Bridging local and global dynamics: a biologically grounded model for cooperative and competitive interactions in the brain

A biologically grounded explanation for why brain regions sometimes sync up and sometimes suppress each other — using only excitatory wiring.

The brain's fMRI signal shows a puzzling pattern: some region pairs move together (positive correlations), while others move in opposition (negative correlations, or anti-correlations). Most existing whole-brain models assume all long-range connections are excitatory and cooperative, which means they simply can't reproduce the anti-correlated part. This paper fixes that by showing the anti-correlations don't require inhibitory long-range wiring at all — they emerge naturally from how excitatory projections land on different cortical layers.

The key insight lives in the laminar neural mass model (LaNMM), a computational model of a cortical column with two oscillating pyramidal populations: P1, a slow alpha-band (~10 Hz) oscillator associated with deep (infragranular) layers, and P2, a fast gamma-band (~40 Hz) oscillator associated with superficial (supragranular) layers. These two oscillators are locally coupled such that when alpha power goes up, gamma power goes down — a negative envelope-envelope coupling (EEC). Because BOLD fMRI tracks metabolic demand, which is driven mostly by gamma-band synaptic activity, this means alpha power anti-correlates with BOLD and gamma power correlates with it. That matches well-established empirical findings.

Now connect two such columns via long-range excitatory projections. If you wire P1→P1 or P2→P2 (homologous connections, same layer to same layer), the alpha envelopes of the two columns synchronize, and their BOLD signals become positively correlated — cooperation. But if you wire P1→P2 (heterologous, deep to superficial), the alpha envelope of one column drives up gamma in the other, which then suppresses that column's own alpha via the local EEC — producing anti-correlated alpha envelopes and negative BOLD correlations — competition. Both effects arise from purely excitatory long-range connections. The sign of the interaction is determined entirely by which layer you target.

To test this at scale, the authors built personalized whole-brain models for 60 subjects. They first inferred subject-specific "signed" connectivity matrices from resting-state fMRI using Ising spin-glass models — a maximum-entropy approach that assigns each region pair a coupling strength that can be positive or negative. Crucially, they tried two variants: one constrained by each subject's diffusion MRI structural connectome (Ising-SC), and one constrained only by a sparsity penalty (Ising-L1, no structural data needed). They then wired up 84 LaNMM units per subject, using positive Ising connections to drive homologous (cooperative) projections and negative connections to drive heterologous (competitive) ones. The Ising-L1 models — inferred from fMRI alone — outperformed both the structural-connectome-only models and the structurally constrained Ising models, reproducing static functional connectivity, dynamic brain state transitions, and the correct alpha/gamma-BOLD relationships. The likely reason: diffusion tractography systematically underestimates interhemispheric and homotopic connections, which are among the strongest functional links in fMRI data.

The practical upshot is twofold. First, the paper provides a mechanistic circuit story — grounded in known cortical anatomy and cross-frequency physiology — for something that previously had only mathematical workarounds. Second, it demonstrates that personalized whole-brain models can be built from fMRI alone, without expensive diffusion MRI, which matters for clinical scalability.