Most cells keep time with an internal circadian clock, a loop of genes such as BMAL1, CLOCK, PER and CRY whose products switch each other on and off roughly every 24 hours. Bulk tumor samples, in which millions of cells are ground up and sequenced together, consistently show this loop "broken": the clock genes no longer co-vary the way they do in healthy tissue. The field has largely read that as oncogenic destruction of the clock. A bulk sample, however, will look identical whether every cell's clock has stopped or every cell's clock is still running but out of step with its neighbours, because out-of-phase oscillations cancel when averaged. A second confound sits on top of the first: the signal mixes cancer cells with fibroblasts and immune cells. Neither question has been answered for human cancer, and the answer changes what drugs or timing interventions would make sense.
We propose a per-sample, per-cell-type estimator of two separate quantities. The first is synchrony, meaning how tightly the cells' clock phases are bunched. The second is off-cycle distortion, meaning how far the clock-gene profile sits from anything a running clock could produce. The method rests on one mathematical fact. Averaging cells that run the same clock at scattered phases shrinks every clock gene's oscillation by the same factor and keeps their relative timing intact. Stopping or rewiring the clock instead distorts the genes unequally. We validate the estimator on time-series and single-cell data where the answer is known. We then test it against live single-cell protein imaging in the MYC-inducible cell system in which MYC was shown to dampen the clock. Finally, we apply it to cancer-cell and stromal compartments of human tumors.
Core Claim
The strongest evidence for widespread cancer clock disruption is the Clock Correlation Distance of Evidence for widespread dysregulation of circadian clock progression in human cancer. That paper compares the pattern of pairwise correlations among 12 clock genes across a cohort with a healthy reference, and finds significant loss in 15 of 20 cancer datasets. It is a cohort-level statistic and cannot say whether any individual tumor's clock is running. tauFisher, a method that predicts circadian time from one sample using within-sample gene-pair differences, showed that this distinction matters. Dermal immune cells looked like they had weak clocks in aggregate. They actually had intact clocks with high cell-to-cell phase heterogeneity. tauFisher worked on healthy skin, used pseudobulk aggregates, and did not turn phase heterogeneity into a per-sample quantity that could be separated from arrest. Our claim is that a single bulk or pseudobulk sample carries enough information to estimate synchrony and off-cycle distortion separately, given a reference limit cycle, which is the closed loop that clock-gene expression traces over 24 hours. We further claim that this decomposition, applied per compartment, will show which tumor compartments have scattered clocks and which have arrested ones.
The mechanism fits in one equation. Let gene $g$ in a running clock follow $m_g + a_g\cos(\theta - \phi_g)$ in log expression, where $\theta$ is the cell's phase, $m_g$ the mean, $a_g$ the amplitude and $\phi_g$ the gene's peak phase from the healthy reference. If a sample's cells have phases drawn from a von Mises distribution, a circular analogue of the Gaussian with mean phase $\mu$ and concentration $\kappa$, then the averaged profile is
$$ \bar{x}_g = m_g + R(\kappa)\, a_g \cos(\mu - \phi_g) + r_g, \qquad R(\kappa) = \frac{I_1(\kappa)}{I_0(\kappa)} \in [0,1), $$
where $I_0, I_1$ are modified Bessel functions, $R$ is the synchrony (mean resultant length), and $r_g$ is the residual that no scattered population of running clocks can produce. This form follows from taking the expectation of the first Fourier harmonic under a von Mises density. Pure desynchrony moves the profile radially toward the cycle's centre with $r \approx 0$. Arrest or rewiring, for example MYC/MIZ1 repressing BMAL1, CLOCK and NPAS2 as in MYC/MIZ1-dependent gene repression inversely coordinates the circadian clock with cell cycle and proliferation, produces gene-specific departures and a large $\|r\|$. With roughly 12 to 17 clock genes, only two degrees of freedom ($\mu$ and $R$) go to the running-clock part, which leaves about ten to detect distortion. To cancel sample-wide scale and composition effects, the model is fit on within-sample gene-pair log differences, borrowing tauFisher's normalization idea.
