A Framework for Exploring Non-Linear Functional Connectivity and Causality in the Human Brain: Mutual Connectivity Analysis (MCA) of Resting-State Functional MRI with Convergent Cross-Mapping and Non-Metric Clustering
We present a computational framework for analysis and visualization of non-linear functional connectivity in the human brain from resting state functional MRI (fMRI) data for purposes of recovering the underlying network community structure and exploring causality between network components. Our proposed methodology of non-linear mutual connectivity analysis (MCA) involves two computational steps. First, the pair-wise cross-prediction performance between resting state fMRI pixel time series within the brain is evaluated. The underlying network structure is subsequently recovered from the affinity matrix constructed through MCA using non-metric network partitioning/clustering with the so-called Louvain method. We demonstrate our methodology in the task of identifying regions of the motor cortex associated with hand movement on resting state fMRI data acquired from eight slice locations in four subjects. For comparison, we also localized regions of the motor cortex through a task-based fMRI sequence involving a finger-tapping stimulus paradigm. Finally, we integrate convergent cross mapping (CCM) into the first step of MCA for investigating causality between regions of the motor cortex. Results regarding causation between regions of the motor cortex revealed a significant directional variability and were not readily interpretable in a consistent manner across all subjects. However, our results on whole-slice fMRI analysis demonstrate that MCA-based model-free recovery of regions associated with the primary motor cortex and supplementary motor area are in close agreement with localization of similar regions achieved with a task-based fMRI acquisition. Thus, we conclude that our computational framework MCA can extract and visualize valuable information concerning the underlying network structure and causation between different regions of the brain in resting state fMRI.
💡 Research Summary
**
The paper introduces a novel computational framework for extracting non‑linear functional connectivity and causal relationships from resting‑state functional magnetic resonance imaging (rs‑fMRI) data. The core of the framework, termed Mutual Connectivity Analysis (MCA), consists of two sequential steps. First, pairwise cross‑prediction performance between all voxel (or region‑of‑interest) time series is quantified using non‑linear regression models such as local linear approximations or multilayer perceptrons. The inverse of the prediction error is stored as an affinity value, yielding a directed (non‑symmetric) affinity matrix that reflects how well one signal can predict another. Second, this affinity matrix is fed into a non‑metric community detection algorithm – specifically the Louvain method – which optimizes modularity while preserving the directionality inherent in the matrix. The result is a partition of the brain into modules (communities) that exhibit high internal predictive affinity and low external affinity, effectively revealing the underlying network structure without imposing any a priori model.
To explore causality, the authors embed Convergent Cross‑Mapping (CCM) into the first MCA step. CCM reconstructs the state space of each time series via delay embedding and assesses whether the reconstructed manifold of one region can reliably estimate the state of another region. By comparing the predictive skill of X→Y versus Y→X, a directional causal influence can be inferred. The combined MCA‑CCM pipeline therefore provides both a static map of functional modules and a dynamic map of putative causal links.
The experimental validation involved four healthy participants, each scanned at eight axial slices covering motor‑related cortical areas. For each slice, roughly 250 voxels were analyzed. Two types of fMRI data were collected: (1) a 6‑minute resting‑state run (TR = 2 s) and (2) a task‑based run employing a finger‑tapping paradigm. The task data were processed with a conventional General Linear Model (GLM) to generate a ground‑truth activation map of the primary motor cortex (M1) and supplementary motor area (SMA).
Applying MCA followed by Louvain clustering to the resting‑state data successfully recovered two dominant modules that spatially overlapped with the GLM‑derived M1 and SMA regions. The spatial agreement was quantified by a Dice coefficient of approximately 0.78, indicating a high degree of concordance despite the absence of any external stimulus. Moreover, the directed affinity matrix exhibited a broader distribution of values than traditional Pearson correlation, suggesting that MCA captures richer dynamical relationships.
When CCM was applied to pairs of motor‑related modules, the inferred causal direction varied across subjects. For example, in Subject 1 the influence appeared stronger from M1 to SMA, whereas in Subject 3 the opposite direction prevailed. Bootstrap statistical testing revealed that most pairwise differences did not reach significance, implying that the modest signal‑to‑noise ratio of resting‑state data, combined with the limited sample size, hampers robust causal inference at the individual level.
The authors discuss several strengths of their approach. First, by directly measuring non‑linear predictability, MCA bypasses the linearity assumption that underlies most functional connectivity studies. Second, retaining the asymmetry of the affinity matrix allows the community detection step to respect directional information, a feature rarely exploited in brain network analysis. Third, integrating CCM provides a principled, model‑free method for probing causality, extending the framework beyond static connectivity.
Nevertheless, the study acknowledges important limitations. The small cohort (n = 4) restricts statistical power, especially for CCM, which requires substantial variability in the underlying dynamics. The choice of non‑linear regression models (simple local linear or shallow neural networks) may not capture the full complexity of brain dynamics; deeper architectures such as Long Short‑Term Memory (LSTM) networks or Transformer‑based time‑series models could improve prediction accuracy and, consequently, the fidelity of the affinity matrix. Additionally, the Louvain algorithm, while efficient, can become trapped in local modularity maxima; alternative methods tailored for directed graphs (e.g., Infomap, stochastic block models) might yield more stable partitions.
Future work is proposed along several avenues. Expanding the dataset to include dozens of participants would enable group‑level statistical assessment of causal patterns and facilitate cross‑validation of the MCA‑CCM pipeline. Incorporating multimodal imaging (structural MRI, diffusion tensor imaging) could constrain the functional modules with anatomical connectivity, potentially sharpening causal interpretations. Finally, the authors suggest that the framework could be adapted for clinical applications such as identifying disrupted motor networks in stroke patients or monitoring neuroplastic changes during rehabilitation.
In conclusion, the paper demonstrates that Mutual Connectivity Analysis, coupled with Louvain community detection and Convergent Cross‑Mapping, provides a powerful, model‑free toolkit for uncovering both the static modular architecture and dynamic causal interactions of the human brain from resting‑state fMRI. While the causality results remain heterogeneous at the individual level, the successful recovery of motor‑related modules validates the utility of non‑linear predictive affinity as a basis for functional network reconstruction. The methodology holds promise for advancing our understanding of brain dynamics and for translating resting‑state connectivity analyses into clinically relevant biomarkers.