Hormonal remyelination signaling has reached a clinical impasse that is fundamentally a measurement problem. Thyroid hormone (T3) is known to act as a physiological switch that drives oligodendrocyte progenitor cells (OPCs) — the resident stem-like cells that rebuild myelin — past a developmental arrest and into mature, myelin-producing oligodendrocytes. A Phase 1b trial of oral liothyronine (synthetic T3) in MS patients proved safe and biologically active in the CNS, with 18 out of 20 participants completing the 24-week study and 46 cerebrospinal fluid proteins significantly altered (19 increased, 27 decreased) A Phase 1b, Open-Label Study to Evaluate the Safety and Tolerability of the Putative Remyelinating Agent, Liothyronine, in Individuals with MS. Yet nobody can say what the drug actually did to cells: the trial readouts were clinical disability scales and bulk fluid proteomics, which average over millions of cells. Meanwhile, Low T3 syndrome occurs in 25% of acute MOGAD exacerbations versus 2.1% of controls, and free T3 correlates with disability (Spearman r = −0.377, p = 0.024) Low T3 syndrome is associated with the severity of myelin oligodendrocyte glycoprotein antibody-associated disease exacerbation — but again, only as a bulk serological association. The cellular mechanism, in tissue, is invisible.
The second direction supplies the missing lens. Bulk sequencing masks cellular heterogeneity: an observed expression change may reflect altered transcription in one cell type, a shift in cell-type proportions, or compensation by a completely different cell population David Adams et al., Genome 2011. Single-cell profiling of human OPCs shows the oligodendroglial lineage is not one compartment but 12–13 transcriptionally distinct subpopulations — early OPCs, cytokine-responsive OPCs, pre-myelinating intermediates, mature oligodendrocytes, and an astrocyte-lineage branch — arranged along a pseudotime trajectory with a genuine bifurcation point: human OPCs can become either myelinating oligodendrocytes or astrocytes Single-Cell Transcriptomic Analysis Reveals Molecular Diversity of Human Oligodendrocyte Progenitor Cells; Single-cell transcriptomic reveals molecular diversity and developmental heterogeneity of human stem cell-derived oligodendrocyte lineage cells. This is the crux: T3 could in principle push cells along the myelinating arm, fail to move them at all, or — worst case — divert them toward reactive astrocytes that actively suppress OPC differentiation Reactive Astrocytes Derived From Human Induced Pluripotent Stem Cells Suppress Oligodendrocyte Precursor Cell Differentiation. Bulk readouts cannot distinguish these outcomes. Nobody has connected the two.
Core Claim
I propose to build and validate a trajectory-resolved deconvolution pipeline that uses the 12–13-subpopulation human OPC single-cell atlas as its reference, and to deploy it on bulk transcriptomes from thyroid-hormone remyelination contexts. The methodological claim is that deconvolution should not treat OPC maturation as flat categorical buckets: instead of a rigid signature matrix over discrete cell types, the reference should be parameterized along the pseudotime continuum, so that a bulk profile is explained as a mixture of lineage positions. The biological claim is testable and binary in its extremes: T3 status (depletion in Low T3 syndrome, supplementation with liothyronine) either shifts deconvolved tissue composition and per-state expression along the oligodendrocyte arm of the bifurcation, or it does not — and if it does not, the field's central remyelination hypothesis needs revision at the cellular level.
This differs from all closest work. The one existing CNS deconvolution study applied CIBERSORTx to aging mouse cortex and achieved approximately 90% validation accuracy against independent single-cell datasets, but collapsed the entire oligodendroglial lineage into a single "OG" compartment, never examined hormonal cues, and never touched demyelinating injury models Deconvolution reveals cell-type-specific transcriptomic changes in the aging mouse brain. The human OPC atlas papers produced the trajectory structure but only in vitro, without hormonal perturbation, and never used their atlas as a deconvolution reference Single-Cell Transcriptomic Analysis Reveals Molecular Diversity of Human Oligodendrocyte Progenitor Cells. The clinical thyroid papers generated the bulk specimens but had no cellular-resolution readout A Phase 1b, Open-Label Study...; Low T3 syndrome.... The intersection is genuinely empty.
