Cell-Resolved Co-Expression Mapping of Circadian Tryptophan Flux

Every tissue is a mixture of different cell types, and each cell type keeps its own internal clock, running slightly out of step with its neighbors. When we measure gene expression or metabolites in a chunk of tissue, we average over all of this. A recent study showed that dietary tryptophan — the amino acid sitting at a biochemical fork between serotonin and the inflammatory kynurenine branch — acts as a timing signal for circadian clocks in both liver and brain, and that its depletion disrupts behavioral rhythms in mice (Tryptophan metabolism is a physiological integrator regulating circadian rhythms). But that study, like nearly all work at this intersection, measured whole ground-up tissue. We currently cannot say which cells in the liver shift tryptophan toward neurotoxic kynurenine products at which times of day, or whether the clock's influence on this metabolic fork is a genuine intracellular program or an artifact of immune cells drifting in and out of the tissue on a diurnal schedule.

This matters because the two leading hypotheses make the same prediction in bulk data and opposite predictions at cell resolution. One says clock disruption reroutes tryptophan flux by changing which enzymes specific cells express and when; the other says it changes which cells are present at each hour, and the "rerouting" is a compositional illusion. A third possibility, demonstrated in skin, is that cells keep oscillating individually but fall out of phase, so the tissue average looks damped (tauFisher predicts circadian time from a single sample of bulk and single-cell pseudobulk transcriptomic data). Bulk analysis cannot distinguish these, and the standard tool for finding coordinated gene programs — weighted gene co-expression network analysis (WGCNA), which clusters genes by how strongly their expression tracks together — actively conflates them.

Core Claim

I propose a pipeline, and the biological question it is built to answer, that no existing work addresses: apply count-based cell-type deconvolution to dense bulk circadian time-course data before constructing co-expression networks, then relate the resulting cell-type-specific module eigengenes — the first principal component of each gene cluster's expression — to time-resolved tryptophan-kynurenine metabolite measurements. The claim is twofold. Methodologically, standard WGCNA on bulk circadian time-series produces false "metabolic modules" whenever cell-type proportions cycle diurnally, and a deconvolution-then-mixed-model pipeline removes these while preserving genuine intracellular modules. Biologically, the circadian gating of the serotonin/kynurenine branch is cell-type-specific — hepatocytes rather than Kupffer cells in liver, and specific neuronal populations rather than astrocytes in the suprachiasmatic nucleus (SCN, the master clock) — and clock disruption via constant darkness reroutes flux by shifting the phase and amplitude of enzyme oscillation within those cells, not primarily by changing tissue composition.

The key existing work to distinguish from is Tryptophan metabolism is a physiological integrator regulating circadian rhythms, which established the tryptophan-clock link with bulk RNA-seq of SCN and liver, isotope tracing, and metabolomics at only two static timepoints (ZT4 and ZT16). It did no deconvolution and no network analysis. On the methodological side, Application of Weighted Gene Co-expression Network Analysis for Data from Paired Design fixed WGCNA's independence violation for paired static biopsies using linear mixed-effects models (LMMs — regression where each subject contributes a random offset), and Count Bridges enable Modeling and Deconvolving Transcriptomic Data provides a generative model that treats RNA-seq reads as integers and recovers cell fractions better than CIBERSORTx or MuSiC. Nobody has connected these: deconvolution has never fed into longitudinal WGCNA, and neither has touched a circadian metabolic time-course.

Research Program

The pipeline has four stages. First, build cell-type reference expression profiles for liver and SCN from public single-cell atlases, anchored by a small number of new single-cell RNA-seq runs at two circadian timepoints — enough to capture the cell types present, not enough to be the main measurement. Second, deconvolve the dense bulk time-series using the Count Bridges expectation-maximization formulation, in which the single-cell count profiles are latent variables and bulk counts follow Poisson birth-death dynamics on the integers. The mixing model is simply

$$Y_g(t) = \sum_{k} c_k(t)\, X_{gk}(t) + \varepsilon,$$

where $Y_g(t)$ is the bulk count of gene $g$ at time $t$, $X_{gk}(t)$ is the expression of gene $g$ in cell type $k$, and $c_k(t)$ is cell type $k$'s proportion at time $t$. Crucially, both $c_k(t)$ and $X_{gk}(t)$ are allowed to vary over the circadian cycle, so diurnal composition swings and intracellular transcriptional oscillation are estimated as separate quantities rather than confounded.

