The left and right halves of the brain fire in step. The standard measure of this, voxel-mirrored homotopic connectivity (VMHC), is the correlation between the resting fMRI signal at a point in one hemisphere and at its mirror point in the other. It shifts in depression, schizophrenia, Alzheimer's disease and lupus, but fMRI cannot say what is happening in the tissue underneath. Transcriptomic deconvolution estimates the fraction of microglia, astrocytes, vascular cells and other types from a bulk RNA sample, as a least-squares unmixing against reference cell profiles. It can say what is in the tissue. Its absolute estimates, however, are biased by cell size and by mismatch between the reference cells and diseased ones. We propose to use the same trick VMHC uses, comparing each location to its mirror twin, on the transcriptomic side. Sample homotopic tissue in both hemispheres of the same animal, and deconvolve the pair jointly so that biases shared across hemispheres cancel. Then ask whether left–right immune asymmetry, rather than absolute immune load, is what shapes homotopic synchrony.
This matters because the one human study tying VMHC to autoimmunity reports something unexpected. In lupus, homotopic connectivity rises in the insula and parahippocampal gyrus while gray matter in the same regions shrinks. Higher insular VMHC tracks worse Trail Making performance, and the authors could only guess at "compensation." We offer a testable alternative. Inflammation that affects both hemispheres in the same way acts as a shared driver and should raise mirror-point correlation. Focal, one-sided inflammation should lower it. Distinguishing these requires a reliable estimate of immune asymmetry, which is exactly what we plan to build.
Core Claim
In brain tissue, the deconvolution quantity that can be estimated most reliably is the hemispheric difference in cell-type fractions within one animal, not the absolute fraction. The difference is also the natural biological partner of VMHC, which is itself a left–right comparison.
The reason is the structure of the known error sources. CSsingle shows that unequal total RNA per cell type skews bulk deconvolution. It corrects this with cell-size coefficients inside an iteratively reweighted least-squares fit and validates the approach in mouse cortex. Cell size, platform-specific gene biases and much of the reference mismatch are properties of the animal, disease stage and assay, not of the hemisphere. In a paired fit they should largely cancel.
We write the bulk profile of hemisphere $h \in \{L, R\}$ at one homotopic region as
$$y_h = \mathrm{diag}(b)\, X\, \mathrm{diag}(s)\, w_h + \varepsilon_h, \qquad w_L = \bar w + \tfrac{1}{2}\delta,\quad w_R = \bar w - \tfrac{1}{2}\delta,$$
where the symbols are defined as follows:
- $X$ is the reference matrix (genes by cell types).
- $s$ is a per-animal vector of cell-size factors.
- $b$ is a per-animal vector of gene-wise multiplicative biases.
- $\bar w$ is the shared composition, constrained non-negative and summing to one.
- $\delta$ is the asymmetry vector, fitted with an L1 penalty so that most cell types are assumed symmetric unless the data say otherwise.
- $\varepsilon_h$ is residual noise.
Because $b$ and $s$ are fitted once per animal from both samples, they no longer compete with $\delta$.
The limits are explicit. Biases that do differ between hemispheres do not cancel. The most important is an activated-microglia state present only on the inflamed side. We therefore add disease-state reference columns, such as disease-associated microglia, rather than claiming the pairing fixes this.
Research Program
The work has three parts, each depending on the previous one.
Estimator and simulation benchmark. We first build the paired estimator and test it in simulation where the truth is known. This establishes whether pairing actually beats deconvolving each hemisphere separately and subtracting.
Animal study with matched imaging, transcriptomics and histology. We then run a mouse study that pairs resting fMRI with bilateral transcriptomics from the same animals. It has three arms:
- A unilateral inflammatory lesion, which produces asymmetric inflammation by design.
- A lupus-prone strain with diffuse neuroinflammation, where inflammation is not imposed on one side and is expected to be roughly symmetric.
- Matched controls.
After scanning, we take bilateral punches of homotopic regions: insula, hippocampal/parahippocampal cortex, somatosensory cortex and cingulate. These are chosen to include the regions highlighted in the lupus study. The punches go to bulk RNA-seq. Adjacent sections are stained for microglia (Iba1), astrocytes (GFAP) and endothelium (CD31) to give histological ground truth for asymmetry.
