Spectrotemporal cues and attention jointly modulate fMRI network topology for sentence and melody perception

Speech and music are two fundamental modes of human communication. Lateralisation of key processes underlying their perception has been related both to the distinct sensitivity to low-level spectrotemporal acoustic features and to top-down attention. However, the interplay between bottom-up and top-down processes needs to be clarified. In the present study, we investigated the contribution of acoustics and attention to melodies or sentences to lateralisation in fMRI functional network topology. We used sung speech stimuli selectively filtered in temporal or spectral modulation domains with crossed and balanced verbal and melodic content. Perception of speech decreased with degradation of temporal information, whereas perception of melodies decreased with spectral degradation. Applying graph theoretical metrics on fMRI connectivity matrices, we found that local clustering, reflecting functional specialisation, linearly increased when spectral or temporal cues crucial for the task goal were incrementally degraded. These effects occurred in a bilateral fronto-temporo-parietal network for processing temporally degraded sentences and in right auditory regions for processing spectrally degraded melodies. In contrast, global topology remained stable across conditions. These findings suggest that lateralisation for speech and music partially depends on an interplay of acoustic cues and task goals under increased attentional demands.


Behavioural data
To test which cutoffs drove the interaction "attention:degradation", we fitted GLMMs for each cutoff separately and compared the resulting full (fixed effects "attention" by "degradation", random effect "participant", random slopes of fixed effects within "participant") and null models as post-hoc tests.

Global efficiency
Table S3: Estimates and confindence intervals for the GLMM for the global efficiency (area under the curve)

Fixed effect Estimate
Std.

Post-hoc tests
As for the global clustering coefficient we fitted separate models with the "attention:degradation" interaction as fixed effects (random effect "participant", random slopes of fixed effects within "participant") for each cutoff and ran fullnull model comparisons as post-hoc tests.

Local clustering coefficient
For the remaining cutoffs of the ROI of the fronto-parietal-temporal subnetwork, no other lower-order effects were significant.
We additionally conducted pairwise comparisons as post-hoc tests (FDR corrected) for each of the ROI at each cutoff to characterise whether they could be specifically associated with processing of sentences or melodies, spectral or temporal processing or their interaction.We did not observe a clear lateralisation according to either of our hypotheses.Instead, regions of the subnetwork were characterised by higher clustering coefficients for specific combinations of attention, degradation and cutoff (see Table S6).
Table S6: Characterisation of ROI of the fronto-parietal-temporal subnetwork (local clustering coefficient) observed.
Characterisation was done using post-hoc tests (R package emmeans, version 1.8.4-1) by comparing conditions pairwise (MS,ST,MT,SS).We interpreted conditions with significantly greater estimated marginal means (FDRcorrected) as driving the overall effect for the given ROI."Ambiguous" refers to ROI that cannot clearly be associated with either condition or factor.Bold: ROI still significant after FDR correction.

Local efficiency
To see which ROI were differentially related to processing of melodies, sentences, spectral or temporal acoustic information, we conducted post-hoc likelihood-ratio tests at each cutoff.This analysis revealed significant interactions of attention and degradation for 226 ROI, mostly in cutoff 5. 12 ROI showed a main effect of attention, eleven a main effect of degradation (see tables in OSF repository https://osf.io/merwk/).We again characterised ROI according to pairwise post-hoc comparisons.Most ROI had a significantly higher local efficiency for MT and SS (the non-crucial conditions) at cutoff 5 (low degradation).Overall, there was no clearly localisable pattern, but ROI that changed in efficiency for specific interactions of acoustical cues and top-down attention were widely distributed across the brain.

