Personality as a Dial on Brain Reconfiguration: A Powered Test of Trait Moderation in the Human Connectome Project

Clinical neuroscience keeps asking which people will respond to a treatment and why. The usual evidence comes from trials of 20 to 60 patients, where any estimate of "this trait changes the treatment effect" is unstable. A clean example is the finding that baseline impulsivity changes how neurofeedback alters a slow brain potential in autism (Impulsivity Moderates the Effect of Neurofeedback Training on the Contingent Negative Variation in Autism Spectrum Disorder). The Human Connectome Project (HCP), a public dataset of roughly 1,000 healthy young adults scanned with structural MRI, diffusion MRI (which traces white-matter wiring), resting and task functional MRI, and a Five-Factor personality questionnaire, contains no treatments. It does contain something structurally similar. Every participant goes through the same within-person manipulations: seven tasks, including emotional faces, monetary gambling, and social animations, each of which pushes the brain away from its resting configuration. Each person's shift from rest to task can be read as an individual "response to a standardized intervention," measured in a sample about twenty times larger than a typical trial.

We propose to use the HCP as a testbed with enough statistical power to answer a specific question. Do enduring personality traits act mainly on how the brain changes under a challenge rather than on its resting state, and does a person's white-matter scaffold limit how far that change can go? The motivating puzzle is a published null result. Using static resting connectivity in 818 HCP participants, Functional Connectome of the Five-Factor Model of Personality found that conscientiousness was the only trait with any reliable network signature, and that neuroticism, extraversion, openness, and agreeableness showed no significant association. Our claim is that this null is partly a measurement choice. Traits act like dials on responses, so they should appear in the person-specific rest-to-task reconfiguration that matches each trait's function, and they need not appear in the resting baseline at all.

Core Claim

We define an individual response $\Delta_{i,k}$ for person $i$ and task $k$ as the change in a preregistered network quantity between rest and task. Examples are amygdala–prefrontal connectivity for the emotion task and striatal–frontal connectivity for gambling. Trait moderation is then the dependence of $\Delta_{i,k}$ on a trait score, after accounting for the person's resting level. Because every participant receives every task, there is no control arm. "Moderation" here means explained heterogeneity in a within-person effect, the same quantity a single-arm trial moderator analysis estimates. Our model is

$$\Delta_{i,k} = \beta_0 + \beta_R R_{i,k} + \beta_T T_i + \beta_C C_i + \beta_{TC}\, T_i C_i + \varepsilon_{i,k},$$

where $R_{i,k}$ is the person's resting value of the same network quantity, $T_i$ is the matched personality trait, and $C_i$ is regional structure–function coupling. Coupling is the correlation between a region's white-matter connection profile and its functional connection profile. It is measured as in Heritability and interindividual variability of regional structure-function coupling, which showed in 941 HCP subjects that it is a highly reliable individual property (test–retest $r = 0.977$). Including $R_{i,k}$ blocks the trivial result in which people with extreme baselines simply regress toward the mean.

The model makes three predictions. First, $\beta_T$ should be reliably nonzero for the trait–task pairings fixed in advance (neuroticism with emotion, extraversion with gambling reward, agreeableness with social cognition), even where the same trait has no association with $R$. Second, $\beta_{TC}$ should be nonzero: weak structure–function coupling in association cortex should widen the range of reconfiguration a trait can express, and tight coupling should narrow it. Third, both effects should be much less stable at trial-sized samples than at HCP scale, by an amount we can quantify.

Research Program

The work has three parts.

The first part builds a reliable response measure. Difference scores are notoriously noisy, so before testing any trait we estimate the test–retest reliability of each $\Delta_{i,k}$ on the HCP retest subset. We retain only network quantities whose reliability supports individual-level inference. Each task's connectivity is computed with rank-based correlation and high-amplitude frame retention, following the design-space results in Defining and Benchmarking a Data-Centric Design Space for Brain Graph Construction, whose code (GeQinwen/DataCentricBrainGraphs) we reuse. Head motion usually differs between rest and task, so it enters every model as a covariate. All splits keep families together, because HCP includes twins and siblings.

The second part tests moderation. We fit the model above per trait–task pair, and then fit a multivariate version in which all five traits and coupling across the 360 regions of the HCP multimodal parcellation predict the full vector of responses. We compare the gain from adding traits to a model that predicts the response from resting state alone. This is the proposal's central contrast. Prior multimodal work supplies the method for the structural side. Integrated Brain Connectivity Analysis with fMRI, DTI, and sMRI Powered by Interpretable Graph Neural Networks (MaskGNN) fuses structural and functional edges with morphometry and coupling node features to predict cognition in HCP-Development. Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome learns region-wise coupling weights between a structural and a functional layer. Both predict static targets: cognitive scores, age, and sex. We reuse their representations but change the target to a within-person change and the question to moderation.