Third, run WGCNA separately within each estimated cell-type profile across the time-course, and model each module eigengene as a harmonic

$$\mathrm{ME}_k(t) = a_k + b_k \sin\!\left(\tfrac{2\pi t}{24} + \phi_k\right),$$

with amplitude $b_k$ and phase $\phi_k$ fitted per cell type per module, sharing the 24-hour period across cell types in the spirit of MOSAIC: A Joint Modeling Methodology for Combined Circadian and Non-Circadian Analysis of Multi-Omics Data, which showed that jointly fitting noisy series with a shared period rescues rhythms that standalone detectors miss. Fourth, associate module eigengenes with isotope-traced flux into serotonin versus kynurenine branches using the mixed-effects module-trait machinery of the paired-design WGCNA paper, with mouse identity as a random effect so that repeated sampling of the same animal across the cycle does not inflate significance.

The biological payoff is a cell-type-attributed map of the tryptophan fork: which cells upregulate IDO/TDO (the enzymes that commit tryptophan to the kynurenine branch), when in the cycle, and whether the accelerated kynurenine synthesis observed under constant darkness in the integrator study lands in hepatocytes or in infiltrating myeloid cells. The same map tells us whether the receptors that gate kynurenine's downstream effects — GPR35, which kynurenic acid uses to block NLRP3 inflammasome activation (Kynurenic acid ameliorates NLRP3 inflammasome activation by blocking calcium mobilization via GPR35), and AHR, which regulates kynurenine-synthesizing enzymes in hepatocytes and astrocytes (Beta-defensin 1, aryl hydrocarbon receptor and plasma kynurenine in major depressive disorder: metabolomics-informed genomics) — oscillate in specific resident cell types, which would explain how peripheral clock desynchronization converts into NLRP3-mediated tissue inflammation.

Baselines

The naive reference point is standard WGCNA run directly on the raw bulk time-series, module eigengenes correlated with metabolite levels by ordinary Pearson correlation. This is what the field effectively does today, and it is the baseline most likely to show the confounding effect: if its top metabolite-correlated modules dissolve under deconvolution into composition-cycling markers, that is the central methodological finding. For a quantitative anchor on what an unconfounded module-trait association looks like, the paired-design WGCNA paper reports, in oral squamous cell carcinoma (GSE45238), a turquoise module association with disease status of $t = -18.68$, $p < 1.97 \times 10^{-20}$, and a hub-gene correlation of $\text{cor} = 0.53$, $p = 8.9 \times 10^{-24}$ (Application of Weighted Gene Co-expression Network Analysis for Data from Paired Design); modules in our setting should clear effect sizes of that order, and associations that survive only in bulk but not in any deconvolved cell type get flagged as compositional artifacts.

The strongest competitor for the rhythm-detection component is tauFisher, which predicts internal circadian time from a single un-timestamped transcriptomic sample using within-sample gene-pair differences; it beat ZeitZeiger and TimeSignatR on 11 of 12 benchmark datasets with prediction error typically within a 2-hour window, and reached roughly 78–83% accuracy on mouse SCN and skin pseudobulk data (tauFisher predicts circadian time from a single sample of bulk and single-cell pseudobulk transcriptomic data). We will run tauFisher on each deconvolved cell-type profile as a phase estimate and compare its per-cell-type time predictions against our fitted $\phi_k$: agreement would independently validate the harmonic fits, and systematic disagreement (per-cell-type times that scatter while the tissue average looks coherent) is exactly the phase-dispersion signature. For the joint harmonic modeling of eigengenes, MOSAIC is the comparator; it rescued up to 45 oscillating proteins in high-noise, low-resolution scenarios (MOSAIC), though it assumes bulk samples and no cell mixture, so it serves as the strongest existing rhythm detector rather than a full rival pipeline.

