Organizational principles of semantic control in the human brain

1076 

Abstract Submission 

KAIXIANG ZHUANG¹, Xinyu Liang¹, Cheng Liu², Theodoros Karapanagiotidis³, Jonathan Smallwood⁴, Elizabeth Jefferies⁵, Deniz Vatansever¹

¹Fudan University, Shanghai, Shanghai, ²Southwest University, Chongqing, Chongqing, ³University of Sussex, Sussex, Sussex, ⁴Queen's University, Ontario, Canada, ⁵University of York, York, United Kingdom

KAIXIANG ZHUANG  
Fudan University
Shanghai, Shanghai

Xinyu Liang  
Fudan University
Shanghai, Shanghai

Cheng Liu  
Southwest University
Chongqing, Chongqing

Theodoros Karapanagiotidis  
University of Sussex
Sussex, Sussex

Jonathan Smallwood  
Queen's University
Ontario, Canada

Elizabeth Jefferies  
University of York
York, United Kingdom

Deniz Vatansever  
Fudan University
Shanghai, Shanghai

The human semantic system affords a multi-dimensional conceptual space through which we ascribe meaning to various words and objects around us. Notably, accessing concepts that are more remotely connected in this space is suggested to require higher levels of demand for semantic control [1]. However, the precise neural signature of semantic control, and its distributed organization within the cortical hierarchy remains unclear. By combining an fMRI-based semantic retrieval task [2], a natural language processing model [3] and multivoxel pattern analysis (MVPA) [4], here we captured a neural signature associated with varying demands for semantic control and charted its distribution within the cortical connectivity gradients [5]. We demonstrate that semantic control requires the engagement of multiple brain networks, dispersed along two principal gradients relevant to different aspects of semantic processing. This offers new insights into how the brain's functional networks are architecturally specialized to support semantic cognition.

A group of 46 healthy young adults (mean = 21.31 years old, 21/25 Word List A to B ratio) completed a 3-alternative forced choice semantic retrieval fMRI task, in which they were probed with a cue word and asked to select the most conceptually associated target word amongst two other distractors across 80 trials. Crucially in each trial, semantic distance between word pairs, defined by the cosine distance within 300-dimensional GloVe vectors, were used to systematically manipulate the demand for semantic control (Fig. 1a-b). Next, we employed thresholded partial least squares T-PLS [6] and GLMsingle derived single-trial beta maps [7] to identify a whole-brain multivariate signature of semantic control (Fig. 1c), which served as a predictive model for semantic distance during semantic retrieval. After a series of rigorous assessments, the feature weights of semantic control signature were projected onto a continuous 2D space of functional connectivity gradients in order to characterize its organizational principles within the large-scale cortical hierarchy.

The identified neural signature of semantic control spanned areas sensitive to both high and low-demand semantic decisions (e.g. left IFG, pMTG, mPFC, PCC), as well as bilateral anterior insula and primary visual cortices (Fig. 1c). In addition to demonstrating high accuracy and generalizability (Fig. 1d-e), the semantic control signature distinguished subtle differences between word pairs (Fig. 1f), and accurately captured the response speed of participants during semantic retrieval (Fig. 1g). Furthermore, the identified semantic control signature exhibited spatial correlations with two out of 10 principal gradients (Fig. 2a). Specifically, we observed clear boundaries in the 2D space constituted by Gradients 6 and 10, which demarcated positive and negative feature weights of the neural signature (Fig. 2b). Based on the NeuroSynth meta-analytic decoding, we further revealed that both gradients were closely related to semantic cognition (Fig. 2c). While Gradient 6 arranged brain regions in a cognitive continuum from "low to high demand semantic cognition", Gradient 10 arranged brain regions from "verbal to visual semantic cognition".

Based on a semantic retrieval task and machine learning approach, we revealed a robust and generalizable neural signature sensitive to varying levels of semantic control demands. Notably, the identified signature was constrained by two functional connectivity gradients, both of which related to different aspects of semantic cognition. Together, our findings demonstrate how disparate regions within the cortical organization unite within a lower dimensional space to facilitate the control of semantic cognition.

Executive Function, Cognitive Control and Decision Making ²

Language Comprehension and Semantics

Long-Term Memory (Episodic and Semantic) ¹

BOLD fMRI

Cognition

Cortex

Data analysis

Experimental Design

FUNCTIONAL MRI

Memory