A Circadian-Phase Null for Consensus Clustering of Patient Transcriptomes

When researchers cluster patients' gene-expression profiles to find disease "subtypes," they usually ignore what time of day each tissue sample was taken. That choice is risky. Thousands of genes rise and fall on a roughly 24-hour clock, so two biopsies from the same disease state, taken at 8 a.m. and 8 p.m., can look as different as two diseases. The best current tool for deciding whether clusters are real is M3C, which repeatedly reclusters resampled data and compares the result against simulated data with no clusters. Its "no structure" reference is a single Gaussian blob. A blob has no clock. As a result, a smooth cycle of time-of-day variation can pass the test as discrete subtypes. In the other direction, real disease groups defined by broken rhythms can be diluted by phase noise and dismissed.

We propose a phase-aware reference for consensus clustering, meaning clustering that is repeated many times on resampled data and then aggregated. The reference datasets contain no disease subtypes but do contain realistic circular variation from circadian phase and continuous variation in rhythm strength. A cluster is declared real only if it is more stable than anything that phase and amplitude alone could produce. We pair this with a phase-conditioned clustering step aimed at recovering disease groups that phase variation hides. Both pieces are tested on synthetic benchmarks whose subtype sizes come from a non-linear rich-get-richer process. Those size distributions run from balanced groups to one dominant subtype with tiny satellites, which is the regime where small real subtypes are hardest to separate from clock-driven artifacts.

Core Claim

The claim is that circadian phase is the largest unmodelled confounder in transcriptomic subtyping. It is removable only if it is put inside the null model rather than regressed out beforehand. The motivating tension comes from two cited papers. M3C builds Monte Carlo Gaussian nulls that keep gene–gene correlation, and with them it re-analyzed a published lupus transcriptome and found no significant clusters (K=1) where seven subtypes had been claimed. Separately, Reversible dysregulation of renal circadian rhythm in lupus nephritis showed in mice that active nephritis collapses rhythmic renal genes from 2,262 to 755 and that remission restores about half of them. Disease in this organ is therefore partly a change in rhythm. A static null cannot represent that, so neither the original seven-subtype call nor the K=1 correction can be taken at face value.

What is new is a null that is circular rather than Gaussian. Under this null, discrete structure must exceed what a continuous phase-by-amplitude manifold produces. Phase-driven pseudo-clusters stop producing false positives, and the null's power and calibration can be characterized explicitly across levels of subtype imbalance.

Research Program

The method has three parts, and the reference construction and the clustering transform are both applied identically to real and simulated data.

The first part estimates phase. For each sample we infer a latent circadian phase $\phi_i \in [0, 2\pi)$ and an amplitude scale $a_i \ge 0$ by fitting a circular latent variable to a small panel of core clock genes (Bmal1, Per1/2, Cry1/2 and their partners). Clock genes are the right anchor because they keep phase-locked relationships with one another even when downstream rhythmic output collapses, as both the lupus study and Widespread dysregulation of the circadian clock in human cancer document. Where wall-clock sampling time exists, we use it only for validation.

The second part builds the phase-aware null. For each gene $g$ we fit a phase-dependent mean with two harmonics:

$$x_{ig} = \mu_g + a_i \sum_{h=1}^{2} \big(\beta_{gh}\cos(h\phi_i) + \gamma_{gh}\sin(h\phi_i)\big) + \varepsilon_{ig}.$$

Here $\mu_g$ is the gene's baseline, $\beta_{gh}$ and $\gamma_{gh}$ are its harmonic coefficients pooled across all samples without disease labels, and $\varepsilon_{ig}$ is a residual whose cross-gene covariance we estimate empirically, as M3C does. Each null dataset is generated by drawing phases and amplitudes jointly from their empirical distribution, so continuous rhythm collapse is preserved. We then rebuild the mean from the fitted harmonics and add Gaussian residuals with the estimated covariance. Each null dataset therefore has exactly one population but the real data's clock structure.

The third part clusters on the same footing. We run consensus clustering on both the raw features and a phase-conditioned version that subtracts the fitted phase term. The transform is applied identically to real and null data, so calibration does not depend on whether the transform is correct. The transform only affects power. Stability is scored with M3C's proportion-of-ambiguous-clustering score (the share of sample pairs whose co-assignment frequency is neither near 0 nor near 1) and its relative cluster stability index, with empirical p-values against the new null. As a second aggregator we use the association-matrix method with local refinement from Statistical Guarantees for Consensus Clustering. That method proves error decays exponentially in the number of resamples and is reported to matter for unbalanced clusters. Its independence assumption holds here because resamples are drawn independently.