For deconvolution itself, the strongest baselines are Count Bridges — which outperformed CIBERSORTx and MuSiC at recovering cell fractions, though no absolute cell-fraction-recovery number is reported for it, and no reported result found on Lacuna for its $W_2$ or Energy Score values on the Onek1k data (Count Bridges) — and BayesPrism/DWLS, which showed the lowest combined false-positive and false-negative rates across 38,000 simulated breast-cancer bulk mixtures, again without an absolute metric quoted in the benchmark study (Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures). All deconvolution methods will be run on identical inputs with the same reference atlas and the same tuning budget, evaluated on simulated pseudo-bulk mixtures with known time-varying composition, where the target is proportion recovery error rather than downstream biology. Beating a baseline means lower proportion error at matched sampling density, or — for the pipeline as a whole — cell-type-specific module-metabolite associations that replicate across independent conditions (light-dark vs. constant darkness, control vs. tryptophan-depleted diet) while the bulk-only pipeline's associations do not.

The decisive ablation is the pipeline with deconvolution switched off: identical data, identical LMM-WGCNA machinery, but networks built on bulk profiles. This isolates the single ingredient the proposal stands or falls on. A second ablation keeps deconvolution but replaces the mixed-effects module-trait step with ordinary Pearson correlation, testing whether the paired-design paper's correction matters in a longitudinal setting and quantifying how many spurious associations the independence violation manufactures.

Why not just…?

Why not just run single-cell RNA-seq across the whole circadian time-course? This is the most serious objection, and the honest answer is that a scaled-down version of it is already part of the design (the anchor-point references). But the full version fails for reasons documented in the file: tauFisher's authors state that dropout and sparsity make single-sample time prediction impossible for individual cells, forcing pseudobulk aggregation anyway — accuracy of roughly 78–83% was achieved on pseudobulk, not single cells (tauFisher) — and enzymatic dissociation of solid metabolic tissues like liver induces immediate-early gene artifacts that corrupt the very metabolic states under study. Dense bulk time-series with deconvolution buys 4-hour or finer temporal resolution at a fraction of the cost; the anchor single-cell runs exist precisely to test whether the deconvolved profiles track the directly measured ones. If the anchors show the deconvolution is badly wrong, the fallback is to expand single-cell coverage, and the proposal's biological question survives even if the cheap method does not.

Why not just use differential expression plus a rhythm detector like JTK_CYCLE or Cosinor? Per-gene testing discards the coordinated-module structure that WGCNA captures and pays a heavy multiple-testing penalty when correlating thousands of genes with metabolite traits (Shen et al.). More concretely, standalone detectors degrade under noise and low temporal resolution, whereas joint modeling with a shared period rescued masked oscillations that per-series methods missed (MOSAIC). We will still run per-gene Cosinor as a sanity check, and the proposal is falsified in this respect if DEG-plus-Cosinor recovers the same cell-type assignments and metabolite associations as the module analysis with less machinery — but a gene-list output cannot by itself distinguish compositional from intracellular effects, which is the central question.

Why not just use off-the-shelf CIBERSORTx or MuSiC for the deconvolution step? Both were directly beaten by the integer-count generative approach on cell-fraction recovery (Count Bridges), and the 38,000-mixture benchmark showed widespread misclassification of cell types as purity shifts, with BayesPrism and DWLS the only robust performers on granular subtypes (Performance of tumour microenvironment deconvolution methods). Because the proposal's downstream claim depends on trustworthy per-cell-type profiles, we include CIBERSORTx, MuSiC, BayesPrism, and Count Bridges as interchangeable front-ends in the simulation benchmark; if a simpler linear tool matches Count Bridges on time-varying mixtures, we will use it and say so — the methodological claim is about the deconvolution-then-network ordering, not about any particular deconvolver.

