People with chronic pain in many body sites are far more likely to develop depression later. Across 431,038 initially depression-free UK Biobank participants, those reporting five or more pain sites had a 3.32-fold higher risk of depression than pain-free people. Mendelian randomization, a method that uses genetic variants as natural experiments to test causal direction, points to pain driving depression rather than the reverse. Blood C-reactive protein (CRP), a marker of systemic inflammation, carries part of the effect (The inflammatory and genetic mechanisms underlying the cumulative effect of co-occurring pain conditions on depression). What happens in the brain between "has widespread pain" and "becomes depressed" is still unmapped in humans. Rodent work points to specific circuits, such as the anterior cingulate cortex (ACC, a frontal region that processes both the hurt and the unpleasantness of pain) projecting to the subthalamic nucleus. Those circuits were found with invasive tools that cannot be used in people.
This note proposes a focused, falsifiable study. Among imaged UK Biobank participants who have chronic pain but no depression at scan time, does a per-person measure of how tightly brain anatomy and brain activity are coupled, region by region, predict who later develops depression? The test is whether it adds anything beyond pain-site count, CRP, and medication and diagnosis history. The answer is useful either way. A positive answer gives a human, non-invasive readout of the pain-to-mood transition. A negative answer tells the multimodal-imaging community that its fused models do not yet earn their cost on this question.
Core Claim
The claim is that the pain-to-depression transition leaves a measurable signature in regional structure–function coupling. Here, coupling means how strongly a region's local anatomy tracks its pattern of functional connections across the brain.
Two prior findings point in opposite directions and motivate this. In unmedicated depression, coupling between anatomy and activity in prefrontal cortex and ACC is higher than in controls (p = 0.03), which the authors read as pathological rigidity (Multimodal Investigation of Network Level Effects Using Intrinsic Functional Connectivity, Anatomical Covariance, and Structure-to-Function Correlations in Unmedicated Major Depressive Disorder). In primary dysmenorrhea, a chronic pelvic pain condition, perigenual ACC to cerebellum connectivity is strongly reorganized while gray matter volume is unchanged (Altered Posterior Cerebellar Lobule Connectivity With Perigenual Anterior Cingulate Cortex in Women With Primary Dysmenorrhea). That is a decoupling-like pattern. Both studies have fewer than 75 people and neither follows anyone forward in time.
Our hypothesis is directional. Among people with multisite pain, those who show depression-like tightening of coupling in ACC and medial prefrontal regions are the ones who convert to depression. Those who show only the pain-like functional reorganization do not. The prediction is therefore an interaction: coupling should predict depression more strongly in people with pain than in pain-free people. That interaction is the comorbidity-specific signal.
The modeling contribution is deliberately narrow. Joint graph convolution (Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome) links a structural and a functional brain graph through learnable per-region coupling weights. Those weights are shared across the whole cohort. They were learned only for age and sex prediction in adolescents. We make the coupling weights subject-specific: a small network infers them from each person's own data, and they serve as an interpretable bottleneck feeding a survival model. This addresses the "rigid latent coupling" limitation the literature itself flags. Because a reviewer could fairly argue that a simple correlation-based coupling score would do as well, the study is built so that this possibility is tested directly rather than assumed away.
Research Program
Population and outcome. The cohort is UK Biobank participants with imaging, no depression diagnosis at or before the imaging visit, and pain-site data. Incident depression after the scan is defined as in Chen et al. The imaging visit came years after baseline, so follow-up from the scan is shorter than the 14.22-year average reported. The file contains no count of incident cases in the imaged subset. Establishing that count and the resulting statistical power is the first deliverable, and the study proceeds only if the expected event count supports detecting a modest gain in discrimination.
Imaging inputs. The primary arm uses T1 structural MRI and resting-state fMRI, which do not require diffusion imaging. Regional morphometry plays the role of the deformation-based volume measure and whole-brain functional connectivity plays the role of the intrinsic connectivity measure in Scheinost et al. Coupling is computed per region per subject.
