Is the Depression Brain Map Partly a Drowsiness Map?

Some studies link brain scans of people with a disorder to the cell types that live in the affected regions. The method takes a map of how patients' brains differ from controls and correlates it, region by region, with gene-expression atlases from post-mortem brains. A recent large study used this approach on major depression and implicated somatostatin interneurons and astrocytes (Convergent molecular, cellular, and neural signatures of major depressive disorder). The functional part of that disease map came from resting-state fMRI, which is a scan taken while people lie still with nothing to do. In such scans, people drift toward sleep. When arousal drops, the scanner records a brain-wide surge in the BOLD signal (the blood-oxygen signal fMRI measures). That surge is strongest in sensory and motor cortex, and it comes with deactivation in cholinergic arousal centres (Subcortical evidence for a contribution of arousal to fMRI studies of brain activity). People with depression sleep badly and feel fatigued. If they also get drowsier in the scanner, part of their "disease map" is really a map of falling asleep. Correlating that map with cell-type atlases would then identify the cells that make a region respond strongly to drowsiness, not the cells involved in depression.

We propose to measure this risk directly and to remove it. The correction borrows the logic of bulk genomics. There, a tissue sample is a mixture of cell types, and tumour purity is modelled explicitly so that signal from contaminating cells is not mistaken for signal from cancer cells. We treat each fMRI scan the same way, as a mixture of trait signal and arousal-state signal. We estimate each subject's in-scanner arousal and model diagnosis, arousal, and their interaction at the subject level. Only after that do we run the cell-type enrichment.

Core Claim

Imaging-transcriptomic enrichments drawn from resting-state functional maps can be produced by group differences in in-scanner arousal alone. The reason is that the cortical pattern of arousal-driven BOLD variance is itself spatially organised, with sensory-motor regions responding most. Any gene or cell-type map that shares that sensory-to-association gradient will correlate with it.

We make three claims, each testable:

  1. A pseudo-diagnosis that only separates drowsier from more alert people will produce "significant" cell-type enrichments under the standard pipeline.
  2. A subject-level model of the form $\text{FC} \sim \text{Group} + \text{Arousal} + \text{Group}\times\text{Arousal}$ removes most of these spurious enrichments, where FC is a regional functional measure.
  3. Simply regressing the arousal pattern out of the finished group map removes them less cleanly, or removes real signal along with them.

The model in the second claim is a direct translation of the purity-interaction model of Cancer expression quantitative trait loci (eQTLs) can be determined from heterogeneous tumor gene expression data by modeling variation in tumor purity. In that paper, adding the interaction term cut a simulated false discovery rate from about 46% to about 3%.

The novelty is the claim that the confound passes from the scan into the cell-type inference, together with a correction placed upstream of the spatial correlation. The MDD study ran spin permutation tests. These tests shuffle a cortical map by randomly rotating it on the sphere, which builds a null distribution that keeps spatial smoothness. They control spatial autocorrelation, but they assume the input map is a pure trait. Nothing in them asks whether the map is a state artifact.

Research Program

Arousal proxy. We first build a per-subject, per-scan arousal estimate that works in cohorts without physiological recordings. Most large repositories, including UK Biobank, lack concurrent EEG, pupillometry or ECG. Following the cortical–subcortical signature in Subcortical evidence for a contribution of arousal to fMRI studies of brain activity, we derive a spatial arousal template from global-signal peaks in Human Connectome Project (HCP) data. The template has positive sensory-motor transients and negative basal-forebrain and midline-thalamus transients. Each subject's arousal score is the time-resolved correlation of their BOLD data with this template, summarised as the fraction of time in low-arousal epochs.

We validate the proxy in two ways. First, against ECG-derived heart-rate variability in the test-retest dataset of Improving the Test-Retest Reliability of Resting State fMRI by Removing the Impact of Sleep, which used RMSSD (a beat-to-beat heart-rate variability measure) over 16-second windows to flag sleepy volumes. Second, against the short-versus-conventional-sleeper contrast in HCP, which Multivariate Heteroscedasticity Models for Functional Brain Connectivity showed produces artifactually stronger connectivity.

Cell-type sensitivity of arousal. Next we ask which cell-type expression maps correlate with the arousal template. We use AHBA expression and the same single-nucleus cell-class signatures and spin tests as the MDD study. This gives the "drowsiness-susceptibility" cell profile: the cell types that would be falsely implicated by any group that is drowsier. It also addresses an open question raised by the arousal paper itself, namely why sensory-motor cortex responds most to arousal drops. We treat that answer as descriptive, not causal.

Rerunning the depression analysis. Third, we redo the MDD functional analyses in UK Biobank and GSP. We compute RSFA (resting-state signal amplitude) and GBC (global brain connectivity) group-effect maps three ways: uncorrected, with the arousal score as an additive covariate, and with the purity-style interaction model. Under the interaction model, the reported disease map is the group effect evaluated at a fixed, matched arousal level, typically the cohort's high-arousal quantile. This mirrors how the cancer eQTL work reads off the genotype effect in pure tumour. For each map we rerun cell-type enrichment. The question is whether somatostatin-interneuron and astrocyte enrichment survives, shrinks, or changes identity.

Baselines

The naive reference is the published pipeline as it stands: an uncorrected functional effect map, AHBA correlation, and spin tests. The strongest reported number on this question is the spatial correlation of $r_s = 0.72$ (binned analysis) between MDD imaging alterations and AHBA expression, from Convergent molecular, cellular, and neural signatures of major depressive disorder. We reimplement this pipeline on the same UK Biobank and GSP subjects, the same parcellation, and the same spin-test count, and confirm that we reproduce the published figure approximately before comparing anything.