There is one honest blind spot. A clock frozen exactly at a state that lies on the healthy cycle is indistinguishable, in a single snapshot, from a perfectly synchronized running clock. We address this empirically, by locating where MYC-arrested profiles actually land, and at the cohort level, by checking whether inferred phases pile up at one point regardless of sampling context.
Research Program
The estimator has three parts.
The first is a reference limit cycle per cell type ($m_g, a_g, \phi_g$). We fit it on healthy time-series data: the 12 circadian benchmark datasets used by tauFisher, the healthy references underlying the Clock Correlation Distance, and per-cell-type pseudobulk from the 72-hour, 4-hourly mouse dermal skin single-cell dataset analysed in tauFisher.
The second is a per-sample fit of $(\mu, R, r)$. We use a likelihood that treats pseudobulk counts as negative binomial, and we place a sparsity penalty on $r$ so that small noise is not called distortion. Calibrated null distributions for $\|r\|$ come from held-out healthy samples.
The third is compartment resolution, handled in two ways. Where tumor single-cell atlases exist, we fit per-patient, per-cell-type pseudobulk directly. This covers the Wu et al. breast atlas of 100,064 cells from 26 patients used in Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures, and the 3CA atlas of 2.7 million cells across 15 cancers used by POPPY. For bulk TCGA/GEO tumors we first obtain per-sample, cell-type-specific expression with BayesPrism, the best performer in that breast deconvolution benchmark, and then fit the model. We validate this two-step route on pseudobulk mixtures built from the same atlas, where per-compartment truth is known.
Transcript-protein discordance enters as a constraint, not as a modelling target. The estimator reads mRNA, but the clock's function lives in proteins. MOSAIC found 18.5% of the oscillating Neurospora proteome cycling without transcript rhythms. A high-stringency blueprint of the human proteome puts mRNA–protein agreement at only 40–60%. An mRNA-level "arrest" call could therefore be wrong if the protein oscillator keeps running. We do not impute proteins. scChord-style translation (scChord) is trained only on surface epitopes, and clock factors are intracellular. Instead we measure clock protein directly in a controlled system and ask whether the mRNA-level verdict agrees with it.
Baselines
The naive reference point is the cohort-level Clock Correlation Distance and its tumor-minus-normal difference from Evidence for widespread dysregulation of circadian clock progression in human cancer. That paper reports significant loss in 15 of 20 cancer datasets at $p<0.005$. We run it on exactly the same cohorts and cell-type pseudobulks with its published 12-gene reference. Because it yields no per-sample quantity, we compare it on the one question it can address: whether a cohort's correlation loss is explained by desynchrony. In synthetic cohorts with known mixtures of scattered and arrested samples, the proposal beats this baseline if our per-sample decomposition predicts the cohort's distance value and also attributes it correctly. The baseline cannot make that attribution.
The strongest existing method is tauFisher (tauFisher). It outperformed ZeitZeiger and TimeSignatR on 11 of 12 benchmarks, typically predicting time within 2 hours, and reached about 78–83% accuracy on mouse SCN and skin pseudobulk. Its bootstrap over cells is the existing tool for detecting phase heterogeneity. We give it the same reference datasets for training and the same pseudobulks. We take the entropy of its circular class posterior, and the spread of its bootstrap phases, as its synchrony score, and its maximum posterior probability as an implicit arrest score. We tune both methods with the same held-out healthy samples. We have found no reported result on Lacuna for any method separating desynchrony from arrest, so the target is set by our own benchmark. To beat tauFisher, we need higher area under the ROC curve for classifying scattered versus arrested pseudobulks, and lower error on $R$, while matching its phase accuracy on synchronized samples.
For compartments, the comparison is tumor-level fitting against BayesPrism-then-fit. Its reference point is truth from atlas pseudobulk. Two ablations isolate the key ingredient. One removes the residual term, so the model can only explain distortion as low synchrony. The other drops the shared-shrinkage constraint, letting each gene's amplitude scale freely. If either ablation matches the full model on the arrest-versus-scatter task, the central equation adds nothing.
Why not just…?
Why not just compute the Clock Correlation Distance separately in high- and low-purity tumors? Stratifying by purity tells you whether cancer cells or stroma drive the correlation loss. It still cannot tell scattering from arrest, because both destroy correlation across a cohort. Our synthetic cohorts with matched correlation loss but different scatter/arrest composition test this directly. If purity-stratified distance values can distinguish them, our method is unnecessary for that use.
