Is Brain Network Gene Coupling Just Cell Composition?

Two widely cited studies report that brain regions which fluctuate together in resting-state fMRI also express similar genes. The imaging data come from functional MRI, a scanner that tracks slow blood-oxygen signals. The gene data come from postmortem tissue in the Allen Human Brain Atlas. The reported correlations are about 0.32 to 0.35. They have been read as evidence that functionally connected regions share molecular programs. Both studies, however, rely on bulk tissue samples: each sample is a blend of neurons, glia and vascular cells. Both also rest on only six postmortem donors. This leaves a simpler explanation open. Connected regions may simply contain similar mixtures of cell types, and cell mix alone would make their bulk expression look alike. Cell mix may also shape the local blood signal that drives the fMRI correlation.

We propose a validation study built for six donors that separates these explanations. For each pair of network nodes, we ask how much of the gene–connectivity association survives three steps. First, we estimate each tissue sample's cell-type proportions and remove their contribution to expression. Second, we control for similarity in regional fMRI signal amplitude, because amplitude alone can change correlation-based connectivity. Third, we require the result to replicate on a donor held out from all fitting. The deliverable is a calibrated test that others can apply to any imaging–transcriptomic claim, together with a verdict on the two published results. It is not a new deconvolution model.

Core Claim

We hypothesize that a large share of the reported coupling between seed-based connectivity and gene expression is compositional. We test this against two specific results. The first is the limbic cortico-striatal result of Gene expression links functional networks across cortex and striatum, which aligned the Yeo cortical and Choi striatal network atlases with Allen Atlas samples and reported r = 0.32 (p ≤ 0.01). The second is the network-heterogeneity result of Functional orderly topography of brain networks associated with gene expression heterogeneity, which reported R ≈ 0.35 (P < 0.01) using 50 Human Connectome Project subjects.

Anderson et al. did check cell types, but only afterward: they ran enrichment tests of their gene lists against external single-cell marker sets and found somatostatin interneuron signal for limbic genes and parvalbumin and oligodendrocyte signal for motor genes. That check cannot tell whether connected regions regulate those genes differently inside each cell, or whether they simply contain more of those cells. Zheng et al.'s housekeeping-gene gradient runs from subcortex to neocortex. That is exactly the axis along which the ratio of neurons to glia changes, so the same ambiguity applies.

What is new here is the combination of three things. We apply composition removal to the connectivity–expression question itself. We add an amplitude term motivated by Disambiguating brain functional connectivity. And we calibrate the whole test against planted ground truth at the actual sample sizes involved: six donors and fMRI cohorts of 16 to 50 subjects.

Research Program

The unit of analysis is a pair of network nodes $(i,j)$. The nodes are seed regions defined by the same Yeo and Choi atlases, so that we reproduce Anderson et al. before intervening. We compute four quantities for each pair. $F_{ij}$ is the seed-to-seed functional connectivity: the Pearson correlation of the regions' fMRI time-series, estimated from Human Connectome Project resting-state data. $G_{ij}$ is the similarity of the regions' Allen Atlas expression profiles. $C_{ij}$ is the similarity of their estimated cell-type proportions. $A_{ij}$ is the similarity of their fMRI signal amplitude. We also record $D_{ij}$, the physical distance between the regions.

The central estimate is the coefficient $\beta_G$ in the model below. We compute it under two conditions. In the first, $G$ uses raw bulk expression. In the second, $G$ uses expression residualized on each sample's estimated composition, which is the "cell-intrinsic" version.

$$F_{ij} = \beta_G G_{ij} + \beta_C C_{ij} + \beta_A A_{ij} + \beta_D D_{ij} + \varepsilon_{ij}$$

The gap between $\beta_G$ with raw expression alone and $\beta_G$ with residualized expression plus the composition term measures how much of the published effect is compositional.

To estimate composition, we use reference-based linear unmixing: each tissue sample is modeled as a weighted sum of pure cell-type expression profiles, and the weights are the cell proportions. The file notes that CIBERSORTx and MuSiC are the standard tools of this kind. The reference profiles come from a human brain single-nucleus atlas. This is the class of reference Anderson et al. used for enrichment, but the file does not name a specific dataset, so the choice is ours and will be reported.

We deliberately do not use count-based models such as Count Bridges, which model integer RNA counts. The Allen Atlas uses microarrays, which produce continuous intensities rather than counts, so that paper's advantage does not transfer.

We do borrow ideas from spatial deconvolution. POLARIS separates layer-specific expression changes from shifts in cell proportion in spatial transcriptomics. GNN-C2L shows that cell composition becomes more different as physical distance between tissue spots grows. That second finding is a warning: composition and distance are entangled, so the model must include both $C$ and $D$. POLARIS assumes that a layer-specific shift is shared across all cell types in a spot. We do not make that assumption. Instead, the residualization is done within each broad cell class wherever the unmixing supports it.