Modularity
To assess whether any ROI would change in their role as core hub within its module as a function of attention, degradation, and cutoff we ran GLMMs for each ROI with using z as a response variable and the fixed factors "attention", "degradation", and "cutoff".We included the three-way interaction of these factors, as well as "participant" as random factor and random slopes of all fixed factors within "participant" (i.e. the same model structure as for Clustering coefficient and Efficiency).The z value indicates classification consistency for each ROI, indicating how often ROIs get co-classified within the same module relative to the other ROI in this module.High consistency values can be interpreted as the node being a core hub of its module, that is as having high within-module centrality [1].
We obtained no significant full-null model comparison for any ROI (FDR corrected), indicating no significant change concerning within-module roles for any ROI.
Exploratorily we also tested whether the condition-and cutoff-specific clustering coefficient of a ROI would predict their within-module functional role as indicated by the z-value, recovering the subnetwork we obtained for the clustering coefficient.We ran GLMMs with the clustering coefficient value, condition and cutoff as fixed effect predictors and participant as random factor (including full random slopes structure).We obtained significant full-null model comparisons and 3-way interactions for 7 ROI (see Table S8).None of those ROI were part to the subnetwork we found for the local clustering coefficient.We aimed at identifying whether the participants' behavioural data could be associated with specific subnetworks as a function of condition and cutoff.To this end we used network-based statistics combined with a GLMM framework.
The network-based statistic (NBS, [2]) is a mass-univariate approach for identifying statistical effects distributed across subnetworks.That is, the NBS can detect effects that comprise several nodes in a graph (so called components).
Recently, the NBS has been combined with the GLMM framework in the R-package "NBS" [3].We used this package (version 0.1.5)on the subject-level functional 358x358 connectivity matrices to obtain subnetworks associated with participants' behavioural data.We changed the original script "nbr_lme_aov" of the package to accept our GLMM structure comprising "cutoff", "attention", "degradation", and "behaviour" as fixed-effects predictors and "participant" as random effect.We included random slopes of "cutoff", "attention", "degradation" within "participant".The NBS is calculated by running a model for each cell of the upper triangle of the connectivity matrix (i.e. each edge in the graph) and testing the resulting model statistics against a permutation distribution.We used 1000 permutations and calculated the full-null model comparison as well as likelhood ratio tests using reduced models.This would allow us to discriminate resulting NBS components by effects.In order to reduce the computational cost we calculated reduced models only if the full-null model comparison was significant.For the same reason we also only calculated lowerorder reduced models only if higher-order models were not significant, since lower-order effects can not be meaningfully interpreted when higher-order effects are present.Since the permutation test is based on the proportion of sampled permutations greater than the observed test statistic, for all models that were not calculated we set the test statistic (χ 2 ) to 0 and the p-value to 1, rendering them as "non-significant".
We obtained significant NBS components for all effects (higher-order and lower-order), however, these almost always comprised all ROI in the network.We could therefore not interpret the results in a meaningful way regarding our hypotheses (see OSF repository https://osf.io/merwk/for detailed results and plots).

Figure S1 :
Figure S1: Global clustering coefficient (area under the curve) as function of condition across degradation intensities (cutoff).Boxplot whiskers indicate first quartile -1.5* interquartile range and third quartile + 1.5 interquartile range, respectively.

Figure S2 :
Figure S2: Global efficiency (area under the curve) as function of condition across degradation intensities (cutoff).

Figure S3 :
Figure S3: ROI for which their clustering coefficient significantly predicted their within-module centrality Z for specific conditions and cutoffs.

Table S1 :
Estimates and confindence intervals for the linear model of the behavioural data

Table S2 :
Estimates and confindence intervals for the GLMM for the global clustering coefficient (area under the curve)

Table S4 :
Results for the significant full-null model comparisons for the local clustering coefficient.

Table S5 :
Signifikant likelihood-ratio tests of the interaction of attention and degradation for ROI of the frontoparietal-temporal subnetwork (local clustering coefficient), by cutoff*

Table S7 :
Module affiliations across subjects and across cutoffs (consensus partition) in left and right for each condition.For the affiliations by cutoff for each condition see OSF repository https://osf.io/merwk/

Table S8 :
ROI for which their clustering coefficient significantly predicted their within-module centrality Z for specific conditions and cutoffs.

Table S9 :
Likelihood-ratio tests for the 3-way interaction (z value ~ clustering coefficient * condition * cutoff) for ROI with significant full-null model comparisons.