Imaging analysis. VMHC is computed on a symmetric mouse template using the standard mirrored-correlation pipeline. Heterotopic connectivity is computed alongside it as a control. We also measure the registration asymmetry of each region, because forcing an asymmetric brain onto a symmetric template is a known VMHC artifact that could masquerade as a biological effect.
The scientific test has two parts. The sign of the VMHC change should differ between the asymmetric-lesion arm (VMHC decrease) and the diffuse lupus arm (VMHC increase). Across regions and animals, the estimated immune asymmetry $|\delta|$ should predict VMHC beyond three competing explanations: gray-matter volume, absolute immune fractions $\bar w$, and registration asymmetry.
Prior work plays three roles here. The lupus study Altered structural and functional homotopic connectivity associated with cognitive changes in SLE motivates the hypothesis. It links VMHC only to serum autoantibodies and has no tissue measurement, and we supply that measurement. The Alzheimer's study Inter-hemispheric functional dysconnectivity mediates the association of corpus callosum degeneration with memory impairment in AD and amnestic MCI constrains the design. It shows that callosal atrophy accounts for much of the variation in VMHC, with homotopic dysconnectivity mediating 41.1% of the atrophy–memory link. We therefore measure callosal integrity on structural MRI and include it as a covariate, so that immune asymmetry is not credited with a white-matter effect.
Two papers supply the method:
- Deconvolution reveals cell-type-specific transcriptomic changes in the aging mouse brain deconvolved mouse cortex into five cell types with CIBERSORTx. It found cortical immune genes spread across microglia, oligodendroglia and vascular cells. This is why we deconvolve rather than score microglia alone.
- CSsingle provides the cell-size correction we generalize to the paired setting.
The homotopic parcellation work Homotopic local-global parcellation of the human cerebral cortex from resting-state functional connectivity uses mirror symmetry as a prior that pushes the two hemispheres to agree. Our use is the opposite. The mirror twin is a matched control, and the quantity of interest is the disagreement between hemispheres. A symmetry-enforcing smoother would erase that signal.
Baselines
Naive reference. The simplest competitor is the between-hemisphere log-ratio of a marker-gene score for each cell type, a single-sample rank score in the spirit of ssGSEA. It needs no reference matrix. If it matches histological asymmetry as well as any deconvolution does, then deconvolution is unnecessary for density asymmetry.
Strongest existing method. This is CSsingle, including its cell-size correction, run separately on each hemisphere's sample, with the two fraction vectors then subtracted. It is the strongest competitor because it already handles cell size and was validated in mouse cortex. The file reports that it outperformed ten conventional algorithms but gives no numeric cortex error; no reported result found on Lacuna for CSsingle brain-cortex error. The closest reported brain figure is CIBERSORTx's roughly 90% validation accuracy against three independent single-cell datasets in the mouse aging study, which we also run as a separate-then-subtract baseline. There is no reported result found on Lacuna for any method's accuracy at estimating hemispheric asymmetry. We create that benchmark ourselves.
Fairness. All methods receive the same reference matrix, including the disease-state columns, the same genes and the same samples. Every tunable setting is chosen from the same grid by cross-validation on simulated data held out from evaluation.
Ablations. Two ablations isolate the key ingredient:
- Pooled bulk, no per-hemisphere asymmetry: the paired model with $\delta$ removed, fitted to the pooled sample.
- Paired design without shared nuisance: the paired model with $b$ and $s$ fitted separately per hemisphere.
The second ablation tests directly whether the gain comes from sharing nuisance terms.
What counts as beating the baselines. In simulation, the paired estimator must show at least 25% lower error on $\delta$ than CSsingle-then-subtract, and fewer false asymmetry calls for rare lineages. In the animals, it must show higher correlation with Iba1, GFAP and CD31 asymmetry than both the marker log-ratio and separate-then-subtract. If the second ablation matches the full model, the shared-nuisance argument fails, and we will report the method as a reparameterization of existing tools.
Why not just…?