Why not just use tauFisher's posterior uncertainty as a synchrony readout? This is the real threat. A flat posterior arises both when the profile sits near the cycle centre (scatter) and when it sits off-cycle (arrest). tauFisher has no explicit off-cycle term, so we expect it to conflate the two. We cannot assume that. The scatter-versus-arrest benchmark in Experiment 1 settles it, and if tauFisher's entropy plus a simple reconstruction error matches our model, we would report that and adopt the simpler tool.
Why not just timestamp individual tumor cells from single-cell RNA? The tauFisher authors needed pseudobulk because dropout breaks within-cell gene-pair comparisons. Tumor biopsies also come at one unknown clock time. Our design sidesteps per-cell timestamps: synchrony is a property of a population and can be read from its average. Experiment 1 includes a per-cell baseline, which fits the same model to individual skin cells and aggregates the inferred phases, to show how much noise that adds.
Why not just measure clock proteins in tumors? That would be ideal, but clinical samples are single destructive snapshots, and single-cell protein panels reach mainly surface markers. Live single-cell imaging in culture, used in Experiment 2, is the feasible protein-level ground truth.
Proposed Experiments
Experiment 1 builds the scatter-versus-arrest benchmark. For scatter, we construct pseudobulks by mixing cells or samples from known time points of the 72-hour skin dataset and the 12 time-series benchmarks with von Mises weights, sweeping $\kappa$ from 0 to high. Skin fibroblasts and immune cells give a natural low-versus-high scatter contrast already reported by tauFisher. For arrest, we perturb clock genes toward non-cycle configurations patterned on MYC/MIZ1 repression of the positive arm, at graded strengths. We also create mixtures of the two. Metrics are error on $R$, circular phase error on synchronized samples, and area under the ROC curve for arrest. Robustness is checked at 5–95% tumor purity in Wu-atlas mixtures, because the purity trap documented in that breast benchmark could leak normal-epithelial clock signal into the cancer compartment.
Experiment 2 is the decisive biological check. It uses the U2OS Tet-on MYC system from MYC/MIZ1-dependent gene repression inversely coordinates the circadian clock with cell cycle and proliferation with single-cell PER2 reporter imaging over several days. Imaging gives per-cell protein amplitude and phase, and therefore true synchrony and arrest, while parallel RNA-seq at a single time point feeds our estimator. If MYC's population dampening is mostly per-cell amplitude loss, the model must call arrest. If it is mostly phase spread, the model must call scatter. Cells whose mRNA profile looks arrested but whose PER2 protein still oscillates measure the discordance risk directly.
Experiment 3 is the application. We apply the validated estimator to per-patient compartments in the breast and 3CA atlases, and through BayesPrism to the 20 TCGA/GEO cohorts of the Clock Correlation Distance study. The question is which share of each cancer's correlation loss is synchrony loss versus distortion, separately for malignant cells, fibroblasts and myeloid cells. The fibroblast result can be checked against the localization of circadian-linked risk genes to fibroblasts in Integrate bulk RNA and single-cell sequencing to identify prognostic genes associated with dietary restriction and circadian rhythm in colorectal cancer and conduct experimental verification. Distortion in malignant cells should rise with MYC expression if the MYC/MIZ1 mechanism operates in patients, which is a falsifiable prediction. If instead inferred phases of high-$R$ tumors pile up at a single cycle state across patients, that points to clocks frozen on-cycle, the one case our snapshot test cannot see. We would report it as such rather than as synchrony.
References
- Evidence for widespread dysregulation of circadian clock progression in human cancer
- tauFisher predicts circadian time from a single sample of bulk and single-cell pseudobulk transcriptomic data
- MYC/MIZ1-dependent gene repression inversely coordinates the circadian clock with cell cycle and proliferation
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- MOSAIC: A Joint Modeling Methodology for Combined Circadian and Non-Circadian Analysis of Multi-Omics Data
- A high-stringency blueprint of the human proteome
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Integrate bulk RNA and single-cell sequencing to identify prognostic genes associated with dietary restriction and circadian rhythm in colorectal cancer and conduct experimental verification