A secondary arm adds diffusion-derived structural connectivity wherever usable diffusion scans exist. This allows a graph-based coupling in the Joint-GCN sense. Node features follow MaskGNN (Integrated Brain Connectivity Analysis with fMRI, DTI, and sMRI Powered by Interpretable Graph Neural Networks): cortical thickness, volume, and intrinsic neural timescales, meaning how slowly each region's fMRI signal decorrelates over time. Intrinsic timescales and coupling were that paper's strongest predictors of cognition.
Parcellation. The parcellation must resolve the hubs that rodent and small human studies implicate. These are the ACC subdivisions (perigenual, dorsal, subgenual), thalamic nuclei, the subthalamic nucleus where fMRI resolution allows, and cerebellar lobules VI, VIIb and IX from a cerebellum-specific atlas. A cortical-only atlas would pool exactly these structures away.
Non-imaging inputs. These are pain-site count, CRP and the other inflammatory markers from Chen et al., and a compressed summary of each person's diagnoses and medications. That summary comes from the graph-embedded topic model (A graph-embedded topic model enables characterization of diverse pain phenotypes among UK Biobank individuals), which turns sparse diagnosis and prescription records into interpretable topic mixtures. Including the topics matters because medications such as antidepressants, opioids and blood-pressure drugs alter both imaging and pain reporting. They would otherwise masquerade as brain signal.
Model. A per-subject coupling network outputs one coupling value per region. Its inputs are the subject's structural features, functional connectivity profile, and imaging-centre indicator. These coupling values, along with the clinical covariates, enter a Cox proportional-hazards head. That head is a standard survival model that estimates how each input shifts the instantaneous risk of developing depression.
Imaging-centre effects are handled two ways, as a covariate and through disentangled harmonization in the style of Learning disentangled representations to harmonize connectome network measures. That method was validated on only two sites and healthy people, so we check that harmonization does not erase the pain-associated coupling differences we are looking for.
Mediation. With the predictive question settled, we fit a joint mediation model with pain-site count as exposure, depression as outcome, and two mediators: CRP and ACC/medial-prefrontal coupling. This asks whether the brain pathway is distinct from, downstream of, or redundant with inflammation. It is the step that links the epidemiological result to brain networks.
Baselines
Do-nothing reference. The naive reference point is the clinical-only Cox model. Its inputs are age, sex, pain-site count, CRP and the other significant inflammatory mediators, the diagnosis-and-medication topics, and imaging centre. This is the strongest competitor that uses no imaging at all. Chen et al. show pain-site count alone carries a large, dose-dependent effect (hazard ratio 3.32 for five or more sites versus none). Any imaging model must beat this on the same participants, the same follow-up window, and the same cross-validation folds.
No reported result was found on Lacuna for discrimination metrics, such as the concordance index, of any model predicting incident depression from pain in UK Biobank. The baseline's discrimination must therefore be measured here, not quoted.
Frozen foundation-model baseline. The strongest existing imaging method is BrainHarmonix (Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens). It fuses structural MRI and resting fMRI into compact learned tokens and was pretrained partly on UK Biobank itself. Its reported transfer results are 66.67% accuracy and 74.88% F1 on ABIDE-II and 64.65% accuracy and 68.75% F1 on ADNI. It has not been evaluated on pain or depression.
We use its frozen embeddings, plus a linear Cox head, plus the same clinical covariates. Tuning matches our model: the same hyperparameter-search budget and folds. Because BrainHarmonix saw UK Biobank scans during pretraining, we confirm that no outcome labels leaked into pretraining.
Simple coupling baseline. This baseline computes one Pearson-correlation coupling score per region, in the manner of Scheinost et al., and enters it into the same Cox model. It isolates whether the learned per-subject coupling adds anything over a hand-computed one.
Cohort-level coupling ablation. This ablation fixes coupling weights at the cohort level, as in the original Joint-GCN. It isolates the proposal's key ingredient, subject specificity.
Criterion for beating a baseline. Beating a baseline means a gain in out-of-fold concordance index whose bootstrap confidence interval excludes zero, together with no loss in calibration. If the simple coupling baseline matches the learned model within its confidence interval, we report the simple measure as the finding.
Why not just…?
