Uncovering Latent Communication Patterns in Brain Networks via Adaptive Flow Routing

Unraveling how macroscopic cognitive phenotypes emerge from microscopic neuronal connectivity remains one of the core pursuits of neuroscience. To this end, researchers typically leverage multi-modal information from structural connectivity (SC) and functional connectivity (FC) to complete downstream tasks. Recent methodologies explore the intricate coupling mechanisms between SC and FC, attempting to fuse their representations at the regional level. However, lacking fundamental neuroscientific insight, these approaches fail to uncover the latent interactions between neural regions underlying these connectomes, and thus cannot explain why SC and FC exhibit dynamic states of both coupling and heterogeneity. In this paper, we formulate multi-modal fusion through the lens of neural communication dynamics and propose the Adaptive Flow Routing Network (AFR-Net), a physics-informed framework that models how structural constraints (SC) give rise to functional communication patterns (FC), enabling interpretable discovery of critical neural pathways. Extensive experiments demonstrate that AFR-Net significantly outperforms state-of-the-art baselines. The code is available at https://anonymous.4open.science/r/DIAL-F0D1.

💡 Research Summary

The paper tackles a central challenge in computational neuroscience: how to explain the emergence of functional connectivity (FC) from the underlying structural connectivity (SC) in a way that is both predictive and biologically interpretable. While many recent studies have combined SC and FC using graph neural networks (GNNs) with simple fusion strategies (concatenation, weighted sum, self‑attention), they largely treat the brain as a static graph and focus on local homophily. Consequently, they cannot account for the observed regional heterogeneity—areas where SC and FC are tightly coupled versus regions where the relationship is weak.

AFR‑Net (Adaptive Flow Routing Network) reframes multimodal brain‑network fusion as a global information‑flow problem grounded in physics. The authors draw an analogy between the structural connectome and an electrical circuit: white‑matter tracts are “resistors” with learnable conductances, and functional correlations represent “demands” that drive current through the network. The framework consists of three tightly coupled modules.

1. Physics‑Informed Graph Construction – Instead of using shortest‑path distances, the model computes effective‑resistance distance (ERD) between every pair of brain regions, capturing how easily information can travel through the whole network. Edge capacities are parameterized by learnable scalar gates c₍ᵢⱼ₎, turning the static SC adjacency into a dynamic flow network whose conductances can be optimized during training.

2. Differentiable Information Flow Solver – The structural Laplacian with learned capacities, L_flow = Bᵀ C B (where B is the incidence matrix and C = diag(c₍ᵢⱼ₎)), is combined with the functional Laplacian L_fc derived from FC. By solving the linear system L_flow V = L_fc · 1 (or an equivalent equilibrium condition), the model obtains node potentials V and edge‑wise flows f = C B V in closed form. The flow magnitude f₍ᵢⱼ₎² · c₍ᵢⱼ₎ corresponds to the “information load” on each structural edge, directly mirroring power dissipation in a resistor network. Because the solution is analytic, gradients can be back‑propagated, allowing the edge capacities to be tuned so that the resulting flow pattern best explains the observed FC.

3. Pattern‑Guided Aggregation – The computed flow intensities are fed into a message‑passing layer as attention weights. Nodes exchange information preferentially along high‑load edges, effectively biasing the GNN toward biologically plausible routes rather than treating all neighboring edges equally. This step bridges the physics‑driven global routing with downstream classification or regression tasks.

The authors evaluate AFR‑Net on four large‑scale datasets, including the Adolescent Brain Cognitive Development (ABCD) study, the Parkinson’s Progression Markers Initiative (PPMI), and two Alzheimer’s disease cohorts. Tasks comprise binary disease diagnosis, multi‑class disease staging, and regression of age or cognitive scores. Baselines span standard GCN/GAT, multimodal fusion models (Cross‑GNN, RH‑BrainFS), and recent flow‑oriented approaches (NeuroPath, EDT‑PA). AFR‑Net consistently outperforms all baselines, achieving 3–7 percentage‑point gains in accuracy and notable improvements in AUC and F1‑score.

Beyond performance, the model yields interpretable “routing patterns.” Visualizing the learned edge capacities and flow magnitudes highlights subnetworks that align with known disease‑related circuits: increased flow through fronto‑hippocampal pathways in Alzheimer’s, reduced basal‑ganglia‑motor cortex routing in Parkinson’s, and developmental strengthening of frontoparietal connections in adolescents. These findings suggest that AFR‑Net does not merely fit the data but discovers neurobiologically meaningful communication routes.

The paper also discusses limitations. First, the learned capacities are abstract parameters; mapping them to concrete anatomical properties (e.g., axonal diameter, myelination) would enhance biological validity. Second, the closed‑form Laplacian solution scales quadratically with the number of edges, posing memory and compute challenges for very high‑resolution parcellations; sparse approximations or distributed solvers are needed. Third, FC is treated as a static correlation matrix, ignoring temporal dynamics that could be incorporated via sliding‑window or dynamic functional connectivity.

Future work is outlined: integrating diffusion‑MRI‑derived conductance measures as priors, extending the resistor model to nonlinear, activity‑dependent conductances, and coupling the framework with time‑varying FC to capture state‑dependent routing.

In summary, AFR‑Net introduces a novel physics‑informed paradigm for multimodal brain‑network analysis. By converting SC‑FC coupling into a differentiable global flow optimization problem, it simultaneously achieves state‑of‑the‑art predictive performance and provides a mechanistic, interpretable map of neural communication pathways—offering a promising tool for both basic neuroscience discovery and clinical biomarker identification.