Neural correlates of semantic similarity and associations with GABA and glutamate in the Left Inferior Frontal Gyrus

Objectives

The neural mechanisms underlying spontaneous semantic activation during the processing of related and unrelated words remain unclear. This study examined whether trial-by-trial variations in relational semantic similarity between prime–target pairs are reflected in neural activation patterns. A second aim was to investigate associations between GABA and Glx concentrations and the organization of neural semantic representations, reflecting potential excitatory–inhibitory (E/I) balance effects. To our knowledge, this is the first study to test whether relational semantic similarity is represented in neural representational space and how it relates to neurochemical markers.

Methods

Forty-six healthy participants (26 female; mean age=22.4±3.5) were included. Data were acquired on a 3T MRI scanner with a 64-channel head coil. During fMRI, participants performed a semantic categorization task on targets preceded by primes. On a separate day, MRS (MEGA-PRESS) was used to measure GABA and Glx levels in the left inferior frontal gyrus (IFG), inferior parietal lobule (IPL), and anterior temporal lobe (ATL). Behavioral data were analyzed using linear mixed-effects models. Trial-wise semantic similarity was quantified via cosine similarity from word embeddings. Representational similarity analysis (RSA) tested whether these variations were reflected in neural pattern dissimilarities. Trial-wise beta estimates were obtained using least-squares separate (LS-S) models, and neural representational dissimilarity matrices (nRDMs) were computed within predefined ROIs. Semantic model RDMs were constructed from absolute differences in relational similarity. Statistical inference used sign-flip permutation tests (10,000 iterations) with family-wise error rate (FWER) correction across ROIs; false discovery rate (FDR) results are reported when FWER significance was not reached. Regression analyses tested linear and quadratic relationships between RSA model fits and neurochemical measures.

Results

Error rates were higher in the incongruent condition (β=1.407, p<0.001), while reaction times did not differ between conditions. However, higher semantic similarity predicted faster responses (β=−0.105, p=0.014). Univariate fMRI analyses showed no significant congruency effects after correction. RSA revealed that relational semantic structure was represented across conditions in the left IFG (pars opercularis) and medial superior frontal gyrus (SFG). Greater neural–semantic correspondence in the incongruent condition was observed at the FDR-corrected level in SFG, temporal regions (MTG, STG, temporal pole), bilateral IFG, middle frontal gyrus, and parahippocampal gyrus. MRS results showed that higher GABA levels in the IFG were associated with reduced neural–semantic correspondence (R²=0.087, β=−0.147, p=0.048), while Glx exhibited a quadratic (inverted-U) relationship, with maximal representational strength at intermediate levels. A similar quadratic pattern was observed for the incongruent > congruent contrast. No significant associations were found in the ATL or IPL.

Discussion

These findings indicate that regions within the semantic control network encode relational structure between word pairs, supporting context-dependent semantic representations. Incongruent trials likely increase control demands, leading to more structured representations. The involvement of IFG and medial frontal regions aligns with their roles in controlled semantic retrieval. Neurochemical results suggest that E/I balance modulates representational structure. Higher GABA may reduce differentiation, while optimal Glx levels enhance representational strength. These results highlight the central organizational role of the IFG in shaping semantic representations under varying contextual demands [1,4].

References

1. Binder, J. R., Desai, R. H., Graves, W. W., & Conant, L. L. (2009). Where is the semantic system? A critical review and meta-analysis of 120 functional neuroimaging studies. Cerebral Cortex (New York, N.Y.: 1991), 19(12), 2767–2796. https://doi.org/10.1093/cercor/bhp055
2. Carota, F., Kriegeskorte, N., Nili, H., & Pulvermüller, F. (2017). Representational Similarity Mapping of Distributional Semantics in Left Inferior Frontal, Middle Temporal, and Motor Cortex. Cerebral Cortex, 27(1), 294–309. https://doi.org/10.1093/cercor/bhw379
3. Fernandino, L., Binder, J. R., Desai, R. H., Pendl, S. L., Humphries, C. J., Gross, W. L., Conant, L. L., & Seidenberg, M. S. (2016). Concept Representation Reflects Multimodal Abstraction: A Framework for Embodied Semantics. Cerebral Cortex (New York, NY), 26(5), 2018–2034. https://doi.org/10.1093/cercor/bhv020
4. Liuzzi, A. G., Aglinskas, A., & Fairhall, S. L. (2020). General and feature-based semantic representations in the semantic network. Scientific Reports, 10(1), 8931. https://doi.org/10.1038/s41598-020-65906-0