Coevolutionary balance of resting-state brain networks in autism

TL;DR

It is suggested that coevolutionary balance offers a compact, interpretable descriptor of altered resting-state network dynamics in autism, which showed a reorganization rather than a uniform loss of balance in intrinsic network organization.

Abstract

Autism spectrum disorder (ASD) involves atypical brain organization, yet the large-scale functional principles underlying these alterations remain incompletely understood. Here we examine whether coevolutionary balance-a network-level energy measure derived from signed interactions and nodal activity states-captures disruptions in resting-state functional connectivity in autistic adults. Using ABIDE I resting-state fMRI data, we constructed whole-brain networks by combining binarized fALFF activity with signed functional correlations and quantified their coevolutionary energy. Compared with matched typically developing adults, the ASD group showed a characteristic redistribution of coevolutionary energy, with more negative global energy but higher (less negative) energy within the default mode network and altered energy in its interactions with dorsal attention and salience networks, indicating a reorganization rather than a uniform loss of balance in intrinsic network organization. These effects replicated across validation analyses with null models designed to disrupt link or node structure. Coevolutionary energy also showed modest but significant associations with ADI-R social and communication scores. Finally, incorporating coevolutionary features into a leakage-safe machine-learning classifier supported above-chance ASD versus typically developing (TD) discrimination on a held-out test set. These findings suggest that coevolutionary balance offers a compact, interpretable descriptor of altered resting-state network dynamics in autism.

Figures (5)

• Figure 1: Overview of the analytical workflow used in this study. Resting state fMRI datasets from the ABIDE I repository were first preprocessed using the CPAC pipeline, including standard steps such as motion correction, spatial normalization, and temporal filtering. The preprocessed volumes were then parcellated into 200 regions of interest using the CC200 atlas, yielding a regional time series for each participant. For every region, the fractional amplitude of low frequency fluctuations (fALFF) was computed, and these values were binarized to define nodal activity states distinguishing relatively high from relatively low intrinsic activity. Pairwise functional connectivity between regions was estimated from the regional BOLD time series and converted into a signed adjacency matrix that captures the sign of the interaction between each pair of regions. Nodal states and signed links were combined within the coevolutionary balance framework to compute coevolutionary energy at different scales, including whole brain energy, energy within canonical resting state networks, and energy associated with inter network interactions. To assess whether the observed energies differed from those expected under structured noise, null networks were generated by randomizing nodal states or link signs while preserving the underlying graph topology, and energy distributions from these null models were compared with empirical values. Group comparisons between autistic and typically developing adults were then carried out on the resulting energy measures, and energy derived features were subsequently used as input to machine learning classifiers to evaluate their ability to distinguish ASD from typical development.
• Figure 2: Whole-brain coevolutionary metrics for autistic (ASD) and typically developing (Control) adults. Each panel shows the distribution of a global measure derived from the signed functional network: (top row, from left to right) total coevolutionary energy, the percentage of links in agreement between nodal activity states and link sign (Agreement), and the percentage of links in disagreement (Disagreement); (bottom row, from left to right) the proportion of imbalanced links connecting nodes with the same activity state (Imbalanced Same), the proportion of imbalanced links connecting nodes with opposite activity states (Imbalanced Opp), and a summary index of overall network polarity (Bipolarity). Boxplots display the median (central line), interquartile range (box), and 1.5 $\times$ IQR whiskers, with individual participants overlaid as jittered points (magenta for ASD, green for controls). Across several metrics, ASD networks tend to show more negative coevolutionary energy (more energetically favorable configurations), higher agreement and fewer imbalanced-same links, together with altered patterns of imbalance relative to controls, indicating a systematic shift in the alignment between regional activity states and signed functional connectivity at the whole-brain level.
• Figure 3: Cortical surface representations showing differences in network-level ASD and TD, with thresholds set at uncorrected and FDR-corrected levels. Each panel has a gray cortical mesh that is just partially see-through, with colored Yeo 7-network pieces on top of it that indicate big differences between groups in mean connectivity (t-statistics). The networks in the top row have uncorrected $p<0.05$ (Dorsal Attention, Salience/Ventral Attention, and Default Mode). Bottom row: networks that stayed after FDR correction ($p_{FDR} at \alpha = 0.05$; Default Mode alone). The columns show the axial, coronal, and sagittal views from left to right.
• Figure 4: Confusion matrix and receiver operating characteristic (ROC) curve for the best-performing classifier on the held-out test set. The panel on the left shows the confusion matrix for the logistic regression model trained on the selected coevolutionary and inter-network features, summarizing true and predicted diagnostic labels (ASD versus TD). The panel on the right depicts the corresponding ROC curve based on the predicted probability of the ASD class, with the diagonal grey line indicating chance performance. The area under the ROC curve (AUC) was approximately 0.70, consistent with above-chance but non-clinical discrimination.
• Figure 5: Pairwise settings in the notion of coevolutionary equilibrium. The blue and orange circles show the states of the nodes ($+1$ and $-1$) (high vs. low fALFF), while the green and red lines show the signs of the links ($+1$ and $-1$) (positive vs. negative FC). Disagreement and agreement lower energy, but unbalanced pairs raise it.