Does Individual Wiring Shape How Brain Activity Moves?

Diffusion MRI gives each person a map of their white-matter wiring, the structural connectome. Functional MRI gives a movie of their brain activity. A decade of work has tried to predict a person's time-averaged pattern of correlated activity (their functional connectivity) from their wiring, and the result is sobering. Deep models beat a "same answer for everyone" group average by only 8% to 11% (Predicting an individual's functional connectivity from their structural connectome). This note proposes that the field may have been asking wiring the wrong question. A physical connection is a potential route for signals, not a summary of correlations. If individual wiring carries individual information, that information should show up most clearly in how activity moves from moment to moment: which routes it takes and what pattern comes next. It should show up less clearly in the time-averaged correlation matrix, which washes those routes out.

We propose to build a time-resolved routing model and test it on a forecasting task. At every fMRI frame, we treat the current activation pattern as electrical current injected into a circuit whose conductances are the subject's own white-matter tracts. We compute how that current flows across each tract, and use those per-tract, per-frame flows to gate message passing in a spatio-temporal graph network that forecasts the next frames. The decisive test is a swap experiment. If forecasts made with a subject's own connectome beat forecasts made with the group-average connectome, with another subject's connectome, or with a randomly rewired one, by clearly more than the 8–11% seen for static prediction, then wiring matters for dynamics in a way static mapping missed. If they do not, that negative result is equally informative and cheap to establish.

Core Claim

Our hypothesis is that individual structural connectivity predicts the dynamics of an individual's activity substantially better than it predicts their time-averaged functional connectivity. We further hypothesize that this signal is best exposed by routing each frame's activity through the connectome as physical flow, not by using the connectome as a plain adjacency matrix. The method has two new ingredients. The first is a frame-by-frame transport computation that is linear in the number of tracts, so it scales past the coarse atlases that existing solvers are limited to. The second is a forecasting objective that makes "does the wiring help?" a falsifiable, subject-level question rather than a downstream classification score.

Where This Sits

The closest work is AFR-Net (Uncovering Latent Communication Patterns in Brain Networks via Adaptive Flow Routing). AFR-Net treats tracts as electrical conductances and time-averaged functional connectivity as signaling demand. It solves Kirchhoff's equilibrium equations to get a load on each tract, and uses those loads to bias a Transformer's attention. It reached 90.38% AUC on Parkinson's diagnosis against 71.88% for the next baseline. That result motivates us, since it suggests physically routed structure carries real signal. AFR-Net also constrains us in two ways. Its authors state that the method is static and undirected, and that its closed-form solve scales as $O(N^3)$, which confines it to 100–400 regions. We keep the transport idea and change three things: when demand is defined (every frame, not once per scan), how the solve is done (a few warm-started sparse iterations), and what the model is trained to do (forecast activity, not classify subjects).

Two other structure-plus-function models also collapse time. Joint-GCN (Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome) stacks a structural and a functional copy of each region into one graph with learnable coupling, but it uses a single scan-averaged functional matrix. BrainFlow (BrainFlow: A Holistic Pathway of Dynamic Neural System on Manifold) learns a continuous flow between connectivity matrices on the curved space of positive-definite matrices. Each task is still one static target matrix, and the method needs very large cohorts.

The dynamic brain models have the opposite gap. STAGIN (Learning Dynamic Graph Representation of Brain Connectome with Spatio-Temporal Attention), DSVB (A Deep Probabilistic Spatiotemporal Framework for Dynamic Graph Representation Learning) and STGC-GCAM (A novel spatiotemporal graph convolutional network framework for functional connectivity biomarkers identification of Alzheimer's disease) model time well. However, they build their graphs from sliding-window correlations or a generic atlas graph, with no tracts at all. STGC-GCAM's authors concede that their model finds damaged regions but cannot identify failing connections. Per-tract, per-frame flows give an edge-level readout by construction.

From video understanding we borrow one design principle. Relational action detectors keep a persistent identity for each tracked entity, linked across frames, and pass messages between entities along sparse, structured edges. That design took AVA action detection from 16.7% to 22.2% mean average precision over a dense 3D-convolution baseline (A Structured Model for Action Detection). Sparse motion graphs further let information skip frames where an entity is temporarily hidden (Can Zhang 0001). In our model, each brain region keeps a recurrent state across frames, so a region whose signal briefly drops out is not removed from the routing graph. The anatomical tracts play the role of the sparse spatial edges.

Research Program

Let $W$ be a subject's symmetric matrix of tract strengths over $N$ regions, $L$ its graph Laplacian, and $x_t \in \mathbb{R}^N$ the mean-centered regional activity at frame $t$, used as zero-sum current injected at each region. Region potentials and tract flows at frame $t$ are

$$ (L + \epsilon I)\,\phi_t = x_t, \qquad f_{ab}(t) = W_{ab}\,\big(\phi_a(t) - \phi_b(t)\big), $$

where $\epsilon$ is a small regularizer that makes the system invertible and $f_{ab}(t)$ is the signed flow on the tract between regions $a$ and $b$.

We do not invert the matrix. Instead we run $K$ conjugate-gradient iterations, each of which costs time proportional to the number of tracts, and warm-start each frame from $\phi_{t-1}$. Because consecutive fMRI frames are highly correlated, few iterations should suffice. This warm start is itself a temporal link: the circuit's state is carried across frames, like an entity's track in video. The flows are differentiable, so the conductances can optionally be fine-tuned from their diffusion-MRI values with a penalty keeping them close.

