When a brain region sees the same picture twice, its fMRI signal usually drops the second time. This effect is called repetition suppression. It is one of the few tools that can read out how neurons inside a single fMRI voxel are tuned, because a region that stays suppressed for a new stimulus must be treating it like the old one. The method has a well-known weak spot. A region may look suppressed only because the regions feeding it were suppressed, even though its own neurons never adapted. This is "inherited adaptation." We propose to measure it directly in humans. Using the Human Connectome Project (HCP), we would route each person's measured suppression through their own white-matter wiring, which is reconstructed from diffusion MRI (a scan that traces fibre bundles by how water diffuses along them). The part of each region's suppression that its anatomical inputs cannot explain would be read as local adaptation.
The idea borrows its logic from a single animal result. Cascaded Effects of Spatial Adaptation in the Early Visual System recorded from mouse thalamus and primary visual cortex. It showed that pure gain loss in the thalamus, pooled through fixed feedforward weights, accounts for about 98% of the variance in cortical adaptation, including tuning shifts that look like local plasticity. Repetition suppression: a means to index neural representations using BOLD? names this cascade as a central unresolved confound for human fMRI adaptation. No one has built the human, whole-brain, anatomy-constrained version. The HCP is the only resource that pairs high-quality diffusion tractography, myelin maps, and repetition-containing task fMRI in about a thousand people.
Core Claim
Suppression measured in a cortical region is the sum of a local component and a component carried in along its anatomical inputs. The carried-in component can be estimated from tractography plus a myelin-based ordering of cortical areas, and it should be category-general. The local component, once separated out, should be category-specific in the places where category tuning is known to live, and the raw, undecomposed measure should blur that specificity. The claim can be falsified in two ways. If a subject's own connectome predicts their suppression pattern no better than someone else's connectome, or no better than a degree-matched rewired graph, the anatomical routing claim fails. If the local residual is no more category-specific than raw suppression, the decomposition adds nothing.
Research Program
The data come from the HCP Young Adult working-memory task. In this task, pictures of faces, places, tools and body parts appear in 0-back and 2-back blocks. The design contains item repeats at the target lag and "lure" repeats at the wrong lag. Lures are the key contrast. They repeat an item without requiring a response, which reduces (but does not remove) the confound between repetition and motor or expectation signals. We would estimate trial-level responses by deconvolution within blocks. The short HCP repetition time of 720 ms makes this feasible at 2.5-second trial spacing. Responses are summarised on the 360-area multimodal parcellation used across the HCP work cited below. For each subject, area $i$ and picture category $c$, the observed suppression $s_{ic}$ is the log ratio of the repeated response to the novel response.
The model is a network autoregression:
$$ s_c = a_c + \beta\, W s_c + \varepsilon_c , $$
The terms are as follows. $s_c$ is the vector of regional suppression for category $c$. $a_c$ is the local (intrinsic) suppression. $W$ is a routing matrix built from the subject's tractography streamline weights. $\beta$ is a single gain shared across categories. $\varepsilon_c$ is noise. Solving the equation gives $s_c=(I-\beta W)^{-1}a_c+\ldots$, so local suppression at a few sites spreads along anatomy.
Diffusion tractography cannot tell feedforward from feedback. We therefore make $W$ directed by weighting each edge according to the difference in T1w/T2w myelin between its two ends, a common proxy for position in the cortical hierarchy. We then compare this directed version against a symmetric one. Two constraints make the problem identifiable. First, $W$ and $\beta$ are shared across the four categories while $a_c$ varies. Second, $a_c$ carries a sparsity penalty, which encodes the assumption that most regions inherit rather than generate adaptation. The penalty is tuned only by held-out prediction.
This yields three modalities in one model: diffusion wiring, myelin hierarchy, and task fMRI dynamics. The routing term is where they meet.
Before touching real data, we simulate adaptation on the actual HCP connectomes with planted local sites. We then check whether the fit recovers them at realistic trial counts. Estimating the dimensionality of neural responses with fMRI Repetition Suppression reports that its suppression-matrix estimates became precise only after averaging about 16 trials. That warns us that per-subject lure counts may be too low, so individual-level claims will rely on split-half reliability across the two HCP runs. Once the decomposition is validated, it can be applied to that paper's suppression-matrix rank idea as a later extension, since an inherited decrement can also inflate or deflate the apparent dimensionality of a downstream region.
Relation to prior work
Three HCP studies supply the methods we reuse, and none touches adaptation. Brain Functional Connectivity Estimation Utilizing Diffusion Kernels on a Structural Connectivity Graph (Neuro-Hotnet) spreads signal over diffusion graphs with hub down-weighting to find motor subnetworks in 308 subjects. We borrow its kernel as one candidate routing matrix, but it models time-averaged covariance and not trial-to-trial decrements. Uncovering Cortical Units of Processing From Multi-Layered Connectomes shows that structure regularizes noisy function in 250 subjects. It gets an AUC of 0.9957 for its joint model against 0.9806 with permuted structure. We adopt its permutation logic for our subject-swap null. Network geometry shapes multi-task representational transformations across human cortex propagates activity patterns through resting functional connectivity across 352 subjects. In that work, connectivity dimensionality predicts representational change at $r=0.44$. This is our closest methodological competitor. It routes through functional rather than anatomical links, and it models condition patterns rather than repetition.