Non-linear preferential attachment enters as the generator for the synthetic benchmarks, not as a model of biology. Planted subtype sizes come from a sequential process in which each new synthetic patient joins existing subtype $c$ with probability proportional to $n_c^{\alpha}$, where $n_c$ is that subtype's current size and $\alpha$ is the attachment exponent. With $\alpha < 1$, sizes stay comparatively even. With $\alpha > 1$, one subtype captures most patients early, the winner-take-all behaviour analysed for influence systems in Finite-time influence systems and the Wisdom of Crowd effect. Sublinear values such as $\alpha \approx 0.43$, the value Joint estimation of preferential attachment and node fitness reports for Facebook once fitness is modelled, give an empirically grounded intermediate setting. One knob thus moves the benchmark smoothly from balanced to extremely imbalanced subtypes, so we can report where each method's power fails rather than a single average.

Baselines

The naive reference point is standard Monti consensus clustering with K chosen by the CDF and proportion-of-ambiguous-clustering heuristics. It is what most subtyping papers run, and it should over-call clusters on phase-structured data. It is run with the same base algorithm, number of resamples, and 80% sample and feature subsampling as every other arm.

The strongest existing method is M3C with its Gaussian, correlation-preserving null. It is the only cited method that formally tests "no clusters" against "some clusters." Its reported advantage is roughly 10% higher accuracy in identifying the true K than standard consensus clustering or the gap statistic in simulation, and on TCGA glioblastoma it selected K=4 where standard consensus clustering suggested K=10 (M3C). No reported result was found on Lacuna for M3C, or any consensus method, on circadian-structured or time-stamped transcriptomes, so that comparison is new ground. M3C gets the same number of null datasets, resamples, and candidate K range as our method. Beating it means significantly lower false-positive rates on phase-only data at matched K-recovery accuracy on planted subtypes. Equal false-positive rates with higher K-recovery accuracy would also count.

The ablations isolate the key ingredient. One keeps the null Gaussian but applies our phase-conditioned transform, which tests whether the null or the transform is doing the work. Another replaces estimated phase with true sampling time, which bounds the loss from phase estimation. A third fixes amplitude at one in the null, which tests whether modelling continuous rhythm collapse matters.

Why not just…?

Why not just regress phase out and run M3C as usual? This is the most serious competitor. If disease does not change rhythms, it may work as well as our method. When disease flattens rhythms, though, a single pooled phase model fitted to flat diseased samples leaves inverted rhythmic residuals. That could create spurious structure, or it could erase the rhythm-loss signal that defines the group. Our "transform without phase-aware null" ablation is exactly this alternative. We will report honestly whether it matches the full method. The synthetic arm with disease-specific amplitude collapse is the setting expected to separate the two.

Why not just record the biopsy time and stratify on it? Most public cohorts, including the cross-sectional resources discussed in the file, lack reliable sampling times. Clock time is also not internal phase: A late eating midpoint is associated with increased risk of diabetic kidney disease and the sleep-regularity cohorts show that behaviour shifts peripheral timing between people. The true-time ablation measures how much is lost by estimating phase instead. If that loss is small, the stratification argument largely evaporates for cohorts that do have timestamps.

Why not just summarize each sample's rhythm with a cosinor fit or a scalar index and cluster on that? A single sample has no time series, so per-sample cosinor fitting is impossible. Collapsing to scalar regularity indices discards the multi-gene structure that subtyping needs. The lupus study needed a non-parametric rhythm detector, RAIN, because diseased waveforms are not sinusoidal. We allow two harmonics for the same reason and will test whether one harmonic suffices.

Proposed Experiments

The first experiment is synthetic. Data are generated from the model above with K in {1, 2, 3, 5}, subtype sizes drawn from the attachment process at $\alpha \in \{0.43, 1, 1.5\}$, and subtypes differing either in mean expression or in rhythm amplitude. We measure the false-positive rate when K=1 (phase only), K-recovery accuracy, and adjusted Rand index against planted labels. The method must hold nominal false-positive rates where M3C does not, and the attachment exponent shows where small subtypes become undetectable.

The second experiment uses the NZB/W F1 renal time series (GEO: GSE131894). Samples were taken every four hours from young, nephritic, and remission mice. We hide the timestamps. The prediction is that standard and Gaussian-null M3C will report clusters aligned with time of day, while the phase-aware method will report disease state (nephritic versus not) as the significant structure. The file does not give per-group sample counts, and power may be limited, so this experiment is a qualitative check rather than a precision estimate.

The third experiment is a calibration check on TCGA glioblastoma, where the method should still recover K=4, and on the lupus expression set M3C re-analyzed. On the lupus data either outcome is informative. Persistent K=1 strengthens M3C's conclusion, while a significant phase-conditioned partition would suggest rhythm-defined groups hidden by phase noise. Phase estimation in tumours may be unreliable given widespread clock dysregulation. We therefore report clock-gene coherence per sample and treat low-coherence samples as a stated limitation, not something to tune away.

The decisive check is the K=1, phase-only synthetic condition. If our null does not cut M3C's false-positive rate there, the proposal fails regardless of the real-data results.