The forecasting network is a graph recurrent model. Each region carries a hidden state. At each frame, region $a$ receives messages from its anatomical neighbours $b$, weighted by learned functions of $|f_{ab}(t)|$ and the sign of $f_{ab}(t)$. It then updates its state and predicts $x_{t+1}, \dots, x_{t+h}$. The training loss is forecast error over horizons $h$ of one to ten frames.

We will also train a variant whose demand is AFR-Net's pairwise windowed correlation in place of the per-frame injection. This tells us whether our redefinition of demand loses the meaning AFR-Net's loads had.

Analyses then ask three questions. Do the learned per-tract flow profiles replicate within a subject across the two HCP resting-state sessions? Do they separate HCP task conditions? Do they localize to specific tracts in clinical contrasts, which is the edge-level readout STGC-GCAM could not provide?

Baselines

The naive reference points are a persistence forecast (the next frame equals the current frame) and a per-region linear autoregressive model fit per subject. Any graph model that does not beat them has learned nothing about inter-regional dynamics. We fit them on the same frames and the same horizons. There is no reported result on Lacuna for fMRI frame forecasting on HCP, so beating these baselines by a margin is established within the study rather than against a published number.

The strongest structure-free dynamic competitor is STAGIN's dense functional-graph attention, adapted to forecasting by swapping its classification head for the same forecasting head we use. It is the strongest because it has shown that dynamic functional graphs alone carry rich information: 88.20% accuracy and 0.9296 AUROC on HCP gender classification, and 99.19% on 7-way HCP task decoding (STAGIN). It receives the same input frames, parcellation, parameter count within 10%, and the same hyperparameter search budget. We report it on gender classification as a sanity check that our reimplementation approaches 88.20%.

The strongest structure-using competitor is AFR-Net recomputed per sliding window. At 68–116 regions the cubic solve is affordable, so this is a fair fight on accuracy, and we also report wall-clock time as resolution grows. For its original diagnostic use, AFR-Net's reported 90.38% AUC on PPMI and 70.62% F1 on ABCD-OCD are the numbers our flow features must match when fed to the same classifier head. For multimodal fusion on NCANDA, the target is Joint-GCN's 84.9% sex-classification accuracy (0.907 AUC).

Three ablations isolate the key ingredient. The first is the same recurrent network with raw tract weights $W$ as a fixed adjacency matrix and no flow computation. The second replaces the subject's own connectome with the group average, another random subject's connectome, or a degree-preserving rewiring of their own. The third varies the number of solver iterations $K$ from one to exact. The individuality claim counts as supported only under a specific condition. Own-connectome forecasts must beat group-connectome forecasts by a relative error reduction clearly larger than the 8–11% static benchmark, with confidence intervals excluding it, across held-out subjects. Forecasts must also identify their source subject above the 10.5% ± 2.5% above-chance matching rate reported for static prediction (Predicting an individual's functional connectivity).

Why not just…?

Why not just use the connectome as the adjacency matrix of an ordinary spatio-temporal graph network? This is the most serious threat, and it is the first ablation, not a strawman. A raw adjacency matrix only passes messages one hop at a time and ignores parallel routes. Flow through a resistive network accounts for all routes at once, which is the redundancy AFR-Net's effective-resistance construction was designed to capture. AFR-Net's large gain over functional-topology baselines suggests the routed quantity matters. If the plain-adjacency model matches ours on forecasting and on the swap test, we will report that the flow computation is unnecessary and that the individuality finding, if any, comes from the forecasting objective alone.

Why not just run AFR-Net on sliding windows? At coarse atlases this is feasible, and we run it. Windows of tens of seconds blur exactly the frame-scale routing we hypothesize carries individual signal, and they inherit the window-length artifacts DSVB's authors note. Cubic cost also rules out finer parcellations. If windowed AFR-Net equals our model at 68–116 regions, our remaining contribution is the scaling and the forecasting test, and we will say so.

Why not just forecast with a structure-free model and skip diffusion MRI entirely? Structure-free dynamic models are strong (STAGIN above). Diffusion MRI adds acquisition and tractography cost, so structure has to earn its place. The swap experiment settles this directly. If shuffled or rewired connectomes forecast as well as a subject's own, wiring adds nothing to dynamics at this resolution. That would be a clean negative result extending the static finding to the time domain.

Why not just improve static structure-to-function prediction? Careful evaluation already shows that line of work saturating at modest individual gains (Predicting an individual's functional connectivity). Our proposal tests a different target rather than a better regressor, so it does not depend on beating that ceiling on its own terms.

Proposed Experiments

The primary experiment uses HCP Young Adult resting-state and task fMRI with matched diffusion tractography, in the Desikan-Killiany (68) and AAL (116) parcellations. Subjects are split so that test subjects are entirely unseen. The model is trained on one resting-state session and tested on the other. We measure forecast error at horizons of one to ten frames, subject identification from forecasts, and the swap test described above.

Secondary experiments check that the flows are useful beyond forecasting. We freeze the per-tract flow statistics and train a light classifier for HCP gender and task decoding, NCANDA sex classification against Joint-GCN, and PPMI diagnosis against AFR-Net. On PPMI we also inspect whether diagnosis-relevant flow differences concentrate on specific tracts. A scaling experiment then measures accuracy and wall-clock time against $K$ and parcellation size, comparing iterative warm-started solves with exact inversion. This tests whether the linear-cost solver keeps exact-solve accuracy where the cubic method becomes impractical.

The single decisive check is the swap test on held-out HCP subjects. Each subject's own connectome is compared against the group average, a random other subject's connectome, and a rewired version of their own. A clear, replicated own-connectome advantage on forecasting, larger than the static 8–11% benchmark, supports the claim. A null result means individual wiring adds little to dynamics at macroscale resolution, which closes off the question efficiently either way.