Research Program
The pipeline has three components. First, a lineage-structured reference: I will encode the Chamling atlas as a smooth map from pseudotime position to expected expression, retaining the bifurcation so that oligodendrocyte-arm and astrocyte-arm positions are separate branches rather than one class. If $b$ is a bulk expression vector, it is modeled as
$$b \;\approx\; \sum_{c} \pi_c\, \mu\big(\tau_c, \text{branch}_c\big) + \epsilon,$$
where $\mu(\tau, \text{branch})$ is the expected expression profile at pseudotime position $\tau$ on a given branch of the lineage tree, $\pi_c$ are mixing weights over sampled lineage positions plus non-glial compartments (neurons, microglia, vascular cells, taken from standard brain references), and $\epsilon$ is residual noise. Second, an ensemble-referencing layer following SCDC, which showed that weighting multiple single-cell references by reconstruction error is what makes deconvolution robust to batch effects between reference and bulk data — the single most common failure mode in this literature SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References. Third, a discrete-count refinement following Count Bridges, which demonstrated that transcriptomic counts live on the integer lattice and are better deconvolved by Poisson birth-death dynamics within an Expectation-Maximization loop than by continuous regression, systematically outperforming CIBERSORTx and MuSiC on proportion recovery Count Bridges enable Modeling and Deconvolving Transcriptomic Data; Count Bridges enable Modeling and Deconvolving Transcriptomic Data. The innovation is confined to one place — replacing flat cell-type signatures with the trajectory-parameterized reference $\mu$ — because that is precisely the ingredient that both existing deconvolution tools lack and that this biological question demands: the open-problem literature explicitly flags that current tools treat cell identities as immutable categories rather than differentiation continua.
The biological program runs the pipeline on three specimen tiers: pseudobulk validation data constructed from the atlas itself; bulk transcriptomes from rodent demyelination models (cuprizone-fed and EAE mice) with T3-depleted, euthyroid, and T3-supplemented arms, which we will generate with bulk RNA-seq precisely because serial cellular sampling of lesions is impractical; and human clinical material — stored CSF and blood from the liothyronine trial cohort where available, plus a MOGAD acute-exacerbation cohort stratified by the Low T3 versus euthyroid grouping already established serologically Low T3 syndrome.... The analytic target is a single interpretable quantity: the deconvolved fraction of lineage mass sitting past the bifurcation on the myelinating arm, per sample, and its association with T3 status.
Proposed Experiments
The validation experiment comes first. I will pseudobulk the published time-course (3,271 cells, differentiation days 77, 89, 104, >90% PDGFRα-pure OPCs Single-Cell Transcriptomic Analysis...) at known mixing ratios and ask whether the pipeline recovers the known day-specific composition. The pre-registered success bar is taken from the strongest published number on this exact task type: SCDC achieved a Pearson correlation of r = 0.991 between deconvolved and true mixing ratios on a controlled three-cell-line mixture SCDC. I do not expect to match 0.991 with 12–13 near-collinear lineage states — that is the point of the ablation below — but recovery above r = 0.85 on the major states, with the bifurcation arms separable, is the minimum for proceeding to tissue.