Why not just run standard WGCNA on the bulk samples without modification? The paired-design paper proved that standard module-trait evaluation on correlated samples introduces statistical inaccuracies that the mixed-effects correction fixes, and that the correction surfaced disease-associated modules standard WGCNA missed (Application of WGCNA for Data from Paired Design). In a circadian time-course every sample from one mouse is correlated, so the violation is total, not marginal. This alternative is retained deliberately — it is the do-nothing baseline above, and the experiment is designed so that if it performs as well as the full pipeline on the simulated mixtures with known ground truth, the added machinery is unjustified and the proposal's methodological half collapses. I consider that unlikely because the simulation can plant exactly the diurnal composition swings that bulk co-expression is known to absorb, but the check is decisive rather than rhetorical.

Proposed Experiments

The reanalysis experiment comes first and costs nothing but compute. The integrator study deposited bulk SCN RNA-seq (GSE186242) and liver RNA-seq (GSE190145) under light-dark and constant-darkness conditions with diet manipulations (Tryptophan metabolism is a physiological integrator regulating circadian rhythms); the Circadiomics portal it references hosts additional mouse rhythmic expression profiles. Deconvolving these existing series against liver and brain single-cell references and running the pipeline on them is a fast test of whether cell-type-specific kynurenine-enzyme modules are even detectable in retrospect, with the caveat that these datasets' timepoint density is coarse — the integrator study's metabolomics itself was only ZT4 and ZT16, and the file reports no expression metrics for either series, so no reported result found on Lacuna for a benchmark number on GSE186242 or GSE190145.

The simulation experiment supplies ground truth. Pseudo-bulk mixtures will be constructed from the tauFisher mouse dermal single-cell time series (single-cell RNA-seq sampled every 4 hours for 72 hours) and from the breast-cancer single-cell atlas used in the deconvolution benchmark (100,064 cells), with planted diurnal composition swings and known per-cell-type harmonic programs. Metrics are cell-fraction recovery error, false-positive metabolic-module rate (modules whose eigengene correlates with a synthetic "metabolite" purely through composition), and phase-recovery error of $\phi_k$ against the planted truth. The pre-registered prediction is that bulk WGCNA's false-positive rate rises steeply with the amplitude of planted composition cycling while the deconvolved pipeline's stays flat.

The new wet-lab experiment is a 24-hour mouse study: liver and SCN harvested every 4 hours across light-dark and constant-darkness conditions, in control and tryptophan-depleted diet groups, with U-[¹³C]-tryptophan isotope tracing to quantify flux into serotonin versus kynurenine branches per timepoint — closing the two-timepoint gap the integrator study itself flagged as its main limitation. Bulk RNA-seq at every timepoint; single-cell RNA-seq at two anchor timepoints per tissue for reference construction and validation. Primary outcomes are the cell-type assignment and phase of IDO/TDO, Gpr35, and Ahr oscillation, and whether tryptophan depletion and constant darkness act on the same cell types or different ones. The 5-HTP biomarker performance of AUC = 0.9518 in diabetic nephropathy (Huangkui capsules regulate tryptophan metabolism) sets a realistic bar for how tight metabolite-trait associations in this pathway can be; module-metabolite associations an order of magnitude weaker would still be interpretable, but weaker than that would not.

A final cross-species check applies the pipeline to the human circadian blood datasets (GSE39445, GSE48113, GSE56931) used in the brain-blood phasing meta-analysis, where human brain clock genes are phase-delayed by roughly 8.5 hours relative to blood (Differential Phasing between Circadian Clocks in the Brain and Peripheral Organs in Humans). Deconvolving blood into its immune subsets and asking whether kynurenine-pathway modules phase-shift with the cell composition of the draw would test whether the known inter-individual phase dispersion in human blood — a standing obstacle for that literature — is partly compositional, and would give the method a translational foothold. The decisive checks, in order: deconvolved profiles match anchor single-cell measurements; simulated composition swings produce bulk false modules that the pipeline removes; and the kynurenine-branch shift under circadian disruption has a specific, replicating cell-type address. Any one of these failing reshapes the conclusion; all three succeeding would establish that the tryptophan-clock relationship is a property of particular cells at particular hours, not of tissues in aggregate.