Two papers constrain the design. Disambiguating brain functional connectivity shows that amplitude changes alone shift correlation estimates. Repetition suppression is exactly such an amplitude change, so any functional connectivity used for routing must come from rest, never from the task blocks being modelled. Caveats and Nuances of Model-Based and Model-Free Representational Connectivity Analysis shows that shared noise inflates model-free coupling from −0.22 to 0.84. For that reason we test anatomy against a subject-swap null rather than trusting raw cross-region agreement. Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome learns region-wise structure–function coupling for prediction on a non-HCP cohort. Its finding that coupling is tight in sensory cortex and loose in transmodal cortex gives a prediction we inherit: the inherited share of suppression should fall along the same gradient.
Baselines
The naive reference is the standard analysis. It treats each region's raw suppression $s_{ic}$ as entirely local, which is equivalent to setting $\beta=0$. Every decomposition claim is measured against it, on category specificity and on held-out prediction of suppression in unseen subjects.
The confound baseline predicts each region's suppression from the global mean decrement plus decrements of cortical neighbours within a fixed geodesic distance. This captures vascular fatigue, smoothing and whole-brain arousal with no anatomy. The proposal must beat it. Otherwise "inheritance" is just spatial autocorrelation.
The strongest existing competitor is activity-flow routing through resting-state functional connectivity, following Network geometry shapes multi-task representational transformations across human cortex. We run it with the same equation, the same parcellation, the same sparsity tuning and the same held-out subjects, with $W$ replaced by each subject's resting functional connectivity. For this specific question, no reported result was found on Lacuna for predicting repetition decrements in HCP. The nearest figure is that paper's $r=0.44$ between connectivity dimensionality and representational transformation. That figure is context, not a target on our metric.
A second structural competitor uses the Neuro-Hotnet heat kernel as $W$. This tests whether our myelin-directed tract weights add anything over an established structural propagation rule. Beating a baseline means higher held-out explained variance of regional suppression across a 5-fold subject split, with bootstrap confidence intervals over subjects.
The ablations isolate the key ingredient. They compare the directed matrix against a symmetric matrix, the subject's own connectome against a randomly assigned other subject's connectome, and both against a degree-preserving rewired graph. The routing claim requires own-connectome > swapped > rewired. The swapped-versus-own gap is the decisive number, because a group-average anatomy effect would not produce it.
Why not just…?
Why not just route through resting functional connectivity, as activity-flow models already do? This is a real threat. Functional connectivity may predict suppression spread as well as tractography does, since it partly reflects the same wiring. Resting functional connectivity also mixes direct links with shared input, and correlations cannot be directed. Our experiment settles this by fitting both routing matrices and their combination. If functional connectivity wins and tractography adds nothing, we report that inheritance follows functional rather than anatomical routes. The decomposition into local and carried-in suppression would still stand, though with a weaker mechanistic reading.
Why not just analyse suppression region by region, or decode with multi-voxel patterns? Per-region analysis is the naive baseline, and it is exactly what the cascade result in Cascaded Effects of Spatial Adaptation in the Early Visual System shows can be misleading. Pattern decoding asks a different question. It fails when tuning is mixed within voxels, a case Estimating the dimensionality of neural responses with fMRI Repetition Suppression constructs explicitly. Neither method says where a decrement came from.
Why not just train a graph neural network on structure and function? A Joint-GCN-style model, as in Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome, could predict suppression maps. It would not yield a separable local term. A simple but tough-to-beat baseline for fMRI time-series classification shows that heavy models often add little over simple ones on fMRI. We include such a network only as an upper bound on how predictable suppression is. If our one-parameter routing model comes close to that bound, interpretability costs little.
Why not just run a dedicated factorial adaptation experiment instead of reusing HCP? A purpose-built design that crosses repetition with expectation would be cleaner, and the HCP task was not built for this. What HCP provides that no small study can is individual tractography and myelin maps at scale, which the subject-swap test needs. We treat lures as the least-confounded contrast and name a dedicated replication as the follow-up if the effect appears.
Proposed Experiments and Decisive Checks
We first verify recovery. On simulations over real HCP connectomes, planted local sites must be recovered well above chance at trial counts matching HCP lures. If this fails, the empirical fits are not interpreted.
Second, on the full HCP working-memory cohort, we compare held-out explained variance across all baselines and ablations, with the own-versus-swapped connectome gap as the primary test.
Third comes the specificity check. The local term $a_c$ in fusiform and parahippocampal areas should show face and place preference, respectively, more sharply than raw suppression. The inherited component in the same areas should be nearly equal across categories, as expected for a signal passed on from early visual cortex. We quantify this with a category-selectivity index, compare raw and decomposed estimates, and test the difference by subject bootstrap.
Fourth, the inherited fraction per region should correlate negatively with position on the sensory-to-transmodal hierarchy, measured by myelin.
Fifth, targets and lures are fitted separately. If the decomposition is driven by response demands rather than repetition, the two fits will disagree sharply.
A positive result on all five would give human fMRI adaptation studies a correction they currently lack. A failure of the swap test alone would be informative too. It would tell the field that inherited adaptation, if present, is not traceable along individual white-matter routes at this resolution.
References
- Cascaded Effects of Spatial Adaptation in the Early Visual System
- Repetition suppression: a means to index neural representations using BOLD?
- Estimating the dimensionality of neural responses with fMRI Repetition Suppression
- Brain Functional Connectivity Estimation Utilizing Diffusion Kernels on a Structural Connectivity Graph
- Uncovering Cortical Units of Processing From Multi-Layered Connectomes
- Network geometry shapes multi-task representational transformations across human cortex
- Disambiguating brain functional connectivity
- Caveats and Nuances of Model-Based and Model-Free Representational Connectivity Analysis
- Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome
- A simple but tough-to-beat baseline for fMRI time-series classification