The biological experiments then test three pre-registered predictions. In the MOGAD cohort, deconvolved myelinating-arm fraction should correlate positively with serum free T3, extending the existing disability correlation (FT3 vs. EDSS: r = −0.377, p = 0.024 Low T3 syndrome...) to a mechanistic endpoint. In the rodent models, T3 supplementation should shift the composition histogram toward post-bifurcation oligodendrocyte positions relative to euthyroid demyelinated controls, and Low T3-mimicking conditions should enrich arrested pre-bifurcation states — or, alternatively, astrocyte-arm positions, which would be an important negative finding given that reactive astrocytes suppress OPC differentiation Reactive Astrocytes Derived From Human Induced Pluripotent Stem Cells.... For the liothyronine trial material, the 46 altered CSF proteins (19 up, 27 down) A Phase 1b... serve as an orthogonal check: deconvolution should attribute the myelin-associated subset to maturing oligodendrocyte states rather than to microglial or endothelial compartments, and misattribution would falsify the pipeline before it falsifies the biology.
Baselines
The "do nothing" reference point is bulk differential expression on the same specimens — the readout the liothyronine trial itself used, treating total CSF protein shifts as the efficacy signal A Phase 1b.... It is the strongest competitor on practicality grounds because it requires no model, no reference, and no assumptions, and it is what every existing hormonal-remyelination study actually reports. It will be run on identical RNA-seq libraries with the same multiple-testing corrections. Beating it means showing a case where bulk differential expression is null or ambiguous (because opposing subpopulation shifts cancel in the average — the exact phenomenon demonstrated by scSet, where pseudobulk lost dynamics that a set-aware model retained, scoring weighted F1 of 0.95 versus 0.77 for pseudobulk on SLE clinical prediction A Diffusion-Based Autoencoder for Learning Patient-Level Representations from Single-Cell Data) while trajectory deconvolution recovers a significant, biologically coherent lineage shift. If bulk DE and deconvolution agree everywhere, the extra machinery has not earned its complexity.
The strongest existing method is classical categorical deconvolution with a flat brain signature matrix, as deployed by the aging-cortex CIBERSORTx study, which validated at approximately 90% overall accuracy against three independent single-cell datasets and recovered 45 of 50 immune-module genes Deconvolution reveals cell-type-specific transcriptomic changes in the aging mouse brain. I will run CIBERSORTx and MuSiC on identical bulk libraries with the same non-glial compartments included, so the only difference is the reference structure and the trajectory parameterization. Beating it means one specific thing: resolving within-lineage transitions that the flat matrix cannot, quantified as the correlation between deconvolved myelinating-arm fraction and independently measured T3 status. If the flat "OG" compartment captures the hormone effect as well as the trajectory reference does, the lineage-resolved claim collapses to a replication of Ren et al. — a real risk I will test head-on, since their 90% accuracy was achieved across broad compartments where our question lives precisely in the fine structure that broad compartments discard.
The key ablation isolates the one new ingredient: the identical pipeline, identical count modeling, identical ensemble weighting, but with the trajectory reference $\mu(\tau, \text{branch})$ replaced by flat per-cluster signatures (12–13 fixed categories, no continuum, no explicit bifurcation geometry). This ablation cleanly separates "more reference states helped" from "the trajectory structure helped," and it also probes the known multicollinearity danger that flat matrices face with closely related transitioning sub-states, which POPPY's ontology-aware Earth Mover's Distance supervision was designed to counter, reporting up to a 714% relative improvement in patient composition correlation on simulated bulk benchmarks Cross-modal transfer learning for mapping bulk transcriptomes at cellular level. All baselines receive the same input libraries, the same tuning budget (one grid per method over its own hyperparameters), and the same held-out validation splits.
Why not just…?
Why not just run single-nucleus RNA-seq on lesion tissue directly? For discovery on banked tissue, this is the right answer, and we will do it for a small anchor set. But it cannot serve the program's core use case. The clinical questions concern living trial participants, where serial brain biopsies are unethical and the only accessible specimens are blood and CSF — exactly the constraint the liothyronine trial operated under A Phase 1b.... Longitudinal monitoring of a 24-week dose escalation across dozens of patients by single-cell sequencing is cost-prohibitive, which is why large historical bulk cohorts exist and large historical single-cell cohorts do not Cross-modal transfer learning for mapping bulk transcriptomes at cellular level. There is also a tissue-specific technical hazard: droplet protocols systematically lose fragile, lipid-rich cells, as demonstrated for mammary adipocytes in the SCDC reference-mismatch analyses SCDC — and lipid-rich differentiating oligodendrocytes are plausibly the same failure mode, which would bias the exact population we care about. The honest division of labor: single-nucleus data on a few anchor samples grounds-truths the deconvolution; deconvolution scales it to cohorts.