The amplitude term addresses a separate concern. Cole et al. showed that a change in Pearson correlation can arise purely from adding uncorrelated signal or adding shared signal, without any change in coupling between regions. We extend that reasoning from changes in connectivity to connectivity itself. If regional fMRI amplitude tracks vascular or glial density, then amplitude could carry the compositional effect into $F$ through a route that bypasses neural coupling entirely.

The core of the validation is leave-one-donor-out estimation. We estimate $\beta_G$ on five donors and test the prediction of $G$-based coupling on the sixth, rotating through all six. On the imaging side, we repeat every fit on disjoint groups of Human Connectome Project subjects, with group sizes of 16 and 50 to match the published cohorts.

To check that the test behaves correctly, we build simulations that imitate the Allen Atlas. We create synthetic bulk samples by mixing reference cell profiles, using the real sampling coordinates and the real six-donor structure. We then plant one of three truths: coupling that is purely compositional, coupling that is purely cell-intrinsic, or a mixture. This tells us whether six donors can distinguish these scenarios at all, and what false-positive rate the test has.

Baselines

The naive reference point is to reproduce the published analyses exactly: raw bulk expression similarity correlated with connectivity, with no covariates. This should recover a value near Anderson et al.'s r = 0.32 for the limbic network and near Zheng et al.'s R ≈ 0.35. If we cannot reproduce these numbers, every later comparison is uninterpretable, so reproduction is the first gate.

The strongest existing method is Anderson et al.'s own cell-type analysis: enrichment of the connectivity-linked genes against external single-cell marker sets. We run it on the same samples, with the same parcellation and the same fMRI subjects. The proposal beats this baseline if composition-aware estimation changes the substantive conclusion in a way enrichment cannot detect. The clearest case would be a marker class that enrichment reports as strongly involved, but whose effect disappears once cell-intrinsic expression is used. A second case would be the reverse: a cell-intrinsic signal that survives composition removal, which enrichment has no way to certify.

The strongest cheap competitor is marker-score regression. Here the average expression of a few canonical marker genes per cell class stands in for the cell proportions, and these scores are regressed out. It gets the same parcellation, the same reference cell classes and the same held-out-donor protocol. Full unmixing only earns its place if it predicts better on the held-out donor, or if it gives better-calibrated error rates in the simulations.

The ablation that isolates our key ingredient removes the composition term and the residualization but keeps the distance and amplitude terms. A second ablation removes only the amplitude term.

For every numerical baseline, the only published figures are the two correlations above. Cole et al. report no numeric statistics in the file, and for POLARIS and Count Bridges we found no reported result on Lacuna.

Why not just…?

Why not just regress out marker-gene scores? This is the real threat, and it may work about as well as full unmixing. The Allen Atlas has too few samples per region for a complex deconvolver to have much room to outperform a simple proxy. We treat marker scores as a full experimental arm. If they match unmixing on held-out-donor prediction and on simulated error rates, we will recommend them. The contribution then becomes the calibrated test and the verdict on the published results, rather than the deconvolution machinery. The main risk for marker scores is that they may fail for rare interneuron classes, whose marker genes are often expressed in more than one cell type. The simulations will show whether that happens.

Why not just control for spatial distance or use spatial-autocorrelation nulls? Distance does explain local similarity. However, Anderson et al.'s limbic result links the nucleus accumbens to orbital and ventromedial prefrontal cortex and the temporal pole, which are far apart. Distance cannot account for that, while shared cell composition could. The model includes distance, and the key comparison is how much $\beta_G$ shrinks when the composition term is added on top of distance.

Why not just replicate in more brains, such as GTEx or Brainspan? Anderson et al. already used these as replication sources. They are also bulk tissue, so they carry the same compositional confound and provide coarser regional coverage. Replicating a confounded association does not remove the confound. Held-out-donor testing combined with planted-truth simulation addresses what those replications cannot.

Proposed Experiments

The first experiment establishes whether the test can work at this sample size. In the simulations, we measure the false-positive rate for a cell-intrinsic effect when the true coupling is purely compositional, and the power to detect a truly intrinsic effect, both with six donors and with fMRI cohorts of 16 and 50 subjects. If the test cannot distinguish the scenarios at six donors, we report that the published claims cannot be settled with Allen Atlas data. That negative result is worth publishing on its own.

The second experiment is the real-data decomposition for the limbic and somato/motor networks from Anderson et al. and the network-heterogeneity analysis from Zheng et al. For each, we report $\beta_G$ with raw expression, $\beta_G$ with cell-intrinsic expression, and the variance explained by composition and amplitude. All four quantities come with leave-one-donor-out intervals.

The decisive check is specified before any real-data analysis. The compositional explanation is supported if composition similarity alone reproduces at least half of the raw limbic correlation on held-out donors, and if the interval for cell-intrinsic $\beta_G$ includes zero. The cell-intrinsic reading survives if cell-intrinsic $\beta_G$ stays significant in every held-out donor and in both fMRI cohort sizes, even after the amplitude term is added. For Zheng et al.'s gradient, a dedicated check asks whether the ratio of neurons to glia, estimated by unmixing, predicts the regional housekeeping-gene fraction as well as their network-entropy measures do.