encoded SI, CI, and II transiently, with an effective duration of approximately 250 ms.
  • Network Structure Predicts Choice: SVM classifiers trained on GC network statistics could predict behavioral choices significantly better than chance. S-timescale networks of low II-peak latency neurons were particularly good at this.
  • Correct vs. Incorrect Decisions: Networks associated with correct decisions (H/C trials) showed more robust link structures and were timescale-invariant, suggesting a constrained and efficient processing. Incorrect decisions (M/F trials) involved larger, more dispersed networks, potentially reflecting broader or less efficient processing.
  • Redundancy: GC-linked neurons exhibited higher information redundancy, especially for II, compared to GC-unlinked neurons. This redundancy was more pronounced for behaviorally relevant information.

Limitations and Future Work

  • Limited Sample Size for GC Analysis: Due to computational constraints and the need for statistical confidence, GC analysis was limited to a maximum of 20 neurons per experiment. This means the inferred networks represent subnetworks rather than the entire cortical area.
  • Inferred Causality: Granger causality infers predictive relationships, not direct causal links. While powerful, it's an inference and not a direct manipulation.
  • Future Directions: The authors suggest that further investigation into how the spatiotemporal structure of functional connectivity in A1 L2/3 forms a basis for sustained representation of task-relevant information is warranted. They also highlight the potential role of inhibitory circuits in mediating information exchange and the importance of understanding how different timescales of interaction contribute to behavior.

Summary updated

Abstract

ABSTRACT During auditory task performance, cortical processing of task-relevant information enables mammals to recognize sensory input and flexibly select behavioral responses. In mouse auditory cortex, small neuronal networks encode behavioral choice during a pure-tone detection task, but it is poorly understood how neuronal networks encode behavioral choice during a pure-tone discrimination task where tones have to be categorized into targets and non-targets. While the interactions between networked neurons are thought to encode behavioral choice, it remains unclear how patterns of neuronal network activity indicate the transmission of task-relevant information within the network. To this end, we trained mice to behaviorally discriminate target vs. non-target pure-tones while we used in vivo 2-photon imaging to record neuronal population activity in primary auditory cortex layer 2/3. We found that during task performance, a specialized subset of neurons transiently encoded intersection information, i.e., sensory information that was used to inform behavioral choice. Granger causality analysis showed that these neurons formed functional networks in which task-relevant information was transmitted sequentially between neurons. Differences in network structure between target and non-target sounds encoded behavioral choice. Correct behavioral choices were associated with shorter timescale communication between neurons. In summary, we find that specialized neuronal populations in auditory cortex form functional networks during auditory task performance whose structures depend on both sensory input and behavioral choice.
BibTeX
@misc{francis2021sequential,
  author = {Nikolas A. Francis and Shoutik Mukherjee and Loren Koçillari and Stefano Panzeri and Behtash Babadi and Patrick O. Kanold},
  title = {{Sequential Transmission of Task-Relevant Information in Cortical Neuronal Networks}},
  year = {2021},
  doi = {10.1101/2021.08.31.458395},
  howpublished = {bioRxiv},
}
Figure