Why not just deconvolve each hemisphere with CSsingle and subtract? This is the real threat. If bias is purely multiplicative and well estimated per sample, subtraction removes much of it anyway. The paired model should help most when samples are shallow, when rare cell types are near zero (the non-negativity bound makes separate estimates noisy there), and when per-sample bias estimates are unstable. The simulation varies exactly these conditions. If separate-then-subtract ties everywhere, we will drop the estimator claim and keep the biological study, which stands on either estimator.
Why not just count microglia on histology? For microglial density, histology is the ground truth, and we use it as such. It measures one marker per section, though. The aging-mouse deconvolution shows that immune activity lives in oligodendroglia and vascular cells as well as microglia, and activation state is not visible from a density count. If marker log-ratios or histology alone predict VMHC as well as the full asymmetry vector, the transcriptomic arm adds nothing. The regression comparison reports that outcome directly.
Why not just correlate human VMHC maps with Allen Human Brain Atlas cell-type maps? This is the obvious intersection, and nobody in the file has done it. The TMM model Trustworthy Enhanced Multi-view Multi-modal Alzheimer's Disease Prediction with Brain-wide Imaging Transcriptomics Data uses atlas expression over 116 regions without deconvolution and without any homotopic measure. The atlas comes from a handful of healthy post-mortem donors who were never scanned. It cannot link one individual's inflammation to that individual's synchrony. Any correlation between two smooth brain-wide maps is also confounded by shared spatial gradients. We will run this analysis as a secondary, hypothesis-generating check using a spatial permutation null, but it cannot test the symmetric-versus-asymmetric prediction.
Why not use whole-brain functional connectivity instead of VMHC? Our tissue variable is a left–right contrast, so the matching imaging variable is the mirror-point correlation. Dense whole-brain graphs also carry high false discovery rates (False Discovery Rates in Biological Networks). We still compute heterotopic connectivity as a control. If immune asymmetry predicts heterotopic change as strongly as homotopic change, the effect is generic network disruption, not homotopy-specific.
Proposed Experiments and Decisive Checks
Experiment 1: simulation benchmark. We build in-silico bilateral mixtures from mouse cortex single-cell references, following the pseudo-spot construction used in SpatialCTD. We inject known asymmetry in homeostatic microglia, activated microglia, astrocytes and vascular cells. We add shared cell-size factors and shared gene biases, and deliberately include reference mismatch confined to one hemisphere. Metrics are:
- error on $\delta$;
- detection area under the ROC curve for truly asymmetric cell types;
- false-positive rate for rare lineages when there is no true asymmetry.
Experiment 2: animal study. There are three arms of roughly twelve mice each, as described above. The outcomes are:
- agreement of estimated asymmetry with histological asymmetry, where the estimator comparison is decided;
- the arm-by-direction test of VMHC change;
- a mixed model predicting regional VMHC from $|\delta|$. The covariates are $\bar w$, gray-matter volume, callosal integrity and registration asymmetry, with animal as a random effect.
Decisive outcomes. The biological hypothesis is supported only if two things hold together. The asymmetric-lesion arm shows reduced VMHC and the lupus arm shows increased VMHC. And $|\delta|$ adds explanatory power beyond the covariates for homotopic but not heterotopic connectivity.
It is falsified in any of three cases:
- Both arms shift VMHC in the same direction.
- Absolute immune load explains VMHC as well as asymmetry does.
- Registration asymmetry or callosal integrity accounts for the effect.
A null in any of these cases would still settle something the lupus imaging literature cannot: whether its "compensatory" hypersynchrony has a detectable immune-asymmetry signature in tissue.
References
- Altered structural and functional homotopic connectivity associated with cognitive changes in SLE
- CSsingle: A Unified Tool for Robust Decomposition of Bulk and Spatial Transcriptomic Data Across Diverse Single-Cell References
- Inter-hemispheric functional dysconnectivity mediates the association of corpus callosum degeneration with memory impairment in AD and amnestic MCI
- Deconvolution reveals cell-type-specific transcriptomic changes in the aging mouse brain
- Homotopic local-global parcellation of the human cerebral cortex from resting-state functional connectivity
- Trustworthy Enhanced Multi-view Multi-modal Alzheimer’s Disease Prediction with Brain-wide Imaging Transcriptomics Data
- False Discovery Rates in Biological Networks
- SpatialCTD: a large-scale TME spatial transcriptomic dataset to evaluate cell type deconvolution for immuno-oncology