The impact of functional connectivity on task information coding

1775 

Abstract Submission 

Takuya Ito¹, John D. Murray²

¹IBM Research, Yorktown Heights, NY, ²Dartmouth College, Hanover, NH

Takuya Ito, PhD  
IBM Research
Yorktown Heights, NY

John D. Murray  
Dartmouth College
Hanover, NH

The brain is a complex system with dynamic network changes. Functional connectivity – the measurement of correlated brain activity – is a commonly-used approach to characterizing brain network changes. Despite the wealth of neuroscience studies that have reported reliable state-dependent changes to macroscale brain network organization (Cole et al., 2014), there is no widely accepted theory or understanding of what these functional connectivity changes are for. However, theoretical work in systems neuroscience (at the level of local spiking neurons) has demonstrated that state-dependent neural correlations can be understood from a neural coding framework (Panzeri et al., 2022). Prior theory posits that noise correlations (NC) -- idiosyncratic with functional connectivity -- can be interpreted only if the underlying signal correlation (SC) -- similarity of task tuning (or task co-activations) between pairs of neural units -- is known. Here we investigate whether the theoretical framework used to study neural coding in neuronal spikes can account for macroscale brain network changes (Ito and Murray, 2022).

We analyzed a publicly available fMRI data set – the Multi-Domain Task Battery – that contained a diverse set of 26 cognitive tasks collected per participant (n=24) (King et al., 2019). Data were collected across four imaging sessions, enabling within-subject test-retest reliability of analyses across task conditions. SCs were estimated by measuring the correlation of mean task activations across tasks (e.g., cross-task co-activations; Fig. 1a-c). NCs, which are statistically analogous to background functional connectivity (Al-Aidroos et al., 2012) or functional connectivity estimated with a beta series regression (Rissman et al., 2004), were measured using the block-to-block covariability (Fig. 1d-g). We measured the signal-noise differential matrix – the element-wise multiplication of task-state NC changes (NC) (Fig. 2e) with the SC matrix – to determine which sets of brain regions were likely to have a positive impact on information decoding based on prior theory (Fig. 2f). To test the hypothesis that negative signal-noise differentials enhanced coding of task information (Fig. 2a-c), we randomly subsampled networks of brain regions with exclusively negative signal-noise differentials. We assessed task coding ability by decoding multitask activations from brain regions with negative (or positive) signal-noise differentials (Fig. 2h,i).

We characterized the SC and NC organization of macroscale brain networks. We found that task-state NC changes did not typically change in the same direction as their underlying SC (e.g., task coactivation structure). In other words, task-state NC changes typically decreased between similarly tuned regions. Crucially, NCs that changed in the opposite direction as their SC (i.e., anti-aligned correlations) improved information coding of these brain regions (Fig. 2h,i). In contrast, NCs that changed in the same direction (aligned NCs) as their SC did not. Interestingly, these aligned NCs were primarily correlation increases, suggesting that most functional correlation increases across fMRI networks actually degrade information coding.

We found that regions with opposite SCs and NCs (i.e., negative signal-noise differentials) could better decode multitask activations. This provides empirical evidence that theories developed to study the neural code in local spiking circuits can be extended to understand the structure of correlated brain activity at the level of whole-brain functional networks. We conclude that state-dependent NCs shape information coding of macroscale functional brain networks, with interpretation of correlation changes requiring knowledge of underlying SCs.

Higher Cognitive Functions Other

Activation (eg. BOLD task-fMRI)

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling ¹

Cognition

Computational Neuroscience

FUNCTIONAL MRI

Machine Learning

Multivariate