Why not just use pain-site count and CRP? This is the most serious threat, and it may win. Chen et al. show these two variables already explain much of the risk, and brain measures could simply be noisier proxies for the same burden. The study is designed so that this outcome is a reportable result rather than a failure. The decisive check is a likelihood-ratio test and concordance-index change from adding coupling to the clinical model. A second, sharper check is the pain-by-coupling interaction. If coupling predicts depression equally in pain-free people, it is a generic depression-vulnerability marker rather than a comorbidity mechanism, and we would say so.
Why not just probe a pretrained foundation model? BrainHarmonix embeddings may carry the same predictive information at no development cost. That is why they are a full baseline rather than a straw man. Their weakness for this question is interpretability. One-dimensional bottleneck tokens cannot tell you whether the signal lives in the perigenual ACC, the cerebellum or the precuneus. Answering that is the scientific point of the study. If the foundation model predicts better while our regional coupling explains nothing, we would conclude that the predictive signal is real but not localized to the hypothesized circuits. We would then use attribution on the foundation model as a fallback analysis.
Why not just compute a simple regional coupling score? The Scheinost-style score is cheap, transparent, and already shown to differ in depression. It may suffice. The comparison is built in as a baseline, and the learned model must beat it on held-out folds. Learned per-subject coupling has two plausible advantages. It can use the structural graph where diffusion data exist, and it can borrow strength across regions under noisy fMRI. The Joint-GCN gain over fusion baselines in its original setting was modest (84.9% versus 83.1% sex-classification accuracy), so we expect any advantage here to be small and treat that expectation as part of the test.
Proposed Experiments
Primary experiment. Five-fold cross-validated survival modeling, stratified by imaging centre and repeated over seeds, compares four models on the identical cohort: clinical-only, plus simple coupling, plus learned per-subject coupling, and plus frozen BrainHarmonix embeddings. Metrics are the concordance index, time-dependent area under the ROC curve at fixed horizons, calibration slope, and net reclassification among people with five or more pain sites, the group with the most clinical stakes.
Specificity experiment. We fit the pain-by-coupling interaction in the full imaged cohort. We also check that coupling does not predict increase in pain-site count over time where repeat assessments exist. Such a result would mean coupling tracks pain spread rather than mood.
Directional check against small studies. We test whether conversion-associated coupling shows the predicted signs. Those are tightening in ACC and medial prefrontal cortex, and alterations in precuneus and in perigenual ACC to cerebellar lobule VI, VIIb and IX connections, which echo the shared Crohn's–depression findings (Shared neurobiological changes in individuals with Crohn's disease and major depressive disorder) and the dysmenorrhea findings. These cohorts have 13 and roughly 35 patients, so they act as sign checks, not training data. If they are accessible, the frozen model would be applied to them unchanged.
Mediation experiment. We estimate the proportions of the pain-to-depression effect carried by CRP, by coupling, and by both jointly. The study also reports the sensitivity of these estimates to unmeasured confounding.
Decisive outcome. The study succeeds if three things hold: learned or simple coupling raises the held-out concordance index over the clinical-only model, the pain interaction is significant, and the effect survives harmonization. It fails informatively if the clinical-only model cannot be beaten. In that case, multisite pain burden and inflammation already capture what current MRI can see of the pain-to-mood transition.
References
- The inflammatory and genetic mechanisms underlying the cumulative effect of co-occurring pain conditions on depression
- Multimodal Investigation of Network Level Effects Using Intrinsic Functional Connectivity, Anatomical Covariance, and Structure-to-Function Correlations in Unmedicated Major Depressive Disorder
- Altered Posterior Cerebellar Lobule Connectivity With Perigenual Anterior Cingulate Cortex in Women With Primary Dysmenorrhea
- Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome
- Integrated brain connectivity analysis with fMRI, DTI, and sMRI powered by interpretable graph neural networks
- A graph-embedded topic model enables characterization of diverse pain phenotypes among UK biobank individuals
- Learning disentangled representations to harmonize connectome network measures
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens
- Shared neurobiological changes in individuals with Crohn’s disease and major depressive disorder