The third part translates the results to the small-sample regime. We repeatedly subsample the HCP to N = 20, 40, and 60 and record how often each moderation coefficient flips sign or loses significance. This gives a concrete error profile for trait-moderation claims at trial scale. We then test whether priors estimated from the full HCP stabilize the small-sample estimates. These are empirical-Bayes shrinkage toward HCP effect sizes and covariance, plus an HCP-pretrained encoder of coupling. The external check is HCP-Development, a different cohort of 5–21 year-olds with the same acquisition. There, 50-person subsamples serve as stand-ins for an independent small trial. The aim of this part is to show whether normative priors make small-trial moderator analyses trustworthy. We do not claim that HCP findings transfer directly to patients.

Baselines

The naive reference point is the group-average response. Every person is predicted to reconfigure by the sample mean $\Delta$ for that task, with zero trait or structural information. The linear and group-template results in From Connectomic to Task-Evoked Fingerprints show this collapses individual differences. Any model must explain held-out variance in $\Delta$ above zero relative to it.

The strongest competitor is rest-only prediction of the task response. Predicting individual task contrasts from resting-state functional connectivity using a surface-based convolutional network (BrainSurfCNN) predicts individual task contrast maps from resting connectivity across 919 HCP subjects. It reaches 100% subject identification in the Language and Working Memory domains, at the level of test–retest scans. If resting connectivity already encodes each person's task response, personality may add nothing. We run a BrainSurfCNN-style predictor and a ridge regression from resting connectivity to $\Delta$ on identical family-aware splits, with the same hyperparameter search budget as our models. Beating it means a statistically reliable increase in held-out explained variance of $\Delta$ when traits and coupling are added. No reported result was found on Lacuna for rest-to-reconfiguration prediction with personality moderators, so the bar is set by our own run.

The multimodal static baseline is MaskGNN, which on HCP-D reports a fluid-cognition RMSE of 14.338, against 17.092 for a linear model and 16.853 for a standard graph network (Integrated Brain Connectivity Analysis with fMRI, DTI, and sMRI Powered by Interpretable Graph Neural Networks). We retrain it with $\Delta$ as its target to measure what multimodal fusion buys without an explicit moderation term. A simpler reference is the time-averaged MLP from A simple but tough-to-beat baseline for fMRI time-series classification, which averages per-timepoint outputs over the scan and matches far more complex models. It runs on the same inputs and budget.

Two ablations isolate the key ingredients. Removing the trait terms tests the dial hypothesis. Removing the trait-by-coupling interaction, or replacing coupling with a permuted version, tests whether white-matter structure constrains the response. A comparison point for the personality side is Riccelli et al.'s static-connectivity generalization error (relative RMSE around 0.15, for conscientiousness).

Why not just…?

Why not just predict the task response from resting connectivity, as BrainSurfCNN does? This is a real threat. If resting connectivity carries all the person-specific information about task responses, traits will add nothing to prediction. The proposal is built to settle exactly this. The decisive comparison is held-out variance in $\Delta$ with and without traits, given resting connectivity. If traits add nothing, we report that the dial hypothesis fails, and that Riccelli's nulls likely reflect a true absence rather than a measurement choice. That would itself be a useful result for trial design.

Why not just regress task activation on personality with no multimodal imaging? That is the traits-only arm of our ablation, and it could well capture most of the effect. The case for adding structure rests on evidence that traits have structural anchors. Agreeableness tracks white-matter integrity (Structural Correlates of Personality Dimensions in Healthy Aging and MCI), and gelotophobia tracks global white-matter efficiency (White-Matter Structural Connectivity Underlying Human Laughter-Related Traits Processing). The interaction term $\beta_{TC}$ is the specific test. If it is null, we drop the multimodal claim.

Why not just use a multimodal foundation model's embeddings? Brain Harmony and BrainCSD learn large-scale joint representations of brain structure and function. They target diagnosis, brain age (BrainCSD MAE 4.04 years), and cognitive scores. Where weights are usable, we include their embeddings as an extra predictor of $\Delta$. They may predict well, but they give no named coefficient for "this trait changes this response through this coupling." That coefficient is what a trial designer needs in order to stratify patients.

Why not just add dynamic connectivity models instead of structure? Time-varying network states predict intelligence beyond static connectivity (Behavioural relevance of spontaneous, transient brain network interactions in fMRI), and neuroticism has been linked to dynamic states (Dynamic phase-locking states and personality in sub-acute mild traumatic brain injury). This competes with our explanation of the null results. We therefore add resting-state dynamic-occupancy features as a covariate. If they absorb the trait effect on $\Delta$, the "dial" may be a resting dynamic trait after all.

Decisive Checks

The proposal succeeds if three things hold. At least two of the three preregistered trait–task pairings must show reliable moderation of $\Delta$ beyond resting connectivity in the HCP young-adult cohort, and this must replicate in HCP-Development. The coupling interaction must survive permutation. HCP-derived priors must measurably reduce sign-flip rates at N = 40 relative to unregularized estimates. If traits add nothing over resting connectivity, that negative result will be stated as plainly as a positive one, because it bears directly on whether small-trial trait-moderation claims are worth pursuing.