Why not just use classical CIBERSORTx or MuSiC? This is the most serious objection, and the flat-matrix baseline above is designed to settle it empirically rather than by argument. Two pieces of evidence say it will fall short here. First, classical tools are batch-fragile: in the diabetic-islet cohort, only SCDC's multi-reference ensemble reproduced the clinically validated beta-cell/HbA1c association (p = 0.0019), while single-reference CIBERSORTx and MuSiC failed or gave contradictory results depending on which reference was chosen SCDC — and our setting crosses a worse boundary, human in-vitro reference to human in-vivo disease bulk. Second, flat signature matrices suffer collinearity among near-identical transitioning states; whether the trajectory parameterization actually fixes this in practice, rather than in principle, is exactly what the flat-reference ablation measures. If the ablation shows no gap, the skeptic wins and I will report it.
Why not just measure bulk MBP or MOG by qPCR, or use MRI? Because the question is not "did myelin increase" but "which cellular path produced the change" — and the marker genes themselves betray the bulk approach. In the human OPC trajectories, PDGFRα mRNA becomes undetectable in mature clusters even while intermediate states and residual protein persist Single-Cell Transcriptomic Analysis..., so a bulk PDGFRα reading cannot distinguish arrested intermediates from completed maturation, and an MBP rise could reflect oligodendrocyte restoration or reactive astrocyte compensation — the same non-autonomous masking demonstrated across bulk-expression studies COVID-19 tissue atlases reveal SARS-CoV-2 pathology and cellular targets. The scSet result makes the cost quantitative: set-level dynamics that pseudobulk averaging destroyed carried clinical prediction from F1 0.77 to 0.95 A Diffusion-Based Autoencoder.... MRI, for its part, cannot report lineage composition at all. These readouts remain useful secondary endpoints in our rodent experiments; they simply cannot adjudicate the bifurcation question.
Decisive Checks
The proposal stands or falls on two pre-registered numbers. First, pseudobulk recovery of the known day-77/89/104 composition from the published atlas, benchmarked against SCDC's r = 0.991 on controlled mixtures SCDC, with the flat-matrix ablation quantifying what the trajectory structure specifically buys. Second, in living cohorts, a significant association between T3 status and the deconvolved myelinating-arm fraction — the first cellular-resolution evidence, for or against, that thyroid hormone acts where two decades of remyelination biology has assumed it acts.
References
- A Phase 1b, Open-Label Study to Evaluate the Safety and Tolerability of the Putative Remyelinating Agent, Liothyronine, in Individuals with MS
- Low T3 syndrome is associated with the severity of myelin oligodendrocyte glycoprotein antibody-associated disease exacerbation
- As we come to the end of 2011, several members of the Genome Biology Editorial Board give their views on the state of play in genomics
- Single-Cell Transcriptomic Analysis Reveals Molecular Diversity of Human Oligodendrocyte Progenitor Cells
- Single-cell transcriptomic reveals molecular diversity and developmental heterogeneity of human stem cell-derived oligodendrocyte lineage cells
- Reactive Astrocytes Derived From Human Induced Pluripotent Stem Cells Suppress Oligodendrocyte Precursor Cell Differentiation
- Deconvolution reveals cell-type-specific transcriptomic changes in the aging mouse brain
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- A Diffusion-Based Autoencoder for Learning Patient-Level Representations from Single-Cell Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- COVID-19 tissue atlases reveal SARS-CoV-2 pathology and cellular targets