The strongest existing correction is global signal regression (GSR), which removes each subject's brain-wide mean time course before computing connectivity. It is the default way practitioners remove arousal-linked global fluctuations. We run it within the identical pipeline.

The second strong competitor is HRV-based volume scrubbing following Improving the Test-Retest Reliability of Resting State fMRI by Removing the Impact of Sleep, which reported a 33.4% gain in connectivity test-retest reliability. That method needs ECG, which UK Biobank lacks. We therefore run it in two ways. On the ECG dataset we use real RMSSD. In the large cohorts we use proxy-defined drowsy epochs, removing volumes at matched retention fractions from 100% down to 40%.

Two ablations isolate the key ingredient, which is the subject-level interaction term. The additive-only covariate model tests whether the interaction matters at all. Post-hoc spatial regression of the arousal template from the finished group map tests whether the correction must happen at the subject level.

The proposal beats these baselines if it meets two conditions. On the pseudo-diagnosis benchmark below, it must cut the rate of spurious cell-type enrichment by more than GSR and scrubbing do. At the same time, it must recover the planted true effects better than any baseline, measured by the area under the recovery curve for planted cell-type associations. All methods use the same subjects, parcellation and permutation budget. The interaction model gets no hyperparameter search beyond choosing the arousal reference level, which is fixed in advance.

Why not just…?

Why not just apply global signal regression? GSR subtracts the very brain-wide component that Subcortical evidence for a contribution of arousal to fMRI studies of brain activity ties to real cholinergic arousal circuitry. Questions and controversies in the study of time-varying functional connectivity in resting fMRI notes that GSR induces anticorrelations and removes neural variance. GSR also removes a uniform global time course. The arousal response is not uniform: it is regionally graded, and the gradient is exactly what contaminates the spatial correlation. We therefore expect a residual sensory-motor bias after GSR. The pseudo-diagnosis benchmark measures this directly. If GSR performs as well as our model, we will report that the cheaper fix is sufficient.

Why not just scrub drowsy volumes using heart-rate variability? Scrubbing works where ECG exists, but the large cohorts that drive imaging-transcriptomics lack it. The reliability gains in the original study held even after discarding up to 60% of volumes. That is acceptable for test-retest reliability, but it is harmful for group contrasts if patients lose more data than controls. Scrubbing also discards rather than models: a group that spends more time drowsy keeps a different, smaller sample of its own resting state. We compare at matched retention. If proxy-based scrubbing matches the interaction model, it is a legitimate and simpler outcome.

Why not just regress the arousal template out of the final disease map? This is the most serious threat, because it costs almost nothing. Its weakness is that it cannot separate two cases: patients being drowsier, and patients having real sensory-motor differences at matched arousal. It removes any disease signal that happens to lie along the arousal gradient. The planted-effect benchmark tests this by placing true group effects that partly align with the template. If spatial partialling recovers them as well as the subject-level model, we will recommend it instead.

Why not just add sleep duration or the arousal score as an additive covariate? The genomics precedent says additive adjustment is not enough when contamination scales with a mixing proportion. Additive models left a false discovery rate of about 46% in Cancer expression quantitative trait loci (eQTLs) can be determined from heterogeneous tumor gene expression data by modeling variation in tumor purity, and A computational method for cell type-specific expression quantitative trait loci mapping using bulk RNA-seq data showed leakage across cell types when mixtures are mis-modelled. Whether the brain case behaves like that analogy is an empirical question, and the additive ablation answers it.

Proposed Experiments

Pseudo-diagnosis test. This is the decisive test. In HCP, we label short sleepers (six hours or less) as "cases" and conventional sleepers (seven to nine hours) as "controls". This contrast should carry no psychiatric pathology but does carry known in-scanner drowsiness (Multivariate Heteroscedasticity Models for Functional Brain Connectivity). We run the full imaging-transcriptomics pipeline on it. If the uncorrected pipeline reports significant, spin-tested enrichment for specific cell classes, the vulnerability is demonstrated. If those classes overlap the depression findings, the vulnerability is directly relevant to them. We then measure how much each correction suppresses the false enrichment.

Semi-synthetic benchmark. To build ground truth, we resample HCP subjects into groups that differ in proxy-measured drowsiness by controlled amounts. We plant known group effects in selected parcels whose expression profile marks a chosen cell class. We report false-enrichment rate and true-enrichment recovery as functions of the drowsiness imbalance. This is the analogue of the 38,000 known-composition mixtures used in Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures.

Depression re-analysis. We first check whether cases score as drowsier on the proxy in UK Biobank and GSP. We then test whether $r_s$ with AHBA and the somatostatin-interneuron and astrocyte enrichments survive the interaction model.

Internal control. The cortical-thickness map from the same study cannot be affected by in-scanner arousal, so its enrichment should not change under any of our corrections. If it does change, the correction is removing structure rather than state, and we will treat the method as failing.

Developmental replication. Within budget, we repeat the pseudo-diagnosis logic in ABCD children, where drowsiness was documented as a major confound of dynamic connectivity (Cognitive and psychiatric relevance of dynamic functional connectivity states in a large (N > 10,000) children population). This tests whether the vulnerability generalises beyond adults.

No reported result was found on Lacuna for arousal-corrected imaging-transcriptomic enrichment, or for imaging-transcriptomics applied to a sleep-only pseudo-diagnosis. Those numbers are what